Add support for SVM atomics in Nouveau

Device private pages are used to represent device memory that is not directly accessible from the CPU. Extra references to a device private page are only used to ensure the struct page itself remains valid whilst waiting for migration entries. Therefore extra references should not prevent device private page migration as this can lead to failures to migrate pages back to the CPU which are fatal to the user process.

Quoted 6 lines · Alistair Popple

Device private pages are used to represent device memory that is not directly accessible from the CPU. Extra references to a device private page are only used to ensure the struct page itself remains valid whilst waiting for migration entries. Therefore extra references should not prevent device private page migration as this can lead to failures to migrate pages back to the CPU which are fatal to the user process.

This should identify the extra references in expected_count, just disabling this protection seems unsafe, ZONE_DEVICE is not so special that the refcount means nothing

Is this a side effect of the extra refcounts that Ralph was trying to get rid of? I'd rather see that work finished :)

Jason

Quoted 11 lines · AEDT Jason Gunthorpe

Quoted 6 lines · Alistair Popple

Device private pages are used to represent device memory that is not directly accessible from the CPU. Extra references to a device private page are only used to ensure the struct page itself remains valid whilst waiting for migration entries. Therefore extra references should not prevent device private page migration as this can lead to failures to migrate pages back to the CPU which are fatal to the user process.

This should identify the extra references in expected_count, just disabling this protection seems unsafe, ZONE_DEVICE is not so special that the refcount means nothing

This is similar to what migarte_vma_check_page() does now. The issue is that a migration wait takes a reference on the device private page so you can end up with one thread stuck waiting for migration whilst the other can't migrate due to the extra refcount.

Given device private pages can't undergo GUP and that it's not possible to differentiate the migration wait refcount from any other refcount we assume any possible extra reference must be from migration wait.

This series adds support to Nouveau for atomic memory operations on OpenCL shared virtual memory (SVM). This is achieved using the atomic PTE bits on the GPU to only permit atomic operations to system memory when a page is not mapped in userspace on the CPU.

This is implemented by adding a mode to migrate_vma_pages() which unmaps and isolates existing pages from the CPU and pins them. The original userspace page table entries are migrated to point to device private pages allocated by the driver. This allows the driver to enable GPU atomic access to the page as it will receive a callback when CPU userspace needs to access it.

In response to this callback the driver revokes the atomic access permission from the GPU and migrates entries to point back to the original page. The original page is unpinned as part of the migration operation which also returns it to the LRU.

Patch 3 contains the bulk of the memory management changes to implement unmap and pin.

Patches 6-9 extend Nouveau to use the new mode to allow system wide atomics for OpenCL SVM to be implemented on Nouveau.

This has been tested using the latest upstream Mesa userspace with a simple OpenCL test program which checks the results of atomic GPU operations on a buffer whilst also writing to the same buffer from the CPU.

Problems yet to be addressed:

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in ZONE_MOVABLE as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can be unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Alistair Popple (9):
mm/migrate.c: Always allow device private pages to migrate mm/migrate.c: Allow pfn flags to be passed to migrate_vma_setup() mm/migrate: Add a unmap and pin migration mode
Documentation: Add unmap and pin to HMM
hmm-tests: Add test for unmap and pin
nouveau/dmem: Only map migrating pages
nouveau/svm: Refactor nouveau_range_fault nouveau/dmem: Add support for multiple page types nouveau/svm: Implement atomic SVM access

 Documentation/vm/hmm.rst                      |  22 +-
 arch/powerpc/kvm/book3s_hv_uvmem.c            |   4 +-
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |   1 +
 drivers/gpu/drm/nouveau/nouveau_dmem.c        | 190 +++++++++++++++---
 drivers/gpu/drm/nouveau/nouveau_dmem.h        |   9 +
 drivers/gpu/drm/nouveau/nouveau_svm.c         | 148 +++++++++++---
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   1 +
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   6 +
 include/linux/migrate.h                       |   2 +
 include/linux/migrate_mode.h                  |   1 +
 lib/test_hmm.c                                | 109 ++++++++--
 lib/test_hmm_uapi.h                           |   1 +
 mm/migrate.c                                  |  82 +++++---
 tools/testing/selftests/vm/hmm-tests.c        |  49 +++++
 14 files changed, 524 insertions(+), 101 deletions(-)

--
2.20.1

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

That would avoid all the hacking around long term pin constraints, because for real unmoveable long term pinned memory we really want to have all these checks. So I think we might be missing some other callbacks to be able to move these pages, instead of abusing longterm pins for lack of better tools.

