Btrfs: fix freeing used extents after removing empty block group
authorFilipe Manana <fdmanana@suse.com>
Wed, 26 Nov 2014 15:28:52 +0000 (15:28 +0000)
committerChris Mason <clm@fb.com>
Wed, 3 Dec 2014 02:19:17 +0000 (18:19 -0800)
commitae0ab003f2840e7ed66ed762ec0d4e298ad1e725
treefecd1fe73a1140f30cfaf47c93195d4938b936a6
parent4f69cb987ecbb6a87b5b1d56d6e9f9b5fee69737
Btrfs: fix freeing used extents after removing empty block group

There's a race between adding a block group to the list of the unused
block groups and removing an unused block group (cleaner kthread) that
leads to freeing extents that are in use or a crash during transaction
commmit. Basically the cleaner kthread, when executing
btrfs_delete_unused_bgs(), might catch the newly added block group to
the list fs_info->unused_bgs and clear the range representing the whole
group from fs_info->freed_extents[] before the task that added the block
group to the list (running update_block_group()) marked the last freed
extent as dirty in fs_info->freed_extents (pinned_extents).

That is:

     CPU 1                                CPU 2

                                  btrfs_delete_unused_bgs()
update_block_group()
   add block group to
   fs_info->unused_bgs
                                    got block group from the list
                                    clear_extent_bits for the whole
                                    block group range in freed_extents[]
   set_extent_dirty for the
   range covering the freed
   extent in freed_extents[]
   (fs_info->pinned_extents)

                                  block group deleted, and a new block
                                  group with the same logical address is
                                  created

                                  reserve space from the new block group
                                  for new data or metadata - the reserved
                                  space overlaps the range specified by
                                  CPU 1 for set_extent_dirty()

                                  commit transaction
                                    find all ranges marked as dirty in
                                    fs_info->pinned_extents, clear them
                                    and add them to the free space cache

Alternatively, if CPU 2 doesn't create a new block group with the same
logical address, we get a crash/BUG_ON at transaction commit when unpining
extent ranges because we can't find a block group for the range marked as
dirty by CPU 1. Sample trace:

[ 2163.426462] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
[ 2163.426640] Modules linked in: btrfs xor raid6_pq dm_thin_pool dm_persistent_data dm_bio_prison dm_bufio crc32c_generic libcrc32c dm_mod nfsd auth_rpc
gss oid_registry nfs_acl nfs lockd fscache sunrpc loop psmouse parport_pc parport i2c_piix4 processor thermal_sys i2ccore evdev button pcspkr microcode serio_raw ext4 crc16 jbd2 mbcache
 sg sr_mod cdrom sd_mod crc_t10dif crct10dif_generic crct10dif_common ata_generic virtio_scsi floppy ata_piix libata e1000 scsi_mod virtio_pci virtio_ring virtio
[ 2163.428209] CPU: 0 PID: 11858 Comm: btrfs-transacti Tainted: G        W      3.17.0-rc5-btrfs-next-1+ #1
[ 2163.428519] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[ 2163.428875] task: ffff88009f2c0650 ti: ffff8801356bc000 task.ti: ffff8801356bc000
[ 2163.429157] RIP: 0010:[<ffffffffa037728e>]  [<ffffffffa037728e>] unpin_extent_range.isra.58+0x62/0x192 [btrfs]
[ 2163.429562] RSP: 0018:ffff8801356bfda8  EFLAGS: 00010246
[ 2163.429802] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
[ 2163.429990] RDX: 0000000041bfffff RSI: 0000000001c00000 RDI: ffff880024307080
[ 2163.430042] RBP: ffff8801356bfde8 R08: 0000000000000068 R09: ffff88003734f118
[ 2163.430042] R10: ffff8801356bfcb8 R11: fffffffffffffb69 R12: ffff8800243070d0
[ 2163.430042] R13: 0000000083c04000 R14: ffff8800751b0f00 R15: ffff880024307000
[ 2163.430042] FS:  0000000000000000(0000) GS:ffff88013f400000(0000) knlGS:0000000000000000
[ 2163.430042] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[ 2163.430042] CR2: 00007ff10eb43fc0 CR3: 0000000004cb8000 CR4: 00000000000006f0
[ 2163.430042] Stack:
[ 2163.430042]  ffff8800243070d0 0000000083c08000 0000000083c07fff ffff88012d6bc800
[ 2163.430042]  ffff8800243070d0 ffff8800751b0f18 ffff8800751b0f00 0000000000000000
[ 2163.430042]  ffff8801356bfe18 ffffffffa037a481 0000000083c04000 0000000083c07fff
[ 2163.430042] Call Trace:
[ 2163.430042]  [<ffffffffa037a481>] btrfs_finish_extent_commit+0xac/0xbf [btrfs]
[ 2163.430042]  [<ffffffffa038c06d>] btrfs_commit_transaction+0x6ee/0x882 [btrfs]
[ 2163.430042]  [<ffffffffa03881f1>] transaction_kthread+0xf2/0x1a4 [btrfs]
[ 2163.430042]  [<ffffffffa03880ff>] ? btrfs_cleanup_transaction+0x3d8/0x3d8 [btrfs]
[ 2163.430042]  [<ffffffff8105966b>] kthread+0xb7/0xbf
[ 2163.430042]  [<ffffffff810595b4>] ? __kthread_parkme+0x67/0x67
[ 2163.430042]  [<ffffffff813ebeac>] ret_from_fork+0x7c/0xb0
[ 2163.430042]  [<ffffffff810595b4>] ? __kthread_parkme+0x67/0x67

So fix this by making update_block_group() first set the range as dirty
in pinned_extents before adding the block group to the unused_bgs list.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c