[Intel-xe] [PATCH v2 8/8] drm/xe: VM LRU bulk move

Matthew Brost matthew.brost at intel.com
Tue May 23 13:38:06 UTC 2023


On Mon, May 22, 2023 at 11:52:43PM -0700, Christopher Snowhill wrote:
> On Mon, May 22, 2023 at 11:38 PM Christopher Snowhill <kode54 at gmail.com> wrote:
> >
> > On Mon, May 22, 2023 at 10:44 PM Matthew Brost <matthew.brost at intel.com> wrote:
> > >
> > > Use the TTM LRU bulk move for BOs tied to a VM. Update the bulk moves
> > > LRU position on every exec.
> > >
> > > v2: Bulk move for compute VMs, remove bulk LRU in xe_gem_object_free
> > >
> > > Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_bo.c       | 24 ++++++++++++++++++++----
> > >  drivers/gpu/drm/xe/xe_bo.h       |  4 ++--
> > >  drivers/gpu/drm/xe/xe_dma_buf.c  |  2 +-
> > >  drivers/gpu/drm/xe/xe_exec.c     |  6 ++++++
> > >  drivers/gpu/drm/xe/xe_vm.c       |  4 ++++
> > >  drivers/gpu/drm/xe/xe_vm_types.h |  3 +++
> > >  6 files changed, 36 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> > > index c82e995df779..22cf65bd63fe 100644
> > > --- a/drivers/gpu/drm/xe/xe_bo.c
> > > +++ b/drivers/gpu/drm/xe/xe_bo.c
> > > @@ -968,6 +968,16 @@ static void xe_ttm_bo_destroy(struct ttm_buffer_object *ttm_bo)
> > >
> > >  static void xe_gem_object_free(struct drm_gem_object *obj)
> > >  {
> > > +       struct xe_bo *bo = gem_to_xe_bo(obj);
> > > +
> > > +       if (bo->vm && !xe_vm_in_fault_mode(bo->vm) && xe_bo_is_user(bo)) {
> > > +               struct ww_acquire_ctx ww;
> > > +
> > > +               xe_bo_lock(bo, &ww, 0, false);
> > > +               ttm_bo_set_bulk_move(&bo->ttm, NULL);
> > > +               xe_bo_unlock(bo, &ww);
> > > +       }
> > > +
> > >         /* Our BO reference counting scheme works as follows:
> > >          *
> > >          * The gem object kref is typically used throughout the driver,
> > > @@ -1081,8 +1091,8 @@ void xe_bo_free(struct xe_bo *bo)
> > >
> > >  struct xe_bo *__xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > >                                     struct xe_gt *gt, struct dma_resv *resv,
> > > -                                   size_t size, enum ttm_bo_type type,
> > > -                                   u32 flags)
> > > +                                   struct ttm_lru_bulk_move *bulk, size_t size,
> > > +                                   enum ttm_bo_type type, u32 flags)
> > >  {
> > >         struct ttm_operation_ctx ctx = {
> > >                 .interruptible = true,
> > > @@ -1149,7 +1159,10 @@ struct xe_bo *__xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > >                 return ERR_PTR(err);
> > >
> > >         bo->created = true;
> > > -       ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
> > > +       if (bulk)
> > > +               ttm_bo_set_bulk_move(&bo->ttm, bulk);
> > > +       else
> > > +               ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
> > >
> > >         return bo;
> > >  }
> > > @@ -1219,7 +1232,10 @@ xe_bo_create_locked_range(struct xe_device *xe,
> > >                 }
> > >         }
> > >
> > > -       bo = __xe_bo_create_locked(xe, bo, gt, vm ? &vm->resv : NULL, size,
> > > +       bo = __xe_bo_create_locked(xe, bo, gt, vm ? &vm->resv : NULL,
> > > +                                  vm && !xe_vm_in_fault_mode(vm) &&
> > > +                                  flags & XE_BO_CREATE_USER_BIT ?
> > > +                                  &vm->lru_bulk_move : NULL, size,
> > >                                    type, flags);
> > >         if (IS_ERR(bo))
> > >                 return bo;
> > > diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h
> > > index 7e111332c35a..f7562012b836 100644
> > > --- a/drivers/gpu/drm/xe/xe_bo.h
> > > +++ b/drivers/gpu/drm/xe/xe_bo.h
> > > @@ -81,8 +81,8 @@ void xe_bo_free(struct xe_bo *bo);
> > >
> > >  struct xe_bo *__xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > >                                     struct xe_gt *gt, struct dma_resv *resv,