Cheers, Daniel

Quoted 28 lines

Alistair Popple (9):
mm/migrate.c: Always allow device private pages to migrate mm/migrate.c: Allow pfn flags to be passed to migrate_vma_setup() mm/migrate: Add a unmap and pin migration mode
Documentation: Add unmap and pin to HMM
hmm-tests: Add test for unmap and pin
nouveau/dmem: Only map migrating pages
nouveau/svm: Refactor nouveau_range_fault nouveau/dmem: Add support for multiple page types nouveau/svm: Implement atomic SVM access

 Documentation/vm/hmm.rst                      |  22 +-
 arch/powerpc/kvm/book3s_hv_uvmem.c            |   4 +-
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |   1 +
 drivers/gpu/drm/nouveau/nouveau_dmem.c        | 190 +++++++++++++++---
 drivers/gpu/drm/nouveau/nouveau_dmem.h        |   9 +
 drivers/gpu/drm/nouveau/nouveau_svm.c         | 148 +++++++++++---
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   1 +
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   6 +
 include/linux/migrate.h                       |   2 +
 include/linux/migrate_mode.h                  |   1 +
 lib/test_hmm.c                                | 109 ++++++++--
 lib/test_hmm_uapi.h                           |   1 +
 mm/migrate.c                                  |  82 +++++---
 tools/testing/selftests/vm/hmm-tests.c        |  49 +++++
 14 files changed, 524 insertions(+), 101 deletions(-)

Quoted 3 lines

--
2.20.1

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed.

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

Yes, I would like to avoid the long term pin constraints as well if possible I just haven't found a solution yet. Are you suggesting it might be possible to add a callback in the page migration logic to specially deal with moving these pages?

Quoted 19 lines · Alistair Popple

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Quoted 4 lines

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed.

The issue is you took the page out of the PTE it belongs to, which makes it orphaned and unlocatable by the rest of the mm?

Ideally this would leave the PTE in place so everything continues to work, just disable CPU access to it.

Maybe some kind of special swap entry?

I also don't much like the use of ZONE_DEVICE here, that should only be used for actual device memory, not as a temporary proxy for CPU pages.. Having two struct pages refer to the same physical memory is pretty ugly.

Quoted 4 lines

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

The fact the page is lost from the MM seems to be the main issue here.

Quoted 4 lines

Yes, I would like to avoid the long term pin constraints as well if possible I just haven't found a solution yet. Are you suggesting it might be possible to add a callback in the page migration logic to specially deal with moving these pages?

How would migration even find the page?

Jason

Either way ZONE_DEVICE for not vram/device memory sounds wrong. Is that really going on here?

Quoted 19 lines · Alistair Popple

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Quoted 4 lines

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed.

The issue is you took the page out of the PTE it belongs to, which makes it orphaned and unlocatable by the rest of the mm?

Ideally this would leave the PTE in place so everything continues to work, just disable CPU access to it.

Maybe some kind of special swap entry?

I also don't much like the use of ZONE_DEVICE here, that should only be used for actual device memory, not as a temporary proxy for CPU pages.. Having two struct pages refer to the same physical memory is pretty ugly.

Quoted 4 lines

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

The fact the page is lost from the MM seems to be the main issue here.

Quoted 4 lines

Yes, I would like to avoid the long term pin constraints as well if possible I just haven't found a solution yet. Are you suggesting it might be possible to add a callback in the page migration logic to specially deal with moving these pages?

