drm/syncobj: Don't leak fences when WAIT_FOR_SUBMIT is set

We attempt to get fences earlier in the hopes that everything will
already have fences and no callbacks will be needed.  If we do succeed
in getting a fence, getting one a second time will result in a duplicate
ref with no unref.  This is causing memory leaks in Vulkan applications
that create a lot of fences; playing for a few hours can, apparently,
bring down the system.

Cc: stable@vger.kernel.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107899
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Sean Paul <seanpaul@chromium.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20180926071703.15257-1-jason.ekstrand@intel.com

diff --git a/drivers/gpu/drm/drm_syncobj.c b/drivers/gpu/drm/drm_syncobj.c
index adb3cb2..759278f 100644 (file)
--- a/drivers/gpu/drm/drm_syncobj.c
+++ b/drivers/gpu/drm/drm_syncobj.c
@@ -97,6 +97,8 @@ static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
 {
        int ret;
 
+       WARN_ON(*fence);
+
        *fence = drm_syncobj_fence_get(syncobj);
        if (*fence)
                return 1;
@@ -743,6 +745,9 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 
        if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
                for (i = 0; i < count; ++i) {
+                       if (entries[i].fence)
+                               continue;
+
                        drm_syncobj_fence_get_or_add_callback(syncobjs[i],
                                                              &entries[i].fence,
                                                              &entries[i].syncobj_cb,