All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives. - OpenGL: Implicit all the way. There are some EGL extensions to enable some forms of explicit sync via sync_file but OpenGL itself is still implicit. - Wayland: Currently depends on implicit sync in the kernel (accessed via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later). - X11: With present, it has these "explicit" fence objects but they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit. - linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync. - linux/amdgpu: Supports sync_file and syncobj but it still implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization. - KMS: Implicit sync all the way. There are no KMS APIs which take explicit sync primitives. - v4l: ??? - gstreamer: ??? - Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with
the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this
space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs. So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured. In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue. Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
Hi Jason,
I've been wrestling with the sync problems in Wayland some time ago, but only with regards to 3D drivers.
The guarantee given by the GL/GLES spec is limited to a single graphics context. If the same buffer is accessed by 2 contexts the outcome is unspecified. The cross-context and cross-process synchronisation is not guaranteed. It happens to work on Mesa, because the read/write locking is implemented in the kernel space, but it didn't work on Broadcom driver, which has read-write interlocks in user space.
A Vulkan client makes it even worse because of conflicting requirements: Vulkan's vkQueuePresentKHR() passes in a number of semaphores but disallows waiting. Wayland WSI requires wl_surface_commit() to be called from vkQueuePresentKHR() which does require a wait, unless a synchronisation primitive representing Vulkan samaphores is passed between Vulkan client and the compositor.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
With regards to V4L2, I believe it could easily work the same way as 3D drivers, i.e. pass a buffer+fence pair to the next stage. The encode always succeeds, but for capture or decode, the main problem is the uncertain outcome, I believe? If we're fine with rendering or displaying an occasional broken frame, then buffer+fence pair would work too. The broken frame will go into the pipeline, but application can drain the pipeline and start over once the capture works again.
To answer some points raised by Laurent (although I'm unfamiliar with the camera drivers):
you don't know until capture complete in which buffer the frame has
been captured Surely you do, you only don't know in advance if the capture will be successful
but if an error occurs during capture, they can be recycled internally
and put to the back of the queue. That would have to change in order to use explicit synchronisation. Every started capture becomes immediately available as a buffer+fence pair. Fence is signalled once the capture is finished (successfully or otherwise). The buffer must not be reused until it's released, possibly with another fence - in that case the buffer must not be reused until the release fence is signalled.
Cheers, Tomek
On Mon, 16 Mar 2020 at 10:20, Laurent Pinchart < laurent.pinchart@ideasonboard.com> wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net
wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most
people
who reply will get at least one mailing list rejection. However,
this
is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the
start
was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you
have
two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client
has
to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via
wl_surface::commit)
to tell the compositor it's done at which point the compositor can
now
texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client
space
and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as
an
API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the
fences
on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product
of
submitting work (3D, compute, etc.) to the kernel and is signaled
when
that work completes. When a sync_file is created, it is guaranteed
by
the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for
the
sync_file to be triggered before executing. A sync_file also
supports
poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has
APIs
to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel
(accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for
explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using
asynchronous buffer
synchronisation is something we do already with GL (even if limited). We
place
GLSync object in the pipeline and attach that on related GstBuffer. We
wait on
these GLSync as late as possible (or superseed the sync if we queue more
work
into the same GL context). That requires a special mode of operation of
course.
We don't usually like making lazy blocking call implicit, as it tends to
cause
random issues. If we need to wait, we think it's better to wait int he
module
that is responsible, so in general, we try to negotiate and fallback
locally
(it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other
lower
level APIs first. We need API that provides us these fence (in or out),
and then
we can consider using them. For V4L2, there was an attempt, but it was a
bit of
a miss-fit. Your proposal could work, need to be tested I guess, but it
does not
solve some of other issues that was discussed. Notably for camera
capture, were
the HW timestamp is capture about at the same time the frame is ready.
But the
timestamp is not part of the paylaod, so you need an entire API
asynchronously
deliver that metadata. It's the biggest pain point I've found, such an
API would
be quite invasive or if made really generic, might just never be adopted
widely
enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured. In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue. Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
There is other elements that would implement fencing, notably kmssink,
but no
one actually dared porting it to atomic KMS, so clearly there is very
little
comunity interest. glimagsink could clearly benifit. Right now if we
import a
DMABuf, and that this DMAbuf is used for render, a implicit fence is
attached,
which we are unaware. Philippe Zabbel is working on a patch, so V4L2
QBUF would
wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which
GStreamer
uses), so then the operation will just fail where it worked before
(breaking
userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal
with
the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems
for
me because I can't really use it unless it's implemented in all of
the
compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they
go
to flip a client buffer to the screen directly, they discover that
KMS
doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't
be
scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides
that
they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece
and
then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then
just
hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've
been
giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined
in
i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it
gets
buffers in from a client that doesn't use the explicit sync
extension,
it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file
into
any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The
WSI
code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
-- Regards,
Laurent Pinchart _______________________________________________ wayland-devel mailing list wayland-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/wayland-devel
Hi Tomek,
On Mon, Mar 16, 2020 at 12:55:27PM +0000, Tomek Bury wrote:
Hi Jason,
I've been wrestling with the sync problems in Wayland some time ago, but only with regards to 3D drivers.
The guarantee given by the GL/GLES spec is limited to a single graphics context. If the same buffer is accessed by 2 contexts the outcome is unspecified. The cross-context and cross-process synchronisation is not guaranteed. It happens to work on Mesa, because the read/write locking is implemented in the kernel space, but it didn't work on Broadcom driver, which has read-write interlocks in user space.
A Vulkan client makes it even worse because of conflicting requirements: Vulkan's vkQueuePresentKHR() passes in a number of semaphores but disallows waiting. Wayland WSI requires wl_surface_commit() to be called from vkQueuePresentKHR() which does require a wait, unless a synchronisation primitive representing Vulkan samaphores is passed between Vulkan client and the compositor.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
With regards to V4L2, I believe it could easily work the same way as 3D drivers, i.e. pass a buffer+fence pair to the next stage. The encode always succeeds, but for capture or decode, the main problem is the uncertain outcome, I believe? If we're fine with rendering or displaying an occasional broken frame, then buffer+fence pair would work too. The broken frame will go into the pipeline, but application can drain the pipeline and start over once the capture works again.
To answer some points raised by Laurent (although I'm unfamiliar with the camera drivers):
you don't know until capture complete in which buffer the frame has been captured
Surely you do, you only don't know in advance if the capture will be successful
You do in kernelspace, but not in userspace at the moment, due to buffer recycling.
but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
That would have to change in order to use explicit synchronisation. Every started capture becomes immediately available as a buffer+fence pair. Fence is signalled once the capture is finished (successfully or otherwise). The buffer must not be reused until it's released, possibly with another fence - in that case the buffer must not be reused until the release fence is signalled.
We could certainly change this at least in some cases, but it would break existing userspace that doesn't expect incorrect frames.
I'm however not sure we could change this behaviour in every case, there may be hardware that can't provide a guarantee on the order in which buffers will be used. I'm aware this wouldn't be compatible with explicit synchronization, and that's my point: camera hardware may not always support explicit synchronization. As long as we can fall back to not using fences then we should be fine.
As long as we can fall back to not using fences then we should be fine.
Buffers written by the camera are trivial because you control what happens - just don't attach fence, so that the capture can be used immediately. For recycled buffers there's an extra bit of work to do because won't be up to camera driver to decide whether the buffer comes back with or without fence.
Hi Tomek,
On Mon, 16 Mar 2020 at 12:55, Tomek Bury tomek.bury@gmail.com wrote:
I've been wrestling with the sync problems in Wayland some time ago, but only with regards to 3D drivers.
The guarantee given by the GL/GLES spec is limited to a single graphics context. If the same buffer is accessed by 2 contexts the outcome is unspecified. The cross-context and cross-process synchronisation is not guaranteed. It happens to work on Mesa, because the read/write locking is implemented in the kernel space, but it didn't work on Broadcom driver, which has read-write interlocks in user space.
GL and GLES are not relevant. What is relevant is EGL, which defines interfaces to make things work on the native platform. EGL doesn't define any kind of synchronisation model for the Wayland, X11, or GBM/KMS platforms - but it's one of the things which has to work. It doesn't say that the implementation must make sure that the requested format is displayable, but you sort of take it for granted that if you ask EGL to display something it will do so.
Synchronisation is one of those mechanisms which is left to the platform to implement under the hood. In the absence of platform support for explicit synchronisation, the synchronisation must be implicit.
A Vulkan client makes it even worse because of conflicting requirements: Vulkan's vkQueuePresentKHR() passes in a number of semaphores but disallows waiting. Wayland WSI requires wl_surface_commit() to be called from vkQueuePresentKHR() which does require a wait, unless a synchronisation primitive representing Vulkan samaphores is passed between Vulkan client and the compositor.
If you are using EGL_WL_bind_wayland_display, then one of the things it is explicitly allowed/expected to do is to create a Wayland protocol interface between client and compositor, which can be used to pass buffer handles and metadata in a platform-specific way. Adding synchronisation is also possible.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
You can add your own explicit synchronisation extension.
In every cross-process and cross-subsystem usecase, synchronisation is obviously required. The two options for this are to implement kernel support for implicit synchronisation (as everyone else has done), or implement generic support for explicit synchronisation (as we have been working on with implementations inside Weston and Exosphere at least), or implement private support for explicit synchronisation, or do nothing and then be surprised at the lack of synchronisation.
Cheers, Daniel
GL and GLES are not relevant. What is relevant is EGL, which defines interfaces to make things work on the native platform.
Yes and no. This is what EGL spec says about sharing a texture between contexts:
"OpenGL and OpenGL ES makes no attempt to synchronize access to texture objects. If a texture object is bound to more than one context, then it is up to the programmer to ensure that the contents of the object are not being changed via one context while another context is using the texture object for rendering. The results of changing a texture object while another context is using it are undefined."
There are similar statements with regards to the lack of synchronisation guarantees for EGL images or between GL and native rendering, etc. But the main thing here is that EGL and Vulkan differ significantly. The eglSwapBuffers() is expected to post an unspecified "back buffer" to the display system using some internal driver magic. EGL driver is then expected to obtain another back buffer at some unspecified point in the future. Vulkan on the other hand is very specific and explicit. The vkQueuePresentKHR() is expected to post a specific vkImage with an explicit set of set of semaphores. Another image is obtained through vkAcquireNextImageKHR() and it's the application's decision whether it wants a fence, a semaphore, both or none with the acquired buffer. The implicit synchronisation doesn't mix well with Vulkan drivers and requires a lot of extra plumbing in the WSI code.
If you are using EGL_WL_bind_wayland_display, then one of the things it is explicitly allowed/expected to do is to create a Wayland protocol interface between client and compositor, which can be used to pass buffer handles and metadata in a platform-specific way. Adding synchronisation is also possible.
Only one-way synchronisation is possible with this mechanism. There's a standard protocol for recycling buffers - wl_buffer_release() so buffer hand-over from the compositor to client remains unsynchronised - see below.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
You can add your own explicit synchronisation extension.
I could but that requires implementing in in the driver and in a number of compositors, therefore a standard extension zwp_linux_explicit_synchronization_v1 is much better choice here than a custom one.
In every cross-process and cross-subsystem usecase, synchronisation is obviously required. The two options for this are to implement kernel support for implicit synchronisation (as everyone else has done),
That would require major changes in driver architecture or a 2nd mechanisms doing the same thing but in kernel space - both are non-starters.
or implement generic support for explicit synchronisation (as we have been working on with implementations inside Weston and Exosphere at least),
The zwp_linux_explicit_synchronization_v1 is a good step forward. I'm using this extension as a main synchronisation mechanism in EGL and Vulkan driver whenever available. I remember that Gustavo Padovan was working on explicit sync support in the display system some time ago. I hope it got merged into kernel by now, but I don't know to what extend it's actually being used.
or implement private support for explicit synchronisation,
If everything else fails, that would be the last resort scenario, but far from ideal and very costly in terms of implementation and maintenance as it would require maintaining custom patches for various 3rd party components or littering them with multiple custom explicit synchronisation schemes.
or do nothing and then be surprised at the lack of synchronisation.
Thank you, but no, thank you :)
Cheers, Tomek
On Mon, Mar 16, 2020 at 10:33 AM Tomek Bury tomek.bury@gmail.com wrote:
GL and GLES are not relevant. What is relevant is EGL, which defines interfaces to make things work on the native platform.
Yes and no. This is what EGL spec says about sharing a texture between contexts:
"OpenGL and OpenGL ES makes no attempt to synchronize access to texture objects. If a texture object is bound to more than one context, then it is up to the programmer to ensure that the contents of the object are not being changed via one context while another context is using the texture object for rendering. The results of changing a texture object while another context is using it are undefined."
There are similar statements with regards to the lack of synchronisation guarantees for EGL images or between GL and native rendering, etc. But the main thing here is that EGL and Vulkan differ significantly. The eglSwapBuffers() is expected to post an unspecified "back buffer" to the display system using some internal driver magic. EGL driver is then expected to obtain another back buffer at some unspecified point in the future. Vulkan on the other hand is very specific and explicit. The vkQueuePresentKHR() is expected to post a specific vkImage with an explicit set of set of semaphores. Another image is obtained through vkAcquireNextImageKHR() and it's the application's decision whether it wants a fence, a semaphore, both or none with the acquired buffer. The implicit synchronisation doesn't mix well with Vulkan drivers and requires a lot of extra plumbing in the WSI code.
Yes, and that (the Vulkan issues in particular) is what I'm trying to fix. :-) (among other things...) Assuming the kernel patch I linked to, your usermode driver could stuff fences in the dma-buf without having that be part of your kernel driver. This assumes, of course, that your kernel driver supports sync_file.
If you are using EGL_WL_bind_wayland_display, then one of the things it is explicitly allowed/expected to do is to create a Wayland protocol interface between client and compositor, which can be used to pass buffer handles and metadata in a platform-specific way. Adding synchronisation is also possible.
Only one-way synchronisation is possible with this mechanism. There's a standard protocol for recycling buffers - wl_buffer_release() so buffer hand-over from the compositor to client remains unsynchronised
- see below.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
You can add your own explicit synchronisation extension.
I could but that requires implementing in in the driver and in a number of compositors, therefore a standard extension zwp_linux_explicit_synchronization_v1 is much better choice here than a custom one.
I think you may be missing what Daniel is saying. Wayland allows you to do basically anything you want within your client and server-side EGL implementations. That could include the server-side EGL sending an event with a fence every single time a flush operation happens in the server-side GL/GLES implementation. (Could be glFlush, glFinish, eglSwapBuffers, or other things). Since wayland protocol events are ordered, the client-side EGL implementation would get the most recent flush event before it got the wl_buffer::release. I fully agree that it's rather cumbersome though.
