Due to the ordering of cmpxchg()/dma_fence_signal() inside node_retire(), we must also use the xchg() as our primary memory barrier to flush the outstanding callbacks after expected completion of the i915_active.
Signed-off-by: Chris Wilson <[email protected]> --- drivers/gpu/drm/i915/selftests/i915_active.c | 29 ++++++++++++++------ 1 file changed, 21 insertions(+), 8 deletions(-) diff --git a/drivers/gpu/drm/i915/selftests/i915_active.c b/drivers/gpu/drm/i915/selftests/i915_active.c index 3a37c67ab6c4..68bbb1580162 100644 --- a/drivers/gpu/drm/i915/selftests/i915_active.c +++ b/drivers/gpu/drm/i915/selftests/i915_active.c @@ -311,20 +311,33 @@ static void spin_unlock_wait(spinlock_t *lock) spin_unlock_irq(lock); } +static void active_flush(struct i915_active *ref, + struct i915_active_fence *active) +{ + struct dma_fence *fence; + + fence = xchg(__active_fence_slot(active), NULL); + if (!fence) + return; + + spin_lock_irq(fence->lock); + __list_del_entry(&active->cb.node); + spin_unlock_irq(fence->lock); /* serialise with fence->cb_list */ + atomic_dec(&ref->count); + + GEM_BUG_ON(!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)); +} + void i915_active_unlock_wait(struct i915_active *ref) { if (i915_active_acquire_if_busy(ref)) { struct active_node *it, *n; + /* Wait for all active callbacks */ rcu_read_lock(); - rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) { - struct dma_fence *f; - - /* Wait for all active callbacks */ - f = rcu_dereference(it->base.fence); - if (f) - spin_unlock_wait(f->lock); - } + active_flush(ref, &ref->excl); + rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) + active_flush(ref, &it->base); rcu_read_unlock(); i915_active_release(ref); -- 2.25.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
