Btrfs: fix race when finishing dev replace leading to transaction abort
authorFilipe Manana <fdmanana@suse.com>
Fri, 20 Nov 2015 10:42:47 +0000 (10:42 +0000)
committerFilipe Manana <fdmanana@suse.com>
Thu, 17 Dec 2015 10:59:46 +0000 (10:59 +0000)
commit50460e37186a2b932eacea24dca804bd1bcd2012
tree707fa5c462e795ecc6078ddafa5916b17c279450
parent1d3a5a82fe724c53c472a18a31fb0bbf33dfaba2
Btrfs: fix race when finishing dev replace leading to transaction abort

During the final phase of a device replace operation, I ran into a
transaction abort that resulted in the following trace:

[23919.655368] WARNING: CPU: 10 PID: 30175 at fs/btrfs/extent-tree.c:9843 btrfs_create_pending_block_groups+0x15e/0x1ab [btrfs]()
[23919.664742] BTRFS: Transaction aborted (error -2)
[23919.665749] Modules linked in: btrfs crc32c_generic xor raid6_pq nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace fscache sunrpc loop fuse parport_pc i2c_piix4 parport psmouse acpi_cpufreq processor i2c_core evdev microcode pcspkr button serio_raw ext4 crc16 jbd2 mbcache sd_mod sg sr_mod cdrom virtio_scsi ata_generic ata_piix virtio_pci floppy virtio_ring libata e1000 virtio scsi_mod [last unloaded: btrfs]
[23919.679442] CPU: 10 PID: 30175 Comm: fsstress Not tainted 4.3.0-rc5-btrfs-next-17+ #1
[23919.682392] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.8.1-0-g4adadbd-20150316_085822-nilsson.home.kraxel.org 04/01/2014
[23919.689151]  0000000000000000 ffff8804020cbb50 ffffffff812566f4 ffff8804020cbb98
[23919.692604]  ffff8804020cbb88 ffffffff8104d0a6 ffffffffa03eea69 ffff88041b678a48
[23919.694230]  ffff88042ac38000 ffff88041b678930 00000000fffffffe ffff8804020cbbf0
[23919.696716] Call Trace:
[23919.698669]  [<ffffffff812566f4>] dump_stack+0x4e/0x79
[23919.700597]  [<ffffffff8104d0a6>] warn_slowpath_common+0x9f/0xb8
[23919.701958]  [<ffffffffa03eea69>] ? btrfs_create_pending_block_groups+0x15e/0x1ab [btrfs]
[23919.703612]  [<ffffffff8104d107>] warn_slowpath_fmt+0x48/0x50
[23919.705047]  [<ffffffffa03eea69>] btrfs_create_pending_block_groups+0x15e/0x1ab [btrfs]
[23919.706967]  [<ffffffffa0402097>] __btrfs_end_transaction+0x84/0x2dd [btrfs]
[23919.708611]  [<ffffffffa0402300>] btrfs_end_transaction+0x10/0x12 [btrfs]
[23919.710099]  [<ffffffffa03ef0b8>] btrfs_alloc_data_chunk_ondemand+0x121/0x28b [btrfs]
[23919.711970]  [<ffffffffa0413025>] btrfs_fallocate+0x7d3/0xc6d [btrfs]
[23919.713602]  [<ffffffff8108b78f>] ? lock_acquire+0x10d/0x194
[23919.714756]  [<ffffffff81086dbc>] ? percpu_down_read+0x51/0x78
[23919.716155]  [<ffffffff8116ef1d>] ? __sb_start_write+0x5f/0xb0
[23919.718918]  [<ffffffff8116ef1d>] ? __sb_start_write+0x5f/0xb0
[23919.724170]  [<ffffffff8116b579>] vfs_fallocate+0x170/0x1ff
[23919.725482]  [<ffffffff8117c1d7>] ioctl_preallocate+0x89/0x9b
[23919.726790]  [<ffffffff8117c5ef>] do_vfs_ioctl+0x406/0x4e6
[23919.728428]  [<ffffffff81171175>] ? SYSC_newfstat+0x25/0x2e
[23919.729642]  [<ffffffff8118574d>] ? __fget_light+0x4d/0x71
[23919.730782]  [<ffffffff8117c726>] SyS_ioctl+0x57/0x79
[23919.731847]  [<ffffffff8147cd97>] entry_SYSCALL_64_fastpath+0x12/0x6f
[23919.733330] ---[ end trace 166ef301a335832a ]---

This is due to a race between device replace and chunk allocation, which
the following diagram illustrates:

         CPU 1                                    CPU 2

 btrfs_dev_replace_finishing()

   at this point
    dev_replace->tgtdev->devid ==
    BTRFS_DEV_REPLACE_DEVID (0ULL)

   ...

   btrfs_start_transaction()
   btrfs_commit_transaction()

                                               btrfs_fallocate()
                                                 btrfs_alloc_data_chunk_ondemand()
                                                   btrfs_join_transaction()
                                                     --> starts a new transaction
                                                   do_chunk_alloc()
                                                     lock fs_info->chunk_mutex
                                                       btrfs_alloc_chunk()
                                                         --> creates extent map for
                                                             the new chunk with
                                                             em->bdev->map->stripes[i]->dev->devid
                                                             == X (X > 0)
                                                         --> extent map is added to
                                                             fs_info->mapping_tree
                                                         --> initial phase of bg A
                                                             allocation completes
                                                     unlock fs_info->chunk_mutex

   lock fs_info->chunk_mutex

   btrfs_dev_replace_update_device_in_mapping_tree()
     --> iterates fs_info->mapping_tree and
         replaces the device in every extent
         map's map->stripes[] with
         dev_replace->tgtdev, which still has
         an id of 0ULL (BTRFS_DEV_REPLACE_DEVID)

                                                   btrfs_end_transaction()
                                                     btrfs_create_pending_block_groups()
                                                       --> starts final phase of
                                                           bg A creation (update device,
                                                           extent, and chunk trees, etc)
                                                       btrfs_finish_chunk_alloc()

                                                         btrfs_update_device()
                                                           --> attempts to update a device
                                                               item with ID == 0ULL
                                                               (BTRFS_DEV_REPLACE_DEVID)
                                                               which is the current ID of
                                                               bg A's
                                                               em->bdev->map->stripes[i]->dev->devid
                                                           --> doesn't find such item
                                                               returns -ENOENT
                                                           --> the device id should have been X
                                                               and not 0ULL

                                                       got -ENOENT from
                                                       btrfs_finish_chunk_alloc()
                                                       and aborts current transaction

   finishes setting up the target device,
   namely it sets tgtdev->devid to the value
   of srcdev->devid, which is X (and X > 0)

   frees the srcdev

   unlock fs_info->chunk_mutex

So fix this by taking the device list mutex when processing the chunk's
extent map stripes to update the device items. This avoids getting the
wrong device id and use-after-free problems if the task finishing a
chunk allocation grabs the replaced device, which is freed while the
dev replace task is holding the device list mutex.

This happened while running fstest btrfs/071.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
fs/btrfs/volumes.c