[Mesa-dev] [PATCH 49/70] i965: Allow syncobjects to hook into the internal fence tracking

Chris Wilson chris at chris-wilson.co.uk
Fri Aug 7 13:13:53 PDT 2015


Since we use fences internally for tracking buffer busyness within
brw_batch.c, we can expose those directly for GL/DRI2 sync objects.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 src/mesa/drivers/dri/i965/brw_batch.c     | 41 +++++++++++++-
 src/mesa/drivers/dri/i965/brw_batch.h     | 23 +++++++-
 src/mesa/drivers/dri/i965/intel_syncobj.c | 92 +++++++++----------------------
 3 files changed, 85 insertions(+), 71 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_batch.c b/src/mesa/drivers/dri/i965/brw_batch.c
index d9386bb..d1f5828 100644
--- a/src/mesa/drivers/dri/i965/brw_batch.c
+++ b/src/mesa/drivers/dri/i965/brw_batch.c
@@ -191,7 +191,7 @@ static void __brw_request_retire(struct brw_request * const rq)
       assert(!__brw_bo_busy(RQ_BO(tmp)) || batch->fini);
       assert(RQ_BO(tmp)->exec == NULL);
 
-      list_for_each_entry_safe(struct __brw_fence, fence, &tmp->fences, link) {
+      list_for_each_entry_safe(struct brw_fence, fence, &tmp->fences, link) {
          struct brw_bo *bo;
 
          assert(fence->rq == tmp);
@@ -667,7 +667,7 @@ static void __brw_batch_grow_exec(struct brw_batch *batch)
    if (new_exec != batch->exec) {
       struct list_head * const list = &batch->next_request->fences;
 
-      list_for_each_entry_rev(struct __brw_fence, fence, list, link) {
+      list_for_each_entry_rev(struct brw_fence, fence, list, link) {
          struct brw_bo *bo;
 
          if (unlikely(fence->signal != (void *)READ_SIGNAL)) {
@@ -1075,7 +1075,7 @@ int brw_batch_flush(struct brw_batch *batch, struct perf_debug *perf)
    }
 
 skip:
-   list_for_each_entry_rev(struct __brw_fence, fence, &rq->fences, link) {
+   list_for_each_entry_rev(struct brw_fence, fence, &rq->fences, link) {
       struct brw_bo *bo;
 
       if (unlikely(fence->signal != (void *)READ_SIGNAL)) {
@@ -1391,6 +1391,41 @@ struct brw_bo *brw_bo_create_from_name(struct brw_batch *batch,
    return bo;
 }
 
+bool brw_batch_insert_fence(struct brw_batch *batch,
+                            struct brw_fence *fence,
+                            void (*signal)(struct brw_fence *),
+                            struct perf_debug *perf)
+{
+   brw_batch_flush(batch, perf);
+
+   fence->rq = batch->requests[batch->ring].mru;
+   if (fence->rq == NULL)
+      return false;
+
+   fence->signal = signal ?: (void *)NO_SIGNAL;
+   list_addtail(&fence->link, &fence->rq->fences);
+   return true;
+}
+
+int brw_fence_wait(struct brw_fence *fence,
+                   int64_t timeout,
+                   struct perf_debug *perf)
+{
+   if (fence->rq == NULL)
+      return 0;
+
+   return __brw_bo_wait(RQ_BO(fence->rq), timeout, perf);
+}
+
+void brw_fence_finish(struct brw_fence *fence)
+{
+   if (fence->rq == NULL)
+      return;
+
+   list_del(&fence->link);
+   fence->rq = NULL;
+}
+
 /*
  * Write a portion of the *linear* buffer using the pointer provided.
  *
diff --git a/src/mesa/drivers/dri/i965/brw_batch.h b/src/mesa/drivers/dri/i965/brw_batch.h
index fee4d80..da88bc2 100644
--- a/src/mesa/drivers/dri/i965/brw_batch.h
+++ b/src/mesa/drivers/dri/i965/brw_batch.h
@@ -65,16 +65,16 @@ enum brw_bo_domain { DOMAIN_NONE, DOMAIN_CPU, DOMAIN_GTT, DOMAIN_GPU };
  * the GPU passes that point, the fence will be signalled. Or you can wait
  * for a fence to complete.
  */
-struct __brw_fence {
+struct brw_fence {
    struct brw_request *rq;
    struct list_head link;
-   void (*signal)(struct __brw_fence *);
+   void (*signal)(struct brw_fence *);
 };
 
 typedef struct brw_bo {
    struct brw_batch *batch;
    struct drm_i915_gem_exec_object2 *exec;
-   struct __brw_fence read, write;
+   struct brw_fence read, write;
 
    unsigned dirty : 1;
    unsigned domain : 2;
@@ -341,6 +341,23 @@ static inline void brw_bo_put(struct brw_bo *bo)
       __brw_bo_free(bo);
 }
 
+bool
+brw_batch_insert_fence(struct brw_batch *batch,
+                       struct brw_fence *fence,
+                       void (*signal)(struct brw_fence *),
+                       struct perf_debug *perf);
+
+static inline bool
+brw_fence_busy(struct brw_fence *fence, struct perf_debug *perf)
+{
+   return __brw_request_busy(fence->rq, BUSY_FLUSH | BUSY_RETIRE, perf);
+}
+
+int brw_fence_wait(struct brw_fence *fence,
+                   int64_t timeout,
+                   struct perf_debug *perf);
+void brw_fence_finish(struct brw_fence *fence);
+
 /* Control batch command insertion and submission to hw */
 MUST_CHECK int brw_batch_begin(struct brw_batch *batch,
                                uint32_t estimated_bytes,
diff --git a/src/mesa/drivers/dri/i965/intel_syncobj.c b/src/mesa/drivers/dri/i965/intel_syncobj.c
index d55cf4b..b1ae15c 100644
--- a/src/mesa/drivers/dri/i965/intel_syncobj.c
+++ b/src/mesa/drivers/dri/i965/intel_syncobj.c
@@ -43,80 +43,30 @@
 #include "brw_context.h"
 #include "intel_reg.h"
 
