From: Coly Li Subject: Re: [PATCH 03/13] bcache: inherit the optimal I/O size Date: Mon, 21 Sep 2020 23:09:48 +0800 Message-ID: <5bcc52dc-ca8f-bbdd-69ef-4b6312e7994a@suse.de> References: <20200921080734.452759-1-hch@lst.de> <20200921080734.452759-4-hch@lst.de> <20200921140010.GA14672@lst.de> Mime-Version: 1.0 Content-Transfer-Encoding: 8bit Return-path: In-Reply-To: <20200921140010.GA14672@lst.de> Content-Language: en-US List-ID: Content-Type: text/plain; charset="us-ascii" To: Christoph Hellwig Cc: Jens Axboe , Song Liu , Hans de Goede , Richard Weinberger , Minchan Kim , Johannes Thumshirn , Justin Sanders , linux-mtd@lists.infradead.org, dm-devel@redhat.com, linux-block@vger.kernel.org, linux-bcache@vger.kernel.org, linux-kernel@vger.kernel.org, drbd-dev@lists.linbit.com, linux-raid@vger.kernel.org, linux-fsdevel@vger.kernel.org, linux-mm@kvack.org, cgroups@vger.kernel.org On 2020/9/21 22:00, Christoph Hellwig wrote: > On Mon, Sep 21, 2020 at 05:54:59PM +0800, Coly Li wrote: >> I am not sure whether virtual bcache device's optimal request size can >> be simply set like this. >> >> Most of time inherit backing device's optimal request size is fine, but >> there are two exceptions, >> - Read request hits on cache device >> - User sets sequential_cuttoff as 0, all writing may go into cache >> device firstly. >> For the above two conditions, all I/Os goes into cache device, using >> optimal request size of backing device might be improper. >> >> Just a guess, is it OK to set the optimal request size of the virtual >> bcache device as the least common multiple of cache device's and backing >> device's optimal request sizes ? > > Well, if the optimal I/O size is wrong, the read ahead size also is > wrong. Can we just drop the setting? > I feel this is something should be fixed. Indeed I overlooked it until you point out the issue now. The optimal request size and read ahead pages hint are necessary, but current initialization is simple. A better way might be dynamically setting them depends on the cache mode and some special configuration. By your inspiration, I want to ACK your original patch although it doesn't work fine for all condition. Then we may know these two settings (ra_pages and queue_io_opt) should be improved for more situations. At lease for most part of the situations they provide proper hints. How do you think of the above idea ? Coly Li