Re: kernel BUG at fs/inode.c:LINE!

From: Dmitry Vyukov <dvyukov@google.com>
To: Al Viro <viro@zeniv.linux.org.uk>
Cc: syzbot <syzbot+5399ed0832693e29f392@syzkaller.appspotmail.com>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>,
	syzkaller-bugs <syzkaller-bugs@googlegroups.com>
Subject: Re: kernel BUG at fs/inode.c:LINE!
Date: Mon, 17 Dec 2018 11:08:31 +0100	[thread overview]
Message-ID: <CACT4Y+Zeyfd693tMa4eTqSWYCBK+55_Lt3tzk2MckFa2KoFGuw@mail.gmail.com> (raw)
In-Reply-To: <20181217072144.GQ2217@ZenIV.linux.org.uk>

On Mon, Dec 17, 2018 at 8:21 AM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Sun, Dec 16, 2018 at 10:11:04PM -0800, syzbot wrote:
> > Hello,
> >
> > syzbot found the following crash on:
> >
> > HEAD commit:    d14b746c6c1c Add linux-next specific files for 20181214
> > git tree:       linux-next
> > console output: https://syzkaller.appspot.com/x/log.txt?x=13706347400000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=1da6d2d18f803140
> > dashboard link: https://syzkaller.appspot.com/bug?extid=5399ed0832693e29f392
> > compiler:       gcc (GCC) 8.0.1 20180413 (experimental)
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=101032b3400000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=16534063400000
> >
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+5399ed0832693e29f392@syzkaller.appspotmail.com
> >
> >  slab_pre_alloc_hook mm/slab.h:423 [inline]
> >  slab_alloc mm/slab.c:3365 [inline]
> >  kmem_cache_alloc+0x2c4/0x730 mm/slab.c:3539
> >  __d_alloc+0xc8/0xb90 fs/dcache.c:1599
> > ------------[ cut here ]------------
> > kernel BUG at fs/inode.c:1566!
> >  d_alloc_anon fs/dcache.c:1698 [inline]
> >  d_make_root+0x43/0xc0 fs/dcache.c:1885
> >  autofs_fill_super+0x6f1/0x1c30 fs/autofs/inode.c:273
>
> Huh?  BUG is in iput(), AFAICS, so the stack trace is rather misreported.

Yes, it's a known problem that kernel is generally incapable of
producing parsable crash reports. I think Tetsuo is working on a
solution, but it takes very large amount of discussions and months of
time.

> iput() can be called by d_make_root(), provided that dentry allocation
> fails.  So the most straightforward interpretation would be that we
> had an allocation failure (injected?), followed by iput() of the inode
> passed to d_make_root().  Which happened to find I_CLEAR in ->i_state
> of that inode somehow, which should be impossible short of seriously
> buggered inode refcounting somewhere - the inode has just been returned
> by new_inode(), which clears i_state, and it would have to have passed
> clear_inode() (i.e. has been through inode eviction) since then...

Yes, there are injected failures in the log.
And the repro says:

#{..."fault":true,"fault_call":1,"fault_nth":36 ...}
mkdir(&(0x7f0000000140)='./file0\x00', 0x0)
mount(0x0, &(0x7f0000000100)='./file0\x00',
&(0x7f0000000180)='autofs\x00', 0x0, 0x0)

Since the repro is single-threaded, I would expect that the bug is
directly on the error handling path of the failure allocation.