drm/nouveau/tmr: fix corruption of the pending list when rescheduling an alarm

At least therm/fantog "attempts" to work around this issue, which could
lead to corruption of the pending alarm list.

Fix it properly by not updating the timestamp without the lock held, or
trying to add an already pending alarm to the pending alarm list....

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Cc: stable@vger.kernel.org

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c
index 4e69662..d898787 100644 (file)
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c
@@ -71,14 +71,17 @@ nvkm_timer_alarm(struct nvkm_timer *tmr, u32 nsec, struct nvkm_alarm *alarm)
        struct nvkm_alarm *list;
        unsigned long flags;
 
-       alarm->timestamp = nvkm_timer_read(tmr) + nsec;
-
-       /* append new alarm to list, in soonest-alarm-first order */
+       /* Remove alarm from pending list.
+        *
+        * This both protects against the corruption of the list,
+        * and implements alarm rescheduling/cancellation.
+        */
        spin_lock_irqsave(&tmr->lock, flags);
-       if (!nsec) {
-               if (!list_empty(&alarm->head))
-                       list_del(&alarm->head);
-       } else {
+       list_del_init(&alarm->head);
+
+       if (nsec) {
+               /* Insert into pending list, ordered earliest to latest. */
+               alarm->timestamp = nvkm_timer_read(tmr) + nsec;
                list_for_each_entry(list, &tmr->alarms, head) {
                        if (list->timestamp > alarm->timestamp)
                                break;