btrfs: Correctly handle empty trees in find_first_clear_extent_bit
authorNikolay Borisov <nborisov@suse.com>
Mon, 27 Jan 2020 09:59:26 +0000 (11:59 +0200)
committerDavid Sterba <dsterba@suse.com>
Fri, 31 Jan 2020 13:01:29 +0000 (14:01 +0100)
commit5750c37523a2c8cbb450b9ef31e21c2ba876b05e
tree4e23463e1ff13a95ab832e3d81d73af496d448a9
parent42ffb0bf584ae5b6b38f72259af1e0ee417ac77f
btrfs: Correctly handle empty trees in find_first_clear_extent_bit

Raviu reported that running his regular fs_trim segfaulted with the
following backtrace:

[  237.525947] assertion failed: prev, in ../fs/btrfs/extent_io.c:1595
[  237.525984] ------------[ cut here ]------------
[  237.525985] kernel BUG at ../fs/btrfs/ctree.h:3117!
[  237.525992] invalid opcode: 0000 [#1] SMP PTI
[  237.525998] CPU: 4 PID: 4423 Comm: fstrim Tainted: G     U     OE     5.4.14-8-vanilla #1
[  237.526001] Hardware name: ASUSTeK COMPUTER INC.
[  237.526044] RIP: 0010:assfail.constprop.58+0x18/0x1a [btrfs]
[  237.526079] Call Trace:
[  237.526120]  find_first_clear_extent_bit+0x13d/0x150 [btrfs]
[  237.526148]  btrfs_trim_fs+0x211/0x3f0 [btrfs]
[  237.526184]  btrfs_ioctl_fitrim+0x103/0x170 [btrfs]
[  237.526219]  btrfs_ioctl+0x129a/0x2ed0 [btrfs]
[  237.526227]  ? filemap_map_pages+0x190/0x3d0
[  237.526232]  ? do_filp_open+0xaf/0x110
[  237.526238]  ? _copy_to_user+0x22/0x30
[  237.526242]  ? cp_new_stat+0x150/0x180
[  237.526247]  ? do_vfs_ioctl+0xa4/0x640
[  237.526278]  ? btrfs_ioctl_get_supported_features+0x30/0x30 [btrfs]
[  237.526283]  do_vfs_ioctl+0xa4/0x640
[  237.526288]  ? __do_sys_newfstat+0x3c/0x60
[  237.526292]  ksys_ioctl+0x70/0x80
[  237.526297]  __x64_sys_ioctl+0x16/0x20
[  237.526303]  do_syscall_64+0x5a/0x1c0
[  237.526310]  entry_SYSCALL_64_after_hwframe+0x49/0xbe

That was due to btrfs_fs_device::aloc_tree being empty. Initially I
thought this wasn't possible and as a percaution have put the assert in
find_first_clear_extent_bit. Turns out this is indeed possible and could
happen when a file system with SINGLE data/metadata profile has a 2nd
device added. Until balance is run or a new chunk is allocated on this
device it will be completely empty.

In this case find_first_clear_extent_bit should return the full range
[0, -1ULL] and let the caller handle this i.e for trim the end will be
capped at the size of actual device.

Link: https://lore.kernel.org/linux-btrfs/izW2WNyvy1dEDweBICizKnd2KDwDiDyY2EYQr4YCwk7pkuIpthx-JRn65MPBde00ND6V0_Lh8mW0kZwzDiLDv25pUYWxkskWNJnVP0kgdMA=@protonmail.com/
Fixes: 45bfcfc168f8 ("btrfs: Implement find_first_clear_extent_bit")
CC: stable@vger.kernel.org # 5.2+
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/extent_io.c
fs/btrfs/tests/extent-io-tests.c