Am 2021-09-01 um 4:29 a.m. schrieb Christoph Hellwig:
On Mon, Aug 30, 2021 at 01:04:43PM -0400, Felix Kuehling wrote:
driver code is not really involved in updating the CPU mappings. Maybe it's something we need to do in the migration helpers.
It looks like I'm totally misunderstanding what you are adding here then. Why do we need any special treatment at all for memory that has normal struct pages and is part of the direct kernel map?
The pages are like normal memory for purposes of mapping them in CPU page tables and for coherent access from the CPU.
That's the user page tables. What about the kernel direct map? If there is a normal kernel struct page backing there really should be no need for the pgmap.
I'm not sure. The physical address ranges are in the UEFI system address map as special-purpose memory. Does Linux create the struct pages and kernel direct map for that without a pgmap call? I didn't see that last time I went digging through that code.
From an application perspective, we want file-backed and anonymous mappings to be able to use DEVICE_PUBLIC pages with coherent CPU access. The goal is to optimize performance for GPU heavy workloads while minimizing the need to migrate data back-and-forth between system memory and device memory.
I don't really understand that part. file backed pages are always allocated by the file system using the pagecache helpers, that is using the page allocator. Anonymouns memory also always comes from the page allocator.
I'm coming at this from my experience with DEVICE_PRIVATE. Both anonymous and file-backed pages should be migrateable to DEVICE_PRIVATE memory by the migrate_vma_* helpers for more efficient access by our GPU. (*) It's part of the basic premise of HMM as I understand it. I would expect the same thing to work for DEVICE_PUBLIC memory.
(*) I believe migrating file-backed pages to DEVICE_PRIVATE doesn't currently work, but that's something I'm hoping to fix at some point.
The pages are special in two ways:
- The memory is managed not by the Linux buddy allocator, but by the GPU driver's TTM memory manager
Why?
It's a system architectural decision based on the access latency to the memory and the expected use cases that we do not want the GPU driver and the Linux buddy allocator and VM subsystem competing for the same device memory.
- We want to migrate data in response to GPU page faults and application hints using the migrate_vma helpers
Why?
Device memory has much higher bandwidth and much lower latency than regular system memory for the GPU to access. It's essential for enabling good GPU application performance. Page-based memory migration enables good performance with more intuitive programming models such as managed/unified memory in HIP or unified shared memory in OpenMP. We do this on our discrete GPUs with DEVICE_PRIVATE memory.
I see DEVICE_PUBLIC as an improved version of DEVICE_PRIVATE that allows the CPU to map the device memory coherently to minimize the need for migrations when CPU and GPU access the same memory concurrently or alternatingly. But we're not going as far as putting that memory entirely under the management of the Linux memory manager and VM subsystem. Our (and HPE's) system architects decided that this memory is not suitable to be used like regular NUMA system memory by the Linux memory manager.
Regards, Felix
On Wed, Sep 01, 2021 at 11:40:43AM -0400, Felix Kuehling wrote:
Am 2021-09-01 um 4:29 a.m. schrieb Christoph Hellwig:
On Mon, Aug 30, 2021 at 01:04:43PM -0400, Felix Kuehling wrote:
driver code is not really involved in updating the CPU mappings. Maybe it's something we need to do in the migration helpers.
It looks like I'm totally misunderstanding what you are adding here then. Why do we need any special treatment at all for memory that has normal struct pages and is part of the direct kernel map?
The pages are like normal memory for purposes of mapping them in CPU page tables and for coherent access from the CPU.
That's the user page tables. What about the kernel direct map? If there is a normal kernel struct page backing there really should be no need for the pgmap.
I'm not sure. The physical address ranges are in the UEFI system address map as special-purpose memory. Does Linux create the struct pages and kernel direct map for that without a pgmap call? I didn't see that last time I went digging through that code.
