blk-cgroup: Fix RCU correctness warning in cfq_init_queue()
authorVivek Goyal <vgoyal@redhat.com>
Thu, 22 Apr 2010 15:54:52 +0000 (11:54 -0400)
committerJens Axboe <jens.axboe@oracle.com>
Thu, 6 May 2010 06:54:00 +0000 (08:54 +0200)
It is necessary to be in an RCU read-side critical section when invoking
css_id(), so this patch adds one to blkiocg_add_blkio_group().  This is
actually a false positive, because this is called at initialization time
and hence always refers to the root cgroup, which cannot go away.

[  103.790505] ===================================================
[  103.790509] [ INFO: suspicious rcu_dereference_check() usage. ]
[  103.790511] ---------------------------------------------------
[  103.790514] kernel/cgroup.c:4432 invoked rcu_dereference_check() without protection!
[  103.790517]
[  103.790517] other info that might help us debug this:
[  103.790519]
[  103.790521]
[  103.790521] rcu_scheduler_active = 1, debug_locks = 1
[  103.790524] 4 locks held by bash/4422:
[  103.790526]  #0:  (&buffer->mutex){+.+.+.}, at: [<ffffffff8114befa>] sysfs_write_file+0x3c/0x144
[  103.790537]  #1:  (s_active#102){.+.+.+}, at: [<ffffffff8114bfa5>] sysfs_write_file+0xe7/0x144
[  103.790544]  #2:  (&q->sysfs_lock){+.+.+.}, at: [<ffffffff812263b1>] queue_attr_store+0x49/0x8f
[  103.790552]  #3:  (&(&blkcg->lock)->rlock){......}, at: [<ffffffff8122e4db>] blkiocg_add_blkio_group+0x2b/0xad
[  103.790560]
[  103.790561] stack backtrace:
[  103.790564] Pid: 4422, comm: bash Not tainted 2.6.34-rc4-blkio-second-crash #81
[  103.790567] Call Trace:
[  103.790572]  [<ffffffff81068f57>] lockdep_rcu_dereference+0x9d/0xa5
[  103.790577]  [<ffffffff8107fac1>] css_id+0x44/0x57
[  103.790581]  [<ffffffff8122e503>] blkiocg_add_blkio_group+0x53/0xad
[  103.790586]  [<ffffffff81231936>] cfq_init_queue+0x139/0x32c
[  103.790591]  [<ffffffff8121f2d0>] elv_iosched_store+0xbf/0x1bf
[  103.790595]  [<ffffffff812263d8>] queue_attr_store+0x70/0x8f
[  103.790599]  [<ffffffff8114bfa5>] ? sysfs_write_file+0xe7/0x144
[  103.790603]  [<ffffffff8114bfc6>] sysfs_write_file+0x108/0x144
[  103.790609]  [<ffffffff810f527f>] vfs_write+0xae/0x10b
[  103.790612]  [<ffffffff81069863>] ? trace_hardirqs_on_caller+0x10c/0x130
[  103.790616]  [<ffffffff810f539c>] sys_write+0x4a/0x6e
[  103.790622]  [<ffffffff81002b5b>] system_call_fastpath+0x16/0x1b
[  103.790625]

Located-by: Miles Lane <miles.lane@gmail.com>
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
block/cfq-iosched.c

index 838834be115b3f341f3bc6e0eba54a5c3fcc424c..5f127cfb2e924baf06f77ae95d1832e13048cee4 100644 (file)
@@ -3694,8 +3694,10 @@ static void *cfq_init_queue(struct request_queue *q)
         * to make sure that cfq_put_cfqg() does not try to kfree root group
         */
        atomic_set(&cfqg->ref, 1);
+       rcu_read_lock();
        blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
                                        0);
+       rcu_read_unlock();
 #endif
        /*
         * Not strictly needed (since RB_ROOT just clears the node and we