How would migration even find the page?

Quoted 6 lines

Quoted 4 lines

Yes, I would like to avoid the long term pin constraints as well if possible I just haven't found a solution yet. Are you suggesting it might be possible to add a callback in the page migration logic to specially deal with moving these pages?

How would migration even find the page?

Migration can scan memory from physical address (isolate_migratepages_range()) So the CPU mapping is not the only path to get to a page.

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed. Obviously I could have just used get_page() but from the perspective of page migration the result is much the same as a pin - a page which can't be moved because of the extra refcount.

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

Quoted 5 lines

That would avoid all the hacking around long term pin constraints, because for real unmoveable long term pinned memory we really want to have all these checks. So I think we might be missing some other callbacks to be able to move these pages, instead of abusing longterm pins for lack of better tools.

Yes, I would like to avoid the long term pin constraints as well if possible I just haven't found a solution yet. Are you suggesting it might be possible to add a callback in the page migration logic to specially deal with moving these pages?

Thanks, Alistair

Quoted 40 lines

Cheers, Daniel

Quoted 28 lines

Alistair Popple (9):
mm/migrate.c: Always allow device private pages to migrate mm/migrate.c: Allow pfn flags to be passed to migrate_vma_setup() mm/migrate: Add a unmap and pin migration mode
Documentation: Add unmap and pin to HMM
hmm-tests: Add test for unmap and pin
nouveau/dmem: Only map migrating pages
nouveau/svm: Refactor nouveau_range_fault nouveau/dmem: Add support for multiple page types nouveau/svm: Implement atomic SVM access

 Documentation/vm/hmm.rst                      |  22 +-
 arch/powerpc/kvm/book3s_hv_uvmem.c            |   4 +-
 drivers/gpu/drm/nouveau/include/nvif/if000c.h |   1 +
 drivers/gpu/drm/nouveau/nouveau_dmem.c        | 190 +++++++++++++++---
 drivers/gpu/drm/nouveau/nouveau_dmem.h        |   9 +
 drivers/gpu/drm/nouveau/nouveau_svm.c         | 148 +++++++++++---
 drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h |   1 +
 .../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c    |   6 +
 include/linux/migrate.h                       |   2 +
 include/linux/migrate_mode.h                  |   1 +
 lib/test_hmm.c                                | 109 ++++++++--
 lib/test_hmm_uapi.h                           |   1 +
 mm/migrate.c                                  |  82 +++++---
 tools/testing/selftests/vm/hmm-tests.c        |  49 +++++
 14 files changed, 524 insertions(+), 101 deletions(-)

Quoted 3 lines

--
2.20.1

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 20 lines · Alistair Popple

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Quoted 6 lines

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed. Obviously I could have just used get_page() but from the perspective of page migration the result is much the same as a pin - a page which can't be moved because of the extra refcount.

long term pin vs short term page reference aren't fully fleshed out. But the rule more or less is:
- short term page reference: _must_ get released in finite time for migration and other things, either because you have a callback, or because it's just for direct I/O, which will complete. This means short term pins will delay migration, but not foul it complete

- long term pin: the page cannot be moved, all migration must fail. Also this will have an impact on COW behaviour for fork (but not sure where those patches are, John Hubbard will know).

So I think for your use case here you want a) short term page reference to make sure it doesn't disappear plus b) callback to make sure migrate isn't blocked.

Breaking ZONE_MOVEABLE with either allowing long term pins or failing migrations because you don't release your short term page reference isn't good.

Quoted 4 lines

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

Yeah you need some other callback for migration on the page directly. it's a bit awkward since there is one already for struct address_space, but that's own by the address_space/page cache, not HMM. So I think we need something else, maybe something for each ZONE_DEVICE?

