gpu: host1x: Rewrite syncpoint interrupt handling

Move from the old, complex intr handling code to a new implementation
based on dma_fences. While there is a fair bit of churn to get there,
the new implementation is much simpler and likely faster as well due
to allowing signaling directly from interrupt context.

Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>

1 files changed, 33 insertions, 63 deletions

diff --git a/drivers/gpu/host1x/fence.c b/drivers/gpu/host1x/fence.c
index df428bcbae69..df5b56692d2c 100644
--- a/drivers/gpu/host1x/fence.c
+++ b/drivers/gpu/host1x/fence.c
@@ -15,22 +15,6 @@
 #include "intr.h"
 #include "syncpt.h"
 
-static DEFINE_SPINLOCK(lock);
-
-struct host1x_syncpt_fence {
-	struct dma_fence base;
-
-	atomic_t signaling;
-
-	struct host1x_syncpt *sp;
-	u32 threshold;
-
-	struct host1x_waitlist *waiter;
-	void *waiter_ref;
-
-	struct delayed_work timeout_work;
-};
-
 static const char *host1x_syncpt_fence_get_driver_name(struct dma_fence *f)
 {
 	return "host1x";
@@ -49,11 +33,12 @@ static struct host1x_syncpt_fence *to_host1x_fence(struct dma_fence *f)
 static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 {
 	struct host1x_syncpt_fence *sf = to_host1x_fence(f);
-	int err;
 
 	if (host1x_syncpt_is_expired(sf->sp, sf->threshold))
 		return false;
 
+	/* One reference for interrupt path, one for timeout path. */
+	dma_fence_get(f);
 	dma_fence_get(f);
 
 	/*
@@ -61,24 +46,13 @@ static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 	 * reference to any fences for which 'enable_signaling' has been
 	 * called (and that have not been signalled).
 	 *
-	 * We provide a userspace API to create arbitrary syncpoint fences,
-	 * so we cannot normally guarantee that all fences get signalled.
+	 * We cannot (for now) normally guarantee that all fences get signalled.
 	 * As such, setup a timeout, so that long-lasting fences will get
 	 * reaped eventually.
 	 */
 	schedule_delayed_work(&sf->timeout_work, msecs_to_jiffies(30000));
 
-	err = host1x_intr_add_action(sf->sp->host, sf->sp, sf->threshold,
-				     HOST1X_INTR_ACTION_SIGNAL_FENCE, f,
-				     sf->waiter, &sf->waiter_ref);
-	if (err) {
-		cancel_delayed_work_sync(&sf->timeout_work);
-		dma_fence_put(f);
-		return false;
-	}
-
-	/* intr framework takes ownership of waiter */
-	sf->waiter = NULL;
+	host1x_intr_add_fence_locked(sf->sp->host, sf);
 
 	/*
 	 * The fence may get signalled at any time after the above call,
@@ -89,37 +63,32 @@ static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 	return true;
 }
 
-static void host1x_syncpt_fence_release(struct dma_fence *f)
-{
-	struct host1x_syncpt_fence *sf = to_host1x_fence(f);
-
-	if (sf->waiter)
-		kfree(sf->waiter);
-
-	dma_fence_free(f);
-}
-
 static const struct dma_fence_ops host1x_syncpt_fence_ops = {
 	.get_driver_name = host1x_syncpt_fence_get_driver_name,
 	.get_timeline_name = host1x_syncpt_fence_get_timeline_name,
 	.enable_signaling = host1x_syncpt_fence_enable_signaling,
-	.release = host1x_syncpt_fence_release,
 };
 
 void host1x_fence_signal(struct host1x_syncpt_fence *f)
 {
-	if (atomic_xchg(&f->signaling, 1))
+	if (atomic_xchg(&f->signaling, 1)) {
+		/*
+		 * Already on timeout path, but we removed the fence before
+		 * timeout path could, so drop interrupt path reference.
+		 */
+		dma_fence_put(&f->base);
 		return;
+	}
 
-	/*
-	 * Cancel pending timeout work - if it races, it will
-	 * not get 'f->signaling' and return.
-	 */
-	cancel_delayed_work_sync(&f->timeout_work);
-
-	host1x_intr_put_ref(f->sp->host, f->sp->id, f->waiter_ref, false);
+	if (cancel_delayed_work(&f->timeout_work)) {
+		/*
+		 * We know that the timeout path will not be entered.
+		 * Safe to drop the timeout path's reference now.
+		 */
+		dma_fence_put(&f->base);
+	}
 
-	dma_fence_signal(&f->base);
+	dma_fence_signal_locked(&f->base);
 	dma_fence_put(&f->base);
 }
 
@@ -129,17 +98,24 @@ static void do_fence_timeout(struct work_struct *work)
 	struct host1x_syncpt_fence *f =
 		container_of(dwork, struct host1x_syncpt_fence, timeout_work);
 
-	if (atomic_xchg(&f->signaling, 1))
+	if (atomic_xchg(&f->signaling, 1)) {
+		/* Already on interrupt path, drop timeout path reference. */
+		dma_fence_put(&f->base);
 		return;
+	}
 
-	/*
-	 * Cancel pending timeout work - if it races, it will
-	 * not get 'f->signaling' and return.
-	 */
-	host1x_intr_put_ref(f->sp->host, f->sp->id, f->waiter_ref, true);
+	if (host1x_intr_remove_fence(f->sp->host, f)) {
+		/*
+		 * Managed to remove fence from queue, so it's safe to drop
+		 * the interrupt path's reference.
+		 */
+		dma_fence_put(&f->base);
+	}
 
 	dma_fence_set_error(&f->base, -ETIMEDOUT);
 	dma_fence_signal(&f->base);
+
+	/* Drop timeout path reference. */
 	dma_fence_put(&f->base);
 }
 
@@ -151,16 +127,10 @@ struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold)
 	if (!fence)
 		return ERR_PTR(-ENOMEM);
 
-	fence->waiter = kzalloc(sizeof(*fence->waiter), GFP_KERNEL);
-	if (!fence->waiter) {
-		kfree(fence);
-		return ERR_PTR(-ENOMEM);
-	}
-
 	fence->sp = sp;
 	fence->threshold = threshold;
 
-	dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &lock,
+	dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &sp->fences.lock,
 		       dma_fence_context_alloc(1), 0);
 
 	INIT_DELAYED_WORK(&fence->timeout_work, do_fence_timeout);