Re: [PATCH] locking/percpu_rwsem: Rewrite to not use rwsem

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Aug 07, 2019 at 11:56:58AM +0200, Oleg Nesterov wrote:

> and either way, with or without 2 queues, what do you think about the code
> below?

Sorry for being so tardy with this thread.. having once again picked up
the patch, I found your email.

> This way the new reader does wake_up() only in the very unlikely case when
> it races with the new writer which sets sem->block = 1 right after
> this_cpu_inc().

Ah, by waiting early, you avoid spurious wakeups when
__percpu_down_read() happens after a successful percpu_down_write().
Nice!

I've made these changes. Now let me go have a play with that second
waitqueue.

> -------------------------------------------------------------------------------
> 
> static inline void percpu_down_read(struct percpu_rw_semaphore *sem)
> {
> 	might_sleep();
> 	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
> 
> 	preempt_disable();
> 
> 	if (likely(rcu_sync_is_idle(&sem->rss)))
> 		__this_cpu_inc(*sem->read_count);
> 	else
> 		__percpu_down_read(sem, false);
> 
> 	preempt_enable();
> }
> 
> static inline void percpu_up_read(struct percpu_rw_semaphore *sem)
> {
> 	rwsem_release(&sem->dep_map, 1, _RET_IP_);
> 
> 	preempt_disable();
> 
> 	if (likely(rcu_sync_is_idle(&sem->rss)))
> 		__this_cpu_dec(*sem->read_count);
> 	else
> 		__percpu_up_read(sem);
> 
> 	preempt_enable();
> }

I like that symmetry, but see below ...

> // both called and return with preemption disabled
> 
> bool __percpu_down_read(struct percpu_rw_semaphore *sem, bool try)
> {
> 
> 	if (atomic_read_acquire(&sem->block)) {
> again:
> 		preempt_enable();
> 		__wait_event(sem->waiters, !atomic_read_acquire(&sem->block));
> 		preempt_disable();
> 	}
> 
> 	__this_cpu_inc(*sem->read_count);
> 
> 	smp_mb();
> 
> 	if (likely(!atomic_read_acquire(&sem->block)))
> 		return true;
> 
> 	__percpu_up_read(sem);
> 
> 	if (try)
> 		return false;
> 
> 	goto again;
> }
> 
> void __percpu_up_read(struct percpu_rw_semaphore *sem)
> {
> 	smp_mb();
> 
> 	__this_cpu_dec(*sem->read_count);
> 
	preempt_enable();
> 	wake_up(&sem->waiters);
	preempt_disable()

and this (sadly) means there's a bunch of back-to-back
preempt_disable()+preempt_enable() calls. Leaving out the
preempt_disable() here makes it ugly again :/

Admittedly, this is PREEMPT_RT only, but given that is >< close to
mainline we'd better get it right.

> }
>