From an application perspective, we want file-backed and anonymous mappings to be able to use DEVICE_PUBLIC pages with coherent CPU access. The goal is to optimize performance for GPU heavy workloads while minimizing the need to migrate data back-and-forth between system memory and device memory.
I don't really understand that part. file backed pages are always allocated by the file system using the pagecache helpers, that is using the page allocator. Anonymouns memory also always comes from the page allocator.
I'm coming at this from my experience with DEVICE_PRIVATE. Both anonymous and file-backed pages should be migrateable to DEVICE_PRIVATE memory by the migrate_vma_* helpers for more efficient access by our GPU. (*) It's part of the basic premise of HMM as I understand it. I would expect the same thing to work for DEVICE_PUBLIC memory.
(*) I believe migrating file-backed pages to DEVICE_PRIVATE doesn't currently work, but that's something I'm hoping to fix at some point.
FWIW, I'd love to see the architecture documents that define how filesystems are supposed to interact with this device private memory. This whole "hand filesystem controlled memory to other devices" is a minefield that is trivial to get wrong iand very difficult to fix - just look at the historical mess that RDMA to/from file backed and/or DAX pages has been.
So, really, from my perspective as a filesystem engineer, I want to see an actual specification for how this new memory type is going to interact with filesystem and the page cache so everyone has some idea of how this is going to work and can point out how it doesn't work before code that simply doesn't work is pushed out into production systems and then merged....
Cheers,
Dave.
On 2021-09-01 6:03 p.m., Dave Chinner wrote:
On Wed, Sep 01, 2021 at 11:40:43AM -0400, Felix Kuehling wrote:
Am 2021-09-01 um 4:29 a.m. schrieb Christoph Hellwig:
On Mon, Aug 30, 2021 at 01:04:43PM -0400, Felix Kuehling wrote:
driver code is not really involved in updating the CPU mappings. Maybe it's something we need to do in the migration helpers.
It looks like I'm totally misunderstanding what you are adding here then. Why do we need any special treatment at all for memory that has normal struct pages and is part of the direct kernel map?
The pages are like normal memory for purposes of mapping them in CPU page tables and for coherent access from the CPU.
That's the user page tables. What about the kernel direct map? If there is a normal kernel struct page backing there really should be no need for the pgmap.
I'm not sure. The physical address ranges are in the UEFI system address map as special-purpose memory. Does Linux create the struct pages and kernel direct map for that without a pgmap call? I didn't see that last time I went digging through that code.
From an application perspective, we want file-backed and anonymous mappings to be able to use DEVICE_PUBLIC pages with coherent CPU access. The goal is to optimize performance for GPU heavy workloads while minimizing the need to migrate data back-and-forth between system memory and device memory.
I don't really understand that part. file backed pages are always allocated by the file system using the pagecache helpers, that is using the page allocator. Anonymouns memory also always comes from the page allocator.
I'm coming at this from my experience with DEVICE_PRIVATE. Both anonymous and file-backed pages should be migrateable to DEVICE_PRIVATE memory by the migrate_vma_* helpers for more efficient access by our GPU. (*) It's part of the basic premise of HMM as I understand it. I would expect the same thing to work for DEVICE_PUBLIC memory.
(*) I believe migrating file-backed pages to DEVICE_PRIVATE doesn't currently work, but that's something I'm hoping to fix at some point.
FWIW, I'd love to see the architecture documents that define how filesystems are supposed to interact with this device private memory. This whole "hand filesystem controlled memory to other devices" is a minefield that is trivial to get wrong iand very difficult to fix - just look at the historical mess that RDMA to/from file backed and/or DAX pages has been.
So, really, from my perspective as a filesystem engineer, I want to see an actual specification for how this new memory type is going to interact with filesystem and the page cache so everyone has some idea of how this is going to work and can point out how it doesn't work before code that simply doesn't work is pushed out into production systems and then merged....
