From: Ted Ts'o Subject: Re: [PATCH 1/3] ext4: s_freeclusters_counter should not tranform to unit of block before assigning to "free_clusters" in ext4_has_free_cluste Date: Sat, 24 Mar 2012 17:05:18 -0400 Message-ID: <20120324210518.GB5883@thunk.org> References: <1331638698-14796-1-git-send-email-hao.bigrat@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-ext4@vger.kernel.org, Robin Dong To: GRobin Dong Return-path: Received: from li9-11.members.linode.com ([67.18.176.11]:60261 "EHLO test.thunk.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753552Ab2CXVFV (ORCPT ); Sat, 24 Mar 2012 17:05:21 -0400 Content-Disposition: inline In-Reply-To: <1331638698-14796-1-git-send-email-hao.bigrat@gmail.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Tue, Mar 13, 2012 at 07:38:16PM +0800, Robin Dong wrote: > Creating 4-byte files until ENOSPC in a delay-allocation and bigalloc ext4 fs and then sync it, the dmseg will report like: > > [ 482.154538] EXT4-fs (sdb6): delayed block allocation failed for inode 1664 at logical offset 0 with max blocks 1 with error -28 > [ 482.154540] EXT4-fs (sdb6): This should not happen!! Data will be lost > > The reason is ext4_has_free_clusters reporting wrong > result. Actually, the unit of sbi->s_dirtyclusters_counter is block, > so we should tranform it to cluster for argument "dirty_clusters", > just like "free_clusters". We have a bigger problem here, which is that this is not the only place where s_dirty_clusters_counter is being used in units of clusters. (See ext4_claim_free_clusters, which when called by mballoc is using units of clusters.) We definitely have brokeness here, but this is not the whole story. We need to take a step back here and decide whether the correct units is clusters or blocks. Ultimately I think it does need to be clusters, because we can't just convert blocks and clusters by using B2C; we could dirty 3 blocks, but if those 3 blocks span two 64-block clusters, what's important is that we have to reserve space for 2 clusters. We can't just calculate "3 >> 6" and assume that we can reserve 0 clusters and be done with it! This is one of the places where I think we need to solve things by having a better data structure for tracking which pages have been subject to delayed allocation, since if we touch another block in a cluster where we've done a delayed allocation, we don't need to bump s_dirtyclusters_counter. However, if this is the first time we've touched a block in a particular cluster, then we *do* need to bump s_dirtyclusters_counter --- and if we need to search all of the pages in the page cache to make this determination, it's going to be painful.... - Ted