Skip to contentSkip to conversation

drm/vc4: Hold a GEM reference while purging userspace BOs

PATCH In review Kazuki Hanai 2 participants 2 messages Aug 30 – Sep 14

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/vc4: Hold a GEM reference while purging userspace BOs

vc4_bo_userspace_cache_purge() removes a BO from the purgeable list and then drops the list lock before taking the BO madvise lock. A concurrent final GEM put can therefore release the BO while the purge path still holds only a raw pointer.

Acquire a non-zero GEM reference while the purgeable list lock still protects the lookup. If final release has already started, drop the list lock so the release path can remove the entry and retry. Put the temporary reference after dropping the madvise lock to avoid recursing into final release while that lock is held.

Fixes b9f19259b84d ("drm/vc4: Add the DRM_IOCTL_VC4_GEM_MADVISE ioctl")
---
 drivers/gpu/drm/vc4/vc4_bo.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
drivers/gpu/drm/vc4/vc4_bo.c +14 -0
index 49ea2ed0996b4..5289e69f3345f 100644
@@ -319,6 +319,19 @@ static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
 		struct drm_gem_object *obj = &bo->base.base;
 		size_t purged_size = 0;
 
+		/*
+		 * The last GEM reference can be dropped concurrently. Its free
+		 * callback removes the BO from this list before releasing it, so
+		 * the list lock makes an unless-zero get safe. Never carry only a
+		 * raw list pointer across the lock drop below.
+		 */
+		if (!kref_get_unless_zero(&obj->refcount)) {
+			mutex_unlock(&vc4->purgeable.lock);
+			cond_resched();
+			mutex_lock(&vc4->purgeable.lock);
+			continue;
+		}
+
 		vc4_bo_remove_from_purgeable_pool_locked(bo);
 
 		/* Release the purgeable lock while we're purging the BO so
@@ -345,6 +358,7 @@ static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
 			vc4_bo_purge(obj);
 		}
 		mutex_unlock(&bo->madv_lock);
+		drm_gem_object_put(obj);
 		mutex_lock(&vc4->purgeable.lock);
 
 		if (purged_size) {
reviewer
#1 lore raw

Hi Kazuki,

Quoted 28 lines · Kazuki Hanai

vc4_bo_userspace_cache_purge() removes a BO from the purgeable list and then drops the list lock before taking the BO madvise lock. A concurrent final GEM put can therefore release the BO while the purge path still holds only a raw pointer.

Acquire a non-zero GEM reference while the purgeable list lock still protects the lookup. If final release has already started, drop the list lock so the release path can remove the entry and retry. Put the temporary reference after dropping the madvise lock to avoid recursing into final release while that lock is held.

Fixes b9f19259b84d ("drm/vc4: Add the DRM_IOCTL_VC4_GEM_MADVISE ioctl")
---
  drivers/gpu/drm/vc4/vc4_bo.c | 14 ++++++++++++++
  1 file changed, 14 insertions(+)
drivers/gpu/drm/vc4/vc4_bo.c +2 -0
index 49ea2ed0996b4..5289e69f3345f 100644
@@ -319,6 +319,19 @@ static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
  		struct drm_gem_object *obj = &bo->base.base;
  		size_t purged_size = 0;
 
+		/*
+		 * The last GEM reference can be dropped concurrently. Its free

drivers/gpu/drm/vc4/vc4_bo.c static void vc4_bo_userspace_cache_purge(struct drm_device *dev) · line 333

Quoted 10 lines
drivers/gpu/drm/vc4/vc4_bo.c +10 -0
+		 * callback removes the BO from this list before releasing it, so
+		 * the list lock makes an unless-zero get safe. Never carry only a
+		 * raw list pointer across the lock drop below.
+		 */
+		if (!kref_get_unless_zero(&obj->refcount)) {
+			mutex_unlock(&vc4->purgeable.lock);
+			cond_resched();
+			mutex_lock(&vc4->purgeable.lock);
+			continue;
+		}

Hum, your fix doesn't look correct to me. First, I don't understand why you are using something like cond_resched() here. Next, with this approach, the BO with refcount=0 will never be removed from the list.

drivers/gpu/drm/vc4/vc4_bo.c static void vc4_bo_userspace_cache_purge(struct drm_device *dev) · line 335

Quoted 2 lines
drivers/gpu/drm/vc4/vc4_bo.c +1 -0
+
  		vc4_bo_remove_from_purgeable_pool_locked(bo);

What you could do instead is calling:

if (!kref_get_unless_zero(&obj->refcount))

      continue;

here.

Best regards,
- Maíra

Quoted 10 lines
drivers/gpu/drm/vc4/vc4_bo.c +1 -0
  		/* Release the purgeable lock while we're purging the BO so
@@ -345,6 +358,7 @@ static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
  			vc4_bo_purge(obj);
  		}
  		mutex_unlock(&bo->madv_lock);
+		drm_gem_object_put(obj);
  		mutex_lock(&vc4->purgeable.lock);
 
  		if (purged_size) {