Re: update-binfmts breaking suspend

[Matthew Wilcox]

Ping?

On Fri, Apr 06, 2018 at 07:43:55AM -0700, Matthew Wilcox wrote:
> Pavel Machek <pavel@xxxxxx> wrote:
> > > Failure is not a hang, as they expect, but... machine locks up, but
> > > does not suspend, and then continues running after a delay..
> > > 
> > > [   35.038766] PM: Syncing filesystems ... done.
> > > [   35.051246] Freezing user space processes ...
> > > [   55.060528] Freezing of tasks failed after 20.009 seconds (1 tasks
> > > refusing to freeze, wq_busy
> > > =0):
> > > [   55.060552] update-binfmts  D    0  2727      1 0x80000004
> > > [   55.060576] Call Trace:
> > > [   55.060600]  __schedule+0x37a/0x7e0
> > > [   55.060618]  schedule+0x29/0x70
> > > [   55.060635]  autofs4_wait+0x359/0x7a0
> > > [   55.060653]  ? wait_woken+0x70/0x70
> > > [   55.060668]  autofs4_mount_wait+0x4a/0xe0
> > > [   55.060684]  ? autofs4_mount_wait+0x4a/0xe0
> > > [   55.060699]  autofs4_d_automount+0xe0/0x200
> > > [   55.060715]  ? autofs4_d_automount+0xe0/0x200
> > > 
> > > Did the rework of freezing start already in -next?
> > 
> > Hmm, so I did git bisect, and it pointed to:
> > 
> > commit 7cb03edf112fea6ead2fcd3c5fd639756d6d114b
> > Author: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
> > Date:   Thu Mar 29 10:15:17 2018 +1100
> > 
> >     autofs4: use wait_event_killable
> > 
> >     This playing with signals to allow only fatal signals appears to
> >     predate
> >         the introduction of wait_event_killable(), and I'm fairly sure
> >     that
> >         wait_event_killable is what was meant to happen here.
> 
> Umm.  I'm not familiar with the freezer.  Help me out here ...
> 
> I see the message coming from here:
> 
>                 pr_err("Freezing of tasks %s after %d.%03d seconds "
>                        "(%d tasks refusing to freeze, wq_busy=%d):\n",
>                        wakeup ? "aborted" : "failed",
>                        elapsed_msecs / 1000, elapsed_msecs % 1000,
>                        todo - wq_busy, wq_busy);
> 
> and then backtracking in that function, I see this:
> 
>                 for_each_process_thread(g, p) {
>                         if (p == current || !freeze_task(p))
>                                 continue;
> 
> in freeze_task(), I see this:
> 
>         if (!(p->flags & PF_KTHREAD))
>                 fake_signal_wake_up(p);
>         else
>                 wake_up_state(p, TASK_INTERRUPTIBLE);
> 
> which does this:
> 
>         if (lock_task_sighand(p, &flags)) {
>                 signal_wake_up(p, 0);
>                 unlock_task_sighand(p, &flags);
>         }
> 
> which does this:
> 
> static inline void signal_wake_up(struct task_struct *t, bool resume)
> {
>         signal_wake_up_state(t, resume ? TASK_WAKEKILL : 0);
> }
> 
> which does this:
> 
> void signal_wake_up_state(struct task_struct *t, unsigned int state)
> {
>         set_tsk_thread_flag(t, TIF_SIGPENDING);
>         /*
>          * TASK_WAKEKILL also means wake it up in the stopped/traced/killable
>          * case. We don't check t->state here because there is a race with it
>          * executing another processor and just now entering stopped state.
>          * By using wake_up_state, we ensure the process will wake up and
>          * handle its death signal.
>          */
>         if (!wake_up_state(t, state | TASK_INTERRUPTIBLE))
>                 kick_process(t);
> }
> 
> Now I don't know why we only wake interruptible tasks here and not killable
> tasks.  I've trawled git history all the way back to 2.6.12-rc2, and the
> reasoning behind signal_wake_up() (as it originally was) is lost to pre-git
> history.
> 
> So ... why do we only wake interruptible tasks on suspend?  Why not wake
> uninterruptible tasks too?
> 
>         if (lock_task_sighand(p, &flags)) {
> -               signal_wake_up(p, 0);
> +               signal_wake_up_state(p, TASK_WAKEKILL);
>                 unlock_task_sighand(p, &flags);
>         }
> 
> or why do we consider tasks waiting uninterruptibly to block freezing?
> Is it because they're (probably) waiting for I/O and we want the I/O
> to complete?