Re: [PATCHSET 0/2] PF_IO_WORKER signal tweaks

[Jens Axboe]

On 3/20/21 4:08 PM, Eric W. Biederman wrote:
> 
> Added criu because I just realized that io_uring (which can open files
> from an io worker thread) looks to require some special handling for
> stopping and freezing processes.  If not in the SIGSTOP case in the
> related cgroup freezer case.
> 
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:
> 
>> On Sat, Mar 20, 2021 at 10:51 AM Linus Torvalds
>> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>>>
>>> Alternatively, make it not use
>>> CLONE_SIGHAND|CLONE_THREAD at all, but that would make it
>>> unnecessarily allocate its own signal state, so that's "cleaner" but
>>> not great either.
>>
>> Thinking some more about that, it would be problematic for things like
>> the resource counters too. They'd be much better shared.
>>
>> Not adding it to the thread list etc might be clever, but feels a bit too scary.
>>
>> So on the whole I think Jens' minor patches to just not have IO helper
>> threads accept signals are probably the right thing to do.
> 
> The way I see it we have two options:
> 
> 1) Don't ask PF_IO_WORKERs to stop do_signal_stop and in
>    task_join_group_stop.
> 
>    The easiest comprehensive implementation looks like just
>    updating task_set_jobctl_pending to treat PF_IO_WORKER
>    as it treats PF_EXITING.
> 
> 2) Have the main loop of the kernel thread test for JOBCTL_STOP_PENDING
>    and call into do_signal_stop.
> 
> It is a wee bit trickier to modify the io_workers to stop, but it does
> not look prohibitively difficult.
> 
> All of the work performed by the io worker is work scheduled via
> io_uring by the process being stopped.
> 
> - Is the amount of work performed by the io worker thread sufficiently
>   negligible that we don't care?
> 
> - Or is the amount of work performed by the io worker so great that it
>   becomes a way for an errant process to escape SIGSTOP?
> 
> As the code is all intermingled with the cgroup_freezer.  I am also
> wondering creating checkpoints needs additional stopping guarantees.

The work done is the same a syscall, basically. So it could be long
running and essentially not doing anything (eg read from a socket, no
data is there), or it's pretty short lived (eg read from a file, just
waiting on DMA).

This is outside of my domain of expertise, which is exactly why I added
you and Linus to make some calls on what the best approach here would
be. My two patches obviously go route #1 in terms of STOP. And fwiw,
I tested this:

> To solve the issue that SIGSTOP is simply broken right now I am totally
> fine with something like:
> 
> diff --git a/kernel/signal.c b/kernel/signal.c
> index ba4d1ef39a9e..cb9acdfb32fa 100644
> --- a/kernel/signal.c
> +++ b/kernel/signal.c
> @@ -288,7 +288,8 @@ bool task_set_jobctl_pending(struct task_struct *task, unsigned long mask)
>  			JOBCTL_STOP_SIGMASK | JOBCTL_TRAPPING));
>  	BUG_ON((mask & JOBCTL_TRAPPING) && !(mask & JOBCTL_PENDING_MASK));
>  
> -	if (unlikely(fatal_signal_pending(task) || (task->flags & PF_EXITING)))
> +	if (unlikely(fatal_signal_pending(task) ||
> +		     (task->flags & (PF_EXITING | PF_IO_WORKER))))
>  		return false;
>  
>  	if (mask & JOBCTL_STOP_SIGMASK)

and can confirm it works fine for me with 2/2 reverted and this applied
instead.

> Which just keeps from creating unstoppable processes today.  I am just
> not convinced that is what we want as a long term solution.

How about we go with either my 2/2 or yours above to at least ensure we
don't leave workers looping as schedule() is a nop with sigpending? If
there's a longer timeline concern that "evading" SIGSTOP is a concern, I
have absolutely no qualms with making the IO threads participate. But
since it seems conceptually simple but with potentially lurking minor
issues, probably not the ideal approach for right now.

-- 
Jens Axboe