Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Fri, 19 Apr 2019 12:26:47 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Waiman Long <longman@redhat.com>
Cc:     Ingo Molnar <mingo@redhat.com>, Will Deacon <will.deacon@arm.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        linux-kernel@vger.kernel.org, x86@kernel.org,
        Davidlohr Bueso <dave@stgolabs.net>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Tim Chen <tim.c.chen@linux.intel.com>,
        huang ying <huang.ying.caritas@gmail.com>
Subject: Re: [PATCH v4 14/16] locking/rwsem: Guard against making count
 negative
Message-ID: <20190419102647.GP7905@worktop.programming.kicks-ass.net>
References: <20190413172259.2740-1-longman@redhat.com>
 <20190413172259.2740-15-longman@redhat.com>
 <20190418135151.GB12232@hirez.programming.kicks-ass.net>
 <d931916b-41f5-f12d-8c38-0cccbc00a301@redhat.com>
 <20190418144036.GE12232@hirez.programming.kicks-ass.net>
 <4cbd3c18-c9c0-56eb-4e01-ee355a69057a@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <4cbd3c18-c9c0-56eb-4e01-ee355a69057a@redhat.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Apr 18, 2019 at 10:54:19AM -0400, Waiman Long wrote:
> On 04/18/2019 10:40 AM, Peter Zijlstra wrote:

> > Having more CPUs than that is not impossible these days.
> >
> 
> Having more than 32k CPUs contending for the same cacheline will be
> horribly slow.

No question about that.

> >> How about disabling preemption before fetch_all and re-enable
> >> it afterward to address the latter concern? 
> > Performance might be an issue, look at what preempt_disable() +
> > preempt_enable() generate for ARM64 for example. That's not particularly
> > pretty.
> 
> That is just for the preempt kernel. Right? Thinking about it some more,
> the above scenario is less likely to happen for CONFIG_PREEMPT_VOLUNTARY
> kernel and the preempt_disable cost will be lower.

Depends a bit on what specific CONFIG knobs are used. IIRC something
like NOHZ_FULL will also select PREEMPT_COUNT, it will just not have the
actual preemption calls in.

> A preempt RT kernel is less likely to run on system with many CPUs
> anyway. We could make that a conifg option as well in a follow-on
> patch and let the distributors decide.

RT has a whole different rwsem implementation anyway, so we don't need
to worry about them.

> >> I have no solution for the first case, though.

> > A cmpxchg() loop can fix this, but that again has performance
> > implications like you mentioned a while back.

I thought of a horrible horrible alternative:

union rwsem_count {
	struct { /* assuming LP64-LE */
		unsigned short  other[3];
		unsigned short  readers;
	};
	unsigned long	value;
};

void down_read(struct rw_semaphore *sem)
{
	union rwsem_count c;
	unsigned short o;

	c.value = atomic_long_read(&sem->count);
	c.readers++;

	if (!c.readers || (c.value & RWSEM_FLAG_WRITER))
		goto slow;

	o = xchg(&((union rwsem_count *)sem)->readers, c.readers);
	if (o != c.readers-1) {
		c.value = atomic_long_fetch_add(&sem->count, o-(c.readers-1));
	} else {
		c.value = atomic_long_read(&sem->count);
		c.readers = o + 1;
	}

	if (!(c.value & RWSEM_FLAG_WRITER))
		return;

slow:
	rwsem_down_read_slow(sem, c.value);
}

It is deterministic in that is has at most 2 unconditional atomic ops,
no cmpxchg loop, and a best case of a single op.