Message ID | 20200522030109.4845-1-chris@chris-wilson.co.uk (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | drm/i915/gem: Avoid waiting inside mmu_notifier_invalidate_range | expand |
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c index 8b0708708671..a398d817543e 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c @@ -126,9 +126,11 @@ userptr_mn_invalidate_range_start(struct mmu_notifier *_mn, } spin_unlock(&mn->lock); - ret = i915_gem_object_unbind(obj, - I915_GEM_OBJECT_UNBIND_ACTIVE | - I915_GEM_OBJECT_UNBIND_BARRIER); + ret = i915_gem_object_wait(obj, + I915_WAIT_ALL, + MAX_SCHEDULE_TIMEOUT); + if (ret == 0) + ret = i915_gem_object_unbind(obj, 0); if (ret == 0) ret = __i915_gem_object_put_pages(obj); i915_gem_object_put(obj);
Since the userptr may be invalidated from inside a reclaim path, we cannot wait on unbinding the object as we may hold any number of locks, including the active reference of the object we are trying to invalidate. E.g. [<0>] __i915_active_wait+0x70/0xc0 [i915] [<0>] i915_vma_unbind+0x2f/0x110 [i915] [<0>] i915_gem_object_unbind+0x17c/0x350 [i915] [<0>] userptr_mn_invalidate_range_start+0xb6/0x140 [i915] [<0>] __mmu_notifier_invalidate_range_start+0x105/0x150 [<0>] try_to_unmap_one+0x7fb/0x900 [<0>] rmap_walk_file+0xe4/0x240 [<0>] try_to_unmap+0x85/0xc0 [<0>] shrink_page_list+0x885/0xe80 [<0>] shrink_inactive_list+0x155/0x260 [<0>] shrink_lruvec+0x4ec/0x5f0 [<0>] shrink_node+0x15b/0x410 [<0>] try_to_free_pages+0x169/0x3a0 [<0>] __alloc_pages_slowpath.constprop.0+0x251/0xa00 [<0>] __alloc_pages_nodemask+0x144/0x170 [<0>] new_slab+0x259/0x280 [<0>] ___slab_alloc.isra.0.constprop.0+0x3dd/0x460 [<0>] __slab_alloc.isra.0.constprop.0+0x9/0x10 [<0>] kmem_cache_alloc+0x11a/0x130 [<0>] alloc_pd+0x17/0x60 [i915] [<0>] __gen8_ppgtt_alloc+0x2d7/0x350 [i915] [<0>] gen8_ppgtt_alloc+0x35/0x70 [i915] [<0>] ppgtt_bind_vma+0x29/0x70 [i915] This situation only occurs when the lru page shrinker tries to sacrifice one the objects it is trying to allocate for. However, we do have to wait for an active object so that we can revoke access to the page range. Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1702 Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)