Re: [Regression 4.7-rc1] btrfs: bugfix: handle FS_IOC32_{GETFLAGS,SETFLAGS,GETVERSION} in btrfs_ioctl

[Jeff Mahoney]

On 1/9/17 6:28 AM, David Sterba wrote:
> On Fri, Jan 06, 2017 at 12:22:34PM -0500, Joseph Salisbury wrote:
>> A kernel bug report was opened against Ubuntu [0].  This bug was fixed
>> by the following commit in v4.7-rc1:
>>
>> commit 4c63c2454eff996c5e27991221106eb511f7db38
>>
>> Author: Luke Dashjr <luke@xxxxxxxxxx>
>> Date:   Thu Oct 29 08:22:21 2015 +0000
>>
>>     btrfs: bugfix: handle FS_IOC32_{GETFLAGS,SETFLAGS,GETVERSION} in
>> btrfs_ioctl
>>
>>
>> However, this commit introduced a new regression.  With this commit
>> applied, "btrfs fi show" no longer works and the btrfs snapshot
>> functionality breaks.
> 
> A plain 32bit kernel with 32bit userspace works fine. The bug seems to
> be on a 64bit kernel with 32bit userspace and the CONFIG_COMPAT compiled
> in. Strace does not show anything special:
> 
> stat64("subv1", 0xffc64fcc)             = -1 ENOENT (No such file or directory)
> statfs64(".", 84, {f_type=BTRFS_SUPER_MAGIC, f_bsize=4096, f_blocks=5110784, f_bfree=4076143, f_bavail=4010951, f_files=0, f_ffree=0, f_fsid={val=[
> 4260464218, 2297804334]}, f_namelen=255, f_frsize=4096, f_flags=ST_VALID|ST_RELATIME}) = 0
> stat64(".", {st_mode=S_IFDIR|0700, st_size=228, ...}) = 0
> stat64(".", {st_mode=S_IFDIR|0700, st_size=228, ...}) = 0
> open(".", O_RDONLY|O_NONBLOCK|O_LARGEFILE|O_DIRECTORY|O_CLOEXEC) = 3
> fstat64(3, {st_mode=S_IFDIR|0700, st_size=228, ...}) = 0
> fstat64(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 1), ...}) = 0
> write(1, "Create subvolume './subv1'\n", 27) = 27
> ioctl(3, BTRFS_IOC_SUBVOL_CREATE, {fd=0, name="subv1"}) = -1 ENOTTY (Inappropriate ioctl for device)
> 
> The value of BTRFS_IOC_SUBVOL_CREATE is same on 32bit and 64bit kernels.
> As it returns ENOTTY, the value is not recognized. A candidate function
> is btrfs_compat_ioctl that checks for just the IOC32 numbers and returns
> -ENOIOCTLCMD otherwise.
> 
> The callchain in fs/compat_ioctl.c:ioctl cheks for the specific callback
> first, if it retunrs -ENOIOCTLCMD then goes to the normal ioctl
> callback, so there's always a point we reach the handler of
> BTRFS_IOC_SUBVOL_CREATE. So I don't see how it could happen.

I got a minute to look at this today.  It doesn't fallback to the normal
ioctl handler.  Since we have ->unlocked_ioctl defined, it checks the
hard-coded compat tables and then fails with -ENOTTY.

Fortunately, the fix is simple.  The default stanza of that switch
statement in btrfs_compat_ioctl needs to be removed.  Then it just
works.  I'll post a fix momentarily.

-Jeff

-- 
Jeff Mahoney
SUSE Labs

Attachment: signature.asc
Description: OpenPGP digital signature