Re: [GIT PULL] signal/exit/ptrace changes for v5.17

["Eric W. Biederman" <ebiederm@xmission.com>]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> On Mon, Jan 17, 2022 at 5:32 PM Eric W. Biederman <ebiederm@xmission.com> wrote:
>>
>> I would like to have a version of pipe_write that sleeps in
>> TASK_KILLABLE.
>
> That would actually be horrible for another reason - now it would
> count towards the load average. That's another difference between
> interruptible waits and non-interruptible ones.
>
> Admittedly it's an entirely arbitrary one, but it's part of the whole
> semantic difference between TASK_INTERRUPTIBLE and
> TASK_UNINTERRUPTIBLE.
>
> You can play with TASK_NOLOAD of course, so it's something that can be
> worked around, but it gets a bit ugly.

Yes.  I don't want to make a change that changes the load average.

>>  I want the I/O wake-ups and I want the SIGKILL wake ups
>> but I don't want any other wake-ups.  Unfortunately the I/O wake-ups in
>> the pipe code are sent with wake_up_interruptible.  So a task sleeping
>> in TASK_KILLABLE won't get them.
>
> Yeah. The code *could* use the non-interruptible 'wake_up()', and
> everything should work - because waking things up too much doesn't
> change semantics, it's just a slight pessimization. Plus the whole
> "nested waitqueues" isn't actually any remotely normal case, so it
> doesn't really matter for performance either.
>
> But I really think it's wrong.
>
> You're trying to work around a problem the wrong way around. If a task
> is dead, and is dumping core, then signals just shouldn't matter in
> the first place, and thus the whole "TASK_INTERRUPTIBLE vs
> TASK_UNINTERRUPTIBLE" really shouldn't be an issue. The fact that it
> is an issue means there's something wrong in signaling, not in the
> pipe code.
>
> So I really think that's where the fix should be - on the signal delivery side.

The actual signaling is shutdown, (except for the special case of
SIGKILL being able to terminate the coredump).  It is io_uring and
anything else that is not a signal that causes signal_pending() to
return true.

I have not found any solution I am happy with yet, I am just
brainstorming.

Part of the problem is that I really don't want to perform process
shutdown and remove evidence of why the process crashed.  So maybe
shutting down io_uring is fine in that case but I don't like that either.

The more I look at all of the interesting corner cases the more I wonder
if the solution isn't to have the coredump code fork a kernel-only
userspace process (like the io_uring threads are kernel-only userspace
threads).  That would at least allow kernel functionality to work like
normal and greatly reduce the chance of weird feature interactions.

Hmm. A special kernel-only thread might even be enough as io_uring would
not be directing task_work at it.

You have been me some good information and I think I just need to sleep
on this problem a bit more to come up with a non-hacky solution.

Eric