From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jens Axboe Subject: Re: [PATCH 0/5] convert sg to use the block layer Date: Tue, 26 Aug 2008 09:56:28 +0200 Message-ID: <20080826075627.GU20055@kernel.dk> References: <1219716654-27827-1-git-send-email-fujita.tomonori@lab.ntt.co.jp> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from [93.163.65.50] ([93.163.65.50]:16177 "EHLO kernel.dk" rhost-flags-FAIL-FAIL-OK-OK) by vger.kernel.org with ESMTP id S1752140AbYHZH43 (ORCPT ); Tue, 26 Aug 2008 03:56:29 -0400 Content-Disposition: inline In-Reply-To: <1219716654-27827-1-git-send-email-fujita.tomonori@lab.ntt.co.jp> Sender: linux-scsi-owner@vger.kernel.org List-Id: linux-scsi@vger.kernel.org To: FUJITA Tomonori Cc: linux-scsi@vger.kernel.org, dougg@torque.net, michaelc@cs.wisc.edu, James.Bottomley@HansenPartnership.com On Tue, Aug 26 2008, FUJITA Tomonori wrote: > This patchset converts sg to use the block layer functions. That is, > sg doesn't use scsi_execute_async() any more. This is a part of the > overdue task to remove scsi_req_map_sg. > > I tested this patchset with sg v3 and the old interface (struct > sg_header) via SG_IO (v3) and the vfs API. > > > Doug, > > 1. I don't demote the sg driver's GFP_ATOMIC to GFP_KERNEL. sg always > uses GFP_ATOMIC as before. > > 2. I don't remove GFP_DMA allocation. As before, sg allocates reserved > pages with GFP_DMA (sfp->low_dma case). > > 3. I keep the reserved buffer per struct sg_fd as before. > > 4. I use high-order page allocation for reserved buffer as before. sg > works well with HBAs that have the limitation of the number of sg > entries. > > 5. I think that you were concern about the overhead of the block layer > functions. But if you look at scsi_execute_async() that sg uses now, > you can find that scsi_execute_async() uses the block layer functions > internally. So the current sg incurs the overhead (if such overhead > exists). > > > Jens, > > I keep the block API changes to a minimum (I might need more changes > for st/osst but I'd like to progress step by step). I did only two > things. > > 1. I add gfp_mask argument to blk_rq_map_user and blk_rq_map_user_iov. > > 2. I introduces struct rq_map_data holding pages. sg puts > pre-allocated pages to it and passes it to bio_copy_user_iov(). The > current users of bio_copy_user_iov simply passes NULL. blk_rq_map_user > and blk_rq_map_user_iov take a pointer to struct rq_map_data and in > the end bio_copy_user_iov gets it. > > > This patchset against the for-linus branch in Jens' tree + the two > patches for 2.6.27: > > http://marc.info/?l=linux-kernel&m=121964251911717&w=2 > http://marc.info/?l=linux-kernel&m=121964241911611&w=2 > > After Jens rebases the for-2.6.28 brach, I'll update this patchset > too. Thanks a lot for doing this work, it's been pending for a long time. I've rebased for-2.6.28 on top of for-linus to ease this integration, so if you could resend the patchset then I can add it. -- Jens Axboe