blk-mq: fix hang issue in blk_queue_enter()
PATCH 1/1 blk-mq: fix hang issue in blk_queue_enter()
There is a race between blk_queue_enter() and block layer's runtime suspend.
CPU0 CPU1
--------------------------------- -------------------------------
blk_pre_runtime_suspend(q) { blk_queue_enter() {
/* q->rpm_status=RPM_ACTIVE */
blk_set_pm_only(q)
/* q->pm_only=1 */
blk_freeze_queue_start(q)
blk_mq_unfreeze_queue(q)
if (percpu_ref_tryget_live()) {
/* pm=0 && q->pm_only=1 */
if (pm || !blk_queue_pm_only(q)) {
} else {
percpu_ref_put()
}
}
wait_event(q->mq_freeze_wq,
(!q->mq_freeze_depth &&
/* q->rpm_status=RPM_ACTIVE
q->pm_only=1 */
(pm || (blk_pm_request_resume(q),
!blk_queue_pm_only(q)))) ||
blk_queue_dying(q))
} spin_lock_irq(&q->queue_lock)
q->rpm_status = RPM_SUSPENDING
spin_unlock_irq(&q->queue_lock)
}
At this point blk_pm_request_resume() missed the chance to resume the queue, so blk_queue_enter() may wait here forever.
The solution is to wake up the mq_freeze_wq after runtime suspend completed, make blk_pm_request_resume() reexamine the q->rpm_status flag.
--- block/blk-pm.c | 2 ++ 1 file changed, 2 insertions(+)
block/blk-pm.c +2 -0
@@ -132,6 +132,8 @@ void blk_post_runtime_suspend(struct request_queue *q, int err)if (err)blk_clear_pm_only (q); +else +wake_up_all (&q->mq_freeze_wq); }EXPORT_SYMBOL (blk_post_runtime_suspend);
Signature · 1 line
2.17.1
Quoted 46 lines · Yang Yang
There is a race between
blk_queue_enter()and block layer's runtime suspend.CPU0 CPU1 --------------------------------- ------------------------------- blk_pre_runtime_suspend(q) { blk_queue_enter() { /* q->rpm_status=RPM_ACTIVE */ blk_set_pm_only(q) /* q->pm_only=1 */ blk_freeze_queue_start(q) blk_mq_unfreeze_queue(q) if (percpu_ref_tryget_live()) { /* pm=0 && q->pm_only=1 */ if (pm || !blk_queue_pm_only(q)) { } else { percpu_ref_put() } } wait_event(q->mq_freeze_wq, (!q->mq_freeze_depth && /* q->rpm_status=RPM_ACTIVE q->pm_only=1 */ (pm || (blk_pm_request_resume(q), !blk_queue_pm_only(q)))) || blk_queue_dying(q)) }
spin_lock_irq(&q->queue_lock)
q->rpm_status=RPM_SUSPENDING
spin_unlock_irq(&q->queue_lock)
}At this point
blk_pm_request_resume()missed the chance to resume the queue, soblk_queue_enter()may wait here forever.
The solution is to wake up themq_freeze_wqafter runtime suspend completed, makeblk_pm_request_resume()reexamine theq->rpm_statusflag.Signed-off-by Yang Yang <[email protected]>--- block/blk-pm.c | 2 ++ 1 file changed, 2 insertions(+)block/blk-pm.c +0
-0
block/blk-pm.c void blk_post_runtime_suspend(struct request_queue *q, int err) · line 138
Quoted 7 lines
block/blk-pm.c +2
-0@@ -132,6 +132,8 @@ void blk_post_runtime_suspend(struct request_queue *q, int err)if (err)blk_clear_pm_only (q); +else +wake_up_all (&q->mq_freeze_wq); }EXPORT_SYMBOL (blk_post_runtime_suspend);
Please verify whether the following patch series also fixes the reported
hang: https://lore.kernel.org/linux-block/[email protected]/T/#t
Thanks,
Bart.