drm/radeon: add infrastructure for advanced ring synchronization v2
authorChristian König <deathsimple@vodafone.de>
Thu, 10 May 2012 13:57:31 +0000 (15:57 +0200)
committerChristian König <deathsimple@vodafone.de>
Thu, 21 Jun 2012 07:38:38 +0000 (09:38 +0200)
v2: BUG_ON not matching rings.

Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
drivers/gpu/drm/radeon/radeon.h
drivers/gpu/drm/radeon/radeon_fence.c

index e2feddd91df5afcacbcafa8759cec3e513cb0b67..33a72dc9af856ab1dc0c12fcd016d060096b9f5d 100644 (file)
@@ -257,8 +257,8 @@ struct radeon_fence_driver {
        uint32_t                        scratch_reg;
        uint64_t                        gpu_addr;
        volatile uint32_t               *cpu_addr;
-       /* seq is protected by ring emission lock */
-       uint64_t                        seq;
+       /* sync_seq is protected by ring emission lock */
+       uint64_t                        sync_seq[RADEON_NUM_RINGS];
        atomic64_t                      last_seq;
        unsigned long                   last_activity;
        bool                            initialized;
@@ -288,6 +288,27 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence);
 void radeon_fence_unref(struct radeon_fence **fence);
 unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring);
+bool radeon_fence_need_sync(struct radeon_fence *fence, int ring);
+void radeon_fence_note_sync(struct radeon_fence *fence, int ring);
+static inline struct radeon_fence *radeon_fence_later(struct radeon_fence *a,
+                                                     struct radeon_fence *b)
+{
+       if (!a) {
+               return b;
+       }
+
+       if (!b) {
+               return a;
+       }
+
+       BUG_ON(a->ring != b->ring);
+
+       if (a->seq > b->seq) {
+               return a;
+       } else {
+               return b;
+       }
+}
 
 /*
  * Tiling registers
index 401d346a05c07f824a53d747a3b055d2414f04cc..7b55625a5e1826dbfb5fcf49546d26e7d48a72ce 100644 (file)
@@ -72,7 +72,7 @@ int radeon_fence_emit(struct radeon_device *rdev,
        }
        kref_init(&((*fence)->kref));
        (*fence)->rdev = rdev;
-       (*fence)->seq = ++rdev->fence_drv[ring].seq;
+       (*fence)->seq = ++rdev->fence_drv[ring].sync_seq[ring];
        (*fence)->ring = ring;
        radeon_fence_ring_emit(rdev, ring, *fence);
        trace_radeon_fence_emit(rdev->ddev, (*fence)->seq);
@@ -449,7 +449,7 @@ int radeon_fence_wait_next_locked(struct radeon_device *rdev, int ring)
         * wait.
         */
        seq = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
-       if (seq >= rdev->fence_drv[ring].seq) {
+       if (seq >= rdev->fence_drv[ring].sync_seq[ring]) {
                /* nothing to wait for, last_seq is
                   already the last emited fence */
                return -ENOENT;
@@ -464,7 +464,7 @@ int radeon_fence_wait_empty_locked(struct radeon_device *rdev, int ring)
         * activity can be scheduled so there won't be concurrent access
         * to seq value.
         */
-       return radeon_fence_wait_seq(rdev, rdev->fence_drv[ring].seq,
+       return radeon_fence_wait_seq(rdev, rdev->fence_drv[ring].sync_seq[ring],
                                     ring, false, false);
 }
 
@@ -492,7 +492,8 @@ unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
         * but it's ok to report slightly wrong fence count here.
         */
        radeon_fence_process(rdev, ring);
-       emitted = rdev->fence_drv[ring].seq - atomic64_read(&rdev->fence_drv[ring].last_seq);
+       emitted = rdev->fence_drv[ring].sync_seq[ring]
+               - atomic64_read(&rdev->fence_drv[ring].last_seq);
        /* to avoid 32bits warp around */
        if (emitted > 0x10000000) {
                emitted = 0x10000000;
@@ -500,6 +501,51 @@ unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
        return (unsigned)emitted;
 }
 
+bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
+{
+       struct radeon_fence_driver *fdrv;
+
+       if (!fence) {
+               return false;
+       }
+
+       if (fence->ring == dst_ring) {
+               return false;
+       }
+
+       /* we are protected by the ring mutex */
+       fdrv = &fence->rdev->fence_drv[dst_ring];
+       if (fence->seq <= fdrv->sync_seq[fence->ring]) {
+               return false;
+       }
+
+       return true;
+}
+
+void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
+{
+       struct radeon_fence_driver *dst, *src;
+       unsigned i;
+
+       if (!fence) {
+               return;
+       }
+
+       if (fence->ring == dst_ring) {
+               return;
+       }
+
+       /* we are protected by the ring mutex */
+       src = &fence->rdev->fence_drv[fence->ring];
+       dst = &fence->rdev->fence_drv[dst_ring];
+       for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+               if (i == dst_ring) {
+                       continue;
+               }
+               dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
+       }
+}
+
 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
 {
        uint64_t index;
@@ -521,7 +567,7 @@ int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
        }
        rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
        rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
-       radeon_fence_write(rdev, rdev->fence_drv[ring].seq, ring);
+       radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
        rdev->fence_drv[ring].initialized = true;
        dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx and cpu addr 0x%p\n",
                 ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr);
@@ -530,10 +576,13 @@ int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
 
 static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
 {
+       int i;
+
        rdev->fence_drv[ring].scratch_reg = -1;
        rdev->fence_drv[ring].cpu_addr = NULL;
        rdev->fence_drv[ring].gpu_addr = 0;
-       rdev->fence_drv[ring].seq = 0;
+       for (i = 0; i < RADEON_NUM_RINGS; ++i)
+               rdev->fence_drv[ring].sync_seq[i] = 0;
        atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
        rdev->fence_drv[ring].last_activity = jiffies;
        rdev->fence_drv[ring].initialized = false;
@@ -579,7 +628,7 @@ static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
        struct drm_info_node *node = (struct drm_info_node *)m->private;
        struct drm_device *dev = node->minor->dev;
        struct radeon_device *rdev = dev->dev_private;
-       int i;
+       int i, j;
 
        for (i = 0; i < RADEON_NUM_RINGS; ++i) {
                if (!rdev->fence_drv[i].initialized)
@@ -588,8 +637,14 @@ static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
                seq_printf(m, "--- ring %d ---\n", i);
                seq_printf(m, "Last signaled fence 0x%016llx\n",
                           (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
-               seq_printf(m, "Last emitted  0x%016llx\n",
-                          rdev->fence_drv[i].seq);
+               seq_printf(m, "Last emitted        0x%016llx\n",
+                          rdev->fence_drv[i].sync_seq[i]);
+
+               for (j = 0; j < RADEON_NUM_RINGS; ++j) {
+                       if (i != j && rdev->fence_drv[j].initialized)
+                               seq_printf(m, "Last sync to ring %d 0x%016llx\n",
+                                          j, rdev->fence_drv[i].sync_seq[j]);
+               }
        }
        return 0;
 }