From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jens Axboe Subject: Re: [PATCH v4] drivers/block/mtip32xx: Adding new driver mtip32xx Date: Tue, 13 Sep 2011 23:03:16 +0200 Message-ID: <4E6FC514.7070401@fusionio.com> References: <22A973199D2C2F46933448F6E7990A3002C80026@ntxboimbx31.micron.com> <4E44E782.7090309@fusionio.com> <2A9BE4FF6209B644B6F8EB62DE6AEA1E07663ED4@ntxfrembx01.micron.com> <20110909085433.GA9593@infradead.org> <20110913124906.GA17855@infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset="ISO-8859-1" Content-Transfer-Encoding: 7bit Return-path: Received: from mx2.fusionio.com ([66.114.96.31]:38476 "EHLO mx2.fusionio.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932702Ab1IMVDW (ORCPT ); Tue, 13 Sep 2011 17:03:22 -0400 In-Reply-To: Sender: linux-ide-owner@vger.kernel.org List-Id: linux-ide@vger.kernel.org To: Eric Seppanen Cc: Christoph Hellwig , "Sam Bradshaw (sbradshaw)" , "alan@lxorguk.ukuu.org.uk" , "linux-ide@vger.kernel.org" , "linux-kernel@vger.kernel.org" , Jeff Garzik , "jmoyer@redhat.com" , "Asai Thambi Samymuthu Pattrayasamy (asamymuthupa) [CONTRACTOR]" On 2011-09-13 18:46, Eric Seppanen wrote: > On Tue, Sep 13, 2011 at 5:49 AM, Christoph Hellwig wrote: >> Btw, there is another _huge_ issue with the driver, and that is the >> lack of any internal queueing. Remember the make_request interface is >> an extremly thin layer (or rather the lack of it) below the filesystem. >> >> So for example if eh_active is non-zero you return -EBUSY to the >> filesystems. That's an error code it a) doesn't recognize and b) >> couldn't handle even if it did. Similarly mtip_hw_get_scatterlist >> simply blocks if no tag is currently available instead of queueing >> it up. > > Just out of curiosity, why is blocking on no-tag-available a bad > thing? How is it any different than the blocking that will occur when > a request queue is full? When the hardware queue depth is bigger than > that of a request queue, what extra benefit does queuing give? The blocking for a free tag is fine, the actual implementation is definitely not optimal (using a rw semaphore with count initialized to the tag depth, ugh). You'll need to block for a free tag in any case, or add a thread to restart things on a free tag. The thread would not help performance. But the -EBUSY is definitely a bug, that needs to be a waiting condition as well. -- Jens Axboe