Re: [PATCH] signal: Allow RT tasks to cache one sigqueue struct

[Oleg Nesterov <oleg@redhat.com>]

On 03/04, Thomas Gleixner wrote:
>
> On Wed, Mar 03 2021 at 16:37, Oleg Nesterov wrote:
> > On 03/03, Sebastian Andrzej Siewior wrote:
> >>
> >> +static struct sigqueue *sigqueue_from_cache(struct task_struct *t)
> >> +{
> >> +	struct sigqueue *q = t->sigqueue_cache;
> >> +
> >> +	if (q && cmpxchg(&t->sigqueue_cache, q, NULL) == q)
> >> +		return q;
> >> +	return NULL;
> >> +}
> >> +
> >> +static bool sigqueue_add_cache(struct task_struct *t, struct sigqueue *q)
> >> +{
> >> +	if (!t->sigqueue_cache && cmpxchg(&t->sigqueue_cache, NULL, q) == NULL)
> >> +		return true;
> >> +	return false;
> >> +}
> >
> > Do we really need cmpxchg? It seems they are always called with
> > spinlock held.
>
> With which spinlock held?
>
> __send_signal()         <- sighand::siglock held
>   __sigqueue_alloc()
>
> alloc_posix_timer()
>   sigqueue_alloc()      <- No lock held
>     __sigqueue_alloc()

In the last case "fromslab" is true, sigqueue_from_cache() won't be called.

> and on the free side we have a bunch of callers which do not hold
> sighand::siglock either.

Where?

Oleg.