| Message ID | 20211005113742.1101-21-christian.koenig@amd.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [01/28] dma-buf: add dma_resv_for_each_fence_unlocked v8 | expand |
Op 05-10-2021 om 13:37 schreef Christian König: > Simplifying the code a bit. > > Signed-off-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/i915/gem/i915_gem_wait.c | 51 +++++------------------- > 1 file changed, 9 insertions(+), 42 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c > index f909aaa09d9c..a13193db1dba 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c > +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c > @@ -37,55 +37,22 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, > unsigned int flags, > long timeout) > { > - struct dma_fence *excl; > - bool prune_fences = false; > - > - if (flags & I915_WAIT_ALL) { > - struct dma_fence **shared; > - unsigned int count, i; > - int ret; > + struct dma_resv_iter cursor; > + struct dma_fence *fence; > > - ret = dma_resv_get_fences(resv, &excl, &count, &shared); > - if (ret) > - return ret; > - > - for (i = 0; i < count; i++) { > - timeout = i915_gem_object_wait_fence(shared[i], > - flags, timeout); > - if (timeout < 0) > - break; > - > - dma_fence_put(shared[i]); > - } > - > - for (; i < count; i++) > - dma_fence_put(shared[i]); > - kfree(shared); > - > - /* > - * If both shared fences and an exclusive fence exist, > - * then by construction the shared fences must be later > - * than the exclusive fence. If we successfully wait for > - * all the shared fences, we know that the exclusive fence > - * must all be signaled. If all the shared fences are > - * signaled, we can prune the array and recover the > - * floating references on the fences/requests. > - */ > - prune_fences = count && timeout >= 0; > - } else { > - excl = dma_resv_get_excl_unlocked(resv); > + dma_resv_iter_begin(&cursor, resv, flags & I915_WAIT_ALL); > + dma_resv_for_each_fence_unlocked(&cursor, fence) { > + timeout = i915_gem_object_wait_fence(fence, flags, timeout); > + if (timeout < 0) > + break; > } > - > - if (excl && timeout >= 0) > - timeout = i915_gem_object_wait_fence(excl, flags, timeout); > - > - dma_fence_put(excl); > + dma_resv_iter_end(&cursor); > > /* > * Opportunistically prune the fences iff we know they have *all* been > * signaled. > */ > - if (prune_fences) > + if (timeout > 0) > dma_resv_prune(resv); > > return timeout; When replying to tvrtko about correctness of the conversion, I just now noticed a logic bug here, the same logic bug also affects dma_resv_wait_timeout. long dma_resv_wait_timeout(struct dma_resv *obj, bool wait_all, bool intr, unsigned long timeout) { long ret = timeout ? timeout : 1; struct dma_resv_iter cursor; struct dma_fence *fence; dma_resv_iter_begin(&cursor, obj, wait_all); dma_resv_for_each_fence_unlocked(&cursor, fence) { ret = dma_fence_wait_timeout(fence, intr, ret); if (ret <= 0) { dma_resv_iter_end(&cursor); return ret; } } dma_resv_iter_end(&cursor); return ret; } It fails to handle the case correctly when timeout = 0, I think the original code probably did. dma_fence_wait_timeout should be called with timeout = 0 explicitly. Fixed code for inner loop: ret = dma_fence_wait_timeout(fence, intr, timeout); if (ret <= 0) break; if (timeout) timeout = ret; This bug also affects i915_gem_object_wait_reservation, so the whole series might need to be respinned, or at least checked, if more wait conversions are affected. ~Maarten
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index f909aaa09d9c..a13193db1dba 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -37,55 +37,22 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, unsigned int flags, long timeout) { - struct dma_fence *excl; - bool prune_fences = false; - - if (flags & I915_WAIT_ALL) { - struct dma_fence **shared; - unsigned int count, i; - int ret; + struct dma_resv_iter cursor; + struct dma_fence *fence; - ret = dma_resv_get_fences(resv, &excl, &count, &shared); - if (ret) - return ret; - - for (i = 0; i < count; i++) { - timeout = i915_gem_object_wait_fence(shared[i], - flags, timeout); - if (timeout < 0) - break; - - dma_fence_put(shared[i]); - } - - for (; i < count; i++) - dma_fence_put(shared[i]); - kfree(shared); - - /* - * If both shared fences and an exclusive fence exist, - * then by construction the shared fences must be later - * than the exclusive fence. If we successfully wait for - * all the shared fences, we know that the exclusive fence - * must all be signaled. If all the shared fences are - * signaled, we can prune the array and recover the - * floating references on the fences/requests. - */ - prune_fences = count && timeout >= 0; - } else { - excl = dma_resv_get_excl_unlocked(resv); + dma_resv_iter_begin(&cursor, resv, flags & I915_WAIT_ALL); + dma_resv_for_each_fence_unlocked(&cursor, fence) { + timeout = i915_gem_object_wait_fence(fence, flags, timeout); + if (timeout < 0) + break; } - - if (excl && timeout >= 0) - timeout = i915_gem_object_wait_fence(excl, flags, timeout); - - dma_fence_put(excl); + dma_resv_iter_end(&cursor); /* * Opportunistically prune the fences iff we know they have *all* been * signaled. */ - if (prune_fences) + if (timeout > 0) dma_resv_prune(resv); return timeout;
Simplifying the code a bit. Signed-off-by: Christian König <christian.koenig@amd.com> --- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 51 +++++------------------- 1 file changed, 9 insertions(+), 42 deletions(-)