From: =?UTF-8?q?Nicolai=20H=C3=A4hnle?= <nicolai.haehnle@amd.com>
Date: Thu, 28 Sep 2017 11:57:32 +0200
Subject: drm/amd/sched: fix deadlock caused by unsignaled fences of deleted
 jobs
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Git-commit: 79867462634836ee5c39a2cdf624719feeb189bd
Patch-mainline: v4.15-rc1
References: FATE#326289 FATE#326079 FATE#326049 FATE#322398 FATE#326166

Highly concurrent Piglit runs can trigger a race condition where a pending
SDMA job on a buffer object is never executed because the corresponding
process is killed (perhaps due to a crash). Since the job's fences were
never signaled, the buffer object was effectively leaked. Worse, the
buffer was stuck wherever it happened to be at the time, possibly in VRAM.

The symptom was user space processes stuck in interruptible waits with
kernel stacks like:

    [<ffffffffbc5e6722>] dma_fence_default_wait+0x112/0x250
    [<ffffffffbc5e6399>] dma_fence_wait_timeout+0x39/0xf0
    [<ffffffffbc5e82d2>] reservation_object_wait_timeout_rcu+0x1c2/0x300
    [<ffffffffc03ce56f>] ttm_bo_cleanup_refs_and_unlock+0xff/0x1a0 [ttm]
    [<ffffffffc03cf1ea>] ttm_mem_evict_first+0xba/0x1a0 [ttm]
    [<ffffffffc03cf611>] ttm_bo_mem_space+0x341/0x4c0 [ttm]
    [<ffffffffc03cfc54>] ttm_bo_validate+0xd4/0x150 [ttm]
    [<ffffffffc03cffbd>] ttm_bo_init_reserved+0x2ed/0x420 [ttm]
    [<ffffffffc042f523>] amdgpu_bo_create_restricted+0x1f3/0x470 [amdgpu]
    [<ffffffffc042f9fa>] amdgpu_bo_create+0xda/0x220 [amdgpu]
    [<ffffffffc04349ea>] amdgpu_gem_object_create+0xaa/0x140 [amdgpu]
    [<ffffffffc0434f97>] amdgpu_gem_create_ioctl+0x97/0x120 [amdgpu]
    [<ffffffffc037ddba>] drm_ioctl+0x1fa/0x480 [drm]
    [<ffffffffc041904f>] amdgpu_drm_ioctl+0x4f/0x90 [amdgpu]
    [<ffffffffbc23db33>] do_vfs_ioctl+0xa3/0x5f0
    [<ffffffffbc23e0f9>] SyS_ioctl+0x79/0x90
    [<ffffffffbc864ffb>] entry_SYSCALL_64_fastpath+0x1e/0xad
    [<ffffffffffffffff>] 0xffffffffffffffff

Note: The correctness of this change depends on the earlier commit
"drm/amd/sched: move adding finish callback to amd_sched_job_begin"

v2: set an error on the finished fence

Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Petr Tesarik <ptesarik@suse.com>
---
 drivers/gpu/drm/amd/scheduler/gpu_scheduler.c |    8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
@@ -227,8 +227,14 @@ void amd_sched_entity_fini(struct amd_gp
 		 */
 		kthread_park(sched->thread);
 		kthread_unpark(sched->thread);
-		while (kfifo_out(&entity->job_queue, &job, sizeof(job)))
+		while (kfifo_out(&entity->job_queue, &job, sizeof(job))) {
+			struct amd_sched_fence *s_fence = job->s_fence;
+			amd_sched_fence_scheduled(s_fence);
+			dma_fence_set_error(&s_fence->finished, -ESRCH);
+			amd_sched_fence_finished(s_fence);
+			dma_fence_put(&s_fence->finished);
 			sched->ops->free_job(job);
+		}
 
 	}
 	kfifo_free(&entity->job_queue);