In every cross-process and cross-subsystem usecase, synchronisation is obviously required. The two options for this are to implement kernel support for implicit synchronisation (as everyone else has done),
That would require major changes in driver architecture or a 2nd mechanisms doing the same thing but in kernel space - both are non-starters.
or implement generic support for explicit synchronisation (as we have been working on with implementations inside Weston and Exosphere at least),
The zwp_linux_explicit_synchronization_v1 is a good step forward. I'm using this extension as a main synchronisation mechanism in EGL and Vulkan driver whenever available. I remember that Gustavo Padovan was working on explicit sync support in the display system some time ago. I hope it got merged into kernel by now, but I don't know to what extend it's actually being used.
It is supported by KMS/atomic. Legacy KMS, however, does not support it.
or implement private support for explicit synchronisation,
If everything else fails, that would be the last resort scenario, but far from ideal and very costly in terms of implementation and maintenance as it would require maintaining custom patches for various 3rd party components or littering them with multiple custom explicit synchronisation schemes.
If you want to see explicit synchronization everywhere, I would very much like to see more developers driving its adoption. I implemented support in the Intel Vulkan driver last week:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4169
Hopefully, that will provide some motivation for other compositors (kwin, gnome-shell, etc.) because they now have a real user of it in an upstream driver for a major desktop platform and not just a few weston examples. However, someone is going to have to drive the actual development in those compositors. I'd be very happy if more people got involved, :-)
--Jason
On Monday, March 16, 2020 5:04 PM, Jason Ekstrand jason@jlekstrand.net wrote:
Hopefully, that will provide some motivation for other compositors (kwin, gnome-shell, etc.) because they now have a real user of it in an upstream driver for a major desktop platform and not just a few weston examples. However, someone is going to have to drive the actual development in those compositors. I'd be very happy if more people got involved, :-)
FWIW, a wlroots pull request is in progress [0]. The plan is first to accept fence FDs from clients, then send them our fences as a second step.
On Tue, Mar 17, 2020 at 3:01 AM Simon Ser contact@emersion.fr wrote:
On Monday, March 16, 2020 5:04 PM, Jason Ekstrand jason@jlekstrand.net wrote:
Hopefully, that will provide some motivation for other compositors (kwin, gnome-shell, etc.) because they now have a real user of it in an upstream driver for a major desktop platform and not just a few weston examples. However, someone is going to have to drive the actual development in those compositors. I'd be very happy if more people got involved, :-)
FWIW, a wlroots pull request is in progress [0]. The plan is first to accept fence FDs from clients, then send them our fences as a second step.
What exactly are the semantics there? Are you going to somehow wait inside wlroots for the buffer to be 100% idle or are you expecting the client to somehow use explicit for sending buffers implicit to wait for idle? If it's the latter, that's not going to work.
--Jason
Hi,
On Mon, 16 Mar 2020 at 15:33, Tomek Bury tomek.bury@gmail.com wrote:
GL and GLES are not relevant. What is relevant is EGL, which defines interfaces to make things work on the native platform.
Yes and no. This is what EGL spec says about sharing a texture between contexts:
Contexts are different though ...
There are similar statements with regards to the lack of synchronisation guarantees for EGL images or between GL and native rendering, etc.
This also isn't about native rendering.
But the main thing here is that EGL and Vulkan differ significantly.
Sure, I totally agree.
The eglSwapBuffers() is expected to post an unspecified "back buffer" to the display system using some internal driver magic. EGL driver is then expected to obtain another back buffer at some unspecified point in the future.
Yes, this is rather the point: EGL doesn't specify platform-related 'black magic' to make things just work, because that's part of the platform implementation details. And, as things stand, on Linux one of those things is implicit synchronisation, unless the desired end state of your driver is no synchronisation.
This thread is a discussion about changing that.
If you are using EGL_WL_bind_wayland_display, then one of the things it is explicitly allowed/expected to do is to create a Wayland protocol interface between client and compositor, which can be used to pass buffer handles and metadata in a platform-specific way. Adding synchronisation is also possible.
Only one-way synchronisation is possible with this mechanism. There's a standard protocol for recycling buffers - wl_buffer_release() so buffer hand-over from the compositor to client remains unsynchronised
- see below.
That's not true; you can post back a sync token every time the client buffer is used by the compositor.
The most troublesome part was Wayland buffer release mechanism, as it only involves a CPU signalling over Wayland IPC, without any 3D driver involvement. The choices were: explicit synchronisation extension or a buffer copy in the compositor (i.e. compositor textures from the copy, so the client can re-write the original), or some implicit synchronisation in kernel space (but that wasn't an option in Broadcom driver).
You can add your own explicit synchronisation extension.
I could but that requires implementing in in the driver and in a number of compositors, therefore a standard extension zwp_linux_explicit_synchronization_v1 is much better choice here than a custom one.
EGL_WL_bind_wayland_display is explicitly designed to allow each driver to implement its own private extensions without modifying compositors. For instance, Mesa adds the `wl_drm` extension, which is used for bidirectional communication between the EGL implementations in the client and compositor address spaces, without modifying either.
In every cross-process and cross-subsystem usecase, synchronisation is obviously required. The two options for this are to implement kernel support for implicit synchronisation (as everyone else has done),
That would require major changes in driver architecture or a 2nd mechanisms doing the same thing but in kernel space - both are non-starters.
OK. As it stands, everyone else has the kernel mechanism (e.g. via dmabuf resv), so in this case if you are reinventing the underlying platform in a proprietary stack, you get to solve the same problems yourselves.
Cheers, Daniel
That's not true; you can post back a sync token every time the client buffer is used by the compositor.
Technically, yes but it's very cumbersome and invasive to the point where it becomes impractical. Explicit sync is much cleaner solution.
For instance, Mesa adds the `wl_drm` extension, which is used for bidirectional communication between the EGL implementations in the client and compositor address spaces, without modifying either.
Broadcom driver adds "wl_nexus" extension which servers similar purpose for both EGL and Vulkan WSI
OK. As it stands, everyone else has the kernel mechanism (e.g. via dmabuf resv), so in this case if you are reinventing the underlying platform in a proprietary stack, you get to solve the same problems yourselves.
That's an important point. In the explicit synchronisation scenario the sync token is passed with the buffer. It becomes irrelevant where the token originated from, as long as it's a commonly used type of token, i.e. dma_fence in kernel space or sync_fd in user space. That allows for greater flexibility and works with and without dma reservation objects.
Cheers, Tomek
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart laurent.pinchart@ideasonboard.com wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Trying to understand. :-)
--Jason
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
-- Regards,
Laurent Pinchart
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them. Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Trying to understand. :-)
So am I :-)
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
On Mon, Mar 16, 2020 at 4:15 PM Laurent Pinchart laurent.pinchart@ideasonboard.com wrote:
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
Yeah, I think I was aware of at least some of that. I would expect (though, again, I don't know) that the hardware which consumes images generally shouldn't have those "packet loss" problems. A video output device might miss vblank but that's something we deal with in display hardware all the time. Codecs, I would hope, are reliable enough that they should more-or-less always succeed. Is this assumption correct?
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Yeah, that makes sense. In that case, there are likely going to be some devices that either need to wait for the actual end-of-frame before handing the buffer back to userspace or will need some sort of out-of-band "ignore this frame, it's corrupted" error. The later sounds fairly painful for userspace to handle correctly. Is this "packet loss" something that all video devices experience or is it mostly cheaper ones?
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
Just to be clear, do you actually use the dma-buf implicit sync stuff today or do all V4L capture devices wait until the full frame is complete before returning anything to userspace?
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them.
Yeah, there's some API design questions to be answered here. It's possible to have an image output API which always provides a sync_file and, depending on the hardware, it may have one of two behaviors:
1. Hand out images before the capture is done and trigger the sync file once that frame's capture is completed 2. Hand out images only after the full frame has been completed and provide an already triggered sync_file
It would also be possible to make whether or not you get a sync_file vs. implicit sync configurable from userspace (it kind-of has to be opt-in since it would be a new UAPI) or to make it depend on the underlying hardware. This potentially makes userspace software more complex which may make it harder to get right. Lots of trade-offs here.
Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Is passing erroneous frames to userspace current behavior? Or are you talking about what a sync_file future looks like?
--Jason
Trying to understand. :-)
So am I :-)
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
-- Regards,
Laurent Pinchart
Le lundi 16 mars 2020 à 23:15 +0200, Laurent Pinchart a écrit :
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
Right now, there is simply no uAPI for supporting asynchronous errors reporting when fences are invovled. That is true for both camera's and CODEC. It's likely what all the attempt was missing, I don't know enough myself to suggest something.
Now, why Stateless video decoders are special is another subject. In CODECs, the decoding and the presentation order may differ. For Stateless kind of CODEC, a bitstream is passed to the HW. We don't know if this bitstream is fully valid, since the it is being parsed and validated by the firmware. It's also firmware job to decide which buffer should be presented first.
In most firmware interface, that information is communicated back all at once when the frame is ready to be presented (which may be quite some time after it was decoded). So indeed, a fence model is not really easy to add, unless the firmware was designed with that model in mind.
Nothing of course would prevent V4L2 framework to generically handle out_fence from other producers. It does not even handle implicit fences at the moment, which is already quite problematic (I've seen glitches on i.MX6/8 and Raspberry Pi 4).
In that specific case, if the fences from etnaviv, vc graphic drivers was exposed, we could solve this issue in userspace. Right now it's implicit, so we rely on all DMABuf driver to have proper support, which is not the case. There is V4L2 support for that coming, but the wait is done synchronously in userspace call that was normally non-blocking. So that is unlikely to fly.
Small note, stateless video decoders don't have this issue. The bitstream is validated by userspace, and userspace controls the "decode" operation. This one would be a good case for bidirectional fencing.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them. Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Also, from a userspace point of view, the synchronization with the "next frame" in V4L2 isn't implicit. We can poll() the device, just like we'd do with a fence FD. What the explicit fence gives, is a unified object we can pass to another driver, or other userspace, so we can delegate the wait.
You refer to performance in few places. In streaming, this is often measure as real-time throughput. Implicit/explicit fences don't really play any role for us in this regard. V4L2 drivers, like m2m drivers, works with buffer queues. So you can queue in advance many buffers on the OUTPUT device side (which is the input of the m2m), and userspace will queue in advance pretty much all free buffers available on the CAPTURE side. The driver is never starved in that model, at the cost of very large memory consumption of course. Maybe a more visual representation would be:
[pending job] -> [M2M Worker] -> [pending results]
So as long as userspace keep the pending job queue non-empty, and that it consumes and give back buffers back to write the results into, the driver will keep running un-interrupted. Performance remains optimal. What isn't optimal is the latency. And what bugs right now is when a DMAbuf implicit out fence is put back into the pending results queue, since the fence is ignored.
Trying to understand. :-)
So am I :-)
Hehe, same here.
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
On Tue, Mar 17, 2020 at 10:33 AM Nicolas Dufresne nicolas@ndufresne.ca wrote:
Le lundi 16 mars 2020 à 23:15 +0200, Laurent Pinchart a écrit :
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote: > All, > > Sorry for casting such a broad net with this one. I'm sure most people > who reply will get at least one mailing list rejection. However, this > is an issue that affects a LOT of components and that's why it's > thorny to begin with. Please pardon the length of this e-mail as > well; I promise there's a concrete point/proposal at the end. > > > Explicit synchronization is the future of graphics and media. At > least, that seems to be the consensus among all the graphics people > I've talked to. I had a chat with one of the lead Android graphics > engineers recently who told me that doing explicit sync from the start > was one of the best engineering decisions Android ever made. It's > also the direction being taken by more modern APIs such as Vulkan. > > > ## What are implicit and explicit synchronization? > > For those that aren't familiar with this space, GPUs, media encoders, > etc. are massively parallel and synchronization of some form is > required to ensure that everything happens in the right order and > avoid data races. Implicit synchronization is when bits of work (3D, > compute, video encode, etc.) are implicitly based on the absolute > CPU-time order in which API calls occur. Explicit synchronization is > when the client (whatever that means in any given context) provides > the dependency graph explicitly via some sort of synchronization > primitives. If you're still confused, consider the following > examples: > > With OpenGL and EGL, almost everything is implicit sync. Say you have > two OpenGL contexts sharing an image where one writes to it and the > other textures from it. The way the OpenGL spec works, the client has > to make the API calls to render to the image before (in CPU time) it > makes the API calls which texture from the image. As long as it does > this (and maybe inserts a glFlush?), the driver will ensure that the > rendering completes before the texturing happens and you get correct > contents. > > Implicit synchronization can also happen across processes. Wayland, > for instance, is currently built on implicit sync where the client > does their rendering and then does a hand-off (via wl_surface::commit) > to tell the compositor it's done at which point the compositor can now > texture from the surface. The hand-off ensures that the client's > OpenGL API calls happen before the server's OpenGL API calls. > > A good example of explicit synchronization is the Vulkan API. There, > a client (or multiple clients) can simultaneously build command > buffers in different threads where one of those command buffers > renders to an image and the other textures from it and then submit > both of them at the same time with instructions to the driver for > which order to execute them in. The execution order is described via > the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore > extension, you can even submit the work which does the texturing > BEFORE the work which does the rendering and the driver will sort it > out. > > The #1 problem with implicit synchronization (which explicit solves) > is that it leads to a lot of over-synchronization both in client space > and in driver/device space. The client has to synchronize a lot more > because it has to ensure that the API calls happen in a particular > order. The driver/device have to synchronize a lot more because they > never know what is going to end up being a synchronization point as an > API call on another thread/process may occur at any time. As we move > to more and more multi-threaded programming this synchronization (on > the client-side especially) becomes more and more painful. > > > ## Current status in Linux > > Implicit synchronization in Linux works via a the kernel's internal > dma_buf and dma_fence data structures. A dma_fence is a tiny object > which represents the "done" status for some bit of work. Typically, > dma_fences are created as a by-product of someone submitting some bit > of work (say, 3D rendering) to the kernel. The dma_buf object has a > set of dma_fences on it representing shared (read) and exclusive > (write) access to the object. When work is submitted which, for > instance renders to the dma_buf, it's queued waiting on all the fences > on the dma_buf and and a dma_fence is created representing the end of > said rendering work and it's installed as the dma_buf's exclusive > fence. This way, the kernel can manage all its internal queues (3D > rendering, display, video encode, etc.) and know which things to > submit in what order. > > For the last few years, we've had sync_file in the kernel and it's > plumbed into some drivers. A sync_file is just a wrapper around a > single dma_fence. A sync_file is typically created as a by-product of > submitting work (3D, compute, etc.) to the kernel and is signaled when > that work completes. When a sync_file is created, it is guaranteed by > the kernel that it will become signaled in finite time and, once it's > signaled, it remains signaled for the rest of time. A sync_file is > represented in UAPIs as a file descriptor and can be used with normal > file APIs such as dup(). It can be passed into another UAPI which > does some bit of queue'd work and the submitted work will wait for the > sync_file to be triggered before executing. A sync_file also supports > poll() if you want to wait on it manually. > > Unfortunately, sync_file is not broadly used and not all kernel GPU > drivers support it. Here's a very quick overview of my understanding > of the status of various components (I don't know the status of > anything in the media world): > > - Vulkan: Explicit synchronization all the way but we have to go > implicit as soon as we interact with a window-system. Vulkan has APIs > to import/export sync_files to/from it's VkSemaphore and VkFence > synchronization primitives. > - OpenGL: Implicit all the way. There are some EGL extensions to > enable some forms of explicit sync via sync_file but OpenGL itself is > still implicit. > - Wayland: Currently depends on implicit sync in the kernel (accessed > via EGL/OpenGL). There is an unstable extension to allow passing > sync_files around but it's questionable how useful it is right now > (more on that later). > - X11: With present, it has these "explicit" fence objects but > they're always a shmfence which lets the X server and client do a > userspace CPU-side hand-off without going over the socket (and > round-tripping through the kernel). However, the only thing that > fence does is order the OpenGL API calls in the client and server and > the real synchronization is still implicit. > - linux/i915/gem: Fully supports using sync_file or syncobj for explicit > sync. > - linux/amdgpu: Supports sync_file and syncobj but it still > implicitly syncs sometimes due to it's internal memory residency > handling which can lead to over-synchronization. > - KMS: Implicit sync all the way. There are no KMS APIs which take > explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
> - v4l: ??? > - gstreamer: ??? > - Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
Right now, there is simply no uAPI for supporting asynchronous errors reporting when fences are invovled. That is true for both camera's and CODEC. It's likely what all the attempt was missing, I don't know enough myself to suggest something.
