On Thu, Nov 27, 2014 at 09:25:15AM +1000, Dave Airlie wrote:
TL;DR summary: I wrote patches. Help me choose the best implementation and interface.
So the i915.ko driver could use some mechanism to run functions after a given number of vblanks. Implementations for this mechanism were already proposed in the past (by Chris Wilson and others), but they were i915-specific instead of a generic drm.ko implementation. We even had patches that make use of this new mechanism.
Since this is very similar to the vblank IOCTL we currently have, but with the caller being a Kernel module instead of user space, I decided to write an implementation that tries to reuse the current IOCTL infrastructure.
In the first patch we add all the necessary code to allow the modules to call the vblank ioctl from Kernel space: they provide a callback function that is going to be called instead of the traditional send_vblank_event(), which means the Kernel callers don't have to deal with the file descriptors and DRM events.
In the second patch we add a simple extension of the feature, to allow the drivers to have their callbacks called in a non-atomic context.
In all the other subsequent patches we rework the underlying code so that the common aspects between the user space vblank IOCTL and the Kernel interface are all stored in a single structure (struct drm_vblank_wait_item), and then anything that is specific to the different users is stored in a structure that contains struct drm_vblank_wait_item. This way, most of the drm.ko code only needs to deal with struct drm_vblank_wait_item, not really knowing how it is wrapped. The advantage of this rework is that it reduces the number of memory allocations needed in some cases (from 2 to 1) and it also reduces the amount of memory used on each allocation.
But while this all sounds nice in the above description, I had the feeling that this adds a lot of complexity and makes the code not-so-beautiful. If we ever need more extensions/options to this feature, we're going to have to untangle the code even more from the IOCTL part. We also have a small "abstraction break" in change introduced in patch 6. And there's the risk that some of the reworks may have added a regression somewhere.
Based on that, I decided to also provide an alternative implementation. In this implementation we add a new dev->vblank_kernel_list instead of reusing dev->vblank_event_list, and add new code to handle that this. This new code is completely based on the code that handles dev->vblank_kernel_list, so there's some duplication here. On the other hand, since the in-Kernel infrastructure is not tied to the IOCTL structure, we can very easily grow and adapt the Kernel code without having to fight against the IOCTL code. And the risk of regressions is greatly reduced.
The second difference of this alternative implementation is the signature of the drm_wait_vblank_kernel() function. While it could be exactly the same as the one used in the other series, I decided to make it different so we can also choose which one to use. In the 7-patch series implementation, all the user needs to do is to allocate the structure, and call the function, properly setting all its arguments. Then the function is responsible for parsing the arguments and populating the structure based on that. On the alternative implementation, the user has to fill the structure with the appropriate arguments, and then call drm_wait_vblank_kernel() passing just the allocated structure as an argument.
If you notice, you will see that these patches add a new debugfs file to i915.ko. This file is just a very simple example user of the new interface, which I used while developing. If you have difficulty understanding the new interface, please also look at the i915/i915_debugfs.c diff. Of course, I plan to exclude this debugfs interface from the final to-be-merged patches.
So, in summary, we have a few things we need to discuss:
- Which implementation to use?
a. Just the 2 first patches of the 7-patch series? Pros: - the Kernel users basically call the vblank IOCTL - the code changes are minimal Cons: - The amount of memory allocations and memory space consumed is not optimal
b. The full 7-patch series? Pros: - The same backend is used for both the Kernel and IOCTL. Cons: - The code gets more complex. - Extending the Kernel interface can get complicated due to the fact that it shares code with the IOCTL interface. - I didn't really like this solution.
c. The alternative implementation I wrote? Pros: - It's independent from the IOCTL code, making further patches easier. Cons: - There is some duplicated code.
d. Something totally different from these alternatives?
- How should the driver interface look like?
a. All the possibilities are passed through the function call, so the drm.ko code needs to set the struct members itself. b. The caller already sets the struct members instead of passing them as parameters to the function. c. Something different?
All these patches are based on top of Daniel Vetter's drm-intel-nightly tree. It's all just an RFC, so don't expect something very well tested.
Flamewars and bikeshedding are welcome!
As we've discussed on irc a bit I still think cleaning up this rathole is beneficial. Getting the event handling and vblank callbacks right just in drivers seems to be really hard, so better to clean this up and unify things.
But like you've said patches are a lot of churn, so do you have a git branch somewhere for the lazy to look at?
Thanks, Daniel
I also think vmwgfx might already have something like this, maybe take a look and see if this is the same use case and can cover that as well.
vmwgfx reuses the drm event logic for fences, so definitely interacts with all this. And Rob has some rfc patches somewhere (as part of atomic) to clean up the drm event handling code and extract the commen prepare/release/send and preclose handling code a bit so that not every driver needs to reinvent that whell (and with bugs).
But I did not spot any vblank callback/worker support in vmwgfx. So I think Paulo's series here is orthogonal. Ofc it would also benefit from an overhaul of the drm event handling code. -Daniel