Re: [PATCH v4 0/4] TTM unlockable restartable LRU list iteration
Hi! On Mon, 2024-03-11 at 14:07 +0100, Thomas Hellström wrote: > On Fri, 2024-03-08 at 13:13 +0530, Somalapuram, Amaranath wrote: > > Patches are tested on AMD platform. > > Repeated stress test on Unigine Heaven, memory full (VRAM + GTT + > > system > > SWAP), then free. > > No errors/warning in kernel log. > > Any suggestion specific tests? > > We are testing locally against Intel Xe CI and Intel i915 CI which > should give rather good coverage. If there are some amdgpu tests that > exercise eviction / swapping also with a lot of local objects (Vulkan > apps?) that would be great. > > Thanks, > Thomas > Any updates on this? FWIW, For patch 3, IMO after looking a bit at other solutions, IMO this is the preferred solution mostly because it is self-contained. In particular if we allow drivers to iterate over the LRU lists with this interface, most likely if we add semantics like "You must block any bulk lru bumping if unlocking the lru_lock" That becomes pretty nasty and will most likely end up incorrect. It might well be that we've traversed well into a bulk move lru sublist before we try to unlock. /Thomas
Re: [PATCH v4 0/4] TTM unlockable restartable LRU list iteration
On Fri, 2024-03-08 at 13:13 +0530, Somalapuram, Amaranath wrote: > Patches are tested on AMD platform. > Repeated stress test on Unigine Heaven, memory full (VRAM + GTT + > system > SWAP), then free. > No errors/warning in kernel log. > Any suggestion specific tests? We are testing locally against Intel Xe CI and Intel i915 CI which should give rather good coverage. If there are some amdgpu tests that exercise eviction / swapping also with a lot of local objects (Vulkan apps?) that would be great. Thanks, Thomas > > Regards, > S.Amarnath > On 3/6/2024 12:31 PM, Thomas Hellström wrote: > > This patch-set is a prerequisite for a standalone TTM shrinker > > and for exhaustive TTM eviction using sleeping dma_resv locks, > > which is the motivation for it. > > > > Currently when unlocking the TTM lru list lock, iteration needs > > to be restarted from the beginning, rather from the next LRU list > > node. This can potentially be a big problem, because if eviction > > or shrinking fails for whatever reason after unlock, restarting > > is likely to cause the same failure over and over again. > > > > There are various schemes to be able to continue the list > > iteration from where we left off. One such scheme used by the > > GEM LRU list traversal is to pull items already considered off > > the LRU list and reinsert them when iteration is done. > > This has the drawback that concurrent list iteration doesn't see > > the complete list (which is bad for exhaustive eviction) and also > > doesn't lend itself well to bulk-move sublists since these will > > be split in the process where items from those lists are > > temporarily pulled from the list and moved to the list tail. > > > > The approach taken here is that list iterators insert themselves > > into the list next position using a special list node. Iteration > > is then using that list node as starting point when restarting. > > Concurrent iterators just skip over the special list nodes. > > > > This is implemented in patch 1 and 2. > > > > For bulk move sublist the approach is the same, but when a bulk > > move sublist is moved to the tail, the iterator is also moved, > > causing us to skip parts of the list. That is undesirable. > > Patch 3 deals with that, and when iterator detects it is > > traversing a sublist, it registers with the ttm_lru_bulk_move > > struct using a linked list, and when that bulk move sublist > > is moved to the tail, any iterator registered with it will > > first be moved to the tail of the sublist. > > This is implemented in patch 3. > > > > The restartable property is used in patch 4 to restart swapout if > > needed, but the main purpose is this paves the way for > > shrinker- and exhaustive eviction. > > > > v2: > > - Rework patch 3 completely. > > v3: > > - Fix a NULL pointer dereference found by Xe CI. > > v4: > > - Remove some leftover code causing build problems. > > > > Cc: Somalapuram Amaranath > > Cc: Christian König > > Cc: > > > > Thomas Hellström (4): > > drm/ttm: Allow TTM LRU list nodes of different types > > drm/ttm: Use LRU hitches > > drm/ttm, drm/amdgpu, drm/xe: Consider hitch moves within bulk > > sublist > > moves > > drm/ttm: Allow continued swapout after -ENOSPC falure > > > > drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 + > > drivers/gpu/drm/ttm/ttm_bo.c | 1 + > > drivers/gpu/drm/ttm/ttm_device.c | 33 +++- > > drivers/gpu/drm/ttm/ttm_resource.c | 228 > > - > > drivers/gpu/drm/xe/xe_vm.c | 4 + > > include/drm/ttm/ttm_device.h | 2 + > > include/drm/ttm/ttm_resource.h | 96 +-- > > 7 files changed, 308 insertions(+), 60 deletions(-) > >
Re: [PATCH v4 0/4] TTM unlockable restartable LRU list iteration
