Re: [PATCH v2] mm/vmscan: fix infinite loop in drop_slab_node

[Matthew Wilcox <willy@infradead.org>]

On Wed, Sep 09, 2020 at 10:47:24PM +0100, Chris Down wrote:
> Vlastimil Babka writes:
> > - Exit also on other signals such as SIGABRT, SIGTERM? If I write to drop_caches
> > and think it's too long, I would prefer to kill it by ctrl-c and not just kill
> 
> Oh dear, fatal_signal_pending() doesn't consider cases with no more
> userspace instructions due to SIG_DFL on TERM/INT etc, that seems misleading
> :-( I had (naively) believed it internally checks the same set as
> TASK_KILLABLE.
> 
> Chuxin, Muchun, can you please make it work using TASK_KILLABLE in a similar
> way to how schedule_timeout_killable and friends do it instead, so that
> other signals will be caught?

You're mistaken.

        if (sig_fatal(p, sig) &&
            !(signal->flags & SIGNAL_GROUP_EXIT) &&
            !sigismember(&t->real_blocked, sig) &&
            (sig == SIGKILL || !p->ptrace)) {
...
                                sigaddset(&t->pending.signal, SIGKILL);

static inline int __fatal_signal_pending(struct task_struct *p)
{
        return unlikely(sigismember(&p->pending.signal, SIGKILL));
}