Now, why Stateless video decoders are special is another subject. In CODECs, the decoding and the presentation order may differ. For Stateless kind of CODEC, a bitstream is passed to the HW. We don't know if this bitstream is fully valid, since the it is being parsed and validated by the firmware. It's also firmware job to decide which buffer should be presented first.
In most firmware interface, that information is communicated back all at once when the frame is ready to be presented (which may be quite some time after it was decoded). So indeed, a fence model is not really easy to add, unless the firmware was designed with that model in mind.
Just to be clear, I think we should do whatever makes sense here and not try to slam sync_file in when it doesn't make sense just because we have it. The more I read on this thread, the less out-fences from video decode sound like they make sense unless we have a really solid plan for async error reporting. It's possible, depending on how many processes are involved in the pipeline, that async error reporting could help reduce latency a bit if it let the kernel report the error directly to the last process in the chain. However, I'm not convinced the potential for userspace programmer error is worth it.. That said, I'm happy to leave that up to the actual video experts. (I just do 3D)
Nothing of course would prevent V4L2 framework to generically handle out_fence from other producers. It does not even handle implicit fences at the moment, which is already quite problematic (I've seen glitches on i.MX6/8 and Raspberry Pi 4).
In that specific case, if the fences from etnaviv, vc graphic drivers was exposed, we could solve this issue in userspace. Right now it's implicit, so we rely on all DMABuf driver to have proper support, which is not the case. There is V4L2 support for that coming, but the wait is done synchronously in userspace call that was normally non-blocking. So that is unlikely to fly.
Yeah... waits in userspace aren't what anyone wants.
Small note, stateless video decoders don't have this issue. The bitstream is validated by userspace, and userspace controls the "decode" operation. This one would be a good case for bidirectional fencing.
Good to know.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them. Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Also, from a userspace point of view, the synchronization with the "next frame" in V4L2 isn't implicit. We can poll() the device, just like we'd do with a fence FD. What the explicit fence gives, is a unified object we can pass to another driver, or other userspace, so we can delegate the wait.
You refer to performance in few places. In streaming, this is often measure as real-time throughput. Implicit/explicit fences don't really play any role for us in this regard. V4L2 drivers, like m2m drivers, works with buffer queues. So you can queue in advance many buffers on the OUTPUT device side (which is the input of the m2m), and userspace will queue in advance pretty much all free buffers available on the CAPTURE side. The driver is never starved in that model, at the cost of very large memory consumption of course. Maybe a more visual representation would be:
[pending job] -> [M2M Worker] -> [pending results]
So as long as userspace keep the pending job queue non-empty, and that it consumes and give back buffers back to write the results into, the driver will keep running un-interrupted. Performance remains optimal. What isn't optimal is the latency. And what bugs right now is when a DMAbuf implicit out fence is put back into the pending results queue, since the fence is ignored.
Yes, that makes sense. In 3D land, we're very concerned about latency. Any time anyone has to stall for anything, it's a potential hitch in someone's game. Being delayed by a single extra frame can be problematic; 2-3 frames puts the gamer at a significant disadvantage. In video, as long as audio and video are in sync and you aren't dropping frames, no one really cares about latency as long as hitting the pause button doesn't take too long.
What concerns me the most, I think is actually the interop issues. You mentioned issues with the raspberry pi. Right now, if someone is rendering frames using a Vulkan driver and trying to pass those on to V4L for encode or to some other api such as VA-API, we don't really have a plan for synchronization. Thanks to dma-buf extensions we at least have most of a plan for sharing the memory and negotiating image layouts (strides, tiling, etc.) but no plan for synchronization at all. The only thing you can do today is to use a VkFence to CPU wait for the 3D rendering to be 100% done and then pass the image on to the encoder.
The more I look over the various hacks we've done over the course of the last 4 years to make window systems work, the less confident I am that I want to expose ANY of them as an official Vulkan extension that we support long-term. The one we do have which I'm reasonably happy to be stuck with is sync_file import/export. That said, it's sounding like V4L doesn't support dma-buf implicit sync at all so maybe CPU waiting with a VkFence is the current state-of-the-art?
--Jason
Trying to understand. :-)
So am I :-)
Hehe, same here.
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
> ## Chicken and egg problems > > Ok, this is where it starts getting depressing. I made the claim > above that Wayland has an explicit synchronization protocol that's of > questionable usefulness. I would claim that basically any bit of > plumbing we do through window systems is currently of questionable > usefulness. Why? > > From my perspective, as a Vulkan driver developer, I have to deal with > the fact that Vulkan is an explicit sync API but Wayland and X11 > aren't. Unfortunately, the Wayland extension solves zero problems for > me because I can't really use it unless it's implemented in all of the > compositors. Until every Wayland compositor I care about my users > being able to use (which is basically all of them) supports the > extension, I have to continue carry around my pile of hacks to keep > implicit sync and Vulkan working nicely together. > > From the perspective of a Wayland compositor (I used to play in this > space), they'd love to implement the new explicit sync extension but > can't. Sure, they could wire up the extension, but the moment they go > to flip a client buffer to the screen directly, they discover that KMS > doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
> So, yes, they can technically > implement the extension assuming the EGL stack they're running on has > the sync_file extensions but any client buffers which come in using > the explicit sync Wayland extension have to be composited and can't be > scanned out directly. As a 3D driver developer, I absolutely don't > want compositors doing that because my users will complain about > performance issues due to the extra blit. > > Ok, so let's say we get KMS wired up with implicit sync. That solves > all our problems, right? It does, right up until someone decides that > they wan to screen capture their Wayland session via some hardware > media encoder that doesn't support explicit sync. Now we have to > plumb it all the way through the media stack, gstreamer, etc. Great, > so let's do that! Oh, but gstreamer won't want to plumb it through > until they're guaranteed that they can use explicit sync when > displaying on X11 or Wayland. Are you seeing the problem? > > To make matters worse, since most things are doing implicit > synchronization today, it's really easy to get your explicit > synchronization wrong and never notice. If you forget to pass a > sync_file into one place (say you never notice KMS doesn't support > them), it will probably work anyway thanks to all the implicit sync > that's going on elsewhere. > > So, clearly, we all need to go write piles of code that we can't > actually properly test until everyone else has written their piece and > then we use explicit sync if and only if all components support it. > Really? We're going to do multiple years of development and then just > hope it works when we finally flip the switch? That doesn't sound > like a good plan to me. > > > ## A proposal: Implicit and explicit sync together > > How to solve all these chicken-and-egg problems is something I've been > giving quite a bit of thought (and talking with many others about) in > the last couple of years. One motivation for this is that we have to > deal with a mismatch in Vulkan. Another motivation is that I'm > becoming increasingly unhappy with the way that synchronization, > memory residency, and command submission are inherently intertwined in > i915 and would like to break things apart. Towards that end, I have > an actual proposal. > > A couple weeks ago, I sent a series of patches to the dri-devel > mailing list which adds a pair of new ioctls to dma-buf which allow > userspace to manually import or export a sync_file from a dma-buf. > The idea is that something like a Wayland compositor can switch to > 100% explicit sync internally once the ioctl is available. If it gets > buffers in from a client that doesn't use the explicit sync extension, > it can pull a sync_file from the dma-buf and use that exactly as it > would a sync_file passed via the explicit sync extension. When it > goes to scan out a user buffer and discovers that KMS doesn't accept > sync_files (or if it tries to use that pesky media encoder no one has > converted), it can take it's sync_file for display and stuff it into > the dma-buf before handing it to KMS. > > Along with the kernel patches, I've also implemented support for this > in the Vulkan WSI code used by ANV and RADV. With those patches, the > only requirement on the Vulkan drivers is that you be able to export > any VkSemaphore as a sync_file and temporarily import a sync_file into > any VkFence or VkSemaphore. As long as that works, the core Vulkan > driver only ever sees explicit synchronization via sync_file. The WSI > code uses these new ioctls to translate the implicit sync of X11 and > Wayland to the explicit sync the Vulkan driver wants. > > I'm hoping (and here's where I want a sanity check) that a simple API > like this will allow us to finally start moving the Linux ecosystem > over to explicit synchronization one piece at a time in a way that's > actually correct. (No Wayland explicit sync with compositors hoping > KMS magically works even though it doesn't have a sync_file API.) > Once some pieces in the ecosystem start moving, there will be > motivation to start moving others and maybe we can actually build the > momentum to get most everything converted. > > For reference, you can find the kernel RFC patches and mesa MR here: > > https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html > > https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037 > > At this point, I welcome your thoughts, comments, objections, and > maybe even help/review. :-)
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Any ideas?
Jacob Lifshay
On Tue, Mar 17, 2020 at 12:13 PM Jacob Lifshay programmerjake@gmail.com wrote:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Yeah... That's going to be a problem. The only way I could see that working is if you created a sync_file that had a timeout associated with it. However, then you run into the issue where you may have corruption if stuff doesn't complete on time. Then again, you're not really dealing with an external unit and so the latency cost of going across the window system protocol probably isn't massively different from the latency cost of triggering the sync_file. Maybe the answer there is to just do everything in-order and not worry about synchronization?
On Tue, Mar 17, 2020 at 12:18:47PM -0500, Jason Ekstrand wrote:
On Tue, Mar 17, 2020 at 12:13 PM Jacob Lifshay programmerjake@gmail.com wrote:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Yeah... That's going to be a problem. The only way I could see that working is if you created a sync_file that had a timeout associated with it. However, then you run into the issue where you may have corruption if stuff doesn't complete on time. Then again, you're not really dealing with an external unit and so the latency cost of going across the window system protocol probably isn't massively different from the latency cost of triggering the sync_file. Maybe the answer there is to just do everything in-order and not worry about synchronization?
vgem does that already (fences with timeout). The corruption issue is also not new, if your shaders take forever real gpus will nick your rendering with a quick reset. Iirc someone (from cros google team maybe) was even looking into making llvmpipe run on top of vgem as a real dri/drm mesa driver. -Daniel
Am Dienstag, den 17.03.2020, 10:12 -0700 schrieb Jacob Lifshay:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Any ideas?
Finite just means "not infinite". If you stop the process that's doing part of the pipeline processing you block the pipeline, you get to keep the pieces in that case. That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
With regard to the process getting killed: whatever you sync primitive is, you need to make sure to signal the fence (possibly with an error condition set) when you are not going to make progress anymore. So whatever your means to creating the sync_fd from your software renderer is, it needs to signal any outstanding fences on the sync_fd when the fd is closed.
Regards, Lucas
On Tue, Mar 17, 2020 at 10:21 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:12 -0700 schrieb Jacob Lifshay:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Any ideas?
Finite just means "not infinite". If you stop the process that's doing part of the pipeline processing you block the pipeline, you get to keep the pieces in that case.
Seems reasonable.
That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
With regard to the process getting killed: whatever you sync primitive is, you need to make sure to signal the fence (possibly with an error condition set) when you are not going to make progress anymore. So whatever your means to creating the sync_fd from your software renderer is, it needs to signal any outstanding fences on the sync_fd when the fd is closed.
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
Jacob
Regards, Lucas
Am Dienstag, den 17.03.2020, 10:59 -0700 schrieb Jacob Lifshay:
On Tue, Mar 17, 2020 at 10:21 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:12 -0700 schrieb Jacob Lifshay:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
Any ideas?
Finite just means "not infinite". If you stop the process that's doing part of the pipeline processing you block the pipeline, you get to keep the pieces in that case.
Seems reasonable.
That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
With regard to the process getting killed: whatever you sync primitive is, you need to make sure to signal the fence (possibly with an error condition set) when you are not going to make progress anymore. So whatever your means to creating the sync_fd from your software renderer is, it needs to signal any outstanding fences on the sync_fd when the fd is closed.
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
Something like that can certainly be lifted for general use if it makes sense. But then with a software renderer I don't really see how fences help you at all. With a software renderer you know exactly when the frame is finished and you can just defer pushing it over to the next pipeline element until that time. You won't gain any parallelism by using fences as the CPU is busy doing the rendering and will not run other stuff concurrently, right?
Regards, Lucas
On Tue, Mar 17, 2020 at 11:14 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:59 -0700 schrieb Jacob Lifshay:
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
Something like that can certainly be lifted for general use if it makes sense. But then with a software renderer I don't really see how fences help you at all. With a software renderer you know exactly when the frame is finished and you can just defer pushing it over to the next pipeline element until that time. You won't gain any parallelism by using fences as the CPU is busy doing the rendering and will not run other stuff concurrently, right?
There definitely may be other hardware and/or processes that can process some stuff concurrently with the main application, such as the compositor and or video encoding processes (for video capture). Additionally, from what I understand, sync_file is the standard way to export and import explicit synchronization between processes and between drivers on Linux, so it seems like a good idea to support it from an interoperability standpoint even if it turns out that there aren't any scheduling/timing benefits.
Jacob
On Tue, Mar 17, 2020 at 7:16 PM Jacob Lifshay programmerjake@gmail.com wrote:
On Tue, Mar 17, 2020 at 11:14 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:59 -0700 schrieb Jacob Lifshay:
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
It is expressly only for debugging and testing. Exposing such an API to userspace would break the finite time guarantees that are relied upon to keep sync_file a secure API.
Something like that can certainly be lifted for general use if it makes sense. But then with a software renderer I don't really see how fences help you at all. With a software renderer you know exactly when the frame is finished and you can just defer pushing it over to the next pipeline element until that time. You won't gain any parallelism by using fences as the CPU is busy doing the rendering and will not run other stuff concurrently, right?
There definitely may be other hardware and/or processes that can process some stuff concurrently with the main application, such as the compositor and or video encoding processes (for video capture). Additionally, from what I understand, sync_file is the standard way to export and import explicit synchronization between processes and between drivers on Linux, so it seems like a good idea to support it from an interoperability standpoint even if it turns out that there aren't any scheduling/timing benefits.
