Re: [PATCH v5 3/3] blk-cgroup: Optimize blkcg_rstat_flush()

From: Waiman Long <longman@redhat.com>
To: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>,
	cgroups@vger.kernel.org, linux-block@vger.kernel.org,
	linux-kernel@vger.kernel.org, Ming Lei <ming.lei@redhat.com>
Subject: Re: [PATCH v5 3/3] blk-cgroup: Optimize blkcg_rstat_flush()
Date: Thu, 2 Jun 2022 13:26:10 -0400	[thread overview]
Message-ID: <42da456d-8f6a-3af0-4cd3-d33a07e3b81e@redhat.com> (raw)
In-Reply-To: <YpjsTNZx4DG+kot4@slm.duckdns.org>

On 6/2/22 12:58, Tejun Heo wrote:
> Hello,
>
> On Thu, Jun 02, 2022 at 09:35:43AM -0400, Waiman Long wrote:
>> @@ -2011,9 +2076,16 @@ void blk_cgroup_bio_start(struct bio *bio)
>>   	}
>>   	bis->cur.ios[rwd]++;
>>   
>> +	if (!READ_ONCE(bis->lnode.next)) {
>> +		struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
>> +
>> +		llist_add(&bis->lnode, lhead);
>> +		percpu_ref_get(&bis->blkg->refcnt);
> Hmm... what guarantees that more than one threads race here? llist assumes
> that there's a single writer for a given llist_node and the ref count would
> be off too, right?

The llist_add() function is atomic. It calls into llist_add_batch() in 
lib/llist.c which uses cmpxchg() to make the change. There is a 
non-atomic version __llist_add() which may be problematic in this case. 
Note that irq is disabled in the u64_stats_update* critical section, 
there shouldn't be a racing thread running in the same cpu. Other cpus 
will modify their own version of lhead. Perhaps the non-atomic version 
can be used here as well.

Cheers,
Longman