Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752813AbbHUUV3 (ORCPT ); Fri, 21 Aug 2015 16:21:29 -0400 Received: from forward-corp1f.mail.yandex.net ([95.108.130.40]:37997 "EHLO forward-corp1f.mail.yandex.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752570AbbHUUV1 (ORCPT ); Fri, 21 Aug 2015 16:21:27 -0400 From: Roman Gushchin To: Linus Torvalds Cc: linux-mm , Linux Kernel Mailing List , Raghavendra K T , Jan Kara , Wu Fengguang , David Rientjes , Andrew Morton In-Reply-To: References: <1440177121-12741-1-git-send-email-klamm@yandex-team.ru> Subject: Re: [PATCH] mm: use only per-device readahead limit MIME-Version: 1.0 Message-Id: <197171440188481@webcorp01e.yandex-team.ru> X-Mailer: Yamail [ http://yandex.ru ] 5.0 Date: Fri, 21 Aug 2015 23:21:21 +0300 Content-Transfer-Encoding: 8bit Content-Type: text/plain; charset=koi8-r Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 1775 Lines: 42 21.08.2015, 21:17, "Linus Torvalds" : > On Fri, Aug 21, 2015 at 10:12 AM, Roman Gushchin wrote: >> ?There are devices, which require custom readahead limit. >> ?For instance, for RAIDs it's calculated as number of devices >> ?multiplied by chunk size times 2. > > So afaik, the default read-ahead size is 128kB, which is actually > smaller than the old 512-page limit. > > Which means that you probably changed "ra_pages" somehow. Is it some > system tool that does that automatically, and if so based on what, > exactly? It's just a raid driver. For instance, drivers/ms/raid5.c:6898 . On my setup I got unexpectedly even slight perfomance increase over O_DIRECT case and over old memory-based readahead limit, as you can see from numbers in the commit message (1.2GB/s vs 1.1 GB/s). So, I like an idea to delegate the readahead limit calculation to the underlying i/o level. > I'm also slightly worried about the fact that now the max read-ahead > may actually be zero, For "normal" readahead nothing changes. Only readahead syscall and madvise(MADV_WILL_NEED) cases are affected. I think, it's ok to do nothing, if readahead was deliberately disabled. > and/or basically infinite (there's a ioctl to > set it that only tests that it's not negative). Does everything react > ok to that? It's an open question, if we have to add some checks to avoid miss-configuration. In any case, we can check the limit on setting rather then adjust them dynamically. -- Roman -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/