Commit 4a7d74f1 authored by Christian König's avatar Christian König Committed by Alex Deucher

drm/amdgpu: signal fences directly in amdgpu_fence_process

Because of the scheduler we need to signal all fences immediately
anyway, so try to avoid the waitqueue overhead.
Signed-off-by: default avatarChristian König <christian.koenig@amd.com>
Acked-by: default avatarAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: default avatarChunming Zhou <david1.zhou@amd.com>
parent f09c2be4
......@@ -392,8 +392,8 @@ struct amdgpu_fence_driver {
struct amdgpu_irq_src *irq_src;
unsigned irq_type;
struct timer_list fallback_timer;
wait_queue_head_t fence_queue;
unsigned num_fences_mask;
spinlock_t lock;
struct fence **fences;
};
......
......@@ -53,8 +53,6 @@ struct amdgpu_fence {
/* RB, DMA, etc. */
struct amdgpu_ring *ring;
uint64_t seq;
wait_queue_t fence_wake;
};
static struct kmem_cache *amdgpu_fence_slab;
......@@ -124,7 +122,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
{
struct amdgpu_device *adev = ring->adev;
struct amdgpu_fence *fence;
struct fence *old, **ptr;
struct fence **ptr;
unsigned idx;
fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
......@@ -134,7 +132,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
fence->seq = ++ring->fence_drv.sync_seq;
fence->ring = ring;
fence_init(&fence->base, &amdgpu_fence_ops,
&ring->fence_drv.fence_queue.lock,
&ring->fence_drv.lock,
adev->fence_context + ring->idx,
fence->seq);
amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
......@@ -145,13 +143,10 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
/* This function can't be called concurrently anyway, otherwise
* emitting the fence would mess up the hardware ring buffer.
*/
old = rcu_dereference_protected(*ptr, 1);
BUG_ON(rcu_dereference_protected(*ptr, 1));
rcu_assign_pointer(*ptr, fence_get(&fence->base));
BUG_ON(old && !fence_is_signaled(old));
fence_put(old);
*f = &fence->base;
return 0;
......@@ -181,11 +176,12 @@ 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;
bool wake = false;
int r;
last_seq = atomic64_read(&ring->fence_drv.last_seq);
do {
last_seq = atomic64_read(&ring->fence_drv.last_seq);
last_emitted = ring->fence_drv.sync_seq;
seq = amdgpu_fence_read(ring);
seq |= last_seq & 0xffffffff00000000LL;
......@@ -195,22 +191,32 @@ void amdgpu_fence_process(struct amdgpu_ring *ring)
}
if (seq <= last_seq || seq > last_emitted)
break;
return;
/* If we loop over we don't want to return without
* checking if a fence is signaled as it means that the
* seq we just read is different from the previous on.
*/
wake = true;
last_seq = seq;
} while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq);
} while (atomic64_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
if (seq < last_emitted)
amdgpu_fence_schedule_fallback(ring);
if (wake)
wake_up_all(&ring->fence_drv.fence_queue);
while (last_seq != seq) {
struct fence *fence, **ptr;
ptr = &drv->fences[++last_seq & drv->num_fences_mask];
/* There is always exactly one thread signaling this fence slot */
fence = rcu_dereference_protected(*ptr, 1);
rcu_assign_pointer(*ptr, NULL);
BUG_ON(!fence);
r = fence_signal(fence);
if (!r)
FENCE_TRACE(fence, "signaled from irq context\n");
else
BUG();
fence_put(fence);
}
}
/**
......@@ -356,8 +362,8 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback,
(unsigned long)ring);
init_waitqueue_head(&ring->fence_drv.fence_queue);
ring->fence_drv.num_fences_mask = num_hw_submission - 1;
spin_lock_init(&ring->fence_drv.lock);
ring->fence_drv.fences = kcalloc(num_hw_submission, sizeof(void *),
GFP_KERNEL);
if (!ring->fence_drv.fences)
......@@ -436,7 +442,6 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
/* no need to trigger GPU reset as we are unloading */
amdgpu_fence_driver_force_completion(adev);
}
wake_up_all(&ring->fence_drv.fence_queue);
amdgpu_irq_put(adev, ring->fence_drv.irq_src,
ring->fence_drv.irq_type);
amd_sched_fini(&ring->sched);
......@@ -568,42 +573,6 @@ static bool amdgpu_fence_is_signaled(struct fence *f)
return false;
}
/**
* amdgpu_fence_check_signaled - callback from fence_queue
*
* this function is called with fence_queue lock held, which is also used
* for the fence locking itself, so unlocked variants are used for
* fence_signal, and remove_wait_queue.
*/
static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key)
{
struct amdgpu_fence *fence;
struct amdgpu_device *adev;
u64 seq;
int ret;
fence = container_of(wait, struct amdgpu_fence, fence_wake);
adev = fence->ring->adev;
/*
* We cannot use amdgpu_fence_process here because we're already
* in the waitqueue, in a call from wake_up_all.
*/
seq = atomic64_read(&fence->ring->fence_drv.last_seq);
if (seq >= fence->seq) {
ret = fence_signal_locked(&fence->base);
if (!ret)
FENCE_TRACE(&fence->base, "signaled from irq context\n");
else
FENCE_TRACE(&fence->base, "was already signaled\n");
__remove_wait_queue(&fence->ring->fence_drv.fence_queue, &fence->fence_wake);
fence_put(&fence->base);
} else
FENCE_TRACE(&fence->base, "pending\n");
return 0;
}
/**
* amdgpu_fence_enable_signaling - enable signalling on fence
* @fence: fence
......@@ -617,17 +586,11 @@ static bool amdgpu_fence_enable_signaling(struct fence *f)
struct amdgpu_fence *fence = to_amdgpu_fence(f);
struct amdgpu_ring *ring = fence->ring;
if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
return false;
fence->fence_wake.flags = 0;
fence->fence_wake.private = NULL;
fence->fence_wake.func = amdgpu_fence_check_signaled;
__add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake);
fence_get(f);
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;
}
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment