[PATCH v3 4/5] drm/amdgpu: use bulk moves for efficient VM LRU handling (v3)
Huang Rui
ray.huang at amd.com
Tue Aug 14 07:24:55 UTC 2018
On Tue, Aug 14, 2018 at 02:45:22PM +0800, Koenig, Christian wrote:
> Am 14.08.2018 um 05:05 schrieb Huang Rui:
> > On Tue, Aug 14, 2018 at 10:26:43AM +0800, Zhang, Jerry wrote:
> >> On 08/13/2018 05:58 PM, Huang Rui wrote:
> >>> I continue to work for bulk moving that based on the proposal by Christian.
> >>>
> >>> Background:
> >>> amdgpu driver will move all PD/PT and PerVM BOs into idle list. Then move all of
> >>> them on the end of LRU list one by one. Thus, that cause so many BOs moved to
> >>> the end of the LRU, and impact performance seriously.
> >>>
> >>> Then Christian provided a workaround to not move PD/PT BOs on LRU with below
> >>> patch:
> >>> "drm/amdgpu: band aid validating VM PTs"
> >>> Commit 0bbf32026cf5ba41e9922b30e26e1bed1ecd38ae
> >>>
> >>> However, the final solution should bulk move all PD/PT and PerVM BOs on the LRU
> >>> instead of one by one.
> >>>
> >>> Whenever amdgpu_vm_validate_pt_bos() is called and we have BOs which need to be
> >>> validated we move all BOs together to the end of the LRU without dropping the
> >>> lock for the LRU.
> >>>
> >>> While doing so we note the beginning and end of this block in the LRU list.
> >>>
> >>> Now when amdgpu_vm_validate_pt_bos() is called and we don't have anything to do,
> >>> we don't move every BO one by one, but instead cut the LRU list into pieces so
> >>> that we bulk move everything to the end in just one operation.
> >>>
> >>> Test data:
> >>> +--------------+-----------------+-----------+---------------------------------------+
> >>> | |The Talos |Clpeak(OCL)|BusSpeedReadback(OCL) |
> >>> | |Principle(Vulkan)| | |
> >>> +------------------------------------------------------------------------------------+
> >>> | | | |0.319 ms(1k) 0.314 ms(2K) 0.308 ms(4K) |
> >>> | Original | 147.7 FPS | 76.86 us |0.307 ms(8K) 0.310 ms(16K) |
> >>> +------------------------------------------------------------------------------------+
> >>> | Orignial + WA| | |0.254 ms(1K) 0.241 ms(2K) |
> >>> |(don't move | 162.1 FPS | 42.15 us |0.230 ms(4K) 0.223 ms(8K) 0.204 ms(16K)|
> >>> |PT BOs on LRU)| | | |
> >>> +------------------------------------------------------------------------------------+
> >>> | Bulk move | 163.1 FPS | 40.52 us |0.244 ms(1K) 0.252 ms(2K) 0.213 ms(4K) |
> >>> | | | |0.214 ms(8K) 0.225 ms(16K) |
> >>> +--------------+-----------------+-----------+---------------------------------------+
> >>>
> >>> After test them with above three benchmarks include vulkan and opencl. We can
> >>> see the visible improvement than original, and even better than original with
> >>> workaround.
> >>>
> >>> v2: move all BOs include idle, relocated, and moved list to the end of LRU and
> >>> put them together.
> >>> v3: remove unused parameter and use list_for_each_entry instead of the one with
> >>> save entry.
> >>>
> >>> Signed-off-by: Christian König <christian.koenig at amd.com>
> >>> Signed-off-by: Huang Rui <ray.huang at amd.com>
> >>> Tested-by: Mike Lothian <mike at fireburn.co.uk>
> >>> ---
> >>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 73 ++++++++++++++++++++++++++--------
> >>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 4 ++
> >>> 2 files changed, 61 insertions(+), 16 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> >>> index 9c84770..ee1af53 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> >>> @@ -268,6 +268,53 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
> >>> }
> >>>
> >>> /**
> >>> + * amdgpu_vm_move_to_lru_tail - move one list of BOs to end of LRU
> >>> + *
> >>> + * @vm: vm providing the BOs
> >>> + * @list: the list that stored BOs
> >>> + *
> >>> + * Move one list of BOs to the end of LRU and update the positions.
> >>> + */
> >>> +static void
> >>> +amdgpu_vm_move_to_lru_tail_by_list(struct amdgpu_vm *vm, struct list_head *list)
> >>> +{
> >>> + struct amdgpu_vm_bo_base *bo_base;
> >>> +
> >>> + list_for_each_entry(bo_base, list, vm_status) {
> >>> + struct amdgpu_bo *bo = bo_base->bo;
> >>> +
> >>> + if (!bo->parent)
> >>> + continue;
> >>> +
> >>> + ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
> >>> + if (bo->shadow)
> >>> + ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
> >>> + &vm->lru_bulk_move);
> >>> + }
> >>> +}
> >>> +
> >>> +/**
> >>> + * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
> >>> + *
> >>> + * @adev: amdgpu device pointer
> >>> + * @vm: vm providing the BOs
> >>> + *
> >>> + * Move all BOs to the end of LRU and remember their positions to put them
> >>> + * together.
