From: "Darrick J. Wong" Subject: Re: [PATCH 2/2] e2fsprogs: Add support for toggling, verifying, and fixing inode checksums Date: Mon, 11 Apr 2011 19:05:29 -0700 Message-ID: <20110412020529.GG24354@tux1.beaverton.ibm.com> References: <20110406224410.GB24354@tux1.beaverton.ibm.com> <20110406224733.GU32706@tux1.beaverton.ibm.com> <20110408192530.GE24354@tux1.beaverton.ibm.com> <001599E2-27BF-48AF-BC4E-DE8B674FF46B@dilger.ca> Reply-To: djwong@us.ibm.com Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: "Theodore Ts'o" , linux-ext4 , linux-kernel To: Andreas Dilger Return-path: Received: from e9.ny.us.ibm.com ([32.97.182.139]:55130 "EHLO e9.ny.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756537Ab1DLCF3 (ORCPT ); Mon, 11 Apr 2011 22:05:29 -0400 Content-Disposition: inline In-Reply-To: <001599E2-27BF-48AF-BC4E-DE8B674FF46B@dilger.ca> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Fri, Apr 08, 2011 at 05:13:13PM -0600, Andreas Dilger wrote: > On 2011-04-08, at 1:25 PM, Darrick J. Wong wrote: > > On Fri, Apr 08, 2011 at 03:14:04AM -0600, Andreas Dilger wrote: > >> Do you have an e2fsck testcase for this code, to show that it detects/fixes > >> inodes with data corruption, and to fix the checksums after the ROCOMPAT flag > >> is set the first time? > > > > Not yet; I suspected that some clarification of exactly that issue was needed. > > It looks to me that in general the checksum will be zero for the "flag is > > enabled but no checksum has yet been provided" case, and nonzero in the "inode > > is corrupt" case. So if e2fsck sees zero it'd first ask to correct the > > checksum, and if it sees nonzero it'll first ask to clear the inode. If the > > user answers no to the first question, e2fsck can then propose the second > > option. > > Seems reasonable, though it is possible that the inode checksums can also > become invalid due to changing the filesystem UUID. This should probably be > handled by tune2fs when the UUID is changed, with an extra prompt if > INODE_CSUM is enabled to indicate that the conversion may take a long time. Yes, sounds reasonable. I guess we'd have to verify all the inode checksums before changing the UUID, change the UUID, and then set new checksums. If the pre-verification step fails, demand e2fsck and don't write anything. > Looking at the checksum algorithm you used, the inode checksum does not > change if the inode is relocated due to resize (i.e. it uses the inode number > and not the underlying block number). This is convenient, and does not > impact the correctness in any way - if the wrong block is read/written then > the inode number used in the checksum will not match either. > > >> With the "ibadness" patch in our tree, the bad checksum should be a > >> significant factor in marking the inode as garbage, but possibly not enough > >> to have it thrown out if there are no other errors in the inode. > > > > Or e2fsck could use that heuristic; which tree is the ibadness patch in? > > Google shows a patch from 2008, but no recent discussion. > > There is a relatively up-to-date version at > http://git.whamcloud.com/?p=tools/e2fsprogs.git;a=blob_plain;f=patches/e2fsprogs-ibadness-counter.patch;hb=8dd11ed9bdf0914d57d78d0c387bd21f747c1d29 Ok, I'll pull that into my tree. > > Something along the lines of: if the inode is not very bad, ask first to fix > > the checksum and second to clear the inode; if the inode seems bad, ask first > > to clear it and second to fix the checksum. > > Yes, that is what I was thinking. The real question is why the checksum > would be bad in the case of no other "badness"? If it is due to the UUID, > that should be handled when the UUID is changed, and if it is due to a > misplaced write (i.e. bad inode number) then it will help us to distinguish > between the "real" inode and the misplaced "bad" inode. Agreed. I discovered another problem is that there seem to be a number of places where e2fsck loads only the first 128 bytes out of an inode, checks it, and then writes out a "corrected" 128 byte inode. Obviously e2fsck needs to be changed to read in the full inode size (whatever that is) and write out the same amount, though this will probably result in a lot of e2fsck code churn. I added the ability for e2fsck to zero out the checksum if it finds a Linux ext* fs and the checksum feature disabled. Mingming was also wondering if we ought to save some rocompat bits and combine all the current checksumming proposals (extent tree, bitmaps) under one rocompat bit? Sounds like a decent idea to me. By the way, I've been uploading my notes about on-disk layout to the wiki: https://ext4.wiki.kernel.org/index.php/Ext4_Disk_Layout (Not sure if it's 100% clueful, but there we go.) --D > > >>> @@ -890,6 +890,11 @@ static struct e2fsck_problem problem_table[] = { > >>> "(size %Is, lblk %r)\n"), > >>> PROMPT_CLEAR, PR_PREEN_OK }, > >>> > >>> + /* Fast symlink has EXTENTS_FL set */ > >>> + { PR_1_INODE_CSUM_INVALID, > >>> + N_("inode %i checksum invalid. "), > >> > >> The comment for each problem should exactly mirror the text that is printed. > >> In this case, you haven't used the abbreviations "@i" and "@n", which would > >> normally make it much harder to search for this error string in the code, but > >> also simplifies the translation of the message. > > > > Oops, comment blooper that was a thinko on my part. What would the @n be for? > > @i is "inode", @n is "invalid", per e2fsck/message.c. > > Cheers, Andreas > > > > > > -- > To unsubscribe from this list: send the line "unsubscribe linux-ext4" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html