> > > -                                   size_t size, enum ttm_bo_type type,
> > > -                                   u32 flags);
> > > +                                   struct ttm_lru_bulk_move *bulk, size_t size,
> > > +                                   enum ttm_bo_type type, u32 flags);
> > >  struct xe_bo *
> > >  xe_bo_create_locked_range(struct xe_device *xe,
> > >                           struct xe_gt *gt, struct xe_vm *vm,
> > > diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
> > > index 9b252cc782b7..975dee1f770f 100644
> > > --- a/drivers/gpu/drm/xe/xe_dma_buf.c
> > > +++ b/drivers/gpu/drm/xe/xe_dma_buf.c
> > > @@ -199,7 +199,7 @@ xe_dma_buf_init_obj(struct drm_device *dev, struct xe_bo *storage,
> > >         int ret;
> > >
> > >         dma_resv_lock(resv, NULL);
> > > -       bo = __xe_bo_create_locked(xe, storage, NULL, resv, dma_buf->size,
> > > +       bo = __xe_bo_create_locked(xe, storage, NULL, resv, NULL, dma_buf->size,
> > >                                    ttm_bo_type_sg, XE_BO_CREATE_SYSTEM_BIT);
> > >         if (IS_ERR(bo)) {
> > >                 ret = PTR_ERR(bo);
> > > diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c
> > > index ff4df00f20a2..b2dcf34af16b 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec.c
> > > +++ b/drivers/gpu/drm/xe/xe_exec.c
> > > @@ -395,6 +395,12 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> > >         xe_sched_job_push(job);
> > >         xe_vm_reactivate_rebind(vm);
> > >
> > > +       if (!err && !xe_vm_no_dma_fences(vm)) {
> > > +               spin_lock(&xe->ttm.lru_lock);
> > > +               ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
> > > +               spin_unlock(&xe->ttm.lru_lock);
> > > +       }
> > > +
> > >  err_repin:
> > >         if (!xe_vm_no_dma_fences(vm))
> > >                 up_read(&vm->userptr.notifier_lock);
> > > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > > index 0398da1ef1e2..a5d65d0325d6 100644
> > > --- a/drivers/gpu/drm/xe/xe_vm.c
> > > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > > @@ -629,6 +629,10 @@ static void preempt_rebind_work_func(struct work_struct *w)
> > >
> > >  #undef retry_required
> > >
> > > +       spin_lock(&vm->xe->ttm.lru_lock);
> > > +       ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
> > > +       spin_unlock(&vm->xe->ttm.lru_lock);
> > > +
> > >         /* Point of no return. */
> > >         arm_preempt_fences(vm, &preempt_fences);
> > >         resume_and_reinstall_preempt_fences(vm);
> > > diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
> > > index fada7896867f..d3e99f22510d 100644
> > > --- a/drivers/gpu/drm/xe/xe_vm_types.h
> > > +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> > > @@ -164,6 +164,9 @@ struct xe_vm {
> > >         /** Protects @rebind_list and the page-table structures */
> > >         struct dma_resv resv;
> > >
> > > +       /** @lru_bulk_move: Bulk LRU move list for this VM's BOs */
> > > +       struct ttm_lru_bulk_move lru_bulk_move;
> > > +
> > >         u64 size;
> > >         struct rb_root vmas;
> > >
> > > --
> > > 2.34.1
> >
> > Something in this patch version, either moving the cleanup to
> > gem_free, or the move_tail later in the patch set, causes kmscube to
> > deadlock on terminate, forcing me to reboot the machine. Reboot got
> > stuck when it tried to terminate my user session, so I had to press
> > the reset button.
>

Any chance you have a dmesg for the deadlock on terminate? Hoping it
gives some clues to the issue.

> Verified, substituting patch 8 of 8 from v1 results in cleanly
> terminating kmscube.

It was Thomas's suggestion to move this from close to free, this is a
fairly old patch and couldn't recall why it was in close rather than
free. Now I think about this, I believe originally this was in free but
moved to close to fix a bug, will try to remind why it needs to be in
close.

Matt


More information about the Intel-xe mailing list