From mboxrd@z Thu Jan 1 00:00:00 1970 From: Christoph Hellwig Subject: Re: [PATCH 12/15] scsi: initial blk-mq support Date: Thu, 6 Feb 2014 08:16:15 -0800 Message-ID: <20140206161615.GA16916@infradead.org> References: <20140205124118.332902571@bombadil.infradead.org> <20140205124154.337539740@bombadil.infradead.org> <52F349F9.7090509@dev.mellanox.co.il> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from bombadil.infradead.org ([198.137.202.9]:57989 "EHLO bombadil.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752307AbaBFQQ1 (ORCPT ); Thu, 6 Feb 2014 11:16:27 -0500 Content-Disposition: inline In-Reply-To: <52F349F9.7090509@dev.mellanox.co.il> Sender: linux-scsi-owner@vger.kernel.org List-Id: linux-scsi@vger.kernel.org To: Sagi Grimberg Cc: Jens Axboe , James Bottomley , Nicholas Bellinger , linux-scsi@vger.kernel.org On Thu, Feb 06, 2014 at 10:38:17AM +0200, Sagi Grimberg wrote: > Both you and Nic offer a single HW queue per sdev. > I'm wandering if that should be the LLD's decision (if chooses to > use multiple queues)? > > Trying to understand how LLDs will fit in a way they exploit > multi-queue and actually > maintain multiple queues. SRP/iSER for example maintain a single > queue per connection > (or session in iSCSI). Now with multi-queue all requests of that > shost will eventually > boil-down to posting on a single queue which might transition the > bottleneck to the LLDs. > > I noticed virtio_scsi implementation is choosing a queue per command > based on current > processor id without any explicit mapping (unless I missed it). > > I guess my question is where do (or should) LLDs plug-in to this mq scheme? Just using blk-mq helps with lock contention and cacheline issues, while being conceptually simple, that's why it's the priority. See the proposal I sent before the patch series for more details. That being said if you have simple enough patches for real multiqueue support I'd be more than happy to carry them along.