> >>> + */
> >>> +static void
> >>> +amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, struct amdgpu_vm *vm)
> >>> +{
> >>> + struct ttm_bo_global *glob = adev->mman.bdev.glob;
> >>> +
> >>> + spin_lock(&glob->lru_lock);
> >>> + amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->idle);
> >>> + amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->relocated);
> >>> + amdgpu_vm_move_to_lru_tail_by_list(vm, &vm->moved);
> >> Moved list is working under vm->moved_lock, so we may hold that as well.
> >> otherwise, to use the same one.
> >> (not sure the detail of history about moved_lock)
> > We actually don't remove them from moved list, just move bo->lru to the end
> > of lru. So here, it doesn't need moved_lock.
> >
> >>> + spin_unlock(&glob->lru_lock);
> >>> +}
> >>> +
> >>> +/**
> >>> * amdgpu_vm_validate_pt_bos - validate the page table BOs
> >>> *
> >>> * @adev: amdgpu device pointer
> >>> @@ -286,6 +333,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> >>> {
> >>> struct ttm_bo_global *glob = adev->mman.bdev.glob;
> >>> struct amdgpu_vm_bo_base *bo_base, *tmp;
> >>> + bool validated = false;
> >>> int r = 0;
> >>>
> >>> list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status) {
> >>> @@ -295,14 +343,9 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> >>> r = validate(param, bo);
> >>> if (r)
> >>> break;
> >>> -
> >>> - spin_lock(&glob->lru_lock);
> >>> - ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
> >>> - if (bo->shadow)
> >>> - ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
> >>> - spin_unlock(&glob->lru_lock);
> >>> }
> >>>
> >>> + validated = true;
> >>> if (bo->tbo.type != ttm_bo_type_kernel) {
> >>> spin_lock(&vm->moved_lock);
> >>> list_move(&bo_base->vm_status, &vm->moved);
> >>> @@ -312,18 +355,16 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> >>> }
> >>> }
> >>>
> >>> - spin_lock(&glob->lru_lock);
> >>> - list_for_each_entry(bo_base, &vm->idle, vm_status) {
> >>> - struct amdgpu_bo *bo = bo_base->bo;
> >>> + if (!validated) {
> >>> + spin_lock(&glob->lru_lock);
> >>> + ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
> >> To confirm
> >> we only do actual bulk move when no evicted or validate failure?
> >>
> > Yes. Because if some BOs are evicted back, they will be moved to
> > moved/relocated list. Then we need update the positions of bulk moving.
>
> Ah, crap that won't work. Jerry pointed out a quite important bug here.
>
> The moved list contains both per-VM as well as independent BOs, so
> walking it and moving everything on the LRU won't work as expected.
>
Our purpose is not to move the independent BOs (shared with other VM),
right?
> Probably better to just walk the idle list after we are done with the
> state machine.
>
If we only walk the idle list here, we probably are not included all the
Per-VM BOs, right? Or walk the idle list after command submission.
Thanks,
Ray
> Christian.
>
> >
> > Thanks,
> > Ray
> >
> >> Regards,
> >> Jerry
> >>
> >>> + spin_unlock(&glob->lru_lock);
> >>> + return 0;
> >>> + }
> >>>
> >>> - if (!bo->parent)
> >>> - continue;
> >>> + memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
> >>>
> >>> - ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
> >>> - if (bo->shadow)
> >>> - ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
> >>> - }
> >>> - spin_unlock(&glob->lru_lock);
> >>> + amdgpu_vm_move_to_lru_tail(adev, vm);
> >>>
> >>> return r;
> >>> }
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> >>> index 67a15d4..92725ac 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> >>> @@ -29,6 +29,7 @@
> >>> #include <linux/rbtree.h>
> >>> #include <drm/gpu_scheduler.h>
> >>> #include <drm/drm_file.h>
> >>> +#include <drm/ttm/ttm_bo_driver.h>
> >>>
> >>> #include "amdgpu_sync.h"
> >>> #include "amdgpu_ring.h"
> >>> @@ -226,6 +227,9 @@ struct amdgpu_vm {
> >>>
> >>> /* Some basic info about the task */
> >>> struct amdgpu_task_info task_info;
> >>> +
> >>> + /* Store positions of group of BOs */
> >>> + struct ttm_lru_bulk_move lru_bulk_move;
> >>> };
> >>>
> >>> struct amdgpu_vm_manager {
> >>>
>
More information about the dri-devel
mailing list