From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Date: Wed, 21 May 2014 07:36:34 +0200 From: Christoph Hellwig To: Ming Lei Cc: Christoph Hellwig , Jens Axboe , Linux Kernel Mailing List , stable Subject: Re: [PATCH] block: mq flush: fix race between IPI handler and mq flush worker Message-ID: <20140521053634.GA4764@lst.de> References: <1400511950-18522-1-git-send-email-tom.leiming@gmail.com> <20140519151824.GA21762@lst.de> <20140520152303.GA20352@lst.de> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org List-ID: On Wed, May 21, 2014 at 01:16:14PM +0800, Ming Lei wrote: > I am wondering if virtio-blk is trivial block driver, :-) It's about as simple as it gets. > > The scsi-mq work that I plant to submit for the next merge window is > > the prime example. > > It depends if one scsi-mq req has to requeue itself with rq->requeue_work > inside its own .softirq_done_fn. If yes, we can't put call_single_data > and requeue_work into one union simply. From you last scsi-mq post, > looks the request may do that if I understand correctly. Requeueing a request from the completion handler is indeed what we'll need with various more complete drivers. > I think the patch is clean and simple, with documenting the special > conflict case clearly too. While I can't say anything against the fact that it fixes the issue it's neither clean nor simple. > Follows current ideas: > 1), this patch with scsi-mq sharing abort_work together? > 2), move requeue_work out of the union inside request > 3), spin_lock_irqsave(&ctx->lock) everywhere and requeue > request directly to ctx without using work I think Jens very much wanted to avoid irq disabling in the I/O path if possible. If we have a separate requeue list with it's separate lock we can avoid that unless we actually have to take requests of that requeue list. I can look into that implementation.