cfq-iosched: fix use-after-free of cfqq
Jens Axboe [Tue, 17 Jan 2012 20:26:11 +0000 (21:26 +0100)]
With the changes in life time management between the cfq IO contexts
and the cfq queues, we now risk having cfqd->active_queue being
freed when cfq_slice_expired() is being called. cfq_preempt_queue()
caches this queue and uses it after calling said function, causing
a use-after-free condition. This triggers the following oops,
when cfqq_type() attempts to dereference it:

BUG: unable to handle kernel paging request at ffff8800746c4f0c
IP: [<ffffffff81266d59>] cfqq_type+0xb/0x20
PGD 18d4063 PUD 1fe15067 PMD 1ffb9067 PTE 80000000746c4160
Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
CPU 3
Modules linked in:

Pid: 1, comm: init Not tainted 3.2.0-josef+ #367 Bochs Bochs
RIP: 0010:[<ffffffff81266d59>]  [<ffffffff81266d59>] cfqq_type+0xb/0x20
RSP: 0018:ffff880079c11778  EFLAGS: 00010046
RAX: 0000000000000000 RBX: ffff880076f3df08 RCX: 0000000000000000
RDX: 0000000000000006 RSI: ffff880074271888 RDI: ffff8800746c4f08
RBP: ffff880079c11778 R08: 0000000000000078 R09: 0000000000000001
R10: 09f911029d74e35b R11: 09f911029d74e35b R12: ffff880076f337f0
R13: ffff8800746c4f08 R14: ffff8800746c4f08 R15: 0000000000000002
FS:  00007f62fd44f700(0000) GS:ffff88007cd80000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: ffff8800746c4f0c CR3: 0000000076c21000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process init (pid: 1, threadinfo ffff880079c10000, task ffff880079c0a040)
Stack:
 ffff880079c117c8 ffffffff812683d8 ffff880079c117a8 ffffffff8125de43
 ffff8800744fcf48 ffff880074b43e98 ffff8800770c8828 ffff880074b43e98
 0000000000000003 0000000000000000 ffff880079c117f8 ffffffff81254149
Call Trace:
 [<ffffffff812683d8>] cfq_insert_request+0x3f5/0x47c
 [<ffffffff8125de43>] ? blk_recount_segments+0x20/0x31
 [<ffffffff81254149>] __elv_add_request+0x1ca/0x200
 [<ffffffff8125aa99>] blk_queue_bio+0x2ef/0x312
 [<ffffffff81258f7b>] generic_make_request+0x9f/0xe0
 [<ffffffff8125907b>] submit_bio+0xbf/0xca
 [<ffffffff81136ec7>] submit_bh+0xdf/0xfe
 [<ffffffff81176d04>] ext3_bread+0x50/0x99
 [<ffffffff811785b3>] dx_probe+0x38/0x291
 [<ffffffff81178864>] ext3_dx_find_entry+0x58/0x219
 [<ffffffff81178ad5>] ext3_find_entry+0xb0/0x406
 [<ffffffff8110c4d5>] ? cache_alloc_debugcheck_after.isra.46+0x14d/0x1a0
 [<ffffffff8110cfbd>] ? kmem_cache_alloc+0xef/0x191
 [<ffffffff8117a330>] ext3_lookup+0x39/0xe1
 [<ffffffff81119461>] d_alloc_and_lookup+0x45/0x6c
 [<ffffffff8111ac41>] do_lookup+0x1e4/0x2f5
 [<ffffffff8111aef6>] link_path_walk+0x1a4/0x6ef
 [<ffffffff8111b557>] path_lookupat+0x59/0x5ea
 [<ffffffff8127406c>] ? __strncpy_from_user+0x30/0x5a
 [<ffffffff8111bce0>] do_path_lookup+0x23/0x59
 [<ffffffff8111cfd6>] user_path_at_empty+0x53/0x99
 [<ffffffff8107b37b>] ? remove_wait_queue+0x51/0x56
 [<ffffffff8111d02d>] user_path_at+0x11/0x13
 [<ffffffff811141f5>] vfs_fstatat+0x3a/0x64
 [<ffffffff8111425a>] vfs_stat+0x1b/0x1d
 [<ffffffff81114359>] sys_newstat+0x1a/0x33
 [<ffffffff81060e12>] ? task_stopped_code+0x42/0x42
 [<ffffffff815d6712>] system_call_fastpath+0x16/0x1b
Code: 89 e6 48 89 c7 e8 fa ca fe ff 85 c0 74 06 4c 89 2b 41 b6 01 5b 44 89 f0 41 5c 41 5d 41 5e 5d c3 55 48 89 e5 66 66 66 66 90 31 c0 <8b> 57 04 f6 c6 01 74 0b 83 e2 20 83 fa 01 19 c0 83 c0 02 5d c3
RIP  [<ffffffff81266d59>] cfqq_type+0xb/0x20
 RSP <ffff880079c11778>
CR2: ffff8800746c4f0c

Get rid of the caching of cfqd->active_queue, and reorder the
check so that it happens before we expire the active queue.

Thanks to Tejun for pin pointing the error location.

Reported-by: Chris Mason <chris.mason@oracle.com>
Tested-by: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/cfq-iosched.c

index 163263d..ee55019 100644 (file)
@@ -3117,18 +3117,17 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
  */
 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
-       struct cfq_queue *old_cfqq = cfqd->active_queue;
-
        cfq_log_cfqq(cfqd, cfqq, "preempt");
-       cfq_slice_expired(cfqd, 1);
 
        /*
         * workload type is changed, don't save slice, otherwise preempt
         * doesn't happen
         */
-       if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
+       if (cfqq_type(cfqd->active_queue) != cfqq_type(cfqq))
                cfqq->cfqg->saved_workload_slice = 0;
 
+       cfq_slice_expired(cfqd, 1);
+
        /*
         * Put the new queue at the front of the of the current list,
         * so we know that it will be selected next.