drm/msm/gem: Unpin objects slightly later

The introduction of "drm/msm/gem: Evict active GEM objects when necessary"
exposes a problem with "drm/msm/gem: Unpin buffers earlier", in that we
need to keep the object pinned in the time the submit is queued up in the
gpu scheduler.  Otherwise the shrinker will see it as a thing that can be
evicted if we wait for it to be signaled.  But if the shrinker path is
waiting on it with the obj lock held, the job cannot be scheduled, as that
also requires briefly grabbing the obj lock, leading to deadlock.  (Not to
mention, we don't want the shrinker to evict an obj queued up in gpu
scheduler.)

Fixes: f371bcc0 ("drm/msm/gem: Unpin buffers earlier")
Fixes: 025d2723 ("drm/msm/gem: Evict active GEM objects when necessary")
Closes: https://gitlab.freedesktop.org/drm/msm/-/issues/19Signed-off-by: NRob Clark <robdclark@chromium.org>
Tested-by: NChia-I Wu <olvaffe@gmail.com>
Patchwork: https://patchwork.freedesktop.org/patch/504528/
Link: https://lore.kernel.org/r/20220923224043.2449152-1-robdclark@gmail.com

drivers/gpu/drm/msm/msm_gem_submit.c

@@ -501,11 +501,11 @@ static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *ob
  */
 static void submit_cleanup(struct msm_gem_submit *submit, bool error)
 {
-	unsigned cleanup_flags = BO_LOCKED | BO_OBJ_PINNED;
+	unsigned cleanup_flags = BO_LOCKED;
 	unsigned i;
 
 	if (error)
-		cleanup_flags |= BO_VMA_PINNED;
+		cleanup_flags |= BO_VMA_PINNED | BO_OBJ_PINNED;
 
 	for (i = 0; i < submit->nr_bos; i++) {
 		struct msm_gem_object *msm_obj = submit->bos[i].obj;

drivers/gpu/drm/msm/msm_ringbuffer.c

@@ -25,7 +25,8 @@ static struct dma_fence *msm_job_run(struct drm_sched_job *job)
 
 		msm_gem_lock(obj);
 		msm_gem_unpin_vma_fenced(submit->bos[i].vma, fctx);
-		submit->bos[i].flags &= ~BO_VMA_PINNED;
+		msm_gem_unpin_locked(obj);
+		submit->bos[i].flags &= ~(BO_VMA_PINNED | BO_OBJ_PINNED);
 		msm_gem_unlock(obj);
 	}