block: avoid use-after-free on throttle data
authorMing Lei <ming.lei@redhat.com>
Fri, 18 Mar 2022 13:01:42 +0000 (21:01 +0800)
committerJens Axboe <axboe@kernel.dk>
Fri, 18 Mar 2022 15:57:56 +0000 (09:57 -0600)
In throtl_pending_timer_fn(), request queue is retrieved from throttle
data. And tg's pending timer is deleted synchronously when releasing the
associated blkg, at that time, throttle data may have been freed since
commit 1059699f87eb ("block: move blkcg initialization/destroy into disk
allocation/release handler") moves freeing q->td to disk_release() from
blk_release_queue(). So use-after-free on q->td in throtl_pending_timer_fn
can be triggered.

Fixes the issue by:

- do nothing in case that disk is released, when there isn't any bio to
  dispatch

- retrieve request queue from blkg instead of throttle data for
non top-level pending timer.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20220318130144.1066064-2-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-throttle.c

index a3b3ebc72dd47a8bb6f32b07572f54cbf1103b2e..880701d0106fb4b17ac60efb29b6253871da01b4 100644 (file)
@@ -1137,12 +1137,22 @@ static void throtl_pending_timer_fn(struct timer_list *t)
        struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
        struct throtl_grp *tg = sq_to_tg(sq);
        struct throtl_data *td = sq_to_td(sq);
-       struct request_queue *q = td->queue;
        struct throtl_service_queue *parent_sq;
+       struct request_queue *q;
        bool dispatched;
        int ret;
 
+       /* throtl_data may be gone, so figure out request queue by blkg */
+       if (tg)
+               q = tg->pd.blkg->q;
+       else
+               q = td->queue;
+
        spin_lock_irq(&q->queue_lock);
+
+       if (!q->root_blkg)
+               goto out_unlock;
+
        if (throtl_can_upgrade(td, NULL))
                throtl_upgrade_state(td);