There are different ways that one can handle interoperability, however. One way is to try and make the software rasterizer look as much like a GPU as possible: lots of threads to make things as asynchronous as possible, "real" implementations of semaphores and fences, etc. Another is to let a SW rasterizer be a SW rasterizer: do everything immediately, thread only so you can exercise all the CPU cores, and minimally implement semaphores and fences well enough to maintain compatibility. If you take the first approach, then we have to solve all these problems with letting userspace create unsignaled sync_files which it will signal later and figure out how to make it safe. If you take the second approach, you'll only ever have to return already signaled sync_files and there's no problem with the sync_file finite time guarantees.
--Jason
On Tue, Mar 17, 2020 at 7:08 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Tue, Mar 17, 2020 at 7:16 PM Jacob Lifshay programmerjake@gmail.com wrote:
On Tue, Mar 17, 2020 at 11:14 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:59 -0700 schrieb Jacob Lifshay:
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
It is expressly only for debugging and testing. Exposing such an API to userspace would break the finite time guarantees that are relied upon to keep sync_file a secure API.
Ok, I was figuring that was probably the case.
Something like that can certainly be lifted for general use if it makes sense. But then with a software renderer I don't really see how fences help you at all. With a software renderer you know exactly when the frame is finished and you can just defer pushing it over to the next pipeline element until that time. You won't gain any parallelism by using fences as the CPU is busy doing the rendering and will not run other stuff concurrently, right?
There definitely may be other hardware and/or processes that can process some stuff concurrently with the main application, such as the compositor and or video encoding processes (for video capture). Additionally, from what I understand, sync_file is the standard way to export and import explicit synchronization between processes and between drivers on Linux, so it seems like a good idea to support it from an interoperability standpoint even if it turns out that there aren't any scheduling/timing benefits.
There are different ways that one can handle interoperability, however. One way is to try and make the software rasterizer look as much like a GPU as possible: lots of threads to make things as asynchronous as possible, "real" implementations of semaphores and fences, etc.
This is basically the route I've picked, though rather than making lots of native threads, I'm planning on having just one thread per core and have a work-stealing scheduler (inspired by Rust's rayon crate) schedule all the individual render/compute jobs, because that allows making a lot more jobs to allow finer load balancing.
Another is to let a SW rasterizer be a SW rasterizer: do everything immediately, thread only so you can exercise all the CPU cores, and minimally implement semaphores and fences well enough to maintain compatibility. If you take the first approach, then we have to solve all these problems with letting userspace create unsignaled sync_files which it will signal later and figure out how to make it safe. If you take the second approach, you'll only ever have to return already signaled sync_files and there's no problem with the sync_file finite time guarantees.
The main issue with doing everything immediately is that a lot of the function calls that games expect to take a very short time (e.g. vkQueueSubmit) would instead take a much longer time, potentially causing problems.
One idea for a safe userspace-backed sync_file is to have a step counter that counts down until the sync_file is ready, where if userspace doesn't tell it to count any steps in a certain amount of time, then the sync_file switches to the error state. This way, it will error shortly after a process deadlocks for some reason, while still having the finite-time guarantee.
When the sync_file is created, the step counter would be set to the number of jobs that the fence is waiting on.
It can also be set to pause the timeout to wait until another sync_file signals, to handle cases where a sync_file is waiting on a userspace process that is waiting on another sync_file.
The main issue is that the kernel would have to make sure that the sync_file graph doesn't have loops, maybe by erroring all sync_files that it finds in the loop.
Does that sound like a good idea?
Jacob
On Wed, Mar 18, 2020 at 12:20 AM Jacob Lifshay programmerjake@gmail.com wrote:
On Tue, Mar 17, 2020 at 7:08 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Tue, Mar 17, 2020 at 7:16 PM Jacob Lifshay programmerjake@gmail.com wrote:
On Tue, Mar 17, 2020 at 11:14 AM Lucas Stach dev@lynxeye.de wrote:
Am Dienstag, den 17.03.2020, 10:59 -0700 schrieb Jacob Lifshay:
I think I found a userspace-accessible way to create sync_files and dma_fences that would fulfill the requirements: https://github.com/torvalds/linux/blob/master/drivers/dma-buf/sw_sync.c
I'm just not sure if that's a good interface to use, since it appears to be designed only for debugging. Will have to check for additional requirements of signalling an error when the process that created the fence is killed.
It is expressly only for debugging and testing. Exposing such an API to userspace would break the finite time guarantees that are relied upon to keep sync_file a secure API.
Ok, I was figuring that was probably the case.
Something like that can certainly be lifted for general use if it makes sense. But then with a software renderer I don't really see how fences help you at all. With a software renderer you know exactly when the frame is finished and you can just defer pushing it over to the next pipeline element until that time. You won't gain any parallelism by using fences as the CPU is busy doing the rendering and will not run other stuff concurrently, right?
There definitely may be other hardware and/or processes that can process some stuff concurrently with the main application, such as the compositor and or video encoding processes (for video capture). Additionally, from what I understand, sync_file is the standard way to export and import explicit synchronization between processes and between drivers on Linux, so it seems like a good idea to support it from an interoperability standpoint even if it turns out that there aren't any scheduling/timing benefits.
There are different ways that one can handle interoperability, however. One way is to try and make the software rasterizer look as much like a GPU as possible: lots of threads to make things as asynchronous as possible, "real" implementations of semaphores and fences, etc.
This is basically the route I've picked, though rather than making lots of native threads, I'm planning on having just one thread per core and have a work-stealing scheduler (inspired by Rust's rayon crate) schedule all the individual render/compute jobs, because that allows making a lot more jobs to allow finer load balancing.
Another is to let a SW rasterizer be a SW rasterizer: do everything immediately, thread only so you can exercise all the CPU cores, and minimally implement semaphores and fences well enough to maintain compatibility. If you take the first approach, then we have to solve all these problems with letting userspace create unsignaled sync_files which it will signal later and figure out how to make it safe. If you take the second approach, you'll only ever have to return already signaled sync_files and there's no problem with the sync_file finite time guarantees.
The main issue with doing everything immediately is that a lot of the function calls that games expect to take a very short time (e.g. vkQueueSubmit) would instead take a much longer time, potentially causing problems.
Do you have any evidence that it will cause problems? What I said above is what switfshader is doing and they're running real apps and I've not heard of it causing any problems. It's also worth noting that you would only really have to stall at sync_file export. You can async as much as you want internally.
One idea for a safe userspace-backed sync_file is to have a step counter that counts down until the sync_file is ready, where if userspace doesn't tell it to count any steps in a certain amount of time, then the sync_file switches to the error state. This way, it will error shortly after a process deadlocks for some reason, while still having the finite-time guarantee.
When the sync_file is created, the step counter would be set to the number of jobs that the fence is waiting on.
It can also be set to pause the timeout to wait until another sync_file signals, to handle cases where a sync_file is waiting on a userspace process that is waiting on another sync_file.
The main issue is that the kernel would have to make sure that the sync_file graph doesn't have loops, maybe by erroring all sync_files that it finds in the loop.
Does that sound like a good idea?
Honestly, I don't think you'll ever be able to sell that to the kernel community. All of the deadlock detection would add massive complexity to the already non-trivial dma_fence infrastructure and for what benefit? So that a software rasterizer can try to pretend to be more like a GPU? You're going to have some very serious perf numbers and/or other proof of necessity if you want to convince the kernel to people to accept that level of complexity/risk. "I designed my software to work this way" isn't going to convince anyone of anything especially when literally every other software rasterizer I'm aware of is immediate and they work just fine.
--Jason
On Tue, Mar 17, 2020 at 11:35 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Wed, Mar 18, 2020 at 12:20 AM Jacob Lifshay programmerjake@gmail.com wrote:
The main issue with doing everything immediately is that a lot of the function calls that games expect to take a very short time (e.g. vkQueueSubmit) would instead take a much longer time, potentially causing problems.
Do you have any evidence that it will cause problems? What I said above is what switfshader is doing and they're running real apps and I've not heard of it causing any problems. It's also worth noting that you would only really have to stall at sync_file export. You can async as much as you want internally.
Ok, seems worth trying out.
One idea for a safe userspace-backed sync_file is to have a step counter that counts down until the sync_file is ready, where if userspace doesn't tell it to count any steps in a certain amount of time, then the sync_file switches to the error state. This way, it will error shortly after a process deadlocks for some reason, while still having the finite-time guarantee.
When the sync_file is created, the step counter would be set to the number of jobs that the fence is waiting on.
It can also be set to pause the timeout to wait until another sync_file signals, to handle cases where a sync_file is waiting on a userspace process that is waiting on another sync_file.
The main issue is that the kernel would have to make sure that the sync_file graph doesn't have loops, maybe by erroring all sync_files that it finds in the loop.
Does that sound like a good idea?
Honestly, I don't think you'll ever be able to sell that to the kernel community. All of the deadlock detection would add massive complexity to the already non-trivial dma_fence infrastructure and for what benefit? So that a software rasterizer can try to pretend to be more like a GPU? You're going to have some very serious perf numbers and/or other proof of necessity if you want to convince the kernel to people to accept that level of complexity/risk. "I designed my software to work this way" isn't going to convince anyone of anything especially when literally every other software rasterizer I'm aware of is immediate and they work just fine.
After some further research, it turns out that it will work to have all the sync_files that a sync_file needs to depend on specified at creation, which forces the dependence graph to be a DAG since you can't depend on a sync_file that isn't yet created, so loops are impossible by design.
Since kernel deadlock detection isn't actually required, just timeouts for the case of halted userspace, does this seem feasable?
I'd guess that it'd require maybe 200-300 lines of code in a self-contained driver similar to the sync_file debugging driver mentioned previously but with the additional timeout code for safety.
Jacob
On 2020-03-17 6:21 p.m., Lucas Stach wrote:
That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
FWIW, the compositor can do this with implicit sync as well, by polling a dma-buf fd for the buffer. (Currently, it has to poll for writable, because waiting for the exclusive fence only isn't enough with amdgpu)
Le mercredi 18 mars 2020 à 11:05 +0100, Michel Dänzer a écrit :
On 2020-03-17 6:21 p.m., Lucas Stach wrote:
That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
FWIW, the compositor can do this with implicit sync as well, by polling a dma-buf fd for the buffer. (Currently, it has to poll for writable, because waiting for the exclusive fence only isn't enough with amdgpu)
That is very interesting, thanks for sharing, could allow fixing some issues in userspace for backward compatibility.
thanks, Nicolas
On Wed, Mar 18, 2020 at 11:05:48AM +0100, Michel Dänzer wrote:
On 2020-03-17 6:21 p.m., Lucas Stach wrote:
That's one of the issues with implicit sync that explicit may solve: a single client taking way too much time to render something can block the whole pipeline up until the display flip. With explicit sync the compositor can just decide to use the last client buffer if the latest buffer isn't ready by some deadline.
FWIW, the compositor can do this with implicit sync as well, by polling a dma-buf fd for the buffer. (Currently, it has to poll for writable, because waiting for the exclusive fence only isn't enough with amdgpu)
Would be great if we don't have to make this recommended uapi, just because amdgpu leaks it's trickery into the wider world. Polling for read really should be enough (and I guess Christian gets to fix up amdgpu more, at least for anything that has a dma-buf attached even if it's not shared with anything !amdgpu.ko). -Daniel
On Tue, 2020-03-17 at 10:12 -0700, Jacob Lifshay wrote:
One related issue with explicit sync using sync_file is that combined CPUs/GPUs (the CPU cores *are* the GPU cores) that do all the rendering in userspace (like llvmpipe but for Vulkan and with extra instructions for GPU tasks) but need to synchronize with other drivers/processes is that there should be some way to create an explicit fence/semaphore from userspace and later signal it. This seems to conflict with the requirement for a sync_file to complete in finite time, since the user process could be stopped or killed.
DRI3 (okay, libxshmfence specifically) uses futexes for this. Would that work for you? IIRC the semantics there are that if the process dies the futex is treated as triggered, which seems like the only sensible thing to do.
