From: Christoph Hellwig Subject: Re: switch loop and target to use ITER_BVEC iov_iter Date: Sun, 25 Jan 2015 14:45:42 +0100 Message-ID: <20150125134542.GB21974@lst.de> References: <1421593624-4462-1-git-send-email-hch@lst.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Al Viro , Jens Axboe , Nicholas Bellinger , Linux FS Devel , target-devel@vger.kernel.org To: Ming Lei Return-path: Content-Disposition: inline In-Reply-To: Sender: target-devel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On Thu, Jan 22, 2015 at 12:11:18PM +0800, Ming Lei wrote: > On Sun, Jan 18, 2015 at 11:07 PM, Christoph Hellwig wrote: > > This series adds two new helpers to easily read from and write to bio_bvecs, > > and switches the loop driver and target file backend to use it. > > > > Using bio_vecs directly avoids the need to kmap individual elements in > > the callers, which is epecially important in the target driver, and also > > gets rid of the horrible splice code abuse hack in the loop driver. > > IMO, from driver or kernel view, submit()/complete() model > is very very common, and is more efficient because unnecessary > context switch and process' creation can be avoided when concurrent > read/write is needed for sake of performance, so I think kernel > aio based API is better. My test result about kernel aio based loop > shows CPU can be saved much with kernel AIO in [2]. > > Currently the kernel AIO based approach for loop(v2 patches, [1] [2]) > only handles direct I/O with ITER_BVEC, but it is quite easy to > cover non-direct I/O by BVEC with supporting current submission. > > For encrypt_type loop, it may be covered by kernel aio interface too. > > I will try to cover the above two loop cases by kernel based AIO in > v3, then the current ->write() and splice code may be removed once > it is done. We'll still need to support the buffered I/O case. Some people may rely on it for their workload performance, and other cases like a filesystem with a small sector size ontop of a loop device on a filesystem or device with a larger sector size won't even work at all using direct I/O. So we'll need something like my series to get started, and if aio proves to win over usin a kernel direct I/O read/write (which I suspect it will, but we'll need numbers) we can use aio for the direct I/O code path.