ext4: fix another off-by-one fsmap error on 1k block filesystems
authorDarrick J. Wong <djwong@kernel.org>
Thu, 16 Feb 2023 18:55:48 +0000 (10:55 -0800)
committerTheodore Ts'o <tytso@mit.edu>
Wed, 8 Mar 2023 01:20:48 +0000 (20:20 -0500)
commitc993799baf9c5861f8df91beb80e1611b12efcbd
tree44c5a9fbd8ad08563e70e1df60e5caad3431aa05
parentc9f62c8b2dbf7240536c0cc9a4529397bb8bf38e
ext4: fix another off-by-one fsmap error on 1k block filesystems

Apparently syzbot figured out that issuing this FSMAP call:

struct fsmap_head cmd = {
.fmh_count = ...;
.fmh_keys = {
{ .fmr_device = /* ext4 dev */, .fmr_physical = 0, },
{ .fmr_device = /* ext4 dev */, .fmr_physical = 0, },
},
...
};
ret = ioctl(fd, FS_IOC_GETFSMAP, &cmd);

Produces this crash if the underlying filesystem is a 1k-block ext4
filesystem:

kernel BUG at fs/ext4/ext4.h:3331!
invalid opcode: 0000 [#1] PREEMPT SMP
CPU: 3 PID: 3227965 Comm: xfs_io Tainted: G        W  O       6.2.0-rc8-achx
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.15.0-1 04/01/2014
RIP: 0010:ext4_mb_load_buddy_gfp+0x47c/0x570 [ext4]
RSP: 0018:ffffc90007c03998 EFLAGS: 00010246
RAX: ffff888004978000 RBX: ffffc90007c03a20 RCX: ffff888041618000
RDX: 0000000000000000 RSI: 00000000000005a4 RDI: ffffffffa0c99b11
RBP: ffff888012330000 R08: ffffffffa0c2b7d0 R09: 0000000000000400
R10: ffffc90007c03950 R11: 0000000000000000 R12: 0000000000000001
R13: 00000000ffffffff R14: 0000000000000c40 R15: ffff88802678c398
FS:  00007fdf2020c880(0000) GS:ffff88807e100000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007ffd318a5fe8 CR3: 000000007f80f001 CR4: 00000000001706e0
Call Trace:
 <TASK>
 ext4_mballoc_query_range+0x4b/0x210 [ext4 dfa189daddffe8fecd3cdfd00564e0f265a8ab80]
 ext4_getfsmap_datadev+0x713/0x890 [ext4 dfa189daddffe8fecd3cdfd00564e0f265a8ab80]
 ext4_getfsmap+0x2b7/0x330 [ext4 dfa189daddffe8fecd3cdfd00564e0f265a8ab80]
 ext4_ioc_getfsmap+0x153/0x2b0 [ext4 dfa189daddffe8fecd3cdfd00564e0f265a8ab80]
 __ext4_ioctl+0x2a7/0x17e0 [ext4 dfa189daddffe8fecd3cdfd00564e0f265a8ab80]
 __x64_sys_ioctl+0x82/0xa0
 do_syscall_64+0x2b/0x80
 entry_SYSCALL_64_after_hwframe+0x46/0xb0
RIP: 0033:0x7fdf20558aff
RSP: 002b:00007ffd318a9e30 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
RAX: ffffffffffffffda RBX: 00000000000200c0 RCX: 00007fdf20558aff
RDX: 00007fdf1feb2010 RSI: 00000000c0c0583b RDI: 0000000000000003
RBP: 00005625c0634be0 R08: 00005625c0634c40 R09: 0000000000000001
R10: 0000000000000000 R11: 0000000000000246 R12: 00007fdf1feb2010
R13: 00005625be70d994 R14: 0000000000000800 R15: 0000000000000000

For GETFSMAP calls, the caller selects a physical block device by
writing its block number into fsmap_head.fmh_keys[01].fmr_device.
To query mappings for a subrange of the device, the starting byte of the
range is written to fsmap_head.fmh_keys[0].fmr_physical and the last
byte of the range goes in fsmap_head.fmh_keys[1].fmr_physical.

IOWs, to query what mappings overlap with bytes 3-14 of /dev/sda, you'd
set the inputs as follows:

fmh_keys[0] = { .fmr_device = major(8, 0), .fmr_physical = 3},
fmh_keys[1] = { .fmr_device = major(8, 0), .fmr_physical = 14},

Which would return you whatever is mapped in the 12 bytes starting at
physical offset 3.

The crash is due to insufficient range validation of keys[1] in
ext4_getfsmap_datadev.  On 1k-block filesystems, block 0 is not part of
the filesystem, which means that s_first_data_block is nonzero.
ext4_get_group_no_and_offset subtracts this quantity from the blocknr
argument before cracking it into a group number and a block number
within a group.  IOWs, block group 0 spans blocks 1-8192 (1-based)
instead of 0-8191 (0-based) like what happens with larger blocksizes.

The net result of this encoding is that blocknr < s_first_data_block is
not a valid input to this function.  The end_fsb variable is set from
the keys that are copied from userspace, which means that in the above
example, its value is zero.  That leads to an underflow here:

blocknr = blocknr - le32_to_cpu(es->s_first_data_block);

The division then operates on -1:

offset = do_div(blocknr, EXT4_BLOCKS_PER_GROUP(sb)) >>
EXT4_SB(sb)->s_cluster_bits;

Leaving an impossibly large group number (2^32-1) in blocknr.
ext4_getfsmap_check_keys checked that keys[0].fmr_physical and
keys[1].fmr_physical are in increasing order, but
ext4_getfsmap_datadev adjusts keys[0].fmr_physical to be at least
s_first_data_block.  This implies that we have to check it again after
the adjustment, which is the piece that I forgot.

Reported-by: syzbot+6be2b977c89f79b6b153@syzkaller.appspotmail.com
Fixes: 4a4956249dac ("ext4: fix off-by-one fsmap error on 1k block filesystems")
Link: https://syzkaller.appspot.com/bug?id=79d5768e9bfe362911ac1a5057a36fc6b5c30002
Cc: stable@vger.kernel.org
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Link: https://lore.kernel.org/r/Y+58NPTH7VNGgzdd@magnolia
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
fs/ext4/fsmap.c