- ajax
Le mardi 17 mars 2020 à 11:27 -0500, Jason Ekstrand a écrit :
On Tue, Mar 17, 2020 at 10:33 AM Nicolas Dufresne nicolas@ndufresne.ca wrote:
Le lundi 16 mars 2020 à 23:15 +0200, Laurent Pinchart a écrit :
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit : > On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote: > > All, > > > > Sorry for casting such a broad net with this one. I'm sure most people > > who reply will get at least one mailing list rejection. However, this > > is an issue that affects a LOT of components and that's why it's > > thorny to begin with. Please pardon the length of this e-mail as > > well; I promise there's a concrete point/proposal at the end. > > > > > > Explicit synchronization is the future of graphics and media. At > > least, that seems to be the consensus among all the graphics people > > I've talked to. I had a chat with one of the lead Android graphics > > engineers recently who told me that doing explicit sync from the start > > was one of the best engineering decisions Android ever made. It's > > also the direction being taken by more modern APIs such as Vulkan. > > > > > > ## What are implicit and explicit synchronization? > > > > For those that aren't familiar with this space, GPUs, media encoders, > > etc. are massively parallel and synchronization of some form is > > required to ensure that everything happens in the right order and > > avoid data races. Implicit synchronization is when bits of work (3D, > > compute, video encode, etc.) are implicitly based on the absolute > > CPU-time order in which API calls occur. Explicit synchronization is > > when the client (whatever that means in any given context) provides > > the dependency graph explicitly via some sort of synchronization > > primitives. If you're still confused, consider the following > > examples: > > > > With OpenGL and EGL, almost everything is implicit sync. Say you have > > two OpenGL contexts sharing an image where one writes to it and the > > other textures from it. The way the OpenGL spec works, the client has > > to make the API calls to render to the image before (in CPU time) it > > makes the API calls which texture from the image. As long as it does > > this (and maybe inserts a glFlush?), the driver will ensure that the > > rendering completes before the texturing happens and you get correct > > contents. > > > > Implicit synchronization can also happen across processes. Wayland, > > for instance, is currently built on implicit sync where the client > > does their rendering and then does a hand-off (via wl_surface::commit) > > to tell the compositor it's done at which point the compositor can now > > texture from the surface. The hand-off ensures that the client's > > OpenGL API calls happen before the server's OpenGL API calls. > > > > A good example of explicit synchronization is the Vulkan API. There, > > a client (or multiple clients) can simultaneously build command > > buffers in different threads where one of those command buffers > > renders to an image and the other textures from it and then submit > > both of them at the same time with instructions to the driver for > > which order to execute them in. The execution order is described via > > the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore > > extension, you can even submit the work which does the texturing > > BEFORE the work which does the rendering and the driver will sort it > > out. > > > > The #1 problem with implicit synchronization (which explicit solves) > > is that it leads to a lot of over-synchronization both in client space > > and in driver/device space. The client has to synchronize a lot more > > because it has to ensure that the API calls happen in a particular > > order. The driver/device have to synchronize a lot more because they > > never know what is going to end up being a synchronization point as an > > API call on another thread/process may occur at any time. As we move > > to more and more multi-threaded programming this synchronization (on > > the client-side especially) becomes more and more painful. > > > > > > ## Current status in Linux > > > > Implicit synchronization in Linux works via a the kernel's internal > > dma_buf and dma_fence data structures. A dma_fence is a tiny object > > which represents the "done" status for some bit of work. Typically, > > dma_fences are created as a by-product of someone submitting some bit > > of work (say, 3D rendering) to the kernel. The dma_buf object has a > > set of dma_fences on it representing shared (read) and exclusive > > (write) access to the object. When work is submitted which, for > > instance renders to the dma_buf, it's queued waiting on all the fences > > on the dma_buf and and a dma_fence is created representing the end of > > said rendering work and it's installed as the dma_buf's exclusive > > fence. This way, the kernel can manage all its internal queues (3D > > rendering, display, video encode, etc.) and know which things to > > submit in what order. > > > > For the last few years, we've had sync_file in the kernel and it's > > plumbed into some drivers. A sync_file is just a wrapper around a > > single dma_fence. A sync_file is typically created as a by-product of > > submitting work (3D, compute, etc.) to the kernel and is signaled when > > that work completes. When a sync_file is created, it is guaranteed by > > the kernel that it will become signaled in finite time and, once it's > > signaled, it remains signaled for the rest of time. A sync_file is > > represented in UAPIs as a file descriptor and can be used with normal > > file APIs such as dup(). It can be passed into another UAPI which > > does some bit of queue'd work and the submitted work will wait for the > > sync_file to be triggered before executing. A sync_file also supports > > poll() if you want to wait on it manually. > > > > Unfortunately, sync_file is not broadly used and not all kernel GPU > > drivers support it. Here's a very quick overview of my understanding > > of the status of various components (I don't know the status of > > anything in the media world): > > > > - Vulkan: Explicit synchronization all the way but we have to go > > implicit as soon as we interact with a window-system. Vulkan has APIs > > to import/export sync_files to/from it's VkSemaphore and VkFence > > synchronization primitives. > > - OpenGL: Implicit all the way. There are some EGL extensions to > > enable some forms of explicit sync via sync_file but OpenGL itself is > > still implicit. > > - Wayland: Currently depends on implicit sync in the kernel (accessed > > via EGL/OpenGL). There is an unstable extension to allow passing > > sync_files around but it's questionable how useful it is right now > > (more on that later). > > - X11: With present, it has these "explicit" fence objects but > > they're always a shmfence which lets the X server and client do a > > userspace CPU-side hand-off without going over the socket (and > > round-tripping through the kernel). However, the only thing that > > fence does is order the OpenGL API calls in the client and server and > > the real synchronization is still implicit. > > - linux/i915/gem: Fully supports using sync_file or syncobj for explicit > > sync. > > - linux/amdgpu: Supports sync_file and syncobj but it still > > implicitly syncs sometimes due to it's internal memory residency > > handling which can lead to over-synchronization. > > - KMS: Implicit sync all the way. There are no KMS APIs which take > > explicit sync primitives. > > Correction: Apparently, I missed some things. If you use atomic, KMS > does have explicit in- and out-fences. Non-atomic users (e.g. X11) > are still in trouble but most Wayland compositors use atomic these > days > > > - v4l: ??? > > - gstreamer: ??? > > - Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
Right now, there is simply no uAPI for supporting asynchronous errors reporting when fences are invovled. That is true for both camera's and CODEC. It's likely what all the attempt was missing, I don't know enough myself to suggest something.
Now, why Stateless video decoders are special is another subject. In CODECs, the decoding and the presentation order may differ. For Stateless kind of CODEC, a bitstream is passed to the HW. We don't know if this bitstream is fully valid, since the it is being parsed and validated by the firmware. It's also firmware job to decide which buffer should be presented first.
In most firmware interface, that information is communicated back all at once when the frame is ready to be presented (which may be quite some time after it was decoded). So indeed, a fence model is not really easy to add, unless the firmware was designed with that model in mind.
Just to be clear, I think we should do whatever makes sense here and not try to slam sync_file in when it doesn't make sense just because we have it. The more I read on this thread, the less out-fences from video decode sound like they make sense unless we have a really solid plan for async error reporting. It's possible, depending on how many processes are involved in the pipeline, that async error reporting could help reduce latency a bit if it let the kernel report the error directly to the last process in the chain. However, I'm not convinced the potential for userspace programmer error is worth it.. That said, I'm happy to leave that up to the actual video experts. (I just do 3D)
Nothing of course would prevent V4L2 framework to generically handle out_fence from other producers. It does not even handle implicit fences at the moment, which is already quite problematic (I've seen glitches on i.MX6/8 and Raspberry Pi 4).
In that specific case, if the fences from etnaviv, vc graphic drivers was exposed, we could solve this issue in userspace. Right now it's implicit, so we rely on all DMABuf driver to have proper support, which is not the case. There is V4L2 support for that coming, but the wait is done synchronously in userspace call that was normally non-blocking. So that is unlikely to fly.
Yeah... waits in userspace aren't what anyone wants.
Small note, stateless video decoders don't have this issue. The bitstream is validated by userspace, and userspace controls the "decode" operation. This one would be a good case for bidirectional fencing.
Good to know.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them. Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Also, from a userspace point of view, the synchronization with the "next frame" in V4L2 isn't implicit. We can poll() the device, just like we'd do with a fence FD. What the explicit fence gives, is a unified object we can pass to another driver, or other userspace, so we can delegate the wait.
You refer to performance in few places. In streaming, this is often measure as real-time throughput. Implicit/explicit fences don't really play any role for us in this regard. V4L2 drivers, like m2m drivers, works with buffer queues. So you can queue in advance many buffers on the OUTPUT device side (which is the input of the m2m), and userspace will queue in advance pretty much all free buffers available on the CAPTURE side. The driver is never starved in that model, at the cost of very large memory consumption of course. Maybe a more visual representation would be:
[pending job] -> [M2M Worker] -> [pending results]
So as long as userspace keep the pending job queue non-empty, and that it consumes and give back buffers back to write the results into, the driver will keep running un-interrupted. Performance remains optimal. What isn't optimal is the latency. And what bugs right now is when a DMAbuf implicit out fence is put back into the pending results queue, since the fence is ignored.
Yes, that makes sense. In 3D land, we're very concerned about latency. Any time anyone has to stall for anything, it's a potential hitch in someone's game. Being delayed by a single extra frame can be problematic; 2-3 frames puts the gamer at a significant disadvantage. In video, as long as audio and video are in sync and you aren't dropping frames, no one really cares about latency as long as hitting the pause button doesn't take too long.
Just a note, there exist low latency use cases for streaming too (sub- frame latency between two devices). But everything I'm ware is downstream. The one I have in mind uses a special AXI feature to synchronize between two HW component, but the implementation is not using either implicit or explicit fence, in fact they didn't bother adding a specific kernel object, you have to know when you use these downstream drivers. We are a bit far from being able to make generic software on top of that.
The use case was less prone to capture error, since instead of a camera, they have SDI or HDMI receiver.
What concerns me the most, I think is actually the interop issues. You mentioned issues with the raspberry pi. Right now, if someone is rendering frames using a Vulkan driver and trying to pass those on to V4L for encode or to some other api such as VA-API, we don't really have a plan for synchronization. Thanks to dma-buf extensions we at least have most of a plan for sharing the memory and negotiating image layouts (strides, tiling, etc.) but no plan for synchronization at
I didn't know there was plan for that, this is nice. Right now every userspace carry this information in a slightly different and incompatible way, translating, extrapolation, etc. It's all very error prone.
all. The only thing you can do today is to use a VkFence to CPU wait for the 3D rendering to be 100% done and then pass the image on to the encoder.
The more I look over the various hacks we've done over the course of the last 4 years to make window systems work, the less confident I am that I want to expose ANY of them as an official Vulkan extension that we support long-term. The one we do have which I'm reasonably happy to be stuck with is sync_file import/export. That said, it's sounding like V4L doesn't support dma-buf implicit sync at all so maybe CPU waiting with a VkFence is the current state-of-the-art?
--Jason
Trying to understand. :-)
So am I :-)
Hehe, same here.
There is other elements that would implement fencing, notably kmssink, but no one actually dared porting it to atomic KMS, so clearly there is very little comunity interest. glimagsink could clearly benifit. Right now if we import a DMABuf, and that this DMAbuf is used for render, a implicit fence is attached, which we are unaware. Philippe Zabbel is working on a patch, so V4L2 QBUF would wait, but waiting in QBUF is not allowed if O_NONBLOCK was set (which GStreamer uses), so then the operation will just fail where it worked before (breaking userspace). If it was an explcit fence, we could handle that in GStreamer cleanly as we do for new APIs.
> > ## Chicken and egg problems > > > > Ok, this is where it starts getting depressing. I made the claim > > above that Wayland has an explicit synchronization protocol that's of > > questionable usefulness. I would claim that basically any bit of > > plumbing we do through window systems is currently of questionable > > usefulness. Why? > > > > From my perspective, as a Vulkan driver developer, I have to deal with > > the fact that Vulkan is an explicit sync API but Wayland and X11 > > aren't. Unfortunately, the Wayland extension solves zero problems for > > me because I can't really use it unless it's implemented in all of the > > compositors. Until every Wayland compositor I care about my users > > being able to use (which is basically all of them) supports the > > extension, I have to continue carry around my pile of hacks to keep > > implicit sync and Vulkan working nicely together. > > > > From the perspective of a Wayland compositor (I used to play in this > > space), they'd love to implement the new explicit sync extension but > > can't. Sure, they could wire up the extension, but the moment they go > > to flip a client buffer to the screen directly, they discover that KMS > > doesn't support any explicit sync APIs. > > As per the above correction, Wayland compositors aren't nearly as bad > off as I initially thought. There may still be weird screen capture > cases but the normal cases of compositing and displaying via > KMS/atomic should be in reasonably good shape. > > > So, yes, they can technically > > implement the extension assuming the EGL stack they're running on has > > the sync_file extensions but any client buffers which come in using > > the explicit sync Wayland extension have to be composited and can't be > > scanned out directly. As a 3D driver developer, I absolutely don't > > want compositors doing that because my users will complain about > > performance issues due to the extra blit. > > > > Ok, so let's say we get KMS wired up with implicit sync. That solves > > all our problems, right? It does, right up until someone decides that > > they wan to screen capture their Wayland session via some hardware > > media encoder that doesn't support explicit sync. Now we have to > > plumb it all the way through the media stack, gstreamer, etc. Great, > > so let's do that! Oh, but gstreamer won't want to plumb it through > > until they're guaranteed that they can use explicit sync when > > displaying on X11 or Wayland. Are you seeing the problem? > > > > To make matters worse, since most things are doing implicit > > synchronization today, it's really easy to get your explicit > > synchronization wrong and never notice. If you forget to pass a > > sync_file into one place (say you never notice KMS doesn't support > > them), it will probably work anyway thanks to all the implicit sync > > that's going on elsewhere. > > > > So, clearly, we all need to go write piles of code that we can't > > actually properly test until everyone else has written their piece and > > then we use explicit sync if and only if all components support it. > > Really? We're going to do multiple years of development and then just > > hope it works when we finally flip the switch? That doesn't sound > > like a good plan to me. > > > > > > ## A proposal: Implicit and explicit sync together > > > > How to solve all these chicken-and-egg problems is something I've been > > giving quite a bit of thought (and talking with many others about) in > > the last couple of years. One motivation for this is that we have to > > deal with a mismatch in Vulkan. Another motivation is that I'm > > becoming increasingly unhappy with the way that synchronization, > > memory residency, and command submission are inherently intertwined in > > i915 and would like to break things apart. Towards that end, I have > > an actual proposal. > > > > A couple weeks ago, I sent a series of patches to the dri-devel > > mailing list which adds a pair of new ioctls to dma-buf which allow > > userspace to manually import or export a sync_file from a dma-buf. > > The idea is that something like a Wayland compositor can switch to > > 100% explicit sync internally once the ioctl is available. If it gets > > buffers in from a client that doesn't use the explicit sync extension, > > it can pull a sync_file from the dma-buf and use that exactly as it > > would a sync_file passed via the explicit sync extension. When it > > goes to scan out a user buffer and discovers that KMS doesn't accept > > sync_files (or if it tries to use that pesky media encoder no one has > > converted), it can take it's sync_file for display and stuff it into > > the dma-buf before handing it to KMS. > > > > Along with the kernel patches, I've also implemented support for this > > in the Vulkan WSI code used by ANV and RADV. With those patches, the > > only requirement on the Vulkan drivers is that you be able to export > > any VkSemaphore as a sync_file and temporarily import a sync_file into > > any VkFence or VkSemaphore. As long as that works, the core Vulkan > > driver only ever sees explicit synchronization via sync_file. The WSI > > code uses these new ioctls to translate the implicit sync of X11 and > > Wayland to the explicit sync the Vulkan driver wants. > > > > I'm hoping (and here's where I want a sanity check) that a simple API > > like this will allow us to finally start moving the Linux ecosystem > > over to explicit synchronization one piece at a time in a way that's > > actually correct. (No Wayland explicit sync with compositors hoping > > KMS magically works even though it doesn't have a sync_file API.) > > Once some pieces in the ecosystem start moving, there will be > > motivation to start moving others and maybe we can actually build the > > momentum to get most everything converted. > > > > For reference, you can find the kernel RFC patches and mesa MR here: > > > > https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html > > > > https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037 > > > > At this point, I welcome your thoughts, comments, objections, and > > maybe even help/review. :-)
On Tue, Mar 17, 2020 at 11:27:28AM -0500, Jason Ekstrand wrote:
On Tue, Mar 17, 2020 at 10:33 AM Nicolas Dufresne nicolas@ndufresne.ca wrote:
Le lundi 16 mars 2020 à 23:15 +0200, Laurent Pinchart a écrit :
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
Right now, there is simply no uAPI for supporting asynchronous errors reporting when fences are invovled. That is true for both camera's and CODEC. It's likely what all the attempt was missing, I don't know enough myself to suggest something.
Now, why Stateless video decoders are special is another subject. In CODECs, the decoding and the presentation order may differ. For Stateless kind of CODEC, a bitstream is passed to the HW. We don't know if this bitstream is fully valid, since the it is being parsed and validated by the firmware. It's also firmware job to decide which buffer should be presented first.
In most firmware interface, that information is communicated back all at once when the frame is ready to be presented (which may be quite some time after it was decoded). So indeed, a fence model is not really easy to add, unless the firmware was designed with that model in mind.
Just to be clear, I think we should do whatever makes sense here and not try to slam sync_file in when it doesn't make sense just because we have it. The more I read on this thread, the less out-fences from video decode sound like they make sense unless we have a really solid plan for async error reporting. It's possible, depending on how many processes are involved in the pipeline, that async error reporting could help reduce latency a bit if it let the kernel report the error directly to the last process in the chain. However, I'm not convinced the potential for userspace programmer error is worth it.. That said, I'm happy to leave that up to the actual video experts. (I just do 3D)
dma_fence has an error state which you can set when things went south. The fence still completes (to guarantee forward progress).
