Re: [PATCH] ext4: fix another off-by-one fsmap error on 1k block filesystems

Tudor Ambarus <tudor.ambarus@xxxxxxxxxx> · Fri, 17 Feb 2023 09:04:49 +0000

On 2/17/23 08:37, Tudor Ambarus wrote:


On 2/17/23 08:10, Tudor Ambarus wrote:
Hi, Darrick,

On 2/16/23 18:55, Darrick J. Wong wrote:
From: Darrick J. Wong <djwong@xxxxxxxxxx>

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.

Fixes: 4a4956249dac ("ext4: fix off-by-one fsmap error on 1k block 
filesystems")
Link: 
https://syzkaller.appspot.com/bug?id=79d5768e9bfe362911ac1a5057a36fc6b5c30002
Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
---
  fs/ext4/fsmap.c |    2 ++
  1 file changed, 2 insertions(+)

diff --git a/fs/ext4/fsmap.c b/fs/ext4/fsmap.c
index 4493ef0c715e..cdf9bfe10137 100644
--- a/fs/ext4/fsmap.c
+++ b/fs/ext4/fsmap.c
@@ -486,6 +486,8 @@ static int ext4_getfsmap_datadev(struct 
super_block *sb,
          keys[0].fmr_physical = bofs;
      if (keys[1].fmr_physical >= eofs)
          keys[1].fmr_physical = eofs - 1;
+    if (keys[1].fmr_physical < keys[0].fmr_physical)
+        return 0;

This is an indirect implication, we can be more straightforward. Also we
should stop the execution when high_key->fmr_physical == bofs. So maybe:

--- a/fs/ext4/fsmap.c
+++ b/fs/ext4/fsmap.c
@@ -479,6 +479,8 @@ static int ext4_getfsmap_datadev(struct 
super_block *sb,
         int error = 0;

         bofs = le32_to_cpu(sbi->s_es->s_first_data_block);
+       if (keys[1].fmr_physical <= bofs)
+               return 0;
         eofs = ext4_blocks_count(sbi->s_es);
         if (keys[0].fmr_physical >= eofs)
                 return 0;

oops, it was too morning for me, this is wrong, as the high key can come 
with zero value. So if we want to keep the check in

or maybe not? xfsprogs-dev/io/fsmap.c initializes h->fmr_physical = end;
and end = -1 for the ext4 case.

From the code I see two magic values that are accepted for fmr_device,
zero and UINT_MAX, so maybe we should check the sanity of the keys's
fmr_physical considering the magic values for fmr_device. I'll stop
burbling and wait for Darrick's feedback.

Cheers,
ta

ext4_getfsmap_datadev we can use your patch, but slightly adapted to
if (keys[1].fmr_physical <= keys[0].fmr_physical)
     return 0;

However I think I prefer the version from below if it makes sense.
Cheers,
ta



But I'm not thrilled about this. Can we stop the execution earlier?
How about something like:

diff --git a/fs/ext4/fsmap.c b/fs/ext4/fsmap.c
index 4493ef0c715e..8c4b1fb5c10a 100644
--- a/fs/ext4/fsmap.c
+++ b/fs/ext4/fsmap.c
@@ -586,10 +586,12 @@ static bool ext4_getfsmap_check_keys(struct 
ext4_fsmap *low_key,
  {
         if (low_key->fmr_device > high_key->fmr_device)
                 return false;
-       if (low_key->fmr_device < high_key->fmr_device)
+       if (low_key->fmr_device != 0 &&
+           low_key->fmr_device < high_key->fmr_device)
                 return true;

-       if (low_key->fmr_physical > high_key->fmr_physical)
+       if (high_key->fmr_physical == 0 ||
+           low_key->fmr_physical > high_key->fmr_physical)
                 return false;
         if (low_key->fmr_physical < high_key->fmr_physical)
                 return true;

You'll notice that my proposal will stop the execution with -EINVAL, as
the key range is invalid IMO. But we can adapt code to return zero if
you like, what I'm interested in is whether this key check is valid for
ext4_getfsmap_logdev() or not. How about the
if (keys[1].fmr_physical == bofs) check, is it valid for
ext4_getfsmap_logdev()?

I'd like that with your guidance to fix this bug, extend the xfs_io
fsmap support to query ext4 fs range and to add an xfstest for keys
validation, to check if stable kernels will be fixed. This would help me
with the fs ramp-up, as I'd like to fix some other fs bugs too. If you
want to handle this by yourself, it's fine too, but please let me know
so that we don't duplicate effort.

Cheers,
ta






