Re: [PATCH] tty/vt: Fix possible deadlock in input_inject_event

[pengyu]

在 2025/9/29 12:54, Dmitry Torokhov 写道:
> Hi,
>
> On Sun, Sep 28, 2025 at 09:08:19PM +0800, pengyu wrote:
> > syzkaller testing revealed a potential deadlock involving keyboard
> > handling:
> >
> > CPU0                       CPU1                      CPU2
> > ----                       ----                      ----
> > read_lock(tasklist_lock);  evdev_write
> >                            input_inject_event     write_lock(tasklist_lock);
> >                           lock(&dev->event_lock);
> >                          read_lock(tasklist_lock);
> > <Interrupt>
> > kbd_bh() / kd_sound_helper()
> > input_inject_event
> > lock(&dev->event_lock); // Deadlock risk
> >
> > The deadlock occurs because:
> > 1. Both kbd_bh and kd_sound_helper run in interrupt context
> > 2. tasklist_lock is interrupt-unsafe
> > 3. When evdev_write holds both dev->event_lock and tasklist_lock,
> >     interrupt context attempts to acquire dev->event_lock create deadlock
> >     risks
> >
> > Convert both kbd_bh and kd_sound_helper to use workqueues. This moves
> > input_inject_event execution to process context, where it's safe to
> > acquire locks that may be held by code using interrupt-unsafe locks.
>
> So if we ignore the input code and instead look at the send_sigio()
> (which input core ends up calling) and do_wait() we see that
> send_sigio() disables interrupts and takes the owner's spinlock
> before taking the tasklist_lock, while do_wait() takes the tasklist_lock
> first, without disabling interrupts. This is root of the issue as far as
> I can tell and no amount of changes to the keyboard handler (which is
> just happens to be in the middle) will not solve for all potential cases
> and code paths.
>
> I believe either do_exit() or send_sigio() have to be changed to fix
> this properly.
>
> Thanks.

Hi,

I noticed that besides do_wait, there are many places in the kernel 
where read_lock(tasklist_lock) is used without disabling interrupts. 
Addressing this solely through tasklist_lock may not fully resolve the 
issue.

This involves tasklist_lock, evdev_write, and various input device 
drivers. The only approach I can think of is to move functions like 
input_[inject]_event in the input drivers out of the interrupt context. 
This could affect many code paths, so I plan to start by modifying the 
keyboard code first.

--
Thanks,
Yu Peng