drm/radeon/kms: sync across multiple rings when doing bo moves v3

We need to synchronize across rings when doing a bo move to make
sure we the buffer is idle if it's in use by a different ring than
the ring doing the move.

v2: fix fence setup for bo moves

v3: add missing ring lock/unlock

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

3 files changed, 39 insertions, 4 deletions

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 3483ed9b38e9..7cb63cd2e738 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -1489,6 +1489,8 @@ struct radeon_device {
         unsigned                debugfs_count;
         /* virtual memory */
         struct radeon_vm_manager        vm_manager;
+        /* ring used for bo copies */
+        u32                             copy_ring;
 };
 
 int radeon_device_init(struct radeon_device *rdev,
diff --git a/drivers/gpu/drm/radeon/radeon_asic.c b/drivers/gpu/drm/radeon/radeon_asic.c
index 123a1969d284..36a6192ce862 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.c
+++ b/drivers/gpu/drm/radeon/radeon_asic.c
@@ -1036,6 +1036,9 @@ int radeon_asic_init(struct radeon_device *rdev)
         else
                 rdev->num_crtc = 2;
 
+        /* set the ring used for bo copies */
+        rdev->copy_ring = RADEON_RING_TYPE_GFX_INDEX;
+
         switch (rdev->family) {
         case CHIP_R100:
         case CHIP_RV100:
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index b0ebaf893aca..1882025a9dc8 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -223,10 +223,10 @@ static int radeon_move_blit(struct ttm_buffer_object *bo,
         struct radeon_device *rdev;
         uint64_t old_start, new_start;
         struct radeon_fence *fence;
-        int r;
+        int r, i;
 
         rdev = radeon_get_rdev(bo->bdev);
-        r = radeon_fence_create(rdev, &fence, RADEON_RING_TYPE_GFX_INDEX);
+        r = radeon_fence_create(rdev, &fence, rdev->copy_ring);
         if (unlikely(r)) {
                 return r;
         }
@@ -255,13 +255,43 @@ static int radeon_move_blit(struct ttm_buffer_object *bo,
                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
                 return -EINVAL;
         }
-        if (!rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready) {
-                DRM_ERROR("Trying to move memory with CP turned off.\n");
+        if (!rdev->ring[rdev->copy_ring].ready) {
+                DRM_ERROR("Trying to move memory with ring turned off.\n");
                 return -EINVAL;
         }
 
         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 
+        /* sync other rings */
+        if (rdev->family >= CHIP_R600) {
+                for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+                        /* no need to sync to our own or unused rings */
+                        if (i == rdev->copy_ring || !rdev->ring[i].ready)
+                                continue;
+
+                        if (!fence->semaphore) {
+                                r = radeon_semaphore_create(rdev, &fence->semaphore);
+                                /* FIXME: handle semaphore error */
+                                if (r)
+                                        continue;
+                        }
+
+                        r = radeon_ring_lock(rdev, &rdev->ring[i], 3);
+                        /* FIXME: handle ring lock error */
+                        if (r)
+                                continue;
+                        radeon_semaphore_emit_signal(rdev, i, fence->semaphore);
+                        radeon_ring_unlock_commit(rdev, &rdev->ring[i]);
+
+                        r = radeon_ring_lock(rdev, &rdev->ring[rdev->copy_ring], 3);
+                        /* FIXME: handle ring lock error */
+                        if (r)
+                                continue;
+                        radeon_semaphore_emit_wait(rdev, rdev->copy_ring, fence->semaphore);
+                        radeon_ring_unlock_commit(rdev, &rdev->ring[rdev->copy_ring]);
+                }
+        }
+
         r = radeon_copy(rdev, old_start, new_start,
                         new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
                         fence);