Re: [PATCH] io_uring: handle signals before letting io-worker exit

[Olivier Langlois <olivier@trillion01.com>]

On Thu, 2021-05-27 at 09:30 -0600, Jens Axboe wrote:
> > > 
> > Jens,
> > 
> > You are 100% correct. In fact, this is the same problem for ALL
> > currently existing and future io threads. Therefore, I start to
> > think
> > that the right place for the fix might be straight into
> > do_exit()...
> 
> That is what I was getting at. To avoid poluting do_exit() with it, I
> think it'd be best to add an io_thread_exit() that simply does:
> 
> void io_thread_exit(void)
> {
>         if (signal_pending(current)) {
>                 struct ksignal ksig;
>                 get_signal(&ksig);
>         }
>         do_exit(0);
> }
> 
> and convert the do_exit() calls in io_uring/io-wq to io_thread_exit()
> instead.
> 
IMHO, that would be an acceptable compromise because it does fix my
problem. However, I am of the opinion that it wouldn't be poluting
do_exit() and would in fact be the right place to do it considering
that create_io_thread() is in kernel and theoritically, anyone can call
it to create an io_thread and would be susceptible to get bitten by the
exact same problem and would have to come up with a similar solution if
it is not addressed directly by the kernel.

Also, since I have submitted the patch, I have made the following
realization:

I got bitten by the problem because of a race condition between the io-
mgr thread and its io-wrks threads for processing their pending SIGKILL
and the proposed patch does correct my problem.

The issue would have most likely been buried by 5.13 io-mgr removal...

BUT, even the proposed patch isn't 100% perfect. AFAIK, it is still
possible, but very unlikely, to get a signal between calling
signal_pending() and do_exit().

It might be possible to implement the solution and be 100% correct all
the time by doing it inside do_exit()... I am currently eyeing
exit_signals() as a potential good site for the patch...