OK. To be clear, that's not part of this patch series. And I have no authority to push anything in this part of the kernel, so you have nothing to fear. ;)
FWIW, we already have the ability to map file-backed system memory pages into device page tables with HMM and interval notifiers. But we cannot currently migrate them to ZONE_DEVICE pages. Beyond that, my understanding of how filesystems and page cache work is rather superficial at this point. I'll keep your name in mind for when I am ready to discuss this in more detail.
Cheers, Felix
Cheers,
Dave.
On Wed, Sep 01, 2021 at 07:07:34PM -0400, Felix Kuehling wrote:
On 2021-09-01 6:03 p.m., Dave Chinner wrote:
On Wed, Sep 01, 2021 at 11:40:43AM -0400, Felix Kuehling wrote:
Am 2021-09-01 um 4:29 a.m. schrieb Christoph Hellwig:
On Mon, Aug 30, 2021 at 01:04:43PM -0400, Felix Kuehling wrote:
> driver code is not really involved in updating the CPU mappings. Maybe > it's something we need to do in the migration helpers. It looks like I'm totally misunderstanding what you are adding here then. Why do we need any special treatment at all for memory that has normal struct pages and is part of the direct kernel map?
The pages are like normal memory for purposes of mapping them in CPU page tables and for coherent access from the CPU.
That's the user page tables. What about the kernel direct map? If there is a normal kernel struct page backing there really should be no need for the pgmap.
I'm not sure. The physical address ranges are in the UEFI system address map as special-purpose memory. Does Linux create the struct pages and kernel direct map for that without a pgmap call? I didn't see that last time I went digging through that code.
From an application perspective, we want file-backed and anonymous mappings to be able to use DEVICE_PUBLIC pages with coherent CPU access. The goal is to optimize performance for GPU heavy workloads while minimizing the need to migrate data back-and-forth between system memory and device memory.
I don't really understand that part. file backed pages are always allocated by the file system using the pagecache helpers, that is using the page allocator. Anonymouns memory also always comes from the page allocator.
I'm coming at this from my experience with DEVICE_PRIVATE. Both anonymous and file-backed pages should be migrateable to DEVICE_PRIVATE memory by the migrate_vma_* helpers for more efficient access by our GPU. (*) It's part of the basic premise of HMM as I understand it. I would expect the same thing to work for DEVICE_PUBLIC memory.
(*) I believe migrating file-backed pages to DEVICE_PRIVATE doesn't currently work, but that's something I'm hoping to fix at some point.
FWIW, I'd love to see the architecture documents that define how filesystems are supposed to interact with this device private memory. This whole "hand filesystem controlled memory to other devices" is a minefield that is trivial to get wrong iand very difficult to fix - just look at the historical mess that RDMA to/from file backed and/or DAX pages has been.
So, really, from my perspective as a filesystem engineer, I want to see an actual specification for how this new memory type is going to interact with filesystem and the page cache so everyone has some idea of how this is going to work and can point out how it doesn't work before code that simply doesn't work is pushed out into production systems and then merged....
OK. To be clear, that's not part of this patch series. And I have no authority to push anything in this part of the kernel, so you have nothing to fear. ;)
I know this isn't part of the series. but this patchset is laying the foundation for future work that will impact subsystems that currently have zero visibility and/or knowledge of these changes. There must be an overall architectural plan for this functionality, regardless of the current state of implementation. It's that overall architectural plan I'm asking about here, because I need to understand that before I can sanely comment on the page cache/filesystem aspect of the proposed functionality...
FWIW, we already have the ability to map file-backed system memory pages into device page tables with HMM and interval notifiers. But we cannot currently migrate them to ZONE_DEVICE pages.
Sure, but sharing page cache pages allocated and managed by the filesystem is not what you are talking about. You're talking about migrating page cache data to completely different memory allocated by a different memory manager that the filesystems currently have no knowledge of or have any way of interfacing with.
