1 files changed, 54 insertions, 47 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 257d72205bb5..3671f9f220bd 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -47,6 +47,9 @@
  * that the the relevant GPU caches have been flushed.
  */
 
+static struct kmem_cache *amdgpu_fence_slab;
+static atomic_t amdgpu_fence_slab_ref = ATOMIC_INIT(0);
+
 /**
  * amdgpu_fence_write - write a fence value
  *
@@ -85,24 +88,6 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
 }
 
 /**
- * amdgpu_fence_schedule_check - schedule lockup check
- *
- * @ring: pointer to struct amdgpu_ring
- *
- * Queues a delayed work item to check for lockups.
- */
-static void amdgpu_fence_schedule_check(struct amdgpu_ring *ring)
-{
-        /*
-         * Do not reset the timer here with mod_delayed_work,
-         * this can livelock in an interaction with TTM delayed destroy.
-         */
-        queue_delayed_work(system_power_efficient_wq,
-                &ring->fence_drv.lockup_work,
-                AMDGPU_FENCE_JIFFIES_TIMEOUT);
-}
-
-/**
  * amdgpu_fence_emit - emit a fence on the requested ring
  *
  * @ring: ring the fence is associated with
@@ -118,7 +103,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner,
         struct amdgpu_device *adev = ring->adev;
 
         /* we are protected by the ring emission mutex */
-        *fence = kmalloc(sizeof(struct amdgpu_fence), GFP_KERNEL);
+        *fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
         if ((*fence) == NULL) {
                 return -ENOMEM;
         }
@@ -132,11 +117,23 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner,
         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
                                (*fence)->seq,
                                AMDGPU_FENCE_FLAG_INT);
-        trace_amdgpu_fence_emit(ring->adev->ddev, ring->idx, (*fence)->seq);
         return 0;
 }
 
 /**
+ * amdgpu_fence_schedule_fallback - schedule fallback check
+ *
+ * @ring: pointer to struct amdgpu_ring
+ *
+ * Start a timer as fallback to our interrupts.
+ */
+static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
+{
+        mod_timer(&ring->fence_drv.fallback_timer,
+                  jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
+}
+
+/**
  * amdgpu_fence_activity - check for fence activity
  *
  * @ring: pointer to struct amdgpu_ring
@@ -202,45 +199,38 @@ static bool amdgpu_fence_activity(struct amdgpu_ring *ring)
         } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq);
 
         if (seq < last_emitted)
-                amdgpu_fence_schedule_check(ring);
+                amdgpu_fence_schedule_fallback(ring);
 
         return wake;
 }
 
 /**
- * amdgpu_fence_check_lockup - check for hardware lockup
+ * amdgpu_fence_process - process a fence
  *
- * @work: delayed work item
+ * @adev: amdgpu_device pointer
+ * @ring: ring index the fence is associated with
  *
- * Checks for fence activity and if there is none probe
- * the hardware if a lockup occured.
+ * Checks the current fence value and wakes the fence queue
+ * if the sequence number has increased (all asics).
  */
-static void amdgpu_fence_check_lockup(struct work_struct *work)
+void amdgpu_fence_process(struct amdgpu_ring *ring)
 {
-        struct amdgpu_fence_driver *fence_drv;
-        struct amdgpu_ring *ring;
-
-        fence_drv = container_of(work, struct amdgpu_fence_driver,
-                                lockup_work.work);
-        ring = fence_drv->ring;
-
         if (amdgpu_fence_activity(ring))
                 wake_up_all(&ring->fence_drv.fence_queue);
 }
 
 /**
- * amdgpu_fence_process - process a fence
+ * amdgpu_fence_fallback - fallback for hardware interrupts
  *
- * @adev: amdgpu_device pointer
- * @ring: ring index the fence is associated with
+ * @work: delayed work item
  *
- * Checks the current fence value and wakes the fence queue
- * if the sequence number has increased (all asics).
+ * Checks for fence activity.
  */
-void amdgpu_fence_process(struct amdgpu_ring *ring)
+static void amdgpu_fence_fallback(unsigned long arg)
 {
-        if (amdgpu_fence_activity(ring))
-                wake_up_all(&ring->fence_drv.fence_queue);
+        struct amdgpu_ring *ring = (void *)arg;
+
+        amdgpu_fence_process(ring);
 }
 
 /**
@@ -290,7 +280,7 @@ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq)
         if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
                 return 0;
 
-        amdgpu_fence_schedule_check(ring);
+        amdgpu_fence_schedule_fallback(ring);
         wait_event(ring->fence_drv.fence_queue, (
                    (signaled = amdgpu_fence_seq_signaled(ring, seq))));
 
@@ -491,9 +481,8 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
         atomic64_set(&ring->fence_drv.last_seq, 0);
         ring->fence_drv.initialized = false;
 
-        INIT_DELAYED_WORK(&ring->fence_drv.lockup_work,
-                        amdgpu_fence_check_lockup);
-        ring->fence_drv.ring = ring;
+        setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback,
+                    (unsigned long)ring);
 
         init_waitqueue_head(&ring->fence_drv.fence_queue);
 
@@ -536,6 +525,13 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
  */
 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
 {
+        if (atomic_inc_return(&amdgpu_fence_slab_ref) == 1) {
+                amdgpu_fence_slab = kmem_cache_create(
+                        "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
+                        SLAB_HWCACHE_ALIGN, NULL);
+                if (!amdgpu_fence_slab)
+                        return -ENOMEM;
+        }
         if (amdgpu_debugfs_fence_init(adev))
                 dev_err(adev->dev, "fence debugfs file creation failed\n");
 
@@ -554,9 +550,12 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
 {
         int i, r;
 
+        if (atomic_dec_and_test(&amdgpu_fence_slab_ref))
+                kmem_cache_destroy(amdgpu_fence_slab);
         mutex_lock(&adev->ring_lock);
         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
                 struct amdgpu_ring *ring = adev->rings[i];
+
                 if (!ring || !ring->fence_drv.initialized)
                         continue;
                 r = amdgpu_fence_wait_empty(ring);
@@ -568,6 +567,7 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
                 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
                                ring->fence_drv.irq_type);
                 amd_sched_fini(&ring->sched);
+                del_timer_sync(&ring->fence_drv.fallback_timer);
                 ring->fence_drv.initialized = false;
         }
         mutex_unlock(&adev->ring_lock);
@@ -751,18 +751,25 @@ static bool amdgpu_fence_enable_signaling(struct fence *f)
         fence->fence_wake.func = amdgpu_fence_check_signaled;
         __add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake);
         fence_get(f);
-        amdgpu_fence_schedule_check(ring);
+        if (!timer_pending(&ring->fence_drv.fallback_timer))
+                amdgpu_fence_schedule_fallback(ring);
         FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
         return true;
 }
 
+static void amdgpu_fence_release(struct fence *f)
+{
+        struct amdgpu_fence *fence = to_amdgpu_fence(f);
+        kmem_cache_free(amdgpu_fence_slab, fence);
+}
+
 const struct fence_ops amdgpu_fence_ops = {
         .get_driver_name = amdgpu_fence_get_driver_name,
         .get_timeline_name = amdgpu_fence_get_timeline_name,
         .enable_signaling = amdgpu_fence_enable_signaling,
         .signaled = amdgpu_fence_is_signaled,
         .wait = fence_default_wait,
-        .release = NULL,
+        .release = amdgpu_fence_release,
 };
 
 /*