elevator: Fix a race in elevator switching and md device initialization
authorTomoki Sekiyama <tomoki.sekiyama@hds.com>
Tue, 15 Oct 2013 22:42:16 +0000 (16:42 -0600)
committerJens Axboe <axboe@kernel.dk>
Fri, 8 Nov 2013 16:00:08 +0000 (09:00 -0700)
The soft lockup below happens at the boot time of the system using dm
multipath and the udev rules to switch scheduler.

[  356.127001] BUG: soft lockup - CPU#3 stuck for 22s! [sh:483]
[  356.127001] RIP: 0010:[<ffffffff81072a7d>]  [<ffffffff81072a7d>] lock_timer_base.isra.35+0x1d/0x50
...
[  356.127001] Call Trace:
[  356.127001]  [<ffffffff81073810>] try_to_del_timer_sync+0x20/0x70
[  356.127001]  [<ffffffff8118b08a>] ? kmem_cache_alloc_node_trace+0x20a/0x230
[  356.127001]  [<ffffffff810738b2>] del_timer_sync+0x52/0x60
[  356.127001]  [<ffffffff812ece22>] cfq_exit_queue+0x32/0xf0
[  356.127001]  [<ffffffff812c98df>] elevator_exit+0x2f/0x50
[  356.127001]  [<ffffffff812c9f21>] elevator_change+0xf1/0x1c0
[  356.127001]  [<ffffffff812caa50>] elv_iosched_store+0x20/0x50
[  356.127001]  [<ffffffff812d1d09>] queue_attr_store+0x59/0xb0
[  356.127001]  [<ffffffff812143f6>] sysfs_write_file+0xc6/0x140
[  356.127001]  [<ffffffff811a326d>] vfs_write+0xbd/0x1e0
[  356.127001]  [<ffffffff811a3ca9>] SyS_write+0x49/0xa0
[  356.127001]  [<ffffffff8164e899>] system_call_fastpath+0x16/0x1b

This is caused by a race between md device initialization by multipathd and
shell script to switch the scheduler using sysfs.

 - multipathd:
   SyS_ioctl -> do_vfs_ioctl -> dm_ctl_ioctl -> ctl_ioctl -> table_load
   -> dm_setup_md_queue -> blk_init_allocated_queue -> elevator_init
    q->elevator = elevator_alloc(q, e); // not yet initialized

 - sh -c 'echo deadline > /sys/$DEVPATH/queue/scheduler':
   elevator_switch (in the call trace above)
    struct elevator_queue *old = q->elevator;
    q->elevator = elevator_alloc(q, new_e);
    elevator_exit(old);                 // lockup! (*)

 - multipathd: (cont.)
    err = e->ops.elevator_init_fn(q);   // init fails; q->elevator is modified

(*) When del_timer_sync() is called, lock_timer_base() will loop infinitely
while timer->base == NULL. In this case, as timer will never initialized,
it results in lockup.

This patch introduces acquisition of q->sysfs_lock around elevator_init()
into blk_init_allocated_queue(), to provide mutual exclusion between
initialization of the q->scheduler and switching of the scheduler.

This should fix this bugzilla:
https://bugzilla.redhat.com/show_bug.cgi?id=902012

Signed-off-by: Tomoki Sekiyama <tomoki.sekiyama@hds.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-core.c
block/elevator.c

index 0c611d8..fce4b93 100644 (file)
@@ -741,9 +741,17 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 
        q->sg_reserved_size = INT_MAX;
 
+       /* Protect q->elevator from elevator_change */
+       mutex_lock(&q->sysfs_lock);
+
        /* init elevator */
-       if (elevator_init(q, NULL))
+       if (elevator_init(q, NULL)) {
+               mutex_unlock(&q->sysfs_lock);
                return NULL;
+       }
+
+       mutex_unlock(&q->sysfs_lock);
+
        return q;
 }
 EXPORT_SYMBOL(blk_init_allocated_queue);
index 2bcbd8c..249cc5e 100644 (file)
@@ -186,6 +186,12 @@ int elevator_init(struct request_queue *q, char *name)
        struct elevator_type *e = NULL;
        int err;
 
+       /*
+        * q->sysfs_lock must be held to provide mutual exclusion between
+        * elevator_switch() and here.
+        */
+       lockdep_assert_held(&q->sysfs_lock);
+
        if (unlikely(q->elevator))
                return 0;