Re: [PATCH v4 02/27] ntsync: Introduce NTSYNC_IOC_WAIT_ALL.

[Elizabeth Figura]

On Wednesday, 17 April 2024 06:37:03 CDT Peter Zijlstra wrote:
> On Mon, Apr 15, 2024 at 08:08:12PM -0500, Elizabeth Figura wrote:
> > -
> >
> > if (atomic_read(&sem->all_hint) > 0) {
> > -
> > spin_lock(&dev->wait_all_lock);
> > -
> > spin_lock_nest_lock(&sem->lock, &dev->wait_all_lock);
> >
> > -
> > prev_count = sem->u.sem.count;
> > -
> > ret = post_sem_state(sem, args);
> > -
> > if (!ret) {
> > -
> > 	try_wake_all_obj(dev, sem);
> > -
> > 	try_wake_any_sem(sem);
> > -
> > }
> >
> > -
> > spin_unlock(&sem->lock);
> > -
> > spin_unlock(&dev->wait_all_lock);
> > -
> >
> > } else {
> > -
> > spin_lock(&sem->lock);
> > -
> > -
> > prev_count = sem->u.sem.count;
> > -
> > ret = post_sem_state(sem, args);
> > -
> > if (!ret)
> > -
> > 	try_wake_any_sem(sem);
> > -
> > -
> > spin_unlock(&sem->lock);
> > -
> >
> > }
> >
> > if (!ret && put_user(prev_count, user_args))
> > ret = -EFAULT;
>
> vs.
> > - /* queue ourselves */
> > -
> > - spin_lock(&dev->wait_all_lock);
> > -
> > - for (i = 0; i < args.count; i++) {
> > -
> > struct ntsync_q_entry *entry = &q->entries[i];
> > -
> > struct ntsync_obj *obj = entry->obj;
> > -
> > -
> > atomic_inc(&obj->all_hint);
> > -
> > -
> > /*
> > -
> >  * obj->all_waiters is protected by dev->wait_all_lock rather
> > -
> >  * than obj->lock, so there is no need to acquire obj->lock
> > -
> >  * here.
> > -
> >  */
> > -
> > list_add_tail(&entry->node, &obj->all_waiters);
> > - }
>
> This looks racy, consider:
>
> atomic_read(all_hints) /* 0 */
> 		spin_lock(wait_all_lock)
> 		atomic_inc(all_hint)	/* 1 */
> 		list_add_tail()
>
> spin_lock(sem->lock)
>    /* try_wake_all_obj() missing */
>
> I've not yet thought about if this is harmful or not, but if not, it
> definitely needs a comment.
>
> Anyway, I need a break, maybe more this evening.

Ach. I wrote this with the idea that the race isn't meaningful, but
looking at it again you're right—there is a harmful race here.

I think it should be fixable by moving the atomic_read inside the lock,
though.