dma-buf/sw_sync: force signal all unsignaled fences on dying timeline
authorDominik Behr <dbehr@chromium.org>
Thu, 7 Sep 2017 19:02:46 +0000 (16:02 -0300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 10 Dec 2017 12:40:44 +0000 (13:40 +0100)
commit ea4d5a270b57fa8d4871f372ca9b97b7697fdfda upstream.

To avoid hanging userspace components that might have been waiting on the
active fences of the destroyed timeline we need to signal with error all
remaining fences on such timeline.

This restore the default behaviour of the Android sw_sync framework, which
Android still relies on. It was broken on the dma fence conversion a few
years ago and never fixed.

v2: Do not bother with cleanup do the list (Chris Wilson)

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dominik Behr <dbehr@chromium.org>
Signed-off-by: Gustavo Padovan <gustavo.padovan@collabora.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170907190246.16425-2-gustavo@padovan.org
Cc: Jisheng Zhang <Jisheng.Zhang@synaptics.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/dma-buf/sw_sync.c

index 38cc7389a6c172c4f22ac89ec3694220a2083dc1..24f83f9eeaedce12904ce8190edd1e5ed0b67e76 100644 (file)
@@ -321,8 +321,16 @@ static int sw_sync_debugfs_open(struct inode *inode, struct file *file)
 static int sw_sync_debugfs_release(struct inode *inode, struct file *file)
 {
        struct sync_timeline *obj = file->private_data;
+       struct sync_pt *pt, *next;
+
+       spin_lock_irq(&obj->lock);
+
+       list_for_each_entry_safe(pt, next, &obj->pt_list, link) {
+               dma_fence_set_error(&pt->base, -ENOENT);
+               dma_fence_signal_locked(&pt->base);
+       }
 
-       smp_wmb();
+       spin_unlock_irq(&obj->lock);
 
        sync_timeline_put(obj);
        return 0;