Quoted 28 lines · Daniel Vetter

Quoted 20 lines · Alistair Popple

Quoted 3 lines · AEDT Daniel Vetter

Quoted 3 lines

Recent changes to pin_user_pages() prevent the creation of pinned pages in ZONE_MOVABLE. This series allows pinned pages to be created in

ZONE_MOVABLE

Quoted 3 lines

Quoted 3 lines

as attempts to migrate may fail which would be fatal to userspace.

In this case migration of the pinned page is unnecessary as the page can

be

Quoted 9 lines

Quoted 4 lines

unpinned at anytime by having the driver revoke atomic permission as it does for the migrate_to_ram() callback. However a method of calling this when memory needs to be moved has yet to be resolved so any discussion is welcome.

Why do we need to pin for gpu atomics? You still have the callback for cpu faults, so you
can move the page as needed, and hence a long-term pin sounds like the wrong approach.

Quoted 6 lines

Technically a real long term unmoveable pin isn't required, because as you say the page can be moved as needed at any time. However I needed some way of stopping the CPU page from being freed once the userspace mappings for it had been removed. Obviously I could have just used get_page() but from the perspective of page migration the result is much the same as a pin - a page which can't be moved because of the extra refcount.

Quoted 6 lines

long term pin vs short term page reference aren't fully fleshed out. But the rule more or less is:
- short term page reference: _must_ get released in finite time for migration and other things, either because you have a callback, or because it's just for direct I/O, which will complete. This means short term pins will delay migration, but not foul it complete

GPU atomic operations to sysmem are hard to categorize, because because application programmers could easily write programs that do a long series of atomic operations. Such a program would be a little weird, but it's hard to rule out.

Quoted 3 lines

- long term pin: the page cannot be moved, all migration must fail. Also this will have an impact on COW behaviour for fork (but not sure where those patches are, John Hubbard will know).

That would be Jason's commit 57efa1fe59576 ("mm/gup: prevent gup_fast from racing with COW during fork"), which is in linux-next 20201216.

Quoted 16 lines

So I think for your use case here you want a) short term page reference to make sure it doesn't disappear plus b) callback to make sure migrate isn't blocked.

Breaking ZONE_MOVEABLE with either allowing long term pins or failing migrations because you don't release your short term page reference isn't good.

Quoted 4 lines

The normal solution of registering an MMU notifier to unpin the page when it needs to be moved also doesn't work as the CPU page tables now point to the device-private page and hence the migration code won't call any invalidate notifiers for the CPU page.

Quoted 6 lines

Yeah you need some other callback for migration on the page directly. it's a bit awkward since there is one already for struct address_space, but that's own by the address_space/page cache, not HMM. So I think we need something else, maybe something for each ZONE_DEVICE?

This direction sounds at least...possible. Using MMU notifiers instead of pins is definitely appealing. I'm not quite clear on the callback idea above, but overall it seems like taking advantage of the ZONE_DEVICE tracking of pages (without having to put anything additional in each struct page), could work.

Additional notes or ideas here are definitely welcome.

Quoted 3 lines

GPU atomic operations to sysmem are hard to categorize, because because application programmers could easily write programs that do a long series of atomic operations. Such a program would be a little weird, but it's hard to rule out.

Yeah, but we can forcefully break this whenever we feel like by revoking the page, moving it, and then reinstating the gpu pte again and let it continue.

If that's no possible then what we need here instead is an mlock() type of thing I think.

This direction sounds at least...possible. Using MMU notifiers instead of pins is definitely appealing. I'm not quite clear on the callback idea above, but overall it seems like taking advantage of the ZONE_DEVICE tracking of pages (without having to put anything additional in each struct page), could work.

Really what you want to do here is leave the CPU page in the VMA and the page tables where it started and deny CPU access to the page. Then all the proper machinery will continue to work.

IMHO "migration" is the wrong idea if the data isn't actually moving.

Quoted 5 lines · Jason Gunthorpe

Really what you want to do here is leave the CPU page in the VMA and the page tables where it started and deny CPU access to the page. Then all the proper machinery will continue to work.

IMHO "migration" is the wrong idea if the data isn't actually moving.

Agreed.