drm/amdgpu: switch back to 32bit hw fences v2
authorChristian König <christian.koenig@amd.com>
Mon, 14 Mar 2016 14:46:06 +0000 (15:46 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 16 Mar 2016 22:00:10 +0000 (18:00 -0400)
We don't need to extend them to 64bits any more, so avoid the extra overhead.

v2: update commit message.

Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu.h
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c

index c91156c7bda132427ca407edce7747fb88861bdb..31fd7ee62e69af4ad39eba47996206910d078b41 100644 (file)
@@ -386,8 +386,8 @@ struct amdgpu_fence_driver {
        uint64_t                        gpu_addr;
        volatile uint32_t               *cpu_addr;
        /* sync_seq is protected by ring emission lock */
-       uint64_t                        sync_seq;
-       atomic64_t                      last_seq;
+       uint32_t                        sync_seq;
+       atomic_t                        last_seq;
        bool                            initialized;
        struct amdgpu_irq_src           *irq_src;
        unsigned                        irq_type;
index da9a155a622c9206b33bdcac7c414864e622d1d9..4303b447efe898145433985079a6cd10273acc81 100644 (file)
@@ -52,7 +52,6 @@ struct amdgpu_fence {
 
        /* RB, DMA, etc. */
        struct amdgpu_ring              *ring;
-       uint64_t                        seq;
 };
 
 static struct kmem_cache *amdgpu_fence_slab;
@@ -104,7 +103,7 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
        if (drv->cpu_addr)
                seq = le32_to_cpu(*drv->cpu_addr);
        else
-               seq = lower_32_bits(atomic64_read(&drv->last_seq));
+               seq = atomic_read(&drv->last_seq);
 
        return seq;
 }
@@ -123,23 +122,22 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
        struct amdgpu_device *adev = ring->adev;
        struct amdgpu_fence *fence;
        struct fence **ptr;
-       unsigned idx;
+       uint32_t seq;
 
        fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
        if (fence == NULL)
                return -ENOMEM;
 
-       fence->seq = ++ring->fence_drv.sync_seq;
+       seq = ++ring->fence_drv.sync_seq;
        fence->ring = ring;
        fence_init(&fence->base, &amdgpu_fence_ops,
                   &ring->fence_drv.lock,
                   adev->fence_context + ring->idx,
-                  fence->seq);
+                  seq);
        amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
-                              fence->seq, AMDGPU_FENCE_FLAG_INT);
+                              seq, AMDGPU_FENCE_FLAG_INT);
 
-       idx = fence->seq & ring->fence_drv.num_fences_mask;
-       ptr = &ring->fence_drv.fences[idx];
+       ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
        /* This function can't be called concurrently anyway, otherwise
         * emitting the fence would mess up the hardware ring buffer.
         */
@@ -177,22 +175,16 @@ static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
 void amdgpu_fence_process(struct amdgpu_ring *ring)
 {
        struct amdgpu_fence_driver *drv = &ring->fence_drv;
-       uint64_t seq, last_seq, last_emitted;
+       uint32_t seq, last_seq;
        int r;
 
        do {
-               last_seq = atomic64_read(&ring->fence_drv.last_seq);
-               last_emitted = ring->fence_drv.sync_seq;
+               last_seq = atomic_read(&ring->fence_drv.last_seq);
                seq = amdgpu_fence_read(ring);
-               seq |= last_seq & 0xffffffff00000000LL;
-               if (seq < last_seq) {
-                       seq &= 0xffffffff;
-                       seq |= last_emitted & 0xffffffff00000000LL;
-               }
 
-       } while (atomic64_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
+       } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
 
-       if (seq < last_emitted)
+       if (seq != ring->fence_drv.sync_seq)
                amdgpu_fence_schedule_fallback(ring);
 
        while (last_seq != seq) {
@@ -279,13 +271,10 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
         * but it's ok to report slightly wrong fence count here.
         */
        amdgpu_fence_process(ring);
-       emitted = ring->fence_drv.sync_seq
-               - atomic64_read(&ring->fence_drv.last_seq);
-       /* to avoid 32bits warp around */
-       if (emitted > 0x10000000)
-               emitted = 0x10000000;
-
-       return (unsigned)emitted;
+       emitted = 0x100000000ull;
+       emitted -= atomic_read(&ring->fence_drv.last_seq);
+       emitted += ACCESS_ONCE(ring->fence_drv.sync_seq);
+       return lower_32_bits(emitted);
 }
 
 /**
@@ -317,7 +306,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
                ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
                ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
        }
-       amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq));
+       amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
        amdgpu_irq_get(adev, irq_src, irq_type);
 
        ring->fence_drv.irq_src = irq_src;
@@ -353,7 +342,7 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
        ring->fence_drv.cpu_addr = NULL;
        ring->fence_drv.gpu_addr = 0;
        ring->fence_drv.sync_seq = 0;
-       atomic64_set(&ring->fence_drv.last_seq, 0);
+       atomic_set(&ring->fence_drv.last_seq, 0);
        ring->fence_drv.initialized = false;
 
        setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback,
@@ -621,9 +610,9 @@ static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
                amdgpu_fence_process(ring);
 
                seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
-               seq_printf(m, "Last signaled fence 0x%016llx\n",
-                          (unsigned long long)atomic64_read(&ring->fence_drv.last_seq));
-               seq_printf(m, "Last emitted        0x%016llx\n",
+               seq_printf(m, "Last signaled fence 0x%08x\n",
+                          atomic_read(&ring->fence_drv.last_seq));
+               seq_printf(m, "Last emitted        0x%08x\n",
                           ring->fence_drv.sync_seq);
        }
        return 0;