gpu: nvgpu: Fix semaphore race condition

A race condition existed in gk20a_channel_semaphore_wait_fd().
In some instances the semaphore underlying the sync_fence being
waited on would have already signaled. This would cause the
subsequent sync_fence_wait_async() call to return 1 and do
nothing. Normally, the sync_fence_wait_async() call would
release the newly created semaphore but in the above case that
would not happen and hang any channel waiting on that semaphore.

To fix this problem if sync_fence_wait_async() returns 1
immediately release the newly created semaphore.

Bug 1604892

Change-Id: I1f5e811695bb099f71b7762835aba4a7e27362ec
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: http://git-master/r/935910
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
GVS: Gerrit_Virtual_Submit

3 files changed, 19 insertions, 6 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
index 952e6e6a..bba18789 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
@@ -456,7 +456,7 @@ static int gk20a_channel_semaphore_wait_fd(
         struct priv_cmd_entry *wait_cmd = NULL;
         struct wait_fence_work *w;
         int written;
-        int err;
+        int err, ret;
         u64 va;
 
         sync_fence = gk20a_sync_fence_fdget(fd);
@@ -490,8 +490,18 @@ static int gk20a_channel_semaphore_wait_fd(
         va = gk20a_semaphore_gpu_va(w->sema, c->vm);
         /* GPU unblocked when when the semaphore value becomes 1. */
         written = add_sema_cmd(wait_cmd->ptr, va, 1, true, false);
+
         WARN_ON(written != wait_cmd->size);
-        sync_fence_wait_async(sync_fence, &w->waiter);
+        ret = sync_fence_wait_async(sync_fence, &w->waiter);
+
+        /*
+         * If the sync_fence has already signaled then the above async_wait
+         * will never trigger. This causes the semaphore release op to never
+         * happen which, in turn, hangs the GPU. That's bad. So let's just
+         * do the semaphore_release right now.
+         */
+        if (ret == 1)
+                gk20a_semaphore_release(w->sema);
 
         /* XXX - this fixes an actual bug, we need to hold a ref to this
            semaphore while the job is in flight. */
diff --git a/drivers/gpu/nvgpu/gk20a/fence_gk20a.c b/drivers/gpu/nvgpu/gk20a/fence_gk20a.c
index 7a3f90e9..70666407 100644
--- a/drivers/gpu/nvgpu/gk20a/fence_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/fence_gk20a.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2014-2015, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2014-2016, NVIDIA CORPORATION.  All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms and conditions of the GNU General Public License,
@@ -154,7 +154,9 @@ struct gk20a_fence *gk20a_fence_from_semaphore(
 
 #ifdef CONFIG_SYNC
         sync_fence = gk20a_sync_fence_create(timeline, semaphore,
-                                             dependency, "fence");
+                                             dependency, "f-gk20a-0x%04llx",
+                                             ((u64)(void *)semaphore->value) &
+                                             0xffff);
         if (!sync_fence)
                 return NULL;
 #endif
diff --git a/drivers/gpu/nvgpu/gk20a/sync_gk20a.c b/drivers/gpu/nvgpu/gk20a/sync_gk20a.c
index 8740f0e2..e01c0e9a 100644
--- a/drivers/gpu/nvgpu/gk20a/sync_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/sync_gk20a.c
@@ -47,8 +47,9 @@ struct gk20a_sync_pt {
         ktime_t                         dep_timestamp;
 
         /*
-         * A spinlock is necessary since there are times when this lock
-         * will be acquired in interrupt context.
+         * Use a spin lock here since it will have better performance
+         * than a mutex - there should be very little contention on this
+         * lock.
          */
         spinlock_t                      lock;
 };