Currently that error code isn't really propagated anywhere (well i915 iirc does something like that since it tracks the depedencies internally in the scheduler). Definitely not at the dma_fence level, since we don't track the dependency graph there at all. We might want to add that, would at least be possible.
If we track the cascading dma_fence error state in the kernel I do think this could work. I'm not sure whether it's actually a good/useful idea still. -Daniel
Am Dienstag, den 17.03.2020, 11:33 -0400 schrieb Nicolas Dufresne:
Le lundi 16 mars 2020 à 23:15 +0200, Laurent Pinchart a écrit :
Hi Jason,
On Mon, Mar 16, 2020 at 10:06:07AM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 5:20 AM Laurent Pinchart wrote:
On Wed, Mar 11, 2020 at 04:18:55PM -0400, Nicolas Dufresne wrote:
(I know I'm going to be spammed by so many mailing list ...)
Le mercredi 11 mars 2020 à 14:21 -0500, Jason Ekstrand a écrit :
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote: > All, > > Sorry for casting such a broad net with this one. I'm sure most people > who reply will get at least one mailing list rejection. However, this > is an issue that affects a LOT of components and that's why it's > thorny to begin with. Please pardon the length of this e-mail as > well; I promise there's a concrete point/proposal at the end. > > > Explicit synchronization is the future of graphics and media. At > least, that seems to be the consensus among all the graphics people > I've talked to. I had a chat with one of the lead Android graphics > engineers recently who told me that doing explicit sync from the start > was one of the best engineering decisions Android ever made. It's > also the direction being taken by more modern APIs such as Vulkan. > > > ## What are implicit and explicit synchronization? > > For those that aren't familiar with this space, GPUs, media encoders, > etc. are massively parallel and synchronization of some form is > required to ensure that everything happens in the right order and > avoid data races. Implicit synchronization is when bits of work (3D, > compute, video encode, etc.) are implicitly based on the absolute > CPU-time order in which API calls occur. Explicit synchronization is > when the client (whatever that means in any given context) provides > the dependency graph explicitly via some sort of synchronization > primitives. If you're still confused, consider the following > examples: > > With OpenGL and EGL, almost everything is implicit sync. Say you have > two OpenGL contexts sharing an image where one writes to it and the > other textures from it. The way the OpenGL spec works, the client has > to make the API calls to render to the image before (in CPU time) it > makes the API calls which texture from the image. As long as it does > this (and maybe inserts a glFlush?), the driver will ensure that the > rendering completes before the texturing happens and you get correct > contents. > > Implicit synchronization can also happen across processes. Wayland, > for instance, is currently built on implicit sync where the client > does their rendering and then does a hand-off (via wl_surface::commit) > to tell the compositor it's done at which point the compositor can now > texture from the surface. The hand-off ensures that the client's > OpenGL API calls happen before the server's OpenGL API calls. > > A good example of explicit synchronization is the Vulkan API. There, > a client (or multiple clients) can simultaneously build command > buffers in different threads where one of those command buffers > renders to an image and the other textures from it and then submit > both of them at the same time with instructions to the driver for > which order to execute them in. The execution order is described via > the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore > extension, you can even submit the work which does the texturing > BEFORE the work which does the rendering and the driver will sort it > out. > > The #1 problem with implicit synchronization (which explicit solves) > is that it leads to a lot of over-synchronization both in client space > and in driver/device space. The client has to synchronize a lot more > because it has to ensure that the API calls happen in a particular > order. The driver/device have to synchronize a lot more because they > never know what is going to end up being a synchronization point as an > API call on another thread/process may occur at any time. As we move > to more and more multi-threaded programming this synchronization (on > the client-side especially) becomes more and more painful. > > > ## Current status in Linux > > Implicit synchronization in Linux works via a the kernel's internal > dma_buf and dma_fence data structures. A dma_fence is a tiny object > which represents the "done" status for some bit of work. Typically, > dma_fences are created as a by-product of someone submitting some bit > of work (say, 3D rendering) to the kernel. The dma_buf object has a > set of dma_fences on it representing shared (read) and exclusive > (write) access to the object. When work is submitted which, for > instance renders to the dma_buf, it's queued waiting on all the fences > on the dma_buf and and a dma_fence is created representing the end of > said rendering work and it's installed as the dma_buf's exclusive > fence. This way, the kernel can manage all its internal queues (3D > rendering, display, video encode, etc.) and know which things to > submit in what order. > > For the last few years, we've had sync_file in the kernel and it's > plumbed into some drivers. A sync_file is just a wrapper around a > single dma_fence. A sync_file is typically created as a by-product of > submitting work (3D, compute, etc.) to the kernel and is signaled when > that work completes. When a sync_file is created, it is guaranteed by > the kernel that it will become signaled in finite time and, once it's > signaled, it remains signaled for the rest of time. A sync_file is > represented in UAPIs as a file descriptor and can be used with normal > file APIs such as dup(). It can be passed into another UAPI which > does some bit of queue'd work and the submitted work will wait for the > sync_file to be triggered before executing. A sync_file also supports > poll() if you want to wait on it manually. > > Unfortunately, sync_file is not broadly used and not all kernel GPU > drivers support it. Here's a very quick overview of my understanding > of the status of various components (I don't know the status of > anything in the media world): > > - Vulkan: Explicit synchronization all the way but we have to go > implicit as soon as we interact with a window-system. Vulkan has APIs > to import/export sync_files to/from it's VkSemaphore and VkFence > synchronization primitives. > - OpenGL: Implicit all the way. There are some EGL extensions to > enable some forms of explicit sync via sync_file but OpenGL itself is > still implicit. > - Wayland: Currently depends on implicit sync in the kernel (accessed > via EGL/OpenGL). There is an unstable extension to allow passing > sync_files around but it's questionable how useful it is right now > (more on that later). > - X11: With present, it has these "explicit" fence objects but > they're always a shmfence which lets the X server and client do a > userspace CPU-side hand-off without going over the socket (and > round-tripping through the kernel). However, the only thing that > fence does is order the OpenGL API calls in the client and server and > the real synchronization is still implicit. > - linux/i915/gem: Fully supports using sync_file or syncobj for explicit > sync. > - linux/amdgpu: Supports sync_file and syncobj but it still > implicitly syncs sometimes due to it's internal memory residency > handling which can lead to over-synchronization. > - KMS: Implicit sync all the way. There are no KMS APIs which take > explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
> - v4l: ??? > - gstreamer: ??? > - Media APIs such as vaapi etc.: ???
GStreamer is consumer for V4L2, VAAPI and other stuff. Using asynchronous buffer synchronisation is something we do already with GL (even if limited). We place GLSync object in the pipeline and attach that on related GstBuffer. We wait on these GLSync as late as possible (or superseed the sync if we queue more work into the same GL context). That requires a special mode of operation of course. We don't usually like making lazy blocking call implicit, as it tends to cause random issues. If we need to wait, we think it's better to wait int he module that is responsible, so in general, we try to negotiate and fallback locally (it's plugin base, so this can be really messy otherwise).
So basically this problem needs to be solved in V4L2, VAAPI and other lower level APIs first. We need API that provides us these fence (in or out), and then we can consider using them. For V4L2, there was an attempt, but it was a bit of a miss-fit. Your proposal could work, need to be tested I guess, but it does not solve some of other issues that was discussed. Notably for camera capture, were the HW timestamp is capture about at the same time the frame is ready. But the timestamp is not part of the paylaod, so you need an entire API asynchronously deliver that metadata. It's the biggest pain point I've found, such an API would be quite invasive or if made really generic, might just never be adopted widely enough.
Another issue is that V4L2 doesn't offer any guarantee on job ordering. When you queue multiple buffers for camera capture for instance, you don't know until capture complete in which buffer the frame has been captured.
Is this a Kernel UAPI issue? Surely the kernel driver knows at the start of frame capture which buffer it's getting written into. I would think that the kernel APIs could be adjusted (if we find good reason to do so!) such that they return earlier and return a (buffer, fence) pair. Am I missing something fundamental about video here?
For cameras I believe we could do that, yes. I was pointing out the issues caused by the current API. For video decoders I'll let Nicolas answer the question, he's way more knowledgeable that I am on that topic.
Right now, there is simply no uAPI for supporting asynchronous errors reporting when fences are invovled. That is true for both camera's and CODEC. It's likely what all the attempt was missing, I don't know enough myself to suggest something.
Now, why Stateless video decoders are special is another subject. In CODECs, the decoding and the presentation order may differ. For Stateless kind of CODEC, a bitstream is passed to the HW. We don't know if this bitstream is fully valid, since the it is being parsed and validated by the firmware. It's also firmware job to decide which buffer should be presented first.
In most firmware interface, that information is communicated back all at once when the frame is ready to be presented (which may be quite some time after it was decoded). So indeed, a fence model is not really easy to add, unless the firmware was designed with that model in mind.
Nothing of course would prevent V4L2 framework to generically handle out_fence from other producers. It does not even handle implicit fences at the moment, which is already quite problematic (I've seen glitches on i.MX6/8 and Raspberry Pi 4).
In that specific case, if the fences from etnaviv, vc graphic drivers was exposed, we could solve this issue in userspace. Right now it's implicit, so we rely on all DMABuf driver to have proper support, which is not the case. There is V4L2 support for that coming, but the wait is done synchronously in userspace call that was normally non-blocking. So that is unlikely to fly.
If it helps to settle this part of the discussion I happily volunteer to fix the V4L2 side to wait for the fences without the need for a synchronous wait in qbuf.
Small note, stateless video decoders don't have this issue. The bitstream is validated by userspace, and userspace controls the "decode" operation. This one would be a good case for bidirectional fencing.
I must admit that V4L is a bit of an odd case since the kernel driver is the producer and not the consumer.
Note that V4L2 can be a consumer too. Video output with V4L2 is less frequent than video capture (but it still exists), and codecs and other memory-to-memory processing devices (colorspace converters, scalers, ...) are both consumers and producers.
In the normal case buffers are processed in sequence, but if an error occurs during capture, they can be recycled internally and put to the back of the queue.
Are those errors something that can happen at any time in the middle of a frame capture? If so, that does make things stickier.
Yes it can. Think of packet loss when capturing from a USB webcam for instance.
Unless I'm mistaken, this problem also exists with stateful codecs. And if you don't know in advance which buffer you will receive from the device, the usefulness of fences becomes very questionable :-)
Yeah, if you really are in a situation where there's no way to know until the full frame capture has been completed which buffer is next, then fences are useless. You aren't in an implicit synchronization setting either; you're in a "full flush" setting. It's arguably worse for performance but perhaps unavoidable?
Probably unavoidable in some cases, but nothing that should get in the way for the discussion at hand: there's no need to migrate away from implicit sync when there's implicit sync in the first place :-)
I think we need to analyse the use cases here, and figure out at least guidelines for userspace, otherwise applications will wonder what behaviour to implement, and we'll end up with a wide variety of them. Even just on the kernel side, some V4L2 capture driver will pass erroneous frames to userspace (thus guaranteeing ordering, but without early notification of errors), some will require the frame automatically, and at least one (uvcvideo) has a module parameter to pick the desired behaviour.
Also, from a userspace point of view, the synchronization with the "next frame" in V4L2 isn't implicit. We can poll() the device, just like we'd do with a fence FD. What the explicit fence gives, is a unified object we can pass to another driver, or other userspace, so we can delegate the wait.
You refer to performance in few places. In streaming, this is often measure as real-time throughput. Implicit/explicit fences don't really play any role for us in this regard. V4L2 drivers, like m2m drivers, works with buffer queues. So you can queue in advance many buffers on the OUTPUT device side (which is the input of the m2m), and userspace will queue in advance pretty much all free buffers available on the CAPTURE side. The driver is never starved in that model, at the cost of very large memory consumption of course. Maybe a more visual representation would be:
[pending job] -> [M2M Worker] -> [pending results]
So as long as userspace keep the pending job queue non-empty, and that it consumes and give back buffers back to write the results into, the driver will keep running un-interrupted. Performance remains optimal. What isn't optimal is the latency. And what bugs right now is when a DMAbuf implicit out fence is put back into the pending results queue, since the fence is ignored.
Trying to understand. :-)
So am I :-)
Hehe, same here.
V4L2 just has no notion of something being done asynchronously, which would require fence. The current protocol is that you only queue buffers into the kernel when they are idle and can be consumed by the HW, so there is no need to wait for anything. This requirement is hard to meet with buffers that are shared with DRM today, as all DRM userspace relies on the kernel attached fences to be respected until explicitly told otherwise.
Also V4L2 only allows to dequeue buffers from the kernel into userspace, which are done from the HW perspective. So the V4L2 userspace interface already has an implicit CPU sync on the buffer.
Regards, Lucas
On Wed, Mar 11, 2020 at 8:21 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
Hi Jason,
thanks for pushing this forward and the comprehensive explanation on what it is about.
My question would be what exactly do you now need from Wayland compositor devs? I understood a Wayland compositor needs to: * do atomic page flips, * support [1].
Is there something else? You described a mechanism to pull out and push in these sync_files to dma-bufs depending on what the client provides and what kind of output the compositor puts the final image onto. That's for now just an idea (plus your wip implementation in Vulkan/kernel) and there is not yet anything that can be done for this specifically in Wayland compositors, or is there?
Thanks Roman
[1] https://gitlab.freedesktop.org/wayland/wayland-protocols/blob/master/unstabl...
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
On Mon, Mar 16, 2020 at 6:39 PM Roman Gilg subdiff@gmail.com wrote:
On Wed, Mar 11, 2020 at 8:21 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
Hi Jason,
thanks for pushing this forward and the comprehensive explanation on what it is about.
My question would be what exactly do you now need from Wayland compositor devs? I understood a Wayland compositor needs to:
- do atomic page flips,
- support [1].
Yup, that's pretty much what's needed. From the looks of https://gitlab.gnome.org/GNOME/mutter/issues/548, it appears that mutter is at least on their way to atomic though it also looks like a long road.
Is there something else? You described a mechanism to pull out and push in these sync_files to dma-bufs depending on what the client provides and what kind of output the compositor puts the final image onto. That's for now just an idea (plus your wip implementation in Vulkan/kernel) and there is not yet anything that can be done for this specifically in Wayland compositors, or is there?
The kernel patches I proposed are mostly a way for explicit-sync APIs such as Vulkan to reasonably interop with the implicit sync world. Right now, for instance, we have no real plan for Vulkan to be able to handle synchronization when talking to a media encoder. We have the modifiers and dma-buf stuff which deals with image layout but synchronization is currently an unsolved problem. If we have a mechanism for sync_file import/export from dma-buf then an app can use implicit sync when talking to VAAPI (as an example) and turn that into sync_files to talk to Vulkan. Better yet, we could plumb VAAPI for explicit sync but that's yet one more thing that needs to support explicit sync.
Make sense?
--Jason
Thanks Roman
[1] https://gitlab.freedesktop.org/wayland/wayland-protocols/blob/master/unstabl...
