lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date:   Mon, 5 Dec 2022 12:48:50 -0500
From:   "Theodore Ts'o" <>
To:     zhanchengbin <>
Cc:     Ye Bin <>,,,,, Ye Bin <>
Subject: Re: [PATCH RFC] ext4:record error information when insert extent
 failed in 'ext4_split_extent_at'

On Fri, Oct 28, 2022 at 09:41:55AM +0800, zhanchengbin wrote:
> There have been a lot of problems here before, but the problem has not been
> fundamentally solved.

Indeed, this only papers only the problem.  The commit description is
pretty good, in that it describes the the root cause of the problem:

> > If 'ext4_ext_insert_extent' return '-ENOMEM' which will not fix 'ex->ee_len' by
> > old length. 'ext4_ext_insert_extent' will trigger extent tree merge, fix like
> > 'ex->ee_len = orig_ex.ee_len' may lead to new issues.
> > To solve above issue, record error messages when 'ext4_ext_insert_extent' return
> > 'err' not equal '(-ENOSPC && -EDQUOT)'. If filesysten is mounted with 'errors=continue'
> > as filesystem is not clean 'fsck' will repair issue. If filesystem is mounted with
> > 'errors=remount-ro' filesystem will be remounted by read-only.

What should actually happen here is that we undo the change in orig_ex
if ext4_ext_insert_extent fails.  As you point out, in an earlier part
of the code path, this gets handled by a "goto fix_extent_len".    

The problem is that it's possible that the shape of the extent tree
may have changed before ext4_insert_extent() fails with the ENOMEM.
So the simplistic fix of just jumping to fix_extent_len isn't going to
work.  But fixing it by much marking the file system is corrupt is a
bit of a cop-out.  Consider that if you were running ext4 in a Huawei
Cloud, and you run into the memory allocation failure, what would you
prefer?  For the individual system call to fail, and propagating the
failure to userspace?  Or to leave the file system corrupted?  (And
then either forcing a reboot so the file system to be fixed, or allow
the system to stumble along, with further unknown unexpected behaviour
from userspace?)

The real correct fix is that ext4_ext_insert_extent() needs to
rollback any changes to the extent tree that was made, *or* it needs
to make sure that the operation will succeed before starting to make
any changes, *or* we need to look up the orig_extent via
orig_ex->ee_block, and then undo the change.

It might be that we can't always reliably rollback the change, or we
might think that the operation will succeed, but then it fail due to
an I/O error.  If it's due to an I/O error, then it's fine to bail and
mark the file system as corrupted.  But if the failure is caused by an
ENOMEM, we should be able to handle this case more gracefully.

Can you look into a better fix?


						- Ted

Powered by blists - more mailing lists