-struct brw_fence {
-   /** The fence waits for completion of this batch. */
-   brw_bo *batch_bo;
-
-   bool signalled;
-};
-
 struct intel_gl_sync_object {
    struct gl_sync_object Base;
    struct brw_fence fence;
 };
 
-static void
-brw_fence_finish(struct brw_fence *fence)
-{
-   brw_bo_put(fence->batch_bo);
-}
-
-static void
-brw_fence_insert(struct brw_context *brw, struct brw_fence *fence)
-{
-   assert(!fence->batch_bo);
-   assert(!fence->signalled);
-
-   brw_mi_flush(brw, brw->batch.ring);
-   fence->batch_bo = brw_bo_get(brw->batch.bo);
-   brw_batch_flush(&brw->batch, PERF_DEBUG(brw, "SyncFence"));
-}
-
-static bool
-brw_fence_has_completed(struct brw_fence *fence)
-{
-   if (fence->signalled)
-      return true;
-
-   if (brw_bo_busy(fence->batch_bo, BUSY_WRITE | BUSY_RETIRE, NULL)) {
-      brw_bo_put(fence->batch_bo);
-      fence->batch_bo = NULL;
-      fence->signalled = true;
-      return true;
-   }
-
-   return false;
-}
-
 /**
  * Return true if the function successfully signals or has already signalled.
  * (This matches the behavior expected from __DRI2fence::client_wait_sync).
  */
 static bool
-brw_fence_client_wait(struct brw_context *brw, struct brw_fence *fence,
-                      uint64_t timeout)
+brw_fence_client_wait(struct brw_context *brw,
+                      struct brw_fence *fence,
+                      uint64_t timeout,
+                      struct perf_debug *perf)
 {
-   if (fence->signalled)
-      return true;
-
-   assert(fence->batch_bo);
-
    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and returns
-    * immediately for timeouts <= 0.  The best we can do is to clamp the
-    * timeout to INT64_MAX.  This limits the maximum timeout from 584 years to
-    * 292 years - likely not a big deal.
+    * immediately for timeout == 0, and indefinitely if timeout is negative.
+    * The best we can do is to clamp the timeout to INT64_MAX.  This limits
+    * the maximum timeout from 584 years to 292 years - likely not a big deal.
     */
    if (timeout > INT64_MAX)
       timeout = INT64_MAX;
 
-   if (drm_intel_gem_bo_wait(fence->batch_bo->base, timeout) != 0)
-      return false;
-
-   fence->signalled = true;
-   brw_bo_put(fence->batch_bo);
-   fence->batch_bo = NULL;
-
-   return true;
+   return brw_fence_wait(fence, timeout, perf) == 0;
 }
 
 static void
@@ -151,13 +101,23 @@ intel_gl_delete_sync_object(struct gl_context *ctx, struct gl_sync_object *s)
 }
 
 static void
+__intel_fence_signal(struct brw_fence *fence)
+{
+   struct intel_gl_sync_object *sync = container_of(fence, sync, fence);
+
+   sync->Base.StatusFlag = 1;
+}
+
+static void
 intel_gl_fence_sync(struct gl_context *ctx, struct gl_sync_object *s,
                     GLenum condition, GLbitfield flags)
 {
    struct brw_context *brw = brw_context(ctx);
    struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s;
 
-   brw_fence_insert(brw, &sync->fence);
+   s->StatusFlag = !brw_batch_insert_fence(&brw->batch,
+                                           &sync->fence, __intel_fence_signal,
+                                           PERF_DEBUG(brw, "FenceSync"));
 }
 
 static void
@@ -167,8 +127,8 @@ intel_gl_client_wait_sync(struct gl_context *ctx, struct gl_sync_object *s,
    struct brw_context *brw = brw_context(ctx);
    struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s;
 
-   if (brw_fence_client_wait(brw, &sync->fence, timeout))
-      s->StatusFlag = 1;
+   brw_fence_client_wait(brw, &sync->fence, timeout,
+                         PERF_DEBUG(brw, "ClientWaitSync"));
 }
 
 static void
@@ -186,8 +146,8 @@ intel_gl_check_sync(struct gl_context *ctx, struct gl_sync_object *s)
 {
    struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s;
 
-   if (brw_fence_has_completed(&sync->fence))
-      s->StatusFlag = 1;
+   brw_fence_busy(&sync->fence,
+                  PERF_DEBUG(brw_context(ctx), "CheckSync"));
 }
 
 void
@@ -211,7 +171,8 @@ intel_dri_create_fence(__DRIcontext *ctx)
    if (!fence)
       return NULL;
 
-   brw_fence_insert(brw, fence);
+   brw_batch_insert_fence(&brw->batch, fence, NULL,
+                          PERF_DEBUG(brw, "DRI2CreateFence"));
 
    return fence;
 }
@@ -232,7 +193,8 @@ intel_dri_client_wait_sync(__DRIcontext *ctx, void *driver_fence, unsigned flags
    struct brw_context *brw = ctx->driverPrivate;
    struct brw_fence *fence = driver_fence;
 
-   return brw_fence_client_wait(brw, fence, timeout);
+   return brw_fence_client_wait(brw, fence, timeout,
+                                PERF_DEBUG(brw, "DRI2ClientFenceWait"));
 }
 
 static void
-- 
2.5.0



More information about the mesa-dev mailing list