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
On Mon, Mar 16, 2020 at 10:37:04PM -0500, Jason Ekstrand wrote:
On Mon, Mar 16, 2020 at 6:39 PM Roman Gilg subdiff@gmail.com wrote:
On Wed, Mar 11, 2020 at 8:21 PM Jason Ekstrand jason@jlekstrand.net wrote:
On Wed, Mar 11, 2020 at 12:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
Correction: Apparently, I missed some things. If you use atomic, KMS does have explicit in- and out-fences. Non-atomic users (e.g. X11) are still in trouble but most Wayland compositors use atomic these days
Hi Jason,
thanks for pushing this forward and the comprehensive explanation on what it is about.
My question would be what exactly do you now need from Wayland compositor devs? I understood a Wayland compositor needs to:
- do atomic page flips,
- support [1].
Yup, that's pretty much what's needed. From the looks of https://gitlab.gnome.org/GNOME/mutter/issues/548, it appears that mutter is at least on their way to atomic though it also looks like a long road.
FWIW, I expect to have atomic KMS to be used when available in the driver on GNOME 3.38. We just released 3.36, so this means it would be the next version. Whether it makes sense to eventually backport to 3.36 we'll see, but wouldn't be impossible. The majority of the work has already been done.
Jonas
Is there something else? You described a mechanism to pull out and push in these sync_files to dma-bufs depending on what the client provides and what kind of output the compositor puts the final image onto. That's for now just an idea (plus your wip implementation in Vulkan/kernel) and there is not yet anything that can be done for this specifically in Wayland compositors, or is there?
The kernel patches I proposed are mostly a way for explicit-sync APIs such as Vulkan to reasonably interop with the implicit sync world. Right now, for instance, we have no real plan for Vulkan to be able to handle synchronization when talking to a media encoder. We have the modifiers and dma-buf stuff which deals with image layout but synchronization is currently an unsolved problem. If we have a mechanism for sync_file import/export from dma-buf then an app can use implicit sync when talking to VAAPI (as an example) and turn that into sync_files to talk to Vulkan. Better yet, we could plumb VAAPI for explicit sync but that's yet one more thing that needs to support explicit sync.
Make sense?
--Jason
Thanks Roman
[1] https://gitlab.freedesktop.org/wayland/wayland-protocols/blob/master/unstabl...
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs.
As per the above correction, Wayland compositors aren't nearly as bad off as I initially thought. There may still be weird screen capture cases but the normal cases of compositing and displaying via KMS/atomic should be in reasonably good shape.
So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand
dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
On Wed, 2020-03-11 at 12:31 -0500, Jason Ekstrand wrote:
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
I'm pretty sure "the only thing that fence does" is an implementation detail. PresentPixmap blocks until the wait-fence signals, but when and how it signals are properties of the fence itself. You could have drm give the client back a fence fd, pass that to xserver to create a fence object, and name that in the PresentPixmap request, and then drm can do whatever it wants to signal the fence.
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't.
I'm quite sure we can give you an explicit-sync X11 API. I think you may already have one.
- ajax
It seems I may have not set the tone I intended with this e-mail... My intention was never to stomp on anyone's favorite window system (Adam, isn't the only one who's seemed a bit miffed). My intention was to try and solve some very real problems that we have with Vulkan and I had the hope that a solution there could be helpful for others.
The problem we have in Vulkan is that we have an inherently explicit sync graphics API and we're trying to strap it onto some inherently implicit sync window systems and kernel interfaces. Our mechanisms for doing so have evolved over the course of the last 4-5 years and it's way better now than it was when we started but it's still pretty bad and very invasive to the driver. My objective is to completely remove the concept of implicit sync from the Vulkan driver eventually.
Also (and this is going further down the rabbit hole), I would like to begin cleaning up our i915 UAPI to better separate memory residency handling, command submission, and synchronization. Eventually (and this may sound crazy to some), I'd like to get to the point where i915 doesn't own any of the synchronization primitives except what it needs to handle memory management internally. Linux graphics UAPI is about 10 years behind Windows in terms of design (roughly equivalent to Win7) and I think it's costing us in terms of latency and CPU overhead. Some of that may just be implementation problems in i915; some of it may be core API design. It's a bit unclear.
Why am I bringing up kernel APIs? Because one of the biggest problems in evolving things is the fact that our kernel APIs are tied to implicit sync on dma-buf. We can't detangle that until we can remove implicit dma-buf signaling from the command execution APIs. This means that we either need to get rid of ALL implicit synchronization from window-system APIs far enough back in time that we don't run the risk of "breaking userspace" or else we need a plan which lets the kernel driver not support implicit sync but make implicit sync work anyway. What I'm proposing with dma-buf sync_file import/export is one such plan.
So, while this may not solve any problems for Wayland compositors as I previously thought (KMS/atomic supports sync_file. Yay!), we still have a very real problem in Vulkan. It's great that Wayland has an explicit sync API but until all compositors have supported it for at least 2 years, I can't assume it's existence and start deleting my old code paths. Currently, it's only implemented in Weston and the ChromeOS compositor; gnome-shell, kwin, and sway are all still 100% implicit sync AFAIK. We also have to deal with X11.
For those who are asking the question in the back of their minds: Yes, I'm trying to solve a userspace problem with kernel code and, no, I don't think that's necessarily the wrong way around. Don't get me wrong; I very much want to solve the problem "properly" but unless we're very sure we can get it solved properly everywhere quickly, a solution which lets us improve our driver kernel APIs independently of misc. Wayland compositors seems advantageous.
On Wed, Mar 11, 2020 at 6:02 PM Adam Jackson ajax@redhat.com wrote:
On Wed, 2020-03-11 at 12:31 -0500, Jason Ekstrand wrote:
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
I'm pretty sure "the only thing that fence does" is an implementation detail.
So I've been told, many times.
PresentPixmap blocks until the wait-fence signals, but when and how it signals are properties of the fence itself. You could have drm give the client back a fence fd, pass that to xserver to create a fence object, and name that in the PresentPixmap request, and then drm can do whatever it wants to signal the fence.
Poking around at things, X11 may not be quite as bad as I thought here. It's not really set up for sync_file for a couple reasons:
1. It only passes the file descriptor in once at xcb_dri3_fence_from_fd rather than re-creating every frame from a new sync_file 2. It only takes a fence on present and doesn't return one in the PRESENT_COMPLETE event
That said, plumbing syncobj in as an extension looks like a real possibility. A syncobj is just a container that holds a pointer to a dma_fence and it has roughly the same CPU signal/reset behavior that's exposed by the SyncFenceFuncsRec struct. There's a few things I'm not sure how to handle:
1. The Sync extension has these trigger funcs which get called when the fence is signalled. I'm not sure how to handle that with syncobj without a thread polling on them somehow. 2. Not all kernel GPU drivers support syncobj; currently it's just i915, amdgpu, and maybe freedreno AFAIK. How do we handle cases such as Intel+Nvidia? 3. I have no idea what kinds of issues we'd run into with plumbing it all through. Hopefully, X is sufficiently abstracted but I really don't know.
Please excuse my trepidation but I've got a bit of PTSD from modifiers. That was the last time I tried to solve a problem with someone writing X11 patches and it's been 2-3 years and it's still not shipping in distros. If said syncobj extension suffers the same fate, it isn't a real solution.
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't.
I'm quite sure we can give you an explicit-sync X11 API. I think you may already have one.
It looks like we at least have a bunch of pieces which can probably be used to build one.
--Jason
On Thu, Mar 12, 2020 at 6:36 PM Jason Ekstrand jason@jlekstrand.net wrote:
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs. So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
<troll> Maybe this is something for the Marketing Department to solve? Sell the extra processing that can be done during such extra blit as a feature?
As a former user of a wide-gamut monitor that has no sRGB mode, and a gamer, I would definitely accept the extra step (color conversion, not "just a blit"!) between the application and the actual output. In fact, I have set up compicc just for this purpose. Games with poisonous oversaturated colors (because none of the game authors care about wide-gamut monitors) are worse than the same games affected by the very small performance penalty due to the conversion.
We just need a Marketing Person to come up with a huge list of other cases where such compositing step is required for correctness, and declare that direct scanout is something that makes no sense in the present day, except possibly on embedded devices. </troll>
Of course the above trolling does not solve the problem related to inability to be sure about the correct API usage.
There is no synchronization between processes (e.g. 3D app and compositor) within X on AMD hw. It works because of some hacks in Mesa.
Marek
On Wed, Mar 11, 2020 at 1:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit
sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs. So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Could you elaborate. If there's something missing from my mental model of how implicit sync works, I'd like to have it corrected. People continue claiming that AMD is somehow special but I have yet to grasp what makes it so. (Not that anyone has bothered to try all that hard to explain it.)
--Jason
On March 13, 2020 21:03:21 Marek Olšák maraeo@gmail.com wrote:
There is no synchronization between processes (e.g. 3D app and compositor) within X on AMD hw. It works because of some hacks in Mesa.
Marek
On Wed, Mar 11, 2020 at 1:31 PM Jason Ekstrand jason@jlekstrand.net wrote: All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs. So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
The synchronization works because the Mesa driver waits for idle (drains the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If the Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of its own command buffer while shaders of the previous process are still running.
If the Mesa driver submits a command buffer internally (because it's full), it doesn't wait, so the GFX pipeline doesn't notice that a command buffer ended and a new one started.
The waiting at the end of command buffers happens only when the flush is external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the GFX pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Marek
On Sun., Mar. 15, 2020, 22:49 Jason Ekstrand, jason@jlekstrand.net wrote:
Could you elaborate. If there's something missing from my mental model of how implicit sync works, I'd like to have it corrected. People continue claiming that AMD is somehow special but I have yet to grasp what makes it so. (Not that anyone has bothered to try all that hard to explain it.)
--Jason
On March 13, 2020 21:03:21 Marek Olšák maraeo@gmail.com wrote:
There is no synchronization between processes (e.g. 3D app and compositor) within X on AMD hw. It works because of some hacks in Mesa.
Marek
On Wed, Mar 11, 2020 at 1:31 PM Jason Ekstrand jason@jlekstrand.net wrote:
All,
Sorry for casting such a broad net with this one. I'm sure most people who reply will get at least one mailing list rejection. However, this is an issue that affects a LOT of components and that's why it's thorny to begin with. Please pardon the length of this e-mail as well; I promise there's a concrete point/proposal at the end.
Explicit synchronization is the future of graphics and media. At least, that seems to be the consensus among all the graphics people I've talked to. I had a chat with one of the lead Android graphics engineers recently who told me that doing explicit sync from the start was one of the best engineering decisions Android ever made. It's also the direction being taken by more modern APIs such as Vulkan.
## What are implicit and explicit synchronization?
For those that aren't familiar with this space, GPUs, media encoders, etc. are massively parallel and synchronization of some form is required to ensure that everything happens in the right order and avoid data races. Implicit synchronization is when bits of work (3D, compute, video encode, etc.) are implicitly based on the absolute CPU-time order in which API calls occur. Explicit synchronization is when the client (whatever that means in any given context) provides the dependency graph explicitly via some sort of synchronization primitives. If you're still confused, consider the following examples:
With OpenGL and EGL, almost everything is implicit sync. Say you have two OpenGL contexts sharing an image where one writes to it and the other textures from it. The way the OpenGL spec works, the client has to make the API calls to render to the image before (in CPU time) it makes the API calls which texture from the image. As long as it does this (and maybe inserts a glFlush?), the driver will ensure that the rendering completes before the texturing happens and you get correct contents.
Implicit synchronization can also happen across processes. Wayland, for instance, is currently built on implicit sync where the client does their rendering and then does a hand-off (via wl_surface::commit) to tell the compositor it's done at which point the compositor can now texture from the surface. The hand-off ensures that the client's OpenGL API calls happen before the server's OpenGL API calls.
A good example of explicit synchronization is the Vulkan API. There, a client (or multiple clients) can simultaneously build command buffers in different threads where one of those command buffers renders to an image and the other textures from it and then submit both of them at the same time with instructions to the driver for which order to execute them in. The execution order is described via the VkSemaphore primitive. With the new VK_KHR_timeline_semaphore extension, you can even submit the work which does the texturing BEFORE the work which does the rendering and the driver will sort it out.
The #1 problem with implicit synchronization (which explicit solves) is that it leads to a lot of over-synchronization both in client space and in driver/device space. The client has to synchronize a lot more because it has to ensure that the API calls happen in a particular order. The driver/device have to synchronize a lot more because they never know what is going to end up being a synchronization point as an API call on another thread/process may occur at any time. As we move to more and more multi-threaded programming this synchronization (on the client-side especially) becomes more and more painful.
## Current status in Linux
Implicit synchronization in Linux works via a the kernel's internal dma_buf and dma_fence data structures. A dma_fence is a tiny object which represents the "done" status for some bit of work. Typically, dma_fences are created as a by-product of someone submitting some bit of work (say, 3D rendering) to the kernel. The dma_buf object has a set of dma_fences on it representing shared (read) and exclusive (write) access to the object. When work is submitted which, for instance renders to the dma_buf, it's queued waiting on all the fences on the dma_buf and and a dma_fence is created representing the end of said rendering work and it's installed as the dma_buf's exclusive fence. This way, the kernel can manage all its internal queues (3D rendering, display, video encode, etc.) and know which things to submit in what order.
For the last few years, we've had sync_file in the kernel and it's plumbed into some drivers. A sync_file is just a wrapper around a single dma_fence. A sync_file is typically created as a by-product of submitting work (3D, compute, etc.) to the kernel and is signaled when that work completes. When a sync_file is created, it is guaranteed by the kernel that it will become signaled in finite time and, once it's signaled, it remains signaled for the rest of time. A sync_file is represented in UAPIs as a file descriptor and can be used with normal file APIs such as dup(). It can be passed into another UAPI which does some bit of queue'd work and the submitted work will wait for the sync_file to be triggered before executing. A sync_file also supports poll() if you want to wait on it manually.
Unfortunately, sync_file is not broadly used and not all kernel GPU drivers support it. Here's a very quick overview of my understanding of the status of various components (I don't know the status of anything in the media world):
- Vulkan: Explicit synchronization all the way but we have to go
implicit as soon as we interact with a window-system. Vulkan has APIs to import/export sync_files to/from it's VkSemaphore and VkFence synchronization primitives.
- OpenGL: Implicit all the way. There are some EGL extensions to
enable some forms of explicit sync via sync_file but OpenGL itself is still implicit.
- Wayland: Currently depends on implicit sync in the kernel (accessed
via EGL/OpenGL). There is an unstable extension to allow passing sync_files around but it's questionable how useful it is right now (more on that later).
- X11: With present, it has these "explicit" fence objects but
they're always a shmfence which lets the X server and client do a userspace CPU-side hand-off without going over the socket (and round-tripping through the kernel). However, the only thing that fence does is order the OpenGL API calls in the client and server and the real synchronization is still implicit.
- linux/i915/gem: Fully supports using sync_file or syncobj for
explicit sync.
- linux/amdgpu: Supports sync_file and syncobj but it still
implicitly syncs sometimes due to it's internal memory residency handling which can lead to over-synchronization.
- KMS: Implicit sync all the way. There are no KMS APIs which take
explicit sync primitives.
- v4l: ???
- gstreamer: ???
- Media APIs such as vaapi etc.: ???
