projects / GitHub / exynos8895 / android_kernel_samsung_universal8895.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 001a648)
Btrfs: fix scrub race leading to use-after-free
authorFilipe Manana <fdmanana@suse.com>
Tue, 27 Jan 2015 23:06:42 +0000 (23:06 +0000)
committerChris Mason <clm@fb.com>
Tue, 3 Feb 2015 03:24:50 +0000 (19:24 -0800)
While running a scrub on a kernel with CONFIG_DEBUG_PAGEALLOC=y, I got
the following trace:

[68127.807663] BUG: unable to handle kernel paging request at ffff8803f8947a50
[68127.807663] IP: [<ffffffff8107da31>] do_raw_spin_lock+0x94/0x122
[68127.807663] PGD 3003067 PUD 43e1f5067 PMD 43e030067 PTE 80000003f8947060
[68127.807663] Oops: 0000 [#1] SMP DEBUG_PAGEALLOC
[68127.807663] Modules linked in: dm_flakey dm_mod crc32c_generic btrfs xor raid6_pq nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace fscache sunrpc loop parport_pc processor parpo
[68127.807663] CPU: 2 PID: 3081 Comm: kworker/u8:5 Not tainted 3.18.0-rc6-btrfs-next-3+ #4
[68127.807663] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[68127.807663] Workqueue: btrfs-btrfs-scrub btrfs_scrub_helper [btrfs]
[68127.807663] task: ffff880101fc5250 ti: ffff8803f097c000 task.ti: ffff8803f097c000
[68127.807663] RIP: 0010:[<ffffffff8107da31>]  [<ffffffff8107da31>] do_raw_spin_lock+0x94/0x122
[68127.807663] RSP: 0018:ffff8803f097fbb8  EFLAGS: 00010093
[68127.807663] RAX: 0000000028dd386c RBX: ffff8803f8947a50 RCX: 0000000028dd3854
[68127.807663] RDX: 0000000000000018 RSI: 0000000000000002 RDI: 0000000000000001
[68127.807663] RBP: ffff8803f097fbd8 R08: 0000000000000004 R09: 0000000000000001
[68127.807663] R10: ffff880102620980 R11: ffff8801f3e8c900 R12: 000000000001d390
[68127.807663] R13: 00000000cabd13c8 R14: ffff8803f8947800 R15: ffff88037c574f00
[68127.807663] FS:  0000000000000000(0000) GS:ffff88043dd00000(0000) knlGS:0000000000000000
[68127.807663] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[68127.807663] CR2: ffff8803f8947a50 CR3: 00000000b6481000 CR4: 00000000000006e0
[68127.807663] Stack:
[68127.807663]  ffffffff823942a8 ffff8803f8947a50 ffff8802a3416f80 0000000000000000
[68127.807663]  ffff8803f097fc18 ffffffff8141e7c0 ffffffff81072948 000000000034f314
[68127.807663]  ffff8803f097fc08 0000000000000292 ffff8803f097fc48 ffff8803f8947a50
[68127.807663] Call Trace:
[68127.807663]  [<ffffffff8141e7c0>] _raw_spin_lock_irqsave+0x4b/0x55
[68127.807663]  [<ffffffff81072948>] ? __wake_up+0x22/0x4b
[68127.807663]  [<ffffffff81072948>] __wake_up+0x22/0x4b
[68127.807663]  [<ffffffffa0392327>] scrub_pending_bio_dec+0x32/0x36 [btrfs]
[68127.807663]  [<ffffffffa0395e70>] scrub_bio_end_io_worker+0x5a3/0x5c9 [btrfs]
[68127.807663]  [<ffffffff810e0c7c>] ? time_hardirqs_off+0x15/0x28
[68127.807663]  [<ffffffff81078106>] ? trace_hardirqs_off_caller+0x4c/0xb9
[68127.807663]  [<ffffffffa0372a7c>] normal_work_helper+0xf1/0x238 [btrfs]
[68127.807663]  [<ffffffffa0372d3d>] btrfs_scrub_helper+0x12/0x14 [btrfs]
[68127.807663]  [<ffffffff810582d2>] process_one_work+0x1e4/0x3b6
[68127.807663]  [<ffffffff81078180>] ? trace_hardirqs_off+0xd/0xf
[68127.807663]  [<ffffffff81058dc9>] worker_thread+0x1fb/0x2a8
[68127.807663]  [<ffffffff81058bce>] ? rescuer_thread+0x219/0x219
[68127.807663]  [<ffffffff8105cd75>] kthread+0xdb/0xe3
[68127.807663]  [<ffffffff8105cc9a>] ? __kthread_parkme+0x67/0x67
[68127.807663]  [<ffffffff8141f1ec>] ret_from_fork+0x7c/0xb0
[68127.807663]  [<ffffffff8105cc9a>] ? __kthread_parkme+0x67/0x67
[68127.807663] Code: 39 c2 75 14 8d 8a 00 00 01 00 89 d0 f0 0f b1 0b 39 d0 0f 84 81 00 00 00 4c 69 2d 27 86 99 00 fa 00 00 00 45 31 e4 4d 39 ec 74 2b <8b> 13 89 d0 c1 e8 10 66 39 c2 75
[68127.807663] RIP  [<ffffffff8107da31>] do_raw_spin_lock+0x94/0x122
[68127.807663]  RSP <ffff8803f097fbb8>
[68127.807663] CR2: ffff8803f8947a50
[68127.807663] ---[ end trace d7045aac00a66cd8 ]---