So I'm asking basic, fundamental questions about how these special device based pages are going to work. How are these pages different to normal pages - does page_lock() still guarantee exclusive access to the page state across all hardware that can access it? If not, what provides access serialisation for pages that are allocated in device memory rather than CPU memory (e.g. for truncate serialisation)? Does the hardware that own these pages raise page faults on the CPU when those pages are accessed/dirtied? How does demand paging in and out of device memory work (i.e. mapping files larger than device memory). How does IO to/from storage work - can the filesystem build normal bios out of these device pages and issue IO on them? Are the additional constraints on IO because p2p DMA is needed to move the data from the storage HBA directly into/out of the GPU memory?
I can think of lots more complex questions about how filesystems are supposed to manage remote device memory in the page cache, but these are just some of the basic things that make file-backed mappings different to anonymous mappings that I need to understand before I can make head or tail of what is being proposed here.....
Beyond that, my understanding of how filesystems and page cache work is rather superficial at this point. I'll keep your name in mind for when I am ready to discuss this in more detail.
If you don't know what the bigger picture is, then who does? Somebody built the design/architecture you are working towards, and they had to communicate it to you somehow. I'm asking for that information to documented and made available to all the people these changes might impact, not whether you personally know how it works....
Cheers,
Dave.
Am 2021-09-01 um 9:14 p.m. schrieb Dave Chinner:
On Wed, Sep 01, 2021 at 07:07:34PM -0400, Felix Kuehling wrote:
On 2021-09-01 6:03 p.m., Dave Chinner wrote:
On Wed, Sep 01, 2021 at 11:40:43AM -0400, Felix Kuehling wrote:
Am 2021-09-01 um 4:29 a.m. schrieb Christoph Hellwig:
On Mon, Aug 30, 2021 at 01:04:43PM -0400, Felix Kuehling wrote:
>> driver code is not really involved in updating the CPU mappings. Maybe >> it's something we need to do in the migration helpers. > It looks like I'm totally misunderstanding what you are adding here > then. Why do we need any special treatment at all for memory that > has normal struct pages and is part of the direct kernel map? The pages are like normal memory for purposes of mapping them in CPU page tables and for coherent access from the CPU.
That's the user page tables. What about the kernel direct map? If there is a normal kernel struct page backing there really should be no need for the pgmap.
I'm not sure. The physical address ranges are in the UEFI system address map as special-purpose memory. Does Linux create the struct pages and kernel direct map for that without a pgmap call? I didn't see that last time I went digging through that code.
From an application perspective, we want file-backed and anonymous mappings to be able to use DEVICE_PUBLIC pages with coherent CPU access. The goal is to optimize performance for GPU heavy workloads while minimizing the need to migrate data back-and-forth between system memory and device memory.
I don't really understand that part. file backed pages are always allocated by the file system using the pagecache helpers, that is using the page allocator. Anonymouns memory also always comes from the page allocator.
I'm coming at this from my experience with DEVICE_PRIVATE. Both anonymous and file-backed pages should be migrateable to DEVICE_PRIVATE memory by the migrate_vma_* helpers for more efficient access by our GPU. (*) It's part of the basic premise of HMM as I understand it. I would expect the same thing to work for DEVICE_PUBLIC memory.
(*) I believe migrating file-backed pages to DEVICE_PRIVATE doesn't currently work, but that's something I'm hoping to fix at some point.
FWIW, I'd love to see the architecture documents that define how filesystems are supposed to interact with this device private memory. This whole "hand filesystem controlled memory to other devices" is a minefield that is trivial to get wrong iand very difficult to fix - just look at the historical mess that RDMA to/from file backed and/or DAX pages has been.
So, really, from my perspective as a filesystem engineer, I want to see an actual specification for how this new memory type is going to interact with filesystem and the page cache so everyone has some idea of how this is going to work and can point out how it doesn't work before code that simply doesn't work is pushed out into production systems and then merged....
OK. To be clear, that's not part of this patch series. And I have no authority to push anything in this part of the kernel, so you have nothing to fear. ;)
I know this isn't part of the series. but this patchset is laying the foundation for future work that will impact subsystems that currently have zero visibility and/or knowledge of these changes.
