diff options
| author | Rob Clark <robdclark@chromium.org> | 2021-07-26 17:43:57 +0300 |
|---|---|---|
| committer | Rob Clark <robdclark@chromium.org> | 2021-07-28 03:53:51 +0300 |
| commit | da3d378dec86348d21c02f311da87fa0742e36f6 (patch) | |
| tree | eb7898ecefdf670af9c5c020c503dee5104c2b61 /drivers/gpu/drm/msm/msm_fence.c | |
| parent | ff1176468d368232b684f75e82563369208bc371 (diff) | |
| download | linux-da3d378dec86348d21c02f311da87fa0742e36f6.tar.xz | |
drm/msm: Let fences read directly from memptrs
Let dma_fence::signaled, etc, read directly from the address that the hw
is writing with updated completed fence seqno, so we can potentially
notice that the fence is signaled sooner.
Plus add some docs.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Link: https://lore.kernel.org/r/20210726144359.2179302-2-robdclark@gmail.com
Signed-off-by: Rob Clark <robdclark@chromium.org>
Diffstat (limited to 'drivers/gpu/drm/msm/msm_fence.c')
| -rw-r--r-- | drivers/gpu/drm/msm/msm_fence.c | 11 |
1 files changed, 9 insertions, 2 deletions
diff --git a/drivers/gpu/drm/msm/msm_fence.c b/drivers/gpu/drm/msm/msm_fence.c index cd59a5918038..b92a9091a1e2 100644 --- a/drivers/gpu/drm/msm/msm_fence.c +++ b/drivers/gpu/drm/msm/msm_fence.c @@ -11,7 +11,8 @@ struct msm_fence_context * -msm_fence_context_alloc(struct drm_device *dev, const char *name) +msm_fence_context_alloc(struct drm_device *dev, volatile uint32_t *fenceptr, + const char *name) { struct msm_fence_context *fctx; @@ -22,6 +23,7 @@ msm_fence_context_alloc(struct drm_device *dev, const char *name) fctx->dev = dev; strncpy(fctx->name, name, sizeof(fctx->name)); fctx->context = dma_fence_context_alloc(1); + fctx->fenceptr = fenceptr; init_waitqueue_head(&fctx->event); spin_lock_init(&fctx->spinlock); @@ -35,7 +37,12 @@ void msm_fence_context_free(struct msm_fence_context *fctx) static inline bool fence_completed(struct msm_fence_context *fctx, uint32_t fence) { - return (int32_t)(fctx->completed_fence - fence) >= 0; + /* + * Note: Check completed_fence first, as fenceptr is in a write-combine + * mapping, so it will be more expensive to read. + */ + return (int32_t)(fctx->completed_fence - fence) >= 0 || + (int32_t)(*fctx->fenceptr - fence) >= 0; } /* legacy path for WAIT_FENCE ioctl: */ |