gpu: host1x: External timeout/cancellation for fences

Currently all fences have a 30 second timeout to ensure they are
cleaned up if the fence never completes otherwise. However, this
one size fits all solution doesn't actually fit in every case,
such as syncpoint waiting where we want to be able to have timeouts
longer than 30 seconds. As such, we want to be able to give control
over fence cancellation to the caller (and maybe eventually get rid
of the internal timeout altogether).

Here we add this cancellation mechanism by essentially adding a
function for entering the timeout path by function call, and changing
the syncpoint wait function to use it.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>

drivers/gpu/drm/tegra/submit.c

@@ -654,7 +654,7 @@ int tegra_drm_ioctl_channel_submit(struct drm_device *drm, void *data,
 	args->syncpt.value = job->syncpt_end;
 
 	if (syncobj) {
-		struct dma_fence *fence = host1x_fence_create(job->syncpt, job->syncpt_end);
+		struct dma_fence *fence = host1x_fence_create(job->syncpt, job->syncpt_end, true);
 		if (IS_ERR(fence)) {
 			err = PTR_ERR(fence);
 			SUBMIT_ERR(context, "failed to create postfence: %d", err);

drivers/gpu/host1x/fence.c

@@ -37,8 +37,7 @@ static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f)
 	if (host1x_syncpt_is_expired(sf->sp, sf->threshold))
 		return false;
 
-	/* One reference for interrupt path, one for timeout path. */
-	dma_fence_get(f);
+	/* Reference for interrupt path. */
 	dma_fence_get(f);
 
 	/*
@@ -46,11 +45,15 @@ 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 cannot (for now) normally guarantee that all fences get signalled.
-	 * As such, setup a timeout, so that long-lasting fences will get
-	 * reaped eventually.
+	 * We cannot currently always guarantee that all fences get signalled
+	 * or cancelled. As such, for such situations, set up a timeout, so
+	 * that long-lasting fences will get reaped eventually.
 	 */
+	if (sf->timeout) {
+		/* Reference for timeout path. */
+		dma_fence_get(f);
 		schedule_delayed_work(&sf->timeout_work, msecs_to_jiffies(30000));
+	}
 
 	host1x_intr_add_fence_locked(sf->sp->host, sf);
 
@@ -80,7 +83,7 @@ void host1x_fence_signal(struct host1x_syncpt_fence *f)
 		return;
 	}
 
-	if (cancel_delayed_work(&f->timeout_work)) {
+	if (f->timeout && 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.
@@ -99,7 +102,8 @@ static void do_fence_timeout(struct work_struct *work)
 		container_of(dwork, struct host1x_syncpt_fence, timeout_work);
 
 	if (atomic_xchg(&f->signaling, 1)) {
-		/* Already on interrupt path, drop timeout path reference. */
+		/* Already on interrupt path, drop timeout path reference if any. */
+		if (f->timeout)
 			dma_fence_put(&f->base);
 		return;
 	}
@@ -114,12 +118,12 @@ static void do_fence_timeout(struct work_struct *work)
 
 	dma_fence_set_error(&f->base, -ETIMEDOUT);
 	dma_fence_signal(&f->base);
-
-	/* Drop timeout path reference. */
+	if (f->timeout)
 		dma_fence_put(&f->base);
 }
 
-struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold)
+struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold,
+				      bool timeout)
 {
 	struct host1x_syncpt_fence *fence;
 
@@ -129,6 +133,7 @@ struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold)
 
 	fence->sp = sp;
 	fence->threshold = threshold;
+	fence->timeout = timeout;
 
 	dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &sp->fences.lock,
 		       dma_fence_context_alloc(1), 0);
@@ -138,3 +143,12 @@ struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold)
 	return &fence->base;
 }
 EXPORT_SYMBOL(host1x_fence_create);
+
+void host1x_fence_cancel(struct dma_fence *f)
+{
+	struct host1x_syncpt_fence *sf = to_host1x_fence(f);
+
+	schedule_delayed_work(&sf->timeout_work, 0);
+	flush_delayed_work(&sf->timeout_work);
+}
+EXPORT_SYMBOL(host1x_fence_cancel);

drivers/gpu/host1x/fence.h

@@ -13,6 +13,7 @@ struct host1x_syncpt_fence {
 
 	struct host1x_syncpt *sp;
 	u32 threshold;
+	bool timeout;
 
 	struct delayed_work timeout_work;
 

drivers/gpu/host1x/hw/channel_hw.c

@@ -325,7 +325,7 @@ static int channel_submit(struct host1x_job *job)
 	 * Create fence before submitting job to HW to avoid job completing
 	 * before the fence is set up.
 	 */
-	job->fence = host1x_fence_create(sp, syncval);
+	job->fence = host1x_fence_create(sp, syncval, true);
 	if (WARN(IS_ERR(job->fence), "Failed to create submit complete fence")) {
 		job->fence = NULL;
 	} else {

drivers/gpu/host1x/syncpt.c

@@ -236,11 +236,13 @@ int host1x_syncpt_wait(struct host1x_syncpt *sp, u32 thresh, long timeout,
 	else if (timeout == 0)
 		return -EAGAIN;
 
-	fence = host1x_fence_create(sp, thresh);
+	fence = host1x_fence_create(sp, thresh, false);
 	if (IS_ERR(fence))
 		return PTR_ERR(fence);
 
 	wait_err = dma_fence_wait_timeout(fence, true, timeout);
+	if (wait_err == 0)
+		host1x_fence_cancel(fence);
 	dma_fence_put(fence);
 
 	if (value)

include/linux/host1x.h

@@ -222,7 +222,9 @@ u32 host1x_syncpt_base_id(struct host1x_syncpt_base *base);
 void host1x_syncpt_release_vblank_reservation(struct host1x_client *client,
 					      u32 syncpt_id);
 
-struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold);
+struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold,
+				      bool timeout);
+void host1x_fence_cancel(struct dma_fence *fence);
 
 /*
  * host1x channel