Re: [PATCH 2/2] locking/qrwlock: Reduce cacheline contention for rwlocks used in interrupt context

From: Peter Zijlstra <peterz@infradead.org>
To: Waiman Long <longman@redhat.com>
Cc: Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>,
	Boqun Feng <boqun.feng@gmail.com>, Arnd Bergmann <arnd@arndb.de>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 2/2] locking/qrwlock: Reduce cacheline contention for rwlocks used in interrupt context
Date: Wed, 11 May 2022 15:34:59 +0200	[thread overview]
Message-ID: <20220511133459.GZ76023@worktop.programming.kicks-ass.net> (raw)
In-Reply-To: <7eff9889-56fe-503c-94ea-376054c6579b@redhat.com>

On Wed, May 11, 2022 at 08:44:55AM -0400, Waiman Long wrote:

> > I'm confused; prior to this change:
> > 
> > 	CPU0			CPU1
> > 
> > 	write_lock_irq(&l)
> > 				read_lock(&l)
> > 				<INRQ>
> > 				  read_lock(&l)
> > 				  ...
> > 
> > was not deadlock, but now it would AFAICT.
> 
> Oh you are right. I missed that scenario in my analysis. My bad.

No worries; I suppose we can also still do something like:

void queued_read_lock_slowpath(struct qrwlock *lock, int cnts)
{
	/*
	 * the big comment
	 */
	if (unlikely(in_interrupt())) {
		/*
		 * If not write-locked, insta-grant the reader
		 */
		if (!(cnts & _QW_LOCKED))
			return;

		/*
		 * otherwise, wait for the writer to go away.
		 */
		atomic_cond_read_acquire(&lock->cnts, !(VAL & _QW_LOCKED));
		return;
	}

	...
}

Which saves one load in some cases... not sure it's worth it though.