Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 18 Jan 2022 15:17:54 -0400
From:   Jason Gunthorpe <jgg@ziepe.ca>
To:     Santosh Shilimkar <santosh.shilimkar@oracle.com>
Cc:     Praveen Kannoju <praveen.kannoju@oracle.com>,
        "David S . Miller" <davem@davemloft.net>,
        "kuba@kernel.org" <kuba@kernel.org>,
        "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
        "linux-rdma@vger.kernel.org" <linux-rdma@vger.kernel.org>,
        "rds-devel@oss.oracle.com" <rds-devel@oss.oracle.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Rama Nichanamatlu <rama.nichanamatlu@oracle.com>,
        Rajesh Sivaramasubramaniom 
        <rajesh.sivaramasubramaniom@oracle.com>
Subject: Re: [PATCH RFC] rds: ib: Reduce the contention caused by the
 asynchronous workers to flush the mr pool
Message-ID: <20220118191754.GG8034@ziepe.ca>
References: <1642517238-9912-1-git-send-email-praveen.kannoju@oracle.com>
 <53D98F26-FC52-4F3E-9700-ED0312756785@oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <53D98F26-FC52-4F3E-9700-ED0312756785@oracle.com>
Precedence: bulk

On Tue, Jan 18, 2022 at 04:48:43PM +0000, Santosh Shilimkar wrote:
> 
> > On Jan 18, 2022, at 6:47 AM, Praveen Kannoju <praveen.kannoju@oracle.com> wrote:
> > 
> > This patch aims to reduce the number of asynchronous workers being spawned
> > to execute the function "rds_ib_flush_mr_pool" during the high I/O
> > situations. Synchronous call path's to this function "rds_ib_flush_mr_pool"
> > will be executed without being disturbed. By reducing the number of
> > processes contending to flush the mr pool, the total number of D state
> > processes waiting to acquire the mutex lock will be greatly reduced, which
> > otherwise were causing DB instance crash as the corresponding processes
> > were not progressing while waiting to acquire the mutex lock.
> > 
> > Signed-off-by: Praveen Kumar Kannoju <praveen.kannoju@oracle.com>
> > —
> > 
> […]
> 
> > diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
> > index 8f070ee..6b640b5 100644
> > +++ b/net/rds/ib_rdma.c
> > @@ -393,6 +393,8 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
> > 	 */
> > 	dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
> > 	dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
> > +	WRITE_ONCE(pool->flush_ongoing, true);
> > +	smp_wmb();
> > 	if (free_all) {
> > 		unsigned long flags;
> > 
> > @@ -430,6 +432,8 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
> > 	atomic_sub(nfreed, &pool->item_count);
> > 
> > out:
> > +	WRITE_ONCE(pool->flush_ongoing, false);
> > +	smp_wmb();
> > 	mutex_unlock(&pool->flush_lock);
> > 	if (waitqueue_active(&pool->flush_wait))
> > 		wake_up(&pool->flush_wait);
> > @@ -507,8 +511,17 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
> > 
> > 	/* If we've pinned too many pages, request a flush */
> > 	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
> > -	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
> > -		queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
> > +	    atomic_read(&pool->dirty_count) >= pool->max_items / 5) {
> > +		smp_rmb();
> You won’t need these explicit barriers since above atomic and write once already
> issue them.

No, they don't. Use smp_store_release() and smp_load_acquire if you
want to do something like this, but I still can't quite figure out if
this usage of unlocked memory accesses makes any sense at all.

Jason