Patches are tested on AMD platform. Repeated stress test on Unigine Heaven, memory full (VRAM + GTT + system SWAP), then free. No errors/warning in kernel log. Any suggestion specific tests? Regards, S.Amarnath On 3/6/2024 12:31 PM, Thomas Hellström wrote: This patch-set is a prerequisite for a standalone TTM shrinker and for exhaustive TTM eviction using sleeping dma_resv locks, which is the motivation for it. Currently when unlocking the TTM lru list lock, iteration needs to be restarted from the beginning, rather from the next LRU list node. This can potentially be a big problem, because if eviction or shrinking fails for whatever reason after unlock, restarting is likely to cause the same failure over and over again. There are various schemes to be able to continue the list iteration from where we left off. One such scheme used by the GEM LRU list traversal is to pull items already considered off the LRU list and reinsert them when iteration is done. This has the drawback that concurrent list iteration doesn't see the complete list (which is bad for exhaustive eviction) and also doesn't lend itself well to bulk-move sublists since these will be split in the process where items from those lists are temporarily pulled from the list and moved to the list tail. The approach taken here is that list iterators insert themselves into the list next position using a special list node. Iteration is then using that list node as starting point when restarting. Concurrent iterators just skip over the special list nodes. This is implemented in patch 1 and 2. For bulk move sublist the approach is the same, but when a bulk move sublist is moved to the tail, the iterator is also moved, causing us to skip parts of the list. That is undesirable. Patch 3 deals with that, and when iterator detects it is traversing a sublist, it registers with the ttm_lru_bulk_move struct using a linked list, and when that bulk move sublist is moved to the tail, any iterator registered with it will first be moved to the tail of the sublist. This is implemented in patch 3. The restartable property is used in patch 4 to restart swapout if needed, but the main purpose is this paves the way for shrinker- and exhaustive eviction. v2: - Rework patch 3 completely. v3: - Fix a NULL pointer dereference found by Xe CI. v4: - Remove some leftover code causing build problems. Cc: Somalapuram Amaranath Cc: Christian König Cc: Thomas Hellström (4): drm/ttm: Allow TTM LRU list nodes of different types drm/ttm: Use LRU hitches drm/ttm, drm/amdgpu, drm/xe: Consider hitch moves within bulk sublist moves drm/ttm: Allow continued swapout after -ENOSPC falure drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 + drivers/gpu/drm/ttm/ttm_bo.c | 1 + drivers/gpu/drm/ttm/ttm_device.c | 33 +++- drivers/gpu/drm/ttm/ttm_resource.c | 228 - drivers/gpu/drm/xe/xe_vm.c | 4 + include/drm/ttm/ttm_device.h | 2 + include/drm/ttm/ttm_resource.h | 96 +-- 7 files changed, 308 insertions(+), 60 deletions(-)
[PATCH v4 0/4] TTM unlockable restartable LRU list iteration
This patch-set is a prerequisite for a standalone TTM shrinker and for exhaustive TTM eviction using sleeping dma_resv locks, which is the motivation for it. Currently when unlocking the TTM lru list lock, iteration needs to be restarted from the beginning, rather from the next LRU list node. This can potentially be a big problem, because if eviction or shrinking fails for whatever reason after unlock, restarting is likely to cause the same failure over and over again. There are various schemes to be able to continue the list iteration from where we left off. One such scheme used by the GEM LRU list traversal is to pull items already considered off the LRU list and reinsert them when iteration is done. This has the drawback that concurrent list iteration doesn't see the complete list (which is bad for exhaustive eviction) and also doesn't lend itself well to bulk-move sublists since these will be split in the process where items from those lists are temporarily pulled from the list and moved to the list tail. The approach taken here is that list iterators insert themselves into the list next position using a special list node. Iteration is then using that list node as starting point when restarting. Concurrent iterators just skip over the special list nodes. This is implemented in patch 1 and 2. For bulk move sublist the approach is the same, but when a bulk move sublist is moved to the tail, the iterator is also moved, causing us to skip parts of the list. That is undesirable. Patch 3 deals with that, and when iterator detects it is traversing a sublist, it registers with the ttm_lru_bulk_move struct using a linked list, and when that bulk move sublist is moved to the tail, any iterator registered with it will first be moved to the tail of the sublist. This is implemented in patch 3. The restartable property is used in patch 4 to restart swapout if needed, but the main purpose is this paves the way for shrinker- and exhaustive eviction. v2: - Rework patch 3 completely. v3: - Fix a NULL pointer dereference found by Xe CI. v4: - Remove some leftover code causing build problems. Cc: Somalapuram Amaranath Cc: Christian König Cc: Thomas Hellström (4): drm/ttm: Allow TTM LRU list nodes of different types drm/ttm: Use LRU hitches drm/ttm, drm/amdgpu, drm/xe: Consider hitch moves within bulk sublist moves drm/ttm: Allow continued swapout after -ENOSPC falure drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 + drivers/gpu/drm/ttm/ttm_bo.c | 1 + drivers/gpu/drm/ttm/ttm_device.c | 33 +++- drivers/gpu/drm/ttm/ttm_resource.c | 228 - drivers/gpu/drm/xe/xe_vm.c | 4 + include/drm/ttm/ttm_device.h | 2 + include/drm/ttm/ttm_resource.h | 96 +-- 7 files changed, 308 insertions(+), 60 deletions(-) -- 2.44.0
