Re: WARNING: lock held when returning to user space in fuse_lock_inode

[Dmitry Vyukov]

On Tue, Jul 17, 2018 at 1:14 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
> On Thu, Jul 12, 2018 at 5:49 PM, syzbot
> <syzbot+3f7b29af1baa9d0a55be@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
>> Hello,
>>
>> syzbot found the following crash on:
>>
>> HEAD commit:    c25c74b7476e Merge tag 'trace-v4.18-rc3-2' of git://git.ke..
>> git tree:       upstream
>> console output: https://syzkaller.appspot.com/x/log.txt?x=177bcec2400000
>> kernel config:  https://syzkaller.appspot.com/x/.config?x=25856fac4e580aa7
>> dashboard link: https://syzkaller.appspot.com/bug?extid=3f7b29af1baa9d0a55be
>> compiler:       gcc (GCC) 8.0.1 20180413 (experimental)
>> syzkaller repro:https://syzkaller.appspot.com/x/repro.syz?x=13aa7678400000
>> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=17492678400000
>>
>> IMPORTANT: if you fix the bug, please add the following tag to the commit:
>> Reported-by: syzbot+3f7b29af1baa9d0a55be@xxxxxxxxxxxxxxxxxxxxxxxxx
>>
>> random: sshd: uninitialized urandom read (32 bytes read)
>> random: sshd: uninitialized urandom read (32 bytes read)
>> random: sshd: uninitialized urandom read (32 bytes read)
>>
>> ================================================
>> WARNING: lock held when returning to user space!
>> 4.18.0-rc4+ #143 Not tainted
>> ------------------------------------------------
>> syz-executor012/4539 is leaving the kernel with locks still held!
>> 1 lock held by syz-executor012/4539:
>>  #0: (____ptrval____) (&fi->mutex){+.+.}, at: fuse_lock_inode+0xaf/0xe0
>> fs/fuse/inode.c:363
>
> False positive.
>
> fi->mutex is definitely not held by the acquiring task when returning
> to userspace.  Maybe syzkaller is confused by the fact that there are
> several interdependent tasks involved with fuse:  the one calling into
> fuse by doing something (looking up ./file0/file0) and the one that
> reads the fuse device (returning with the LOOKUP request for "file0").
> The second one will return with that lock held, but it's not the one
> that acquired it, so there's no bug at all here.

Hi Miklos,

syzkaller is unrelated here. That's what kernel self-detects and
prints. So either way there is something to fix in kernel here: either
fuse or lockdep.

+Alistair did some analysis offline, hope you don't mind if I repost
your description:
===
Just from reading the code, I think I can see how this happens. Fuse
is wrapping its inode mutex with a check for "parallel_dirops", which
is set up in process_init_reply(). The FUSE_PARALLEL_DIROPS appears to
always be set, in fuse_send_init(), but its initial state is to be
disabled. So if the mutex gets taken, and it'll never be unlocked if
the initial command is flushed by fuse_readdir()'s use of
fuse_lock_inode().
===