## Chicken and egg problems
Ok, this is where it starts getting depressing. I made the claim above that Wayland has an explicit synchronization protocol that's of questionable usefulness. I would claim that basically any bit of plumbing we do through window systems is currently of questionable usefulness. Why?
From my perspective, as a Vulkan driver developer, I have to deal with the fact that Vulkan is an explicit sync API but Wayland and X11 aren't. Unfortunately, the Wayland extension solves zero problems for me because I can't really use it unless it's implemented in all of the compositors. Until every Wayland compositor I care about my users being able to use (which is basically all of them) supports the extension, I have to continue carry around my pile of hacks to keep implicit sync and Vulkan working nicely together.
From the perspective of a Wayland compositor (I used to play in this space), they'd love to implement the new explicit sync extension but can't. Sure, they could wire up the extension, but the moment they go to flip a client buffer to the screen directly, they discover that KMS doesn't support any explicit sync APIs. So, yes, they can technically implement the extension assuming the EGL stack they're running on has the sync_file extensions but any client buffers which come in using the explicit sync Wayland extension have to be composited and can't be scanned out directly. As a 3D driver developer, I absolutely don't want compositors doing that because my users will complain about performance issues due to the extra blit.
Ok, so let's say we get KMS wired up with implicit sync. That solves all our problems, right? It does, right up until someone decides that they wan to screen capture their Wayland session via some hardware media encoder that doesn't support explicit sync. Now we have to plumb it all the way through the media stack, gstreamer, etc. Great, so let's do that! Oh, but gstreamer won't want to plumb it through until they're guaranteed that they can use explicit sync when displaying on X11 or Wayland. Are you seeing the problem?
To make matters worse, since most things are doing implicit synchronization today, it's really easy to get your explicit synchronization wrong and never notice. If you forget to pass a sync_file into one place (say you never notice KMS doesn't support them), it will probably work anyway thanks to all the implicit sync that's going on elsewhere.
So, clearly, we all need to go write piles of code that we can't actually properly test until everyone else has written their piece and then we use explicit sync if and only if all components support it. Really? We're going to do multiple years of development and then just hope it works when we finally flip the switch? That doesn't sound like a good plan to me.
## A proposal: Implicit and explicit sync together
How to solve all these chicken-and-egg problems is something I've been giving quite a bit of thought (and talking with many others about) in the last couple of years. One motivation for this is that we have to deal with a mismatch in Vulkan. Another motivation is that I'm becoming increasingly unhappy with the way that synchronization, memory residency, and command submission are inherently intertwined in i915 and would like to break things apart. Towards that end, I have an actual proposal.
A couple weeks ago, I sent a series of patches to the dri-devel mailing list which adds a pair of new ioctls to dma-buf which allow userspace to manually import or export a sync_file from a dma-buf. The idea is that something like a Wayland compositor can switch to 100% explicit sync internally once the ioctl is available. If it gets buffers in from a client that doesn't use the explicit sync extension, it can pull a sync_file from the dma-buf and use that exactly as it would a sync_file passed via the explicit sync extension. When it goes to scan out a user buffer and discovers that KMS doesn't accept sync_files (or if it tries to use that pesky media encoder no one has converted), it can take it's sync_file for display and stuff it into the dma-buf before handing it to KMS.
Along with the kernel patches, I've also implemented support for this in the Vulkan WSI code used by ANV and RADV. With those patches, the only requirement on the Vulkan drivers is that you be able to export any VkSemaphore as a sync_file and temporarily import a sync_file into any VkFence or VkSemaphore. As long as that works, the core Vulkan driver only ever sees explicit synchronization via sync_file. The WSI code uses these new ioctls to translate the implicit sync of X11 and Wayland to the explicit sync the Vulkan driver wants.
I'm hoping (and here's where I want a sanity check) that a simple API like this will allow us to finally start moving the Linux ecosystem over to explicit synchronization one piece at a time in a way that's actually correct. (No Wayland explicit sync with compositors hoping KMS magically works even though it doesn't have a sync_file API.) Once some pieces in the ecosystem start moving, there will be motivation to start moving others and maybe we can actually build the momentum to get most everything converted.
For reference, you can find the kernel RFC patches and mesa MR here:
https://lists.freedesktop.org/archives/dri-devel/2020-March/258833.html
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4037
At this point, I welcome your thoughts, comments, objections, and maybe even help/review. :-)
--Jason Ekstrand _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle (drains the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If the Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of its own command buffer while shaders of the previous process are still running.
If the Mesa driver submits a command buffer internally (because it's full), it doesn't wait, so the GFX pipeline doesn't notice that a command buffer ended and a new one started.
The waiting at the end of command buffers happens only when the flush is external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the GFX pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to be done for explicit fences as well, doesn't it?
On Mon, Mar 16, 2020 at 5:57 AM Michel Dänzer michel@daenzer.net wrote:
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle (drains the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If the Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of its own command buffer while shaders of the previous process are still
running.
If the Mesa driver submits a command buffer internally (because it's
full),
it doesn't wait, so the GFX pipeline doesn't notice that a command buffer ended and a new one started.
The waiting at the end of command buffers happens only when the flush is external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the
GFX
pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to be done for explicit fences as well, doesn't it?
No. Explicit fences don't require userspace to wait for idle in the command buffer. Fences are signalled when the last draw is complete and caches are flushed. Before that happens, any command buffer that is not dependent on the fence can start execution. There is never a need for the GPU to be idle if there is enough independent work to do.
Marek
On 2020-03-16 7:33 p.m., Marek Olšák wrote:
On Mon, Mar 16, 2020 at 5:57 AM Michel Dänzer michel@daenzer.net wrote:
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle (drains the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If the Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of its own command buffer while shaders of the previous process are still
running.
If the Mesa driver submits a command buffer internally (because it's
full),
it doesn't wait, so the GFX pipeline doesn't notice that a command buffer ended and a new one started.
The waiting at the end of command buffers happens only when the flush is external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the
GFX
pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to be done for explicit fences as well, doesn't it?
No. Explicit fences don't require userspace to wait for idle in the command buffer. Fences are signalled when the last draw is complete and caches are flushed. Before that happens, any command buffer that is not dependent on the fence can start execution. There is never a need for the GPU to be idle if there is enough independent work to do.
I don't think explicit fences in the context of this discussion imply using that different fence signalling mechanism though. My understanding is that the API proposed by Jason allows implicit fences to be used as explicit ones and vice versa, so presumably they have to use the same signalling mechanism.
Anyway, maybe the different fence signalling mechanism you describe could be used by the amdgpu kernel driver in general, then Mesa could drop the waits for idle and get the benefits with implicit sync as well?
On Tue., Mar. 17, 2020, 06:02 Michel Dänzer, michel@daenzer.net wrote:
On 2020-03-16 7:33 p.m., Marek Olšák wrote:
On Mon, Mar 16, 2020 at 5:57 AM Michel Dänzer michel@daenzer.net
wrote:
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle
(drains
the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If
the
Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of
its
own command buffer while shaders of the previous process are still
running.
If the Mesa driver submits a command buffer internally (because it's
full),
it doesn't wait, so the GFX pipeline doesn't notice that a command
buffer
ended and a new one started.
The waiting at the end of command buffers happens only when the flush
is
external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the
GFX
pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to be done for explicit fences as well, doesn't it?
No. Explicit fences don't require userspace to wait for idle in the
command
buffer. Fences are signalled when the last draw is complete and caches
are
flushed. Before that happens, any command buffer that is not dependent on the fence can start execution. There is never a need for the GPU to be
idle
if there is enough independent work to do.
I don't think explicit fences in the context of this discussion imply using that different fence signalling mechanism though. My understanding is that the API proposed by Jason allows implicit fences to be used as explicit ones and vice versa, so presumably they have to use the same signalling mechanism.
Anyway, maybe the different fence signalling mechanism you describe could be used by the amdgpu kernel driver in general, then Mesa could drop the waits for idle and get the benefits with implicit sync as well?
Yes. If there is any waiting, or should be done in the GPU scheduler, not in the command buffer, so that independent command buffers can use the GFX queue.
Marek
-- Earthling Michel Dänzer | https://redhat.com Libre software enthusiast | Mesa and X developer
On Tue, Mar 17, 2020 at 11:01:57AM +0100, Michel Dänzer wrote:
On 2020-03-16 7:33 p.m., Marek Olšák wrote:
On Mon, Mar 16, 2020 at 5:57 AM Michel Dänzer michel@daenzer.net wrote:
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle (drains the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning the command buffer only starts draws and doesn't wait for completion. If the Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of its own command buffer while shaders of the previous process are still
running.
If the Mesa driver submits a command buffer internally (because it's
full),
it doesn't wait, so the GFX pipeline doesn't notice that a command buffer ended and a new one started.
The waiting at the end of command buffers happens only when the flush is external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until the
GFX
pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to be done for explicit fences as well, doesn't it?
No. Explicit fences don't require userspace to wait for idle in the command buffer. Fences are signalled when the last draw is complete and caches are flushed. Before that happens, any command buffer that is not dependent on the fence can start execution. There is never a need for the GPU to be idle if there is enough independent work to do.
I don't think explicit fences in the context of this discussion imply using that different fence signalling mechanism though. My understanding is that the API proposed by Jason allows implicit fences to be used as explicit ones and vice versa, so presumably they have to use the same signalling mechanism.
Anyway, maybe the different fence signalling mechanism you describe could be used by the amdgpu kernel driver in general, then Mesa could drop the waits for idle and get the benefits with implicit sync as well?
Yeah, this is entirely about the programming model visible to userspace. There shouldn't be any impact on the driver's choice of a top vs. bottom of the gpu pipeline used for synchronization, that's entirely up to what you're hw/driver/scheduler can pull off.
Doing a full gfx pipeline flush for shared buffers, when your hw can do be, sounds like an issue to me that's not related to this here at all. It might be intertwined with amdgpu's special interpretation of dma_resv fences though, no idea. We might need to revamp all that. But for a userspace client that does nothing fancy (no multiple render buffer targets in one bo, or vk style "I write to everything all the time, perhaps" stuff) there should be 0 perf difference between implicit sync through dma_resv and explicit sync through sync_file/syncobj/dma_fence directly.
If there is I'd consider that a bit a driver bug. -Daniel
On Thu., Mar. 19, 2020, 06:51 Daniel Vetter, daniel@ffwll.ch wrote:
On Tue, Mar 17, 2020 at 11:01:57AM +0100, Michel Dänzer wrote:
On 2020-03-16 7:33 p.m., Marek Olšák wrote:
On Mon, Mar 16, 2020 at 5:57 AM Michel Dänzer michel@daenzer.net
wrote:
On 2020-03-16 4:50 a.m., Marek Olšák wrote:
The synchronization works because the Mesa driver waits for idle
(drains
the GFX pipeline) at the end of command buffers and there is only 1 graphics queue, so everything is ordered.
The GFX pipeline runs asynchronously to the command buffer, meaning
the
command buffer only starts draws and doesn't wait for completion. If
the
Mesa driver didn't wait at the end of the command buffer, the command buffer would finish and a different process could start execution of
its
own command buffer while shaders of the previous process are still
running.
If the Mesa driver submits a command buffer internally (because it's
full),
it doesn't wait, so the GFX pipeline doesn't notice that a command
buffer
ended and a new one started.
The waiting at the end of command buffers happens only when the
flush is
external (Swap buffers, glFlush).
It's a performance problem, because the GFX queue is blocked until
the
GFX
pipeline is drained at the end of every frame at least.
So explicit fences for SwapBuffers would help.
Not sure what difference it would make, since the same thing needs to
be
done for explicit fences as well, doesn't it?
No. Explicit fences don't require userspace to wait for idle in the
command
buffer. Fences are signalled when the last draw is complete and caches
are
flushed. Before that happens, any command buffer that is not dependent
on
the fence can start execution. There is never a need for the GPU to be
idle
if there is enough independent work to do.
I don't think explicit fences in the context of this discussion imply using that different fence signalling mechanism though. My understanding is that the API proposed by Jason allows implicit fences to be used as explicit ones and vice versa, so presumably they have to use the same signalling mechanism.
Anyway, maybe the different fence signalling mechanism you describe could be used by the amdgpu kernel driver in general, then Mesa could drop the waits for idle and get the benefits with implicit sync as well?
Yeah, this is entirely about the programming model visible to userspace. There shouldn't be any impact on the driver's choice of a top vs. bottom of the gpu pipeline used for synchronization, that's entirely up to what you're hw/driver/scheduler can pull off.
Doing a full gfx pipeline flush for shared buffers, when your hw can do be, sounds like an issue to me that's not related to this here at all. It might be intertwined with amdgpu's special interpretation of dma_resv fences though, no idea. We might need to revamp all that. But for a userspace client that does nothing fancy (no multiple render buffer targets in one bo, or vk style "I write to everything all the time, perhaps" stuff) there should be 0 perf difference between implicit sync through dma_resv and explicit sync through sync_file/syncobj/dma_fence directly.
If there is I'd consider that a bit a driver bug.
Last time I checked, there was no fence sync in gnome shell and compiz after an app passes a buffer to it. So drivers have to invent hacks to work around it and decrease performance. It's not a driver bug.
Implicit sync really means that apps and compositors don't sync, so the driver has to guess when it should sync.
Marek
-Daniel
-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
On 2020-03-19 8:54 p.m., Marek Olšák wrote:
On Thu., Mar. 19, 2020, 06:51 Daniel Vetter, daniel@ffwll.ch wrote:
Yeah, this is entirely about the programming model visible to userspace. There shouldn't be any impact on the driver's choice of a top vs. bottom of the gpu pipeline used for synchronization, that's entirely up to what you're hw/driver/scheduler can pull off.
Doing a full gfx pipeline flush for shared buffers, when your hw can do be, sounds like an issue to me that's not related to this here at all. It might be intertwined with amdgpu's special interpretation of dma_resv fences though, no idea. We might need to revamp all that. But for a userspace client that does nothing fancy (no multiple render buffer targets in one bo, or vk style "I write to everything all the time, perhaps" stuff) there should be 0 perf difference between implicit sync through dma_resv and explicit sync through sync_file/syncobj/dma_fence directly.
If there is I'd consider that a bit a driver bug.
Last time I checked, there was no fence sync in gnome shell and compiz after an app passes a buffer to it.
They are not required (though encouraged) to do that.
So drivers have to invent hacks to work around it and decrease performance. It's not a driver bug.
Implicit sync really means that apps and compositors don't sync, so the driver has to guess when it should sync.
Making implicit sync work correctly is ultimately the kernel driver's responsibility. It sounds like radeonsi is having to work around the amdgpu/radeon kernel driver(s) not fully living up to this responsibility.
dri-devel@lists.freedesktop.org
-
Adam Jackson -
Alexander E. Patrakov -
Daniel Stone -
Daniel Vetter -
Jacob Lifshay -
Jason Ekstrand -
Jonas Ådahl -
Laurent Pinchart -
Lucas Stach -
Marek Olšák -
Michel Dänzer -
Nicolas Dufresne -
Roman Gilg -
Simon Ser -
Tomek Bury