public inbox for [email protected]
 help / color / mirror / Atom feed
From: [email protected] (Eric W. Biederman)
To: Stefan Metzmacher <[email protected]>
Cc: Jens Axboe <[email protected]>,
	[email protected], [email protected],
	[email protected]
Subject: Re: [PATCH 09/18] io-wq: fork worker threads from original task
Date: Fri, 05 Mar 2021 13:16:50 -0600	[thread overview]
Message-ID: <[email protected]> (raw)
In-Reply-To: <[email protected]> (Stefan Metzmacher's message of "Thu, 4 Mar 2021 17:13:56 +0100")

Stefan Metzmacher <[email protected]> writes:

> Am 04.03.21 um 14:19 schrieb Stefan Metzmacher:
>> Hi Jens,
>> 
>>>> Can you please explain why CLONE_SIGHAND is used here?
>>>
>>> We can't have CLONE_THREAD without CLONE_SIGHAND... The io-wq workers
>>> don't really care about signals, we don't use them internally.
>> 
>> I'm 100% sure, but I heard rumors that in some situations signals get
>> randomly delivered to any thread of a userspace process.
>
> Ok, from task_struct:
>
>         /* Signal handlers: */
>         struct signal_struct            *signal;
>         struct sighand_struct __rcu             *sighand;
>         sigset_t                        blocked;
>         sigset_t                        real_blocked;
>         /* Restored if set_restore_sigmask() was used: */
>         sigset_t                        saved_sigmask;
>         struct sigpending               pending;
>
> The signal handlers are shared, but 'blocked' is per thread/task.

Doing something so that wants_signal won't try and route
a signal to a PF_IO_WORKER seems sensible.

Either blocking the signal or modifying wants_signal.

>> My fear was that the related logic may select a kernel thread if they
>> share the same signal handlers.
>
> I found the related logic in the interaction between
> complete_signal() and wants_signal().
>
> static inline bool wants_signal(int sig, struct task_struct *p)
> {
>         if (sigismember(&p->blocked, sig))
>                 return false;
>
> ...
>
> Would it make sense to set up task->blocked to block all signals?
>
> Something like this:
>
> --- a/fs/io-wq.c
> +++ b/fs/io-wq.c
> @@ -611,15 +611,15 @@ pid_t io_wq_fork_thread(int (*fn)(void *), void *arg)
>  {
>         unsigned long flags = CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|
>                                 CLONE_IO|SIGCHLD;
> -       struct kernel_clone_args args = {
> -               .flags          = ((lower_32_bits(flags) | CLONE_VM |
> -                                   CLONE_UNTRACED) & ~CSIGNAL),
> -               .exit_signal    = (lower_32_bits(flags) & CSIGNAL),
> -               .stack          = (unsigned long)fn,
> -               .stack_size     = (unsigned long)arg,
> -       };
> +       sigset_t mask, oldmask;
> +       pid_t pid;
>
> -       return kernel_clone(&args);
> +       sigfillset(&mask);
> +       sigprocmask(SIG_BLOCK, &mask, &oldmask);
> +       pid = kernel_thread(fn, arg, flags);
> +       sigprocmask(SIG_SETMASK, &oldmask, NULL);
> +
> +       return ret;
>  }
>
> I think using kernel_thread() would be a good simplification anyway.

I have a memory of kernel_thread having a built in assumption that it is
being called from a kthreadd, but I am not seeing it now so that would
be a nice simplification if we can do that.

> sig_task_ignored() has some PF_IO_WORKER logic.
>
> Or is there any PF_IO_WORKER related logic that prevents
> an io_wq thread to be excluded in complete_signal().
>
> Or PF_IO_WORKER would teach kernel_clone to ignore CLONE_SIGHAND
> and create a fresh handler and alter the copy_signal() and copy_sighand()
> checks...

I believe it is desirable for SIGKILL to the process to kill all of it's
PF_IO_WORKERS as well.

All that wants_signal allows/prevents is a wake up to request the task
to call get_signal.  No matter what complete_signal suggests any thread
can still dequeue the signal and process it.

It probably makes sense to block everything except SIGKILL (and
SIGSTOP?) in task_thread so that wants_signal doesn't fail to wake up an
ordinary thread that could handle the signal when the signal arrives.