This is due to a race that can happen in a very tiny time window and is
illustrated by the following sequence diagram:

         CPU 1                                                     CPU 2

                                                                btrfs_scrub_dev()
scrub_bio_end_io_worker()
   scrub_pending_bio_dec()
       atomic_dec(&sctx->bios_in_flight)
                                                                   wait sctx->bios_in_flight == 0
                                                                   wait sctx->workers_pending == 0
                                                                   mutex_lock(&fs_info->scrub_lock)
                                                                   (...)
                                                                   mutex_lock(&fs_info->scrub_lock)
                                                                   scrub_free_ctx(sctx)
                                                                      kfree(sctx)
       wake_up(&sctx->list_wait)
          __wake_up()
              spin_lock_irqsave(&sctx->list_wait->lock, flags)

Another variation of this scenario that results in the same use-after-free
issue is:

         CPU 1                                                     CPU 2

                                                                btrfs_scrub_dev()
                                                                   wait sctx->bios_in_flight == 0
scrub_bio_end_io_worker()
   scrub_pending_bio_dec()
       __wake_up(&sctx->list_wait)
          spin_lock_irqsave(&sctx->list_wait->lock, flags)
          default_wake_function()
              wake up task at CPU 2
                                                                   wait sctx->workers_pending == 0
                                                                   mutex_lock(&fs_info->scrub_lock)
                                                                   (...)
                                                                   mutex_lock(&fs_info->scrub_lock)
                                                                   scrub_free_ctx(sctx)
                                                                      kfree(sctx)
          spin_unlock_irqrestore(&sctx->list_wait->lock, flags)

Fix this by holding the scrub lock while doing the wakeup.

This isn't a recent regression, the issue as been around since the scrub
feature was added (2011, commit a2de733c78fa7af51ba9670482fa7d392aa67c57).

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/scrub.c patch | blob | blame | history

diff --git a/fs/btrfs/scrub.c b/fs/btrfs/scrub.c
index 8af7372238fc13393772000189b9a083ba8f9f7e..6d6e155c8c8bb0773971e22d403aaa6b623cf34f 100644 (file)
--- a/fs/btrfs/scrub.c
+++ b/fs/btrfs/scrub.c
@@ -306,8 +306,17 @@ static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 
 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 {
+       struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
+
+       /*
+        * Hold the scrub_lock while doing the wakeup to ensure the
+        * sctx (and its wait queue list_wait) isn't destroyed/freed
+        * during the wakeup.
+        */
+       mutex_lock(&fs_info->scrub_lock);
        atomic_dec(&sctx->bios_in_flight);
        wake_up(&sctx->list_wait);
+       mutex_unlock(&fs_info->scrub_lock);
 }
 
 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
@@ -379,10 +388,15 @@ static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
        mutex_lock(&fs_info->scrub_lock);
        atomic_dec(&fs_info->scrubs_running);
        atomic_dec(&fs_info->scrubs_paused);
-       mutex_unlock(&fs_info->scrub_lock);
        atomic_dec(&sctx->workers_pending);
        wake_up(&fs_info->scrub_pause_wait);
+       /*
+        * Hold the scrub_lock while doing the wakeup to ensure the
+        * sctx (and its wait queue list_wait) isn't destroyed/freed
+        * during the wakeup.
+        */
        wake_up(&sctx->list_wait);
+       mutex_unlock(&fs_info->scrub_lock);
 }
 
 static void scrub_free_csums(struct scrub_ctx *sctx)