Received-SPF: pass (google.com: domain of linux-ext4+bounces-1038-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Tue, 30 Jan 2024 23:58:22 -0500
From: "Theodore Ts'o" <tytso@mit.edu>
To: Dave Chinner <david@fromorbit.com>
Cc: syzbot <syzbot+cdee56dbcdf0096ef605@syzkaller.appspotmail.com>,
        adilger.kernel@dilger.ca, chandan.babu@oracle.com, jack@suse.com,
        linux-ext4@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-xfs@vger.kernel.org,
        syzkaller-bugs@googlegroups.com
Subject: Re: current->journal_info got nested! (was Re: [syzbot] [xfs?]
 [ext4?] general protection fault in jbd2__journal_start)
Message-ID: <20240131045822.GA2356784@mit.edu>
References: <000000000000e98460060fd59831@google.com>
 <000000000000d6e06d06102ae80b@google.com>
 <ZbmILkfdGks57J4a@dread.disaster.area>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ZbmILkfdGks57J4a@dread.disaster.area>

On Wed, Jan 31, 2024 at 10:37:18AM +1100, Dave Chinner wrote:
> It should be obvious what has happened now -
> current->journal_info is not null, so ext4 thinks it owns the
> structure attached there and panics when it finds that it isn't an
> ext4 journal handle being held there.
> 
> I don't think there are any clear rules as to how filesystems can
> and can't use current->journal_info. In general, a task can't jump
> from one filesystem to another inside a transaction context like
> this, so there's never been a serious concern about nested
> current->journal_info assignments like this in the past.
> 
> XFS is doing nothing wrong - we're allowed to define transaction
> contexts however we want and use current->journal_info in this way.
> However, we have to acknowledge that ext4 has also done nothing
> wrong by assuming current->journal_info should below to it if it is
> not null. Indeed, XFS does the same thing.

Nice analysis.  Fundamentally the current usage of
current->journal_info assumes that a process would only be calling
into one file system at a time.  But obviously that's not going to be
true in the case of one file system writing to memory which then
triggers a page fault.

As far as other potential avenues that could cause this kind of
nesting, the other one which comes to mind might be sendfile(2) --
although in general the reader side won't trigger a transaction since
the atime update tends to be done lazily.

> The question here is what to do about this? The obvious solution is
> to have save/restore semantics in the filesystem code that
> sets/clears current->journal_info, and then filesystems can also do
> the necessary "recursion into same filesystem" checks they need to
> ensure that they aren't nesting transactions in a way that can
> deadlock.
> 
> Maybe there are other options - should filesystems even be allowed to
> trigger page faults when they have set current->journal_info?

Hmm, could XFS pre-fault target memory buffer for the bulkstat output
before starting its transaction?  Alternatively, ext4 could do a save
of current->journal_info before starting to process the page fault,
and restore it when it is done.  Both of these seem a bit hacky, and
the question is indeed, are there other avenues that might cause the
transaction context nesting, such that a more general solution is
called for?

						- Ted