Re: [PATCH v2 1/2] fcntl: fix potential deadlocks for &fown_struct.lock

[Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com>]

On 7/7/21 2:05 pm, Greg KH wrote:
> On Wed, Jul 07, 2021 at 10:35:47AM +0800, Desmond Cheong Zhi Xi wrote:
>> Syzbot reports a potential deadlock in do_fcntl:
>>
>> ========================================================
>> WARNING: possible irq lock inversion dependency detected
>> 5.12.0-syzkaller #0 Not tainted
>> --------------------------------------------------------
>> syz-executor132/8391 just changed the state of lock:
>> ffff888015967bf8 (&f->f_owner.lock){.+..}-{2:2}, at: f_getown_ex fs/fcntl.c:211 [inline]
>> ffff888015967bf8 (&f->f_owner.lock){.+..}-{2:2}, at: do_fcntl+0x8b4/0x1200 fs/fcntl.c:395
>> but this lock was taken by another, HARDIRQ-safe lock in the past:
>>   (&dev->event_lock){-...}-{2:2}
>>
>> and interrupts could create inverse lock ordering between them.
>>
>> other info that might help us debug this:
>> Chain exists of:
>>    &dev->event_lock --> &new->fa_lock --> &f->f_owner.lock
>>
>>   Possible interrupt unsafe locking scenario:
>>
>>         CPU0                    CPU1
>>         ----                    ----
>>    lock(&f->f_owner.lock);
>>                                 local_irq_disable();
>>                                 lock(&dev->event_lock);
>>                                 lock(&new->fa_lock);
>>    <Interrupt>
>>      lock(&dev->event_lock);
>>
>>   *** DEADLOCK ***
>>
>> This happens because there is a lock hierarchy of
>> &dev->event_lock --> &new->fa_lock --> &f->f_owner.lock
>> from the following call chain:
>>
>>    input_inject_event():
>>      spin_lock_irqsave(&dev->event_lock,...);
>>      input_handle_event():
>>        input_pass_values():
>>          input_to_handler():
>>            evdev_events():
>>              evdev_pass_values():
>>                spin_lock(&client->buffer_lock);
>>                __pass_event():
>>                  kill_fasync():
>>                    kill_fasync_rcu():
>>                      read_lock(&fa->fa_lock);
>>                      send_sigio():
>>                        read_lock_irqsave(&fown->lock,...);
>>
>> However, since &dev->event_lock is HARDIRQ-safe, interrupts have to be
>> disabled while grabbing &f->f_owner.lock, otherwise we invert the lock
>> hierarchy.
>>
>> Hence, we replace calls to read_lock/read_unlock on &f->f_owner.lock,
>> with read_lock_irq/read_unlock_irq.
>>
>> Here read_lock_irq/read_unlock_irq should be safe to use because the
>> functions f_getown_ex and f_getowner_uids are only called from
>> do_fcntl, and f_getown is only called from do_fnctl and
>> sock_ioctl. do_fnctl itself is only called from syscalls.
>>
>> For sock_ioctl, the chain is
>>    compat_sock_ioctl():
>>      compat_sock_ioctl_trans():
>>        sock_ioctl()
>>
>> And interrupts are not disabled on either path. We assert this
>> assumption with WARN_ON_ONCE(irqs_disabled()). This check is also
>> inserted into another use of write_lock_irq in f_modown.
>>
>> Reported-and-tested-by: syzbot+e6d5398a02c516ce5e70@syzkaller.appspotmail.com
>> Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
>> ---
>>   fs/fcntl.c | 17 +++++++++++------
>>   1 file changed, 11 insertions(+), 6 deletions(-)
>>
>> diff --git a/fs/fcntl.c b/fs/fcntl.c
>> index dfc72f15be7f..262235e02c4b 100644
>> --- a/fs/fcntl.c
>> +++ b/fs/fcntl.c
>> @@ -88,6 +88,7 @@ static int setfl(int fd, struct file * filp, unsigned long arg)
>>   static void f_modown(struct file *filp, struct pid *pid, enum pid_type type,
>>                        int force)
>>   {
>> +	WARN_ON_ONCE(irqs_disabled());
> 
> If this triggers, you just rebooted the box :(
> 
> Please never do this, either properly handle the problem and return an
> error, or do not check for this.  It is not any type of "fix" at all,
> and at most, a debugging aid while you work on the root problem.
> 
> thanks,
> 
> greg k-h
> 

Hi Greg,

Thanks for the feedback. My bad, I was under the impression that 
WARN_ON_ONCE could be used to document assumptions for other developers, 
but I'll stick to using it for debugging in the future.

I think then in this case it would be best to keep the reasoning for why 
the *_irq() locks are safe to use in the commit message. I'll update the 
patch accordingly.

Best wishes,
Desmond