I don't think this patchset is the foundation. Jerome Glisse's work on HMM is, which was merged 4 years ago and is being used by multiple drivers now, with the AMD GPU driver being a fairly recent addition.
There must be an overall architectural plan for this functionality, regardless of the current state of implementation. It's that overall architectural plan I'm asking about here, because I need to understand that before I can sanely comment on the page cache/filesystem aspect of the proposed functionality...
The overall HMM and ZONE_DEVICE architecture is documented to some extent in Documentation/vm/hmm.rst, though it may not go into the level of detail you are looking for.
FWIW, we already have the ability to map file-backed system memory pages into device page tables with HMM and interval notifiers. But we cannot currently migrate them to ZONE_DEVICE pages.
Sure, but sharing page cache pages allocated and managed by the filesystem is not what you are talking about. You're talking about migrating page cache data to completely different memory allocated by a different memory manager that the filesystems currently have no knowledge of or have any way of interfacing with.
This is not part of the current patch series. It is only my intention to look into ways to migrate file-backed pages to ZONE_DEVICE memory in the future.
So I'm asking basic, fundamental questions about how these special device based pages are going to work. How are these pages different to normal pages - does page_lock() still guarantee exclusive access to the page state across all hardware that can access it?
Yes. The migration API guarantees that pages are locked during the migration. The driver code doesn't touch the page state itself. It only uses the migrate_vma_* helpers to deal with that.
This is not new or changed by this patch series.
If not, what provides access serialisation for pages that are allocated in device memory rather than CPU memory (e.g. for truncate serialisation)? Does the hardware that own these pages raise page faults on the CPU when those pages are accessed/dirtied?
Yes. This is done by the hmm_range_fault API, which the driver calls in order to populate its device page tables. It is synchronized with any mapping changes through mmu_interval_notifiers.
This is not new or changed by this patch series.
How does demand paging in and out of device memory work (i.e. mapping files larger than device memory).
That depends on how the device driver handles device page faults. The AMD GPU driver can handle recoverable device page faults and update the device page table on demand with updated pfns from hmm_range_fault.
This is not new or changed by this patch series.
How does IO to/from storage work - can the filesystem build normal bios out of these device pages and issue IO on them?
DEVICE_PUBLIC pages introduced by this patch series, are CPU and peer-accessible like normal system memory.
DEVICE_PRIVATE pages are not CPU or peer-accessible. Any access to them would go through the CPU page fault path and cause a dev_pagemap_ops.migrate_to_ram callback into the AMD GPU driver to unmap the memory from the GPU and migrate it back to system memory.
Are the additional constraints on IO because p2p DMA is needed to move the data from the storage HBA directly into/out of the GPU memory?
I can think of lots more complex questions about how filesystems are supposed to manage remote device memory in the page cache, but these are just some of the basic things that make file-backed mappings different to anonymous mappings that I need to understand before I can make head or tail of what is being proposed here.....
Beyond that, my understanding of how filesystems and page cache work is rather superficial at this point. I'll keep your name in mind for when I am ready to discuss this in more detail.
If you don't know what the bigger picture is, then who does? Somebody built the design/architecture you are working towards, and they had to communicate it to you somehow. I'm asking for that information to documented and made available to all the people these changes might impact, not whether you personally know how it works....
This patch series builds on top of existing HMM work with major contributions from several people on this thread: Jerome Glisse, Jason Gunthorpe, Christoph Hellwig, Ralph Campbell.
Beyond the reintroduction of DEVICE_PUBLIC memory in this patch series I'm not looking to invent a major new design here. Immediate future work is more about chipping away on a few remaining limitations of the implementation, with respect to migration of file-backed pages and maybe transparent huge pages.
Regards, Felix
Cheers,
Dave.
dri-devel@lists.freedesktop.org
-
Dave Chinner -
Felix Kuehling