Eric


  parent reply	other threads:[~2021-03-05 19:17 UTC|newest]

Thread overview: 49+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-02-19 17:09 [PATCHSET RFC 0/18] Remove kthread usage from io_uring Jens Axboe
2021-02-19 17:09 ` [PATCH 01/18] io_uring: remove the need for relying on an io-wq fallback worker Jens Axboe
2021-02-19 20:25   ` Eric W. Biederman
2021-02-19 20:37     ` Jens Axboe
2021-02-22 13:46   ` Pavel Begunkov
2021-02-19 17:09 ` [PATCH 02/18] io-wq: don't create any IO workers upfront Jens Axboe
2021-02-19 17:09 ` [PATCH 03/18] io_uring: disable io-wq attaching Jens Axboe
2021-02-19 17:09 ` [PATCH 04/18] io-wq: get rid of wq->use_refs Jens Axboe
2021-02-19 17:09 ` [PATCH 05/18] io_uring: tie async worker side to the task context Jens Axboe
2021-02-20  8:11   ` Hao Xu
2021-02-20 14:38     ` Jens Axboe
2021-02-21  9:16       ` Hao Xu
2021-02-19 17:09 ` [PATCH 06/18] io-wq: don't pass 'wqe' needlessly around Jens Axboe
2021-02-19 17:09 ` [PATCH 07/18] arch: setup PF_IO_WORKER threads like PF_KTHREAD Jens Axboe
2021-02-19 22:21   ` Eric W. Biederman
2021-02-19 23:26     ` Jens Axboe
2021-02-19 17:10 ` [PATCH 08/18] kernel: treat PF_IO_WORKER like PF_KTHREAD for ptrace/signals Jens Axboe
2021-02-19 17:10 ` [PATCH 09/18] io-wq: fork worker threads from original task Jens Axboe
2021-03-04 12:23   ` Stefan Metzmacher
2021-03-04 13:05     ` Jens Axboe
2021-03-04 13:19       ` Stefan Metzmacher
2021-03-04 16:13         ` Stefan Metzmacher
2021-03-04 16:42           ` Jens Axboe
2021-03-04 17:09             ` Stefan Metzmacher
2021-03-04 17:32               ` Jens Axboe
2021-03-04 18:19                 ` Jens Axboe
2021-03-04 18:56                   ` Linus Torvalds
2021-03-04 19:19                     ` Jens Axboe
2021-03-04 19:46                       ` Linus Torvalds
2021-03-04 19:54                         ` Jens Axboe
2021-03-04 20:00                           ` Jens Axboe
2021-03-04 20:23                             ` Jens Axboe
2021-03-04 20:50                           ` Linus Torvalds
2021-03-04 20:54                             ` Jens Axboe
2021-03-05 19:16           ` Eric W. Biederman [this message]
2021-03-05 19:00       ` Eric W. Biederman
2021-02-19 17:10 ` [PATCH 10/18] io-wq: worker idling always returns false Jens Axboe
2021-02-19 17:10 ` [PATCH 11/18] io_uring: remove any grabbing of context Jens Axboe
2021-02-19 17:10 ` [PATCH 12/18] io_uring: remove io_identity Jens Axboe
2021-02-19 17:10 ` [PATCH 13/18] io-wq: only remove worker from free_list, if it was there Jens Axboe
2021-02-19 17:10 ` [PATCH 14/18] io-wq: make io_wq_fork_thread() available to other users Jens Axboe
2021-02-19 17:10 ` [PATCH 15/18] io_uring: move SQPOLL thread io-wq forked worker Jens Axboe
2021-02-19 17:10 ` [PATCH 16/18] Revert "proc: don't allow async path resolution of /proc/thread-self components" Jens Axboe
2021-02-19 17:10 ` [PATCH 17/18] Revert "proc: don't allow async path resolution of /proc/self components" Jens Axboe
2021-02-19 17:10 ` [PATCH 18/18] net: remove cmsg restriction from io_uring based send/recvmsg calls Jens Axboe
2021-02-19 23:44 ` [PATCHSET RFC 0/18] Remove kthread usage from io_uring Stefan Metzmacher
2021-02-19 23:51   ` Jens Axboe
2021-02-21  5:04 ` Linus Torvalds
2021-02-21 21:22   ` Jens Axboe

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox