diff options
author | Chris Wilson <chris@chris-wilson.co.uk> | 2019-10-04 14:39:58 +0100 |
---|---|---|
committer | Chris Wilson <chris@chris-wilson.co.uk> | 2019-10-04 15:39:02 +0100 |
commit | 2850748ef8763ab46958e43a4d1c445f29eeb37d (patch) | |
tree | ba213c8039460e33090e9d8bb39b91d102a3a85d /drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | |
parent | 11331125e1480ff786be9d2051301401b652bbe1 (diff) | |
download | linux-2850748ef8763ab46958e43a4d1c445f29eeb37d.tar.bz2 |
drm/i915: Pull i915_vma_pin under the vm->mutex
Replace the struct_mutex requirement for pinning the i915_vma with the
local vm->mutex instead. Note that the vm->mutex is tainted by the
shrinker (we require unbinding from inside fs-reclaim) and so we cannot
allocate while holding that mutex. Instead we have to preallocate
workers to do allocate and apply the PTE updates after we have we
reserved their slot in the drm_mm (using fences to order the PTE writes
with the GPU work and with later unbind).
In adding the asynchronous vma binding, one subtle requirement is to
avoid coupling the binding fence into the backing object->resv. That is
the asynchronous binding only applies to the vma timeline itself and not
to the pages as that is a more global timeline (the binding of one vma
does not need to be ordered with another vma, nor does the implicit GEM
fencing depend on a vma, only on writes to the backing store). Keeping
the vma binding distinct from the backing store timelines is verified by
a number of async gem_exec_fence and gem_exec_schedule tests. The way we
do this is quite simple, we keep the fence for the vma binding separate
and only wait on it as required, and never add it to the obj->resv
itself.
Another consequence in reducing the locking around the vma is the
destruction of the vma is no longer globally serialised by struct_mutex.
A natural solution would be to add a kref to i915_vma, but that requires
decoupling the reference cycles, possibly by introducing a new
i915_mm_pages object that is own by both obj->mm and vma->pages.
However, we have not taken that route due to the overshadowing lmem/ttm
discussions, and instead play a series of complicated games with
trylocks to (hopefully) ensure that only one destruction path is called!
v2: Add some commentary, and some helpers to reduce patch churn.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191004134015.13204-4-chris@chris-wilson.co.uk
Diffstat (limited to 'drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c')
-rw-r--r-- | drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | 43 |
1 files changed, 10 insertions, 33 deletions
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index 228ce24ea280..42c15e8bf166 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -698,7 +698,9 @@ static int eb_reserve(struct i915_execbuffer *eb) case 1: /* Too fragmented, unbind everything and retry */ + mutex_lock(&eb->context->vm->mutex); err = i915_gem_evict_vm(eb->context->vm); + mutex_unlock(&eb->context->vm->mutex); if (err) return err; break; @@ -972,7 +974,9 @@ static void reloc_cache_reset(struct reloc_cache *cache) ggtt->vm.clear_range(&ggtt->vm, cache->node.start, cache->node.size); + mutex_lock(&ggtt->vm.mutex); drm_mm_remove_node(&cache->node); + mutex_unlock(&ggtt->vm.mutex); } else { i915_vma_unpin((struct i915_vma *)cache->node.mm); } @@ -1047,11 +1051,13 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, PIN_NOEVICT); if (IS_ERR(vma)) { memset(&cache->node, 0, sizeof(cache->node)); + mutex_lock(&ggtt->vm.mutex); err = drm_mm_insert_node_in_range (&ggtt->vm.mm, &cache->node, PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE, 0, ggtt->mappable_end, DRM_MM_INSERT_LOW); + mutex_unlock(&ggtt->vm.mutex); if (err) /* no inactive aperture space, use cpu reloc */ return NULL; } else { @@ -1416,7 +1422,7 @@ eb_relocate_entry(struct i915_execbuffer *eb, if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION && IS_GEN(eb->i915, 6)) { err = i915_vma_bind(target, target->obj->cache_level, - PIN_GLOBAL); + PIN_GLOBAL, NULL); if (WARN_ONCE(err, "Unexpected failure to bind target VMA!")) return err; @@ -2142,35 +2148,6 @@ static struct i915_request *eb_throttle(struct intel_context *ce) return i915_request_get(rq); } -static int -__eb_pin_context(struct i915_execbuffer *eb, struct intel_context *ce) -{ - int err; - - if (likely(atomic_inc_not_zero(&ce->pin_count))) - return 0; - - err = mutex_lock_interruptible(&eb->i915->drm.struct_mutex); - if (err) - return err; - - err = __intel_context_do_pin(ce); - mutex_unlock(&eb->i915->drm.struct_mutex); - - return err; -} - -static void -__eb_unpin_context(struct i915_execbuffer *eb, struct intel_context *ce) -{ - if (likely(atomic_add_unless(&ce->pin_count, -1, 1))) - return; - - mutex_lock(&eb->i915->drm.struct_mutex); - intel_context_unpin(ce); - mutex_unlock(&eb->i915->drm.struct_mutex); -} - static int __eb_pin_engine(struct i915_execbuffer *eb, struct intel_context *ce) { struct intel_timeline *tl; @@ -2190,7 +2167,7 @@ static int __eb_pin_engine(struct i915_execbuffer *eb, struct intel_context *ce) * GGTT space, so do this first before we reserve a seqno for * ourselves. */ - err = __eb_pin_context(eb, ce); + err = intel_context_pin(ce); if (err) return err; @@ -2234,7 +2211,7 @@ err_exit: intel_context_exit(ce); intel_context_timeline_unlock(tl); err_unpin: - __eb_unpin_context(eb, ce); + intel_context_unpin(ce); return err; } @@ -2247,7 +2224,7 @@ static void eb_unpin_engine(struct i915_execbuffer *eb) intel_context_exit(ce); mutex_unlock(&tl->mutex); - __eb_unpin_context(eb, ce); + intel_context_unpin(ce); } static unsigned int |