From mboxrd@z Thu Jan 1 00:00:00 1970 From: Hans Reiser Subject: Re: Reiserfs munched directory on power failure Date: Mon, 16 Aug 2004 00:45:36 -0700 Message-ID: <41206620.3090504@namesys.com> References: <16672.22947.218114.492596@crowfix.com> Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Return-path: list-help: list-unsubscribe: list-post: Errors-To: flx@namesys.com In-Reply-To: <16672.22947.218114.492596@crowfix.com> List-Id: Content-Type: text/plain; charset="us-ascii"; format="flowed" To: Felix Finch Cc: reiserfs-list@namesys.com What you experienced is what metadata journaling does when you crash in the middle of a write. The ordered writes option should prevent your problem at only a modest performance loss. Reiser4 is fully atomic and higher performance, I encourage you to use it. Hans Felix Finch wrote: >I am running Linux 2.4.26 SMP with reiserfsprogs 3.6.4. It is a dual >Athlon with 1G of RAM. I have had losses three times in reiserfs >partitions when it loses power. My mostly static partitions are ext3, >all others are reiserfs. This is Slackware 9.0. > >This poor machine is not behaving very well. I added two disks to it, >the power supply called it quits and apparently left some lingering >resentment with the motherboard, maybe the processors, and now it >shuts itself off for no reason every once in a while. Sometimes it >runs for a couple of minutes, other times for several days. Not heat >related, I think, because I wait for the disks to spin down, hit the >power button, and it comes right back up. It has failed on cold >mornings and hot afternoons. The only common denominator is that it >has only failed when I am at the keyboard; I have never found it >turned off when I wake up, or when I come home. > >I am awaiting parts to build a new machine. Bummer! Neato! :-) > >I tell you all this in case you decide it makes this bug report >unreliable. I am not certain of that, myself, because other than >losing unsaved edits, the only other problems are from the Reiserfs >partitions. > >I have had three reiserfs failures. > >One power failure occurred within a few seconds of opening a text file >in emacs. I do not remember if I had made any changes, but if I >remember emacs correctly, it doesn't do anything with the original >file until you tell it to save. At any rate, soon after opening the >file, power went bye-bye, and when I brought it back up, that file's >2000 lines of Perl code had been replaced by about half the GPL. >Luckily subversion had the original, so I did not lose anything. No >other files were damaged that I could tell. > >The other two were similar to each other, an hour or two apart. I had >just visited a site with mozilla when power failed. When it came back >up, and I restarted the browser, it had lost all my bookmarks and some >of my preferences. I regret that I cannot say with detail what files >had gone walkabout, since I did not notice the corruption until I >brought mozilla up, and mozilla had (re)written new files by the time >I noticed. I recovered everything from backup, and it hasn't lost >power while surfing the web since. > >Mozilla does have LAST VISIT times in the bookmarks file, so it may >well have been modifying the file when power failed. I don't know >what it would have been changing in the preferences file. > >I don't have any idea why the Perl code was replaced by the GPL, since >I haven't done anything to the GPL file since I first vreated it many >moons ago, and I had not written any changes to the file. > >The mozilla surprise was spookier, in some ways, but also more >understandable if mozilla had just written new versions and the >reiserfs journal is metadata only. > >Once I get the new machine built and it has replaced the old machine, >I would be glad to help debug this, if it would help. Installing a >new 2.4 kernel is no problem, but 2.6 is perhaps too much hassle; I >believe I would have to upgrade some utilities also. But I can >certainly turn on extra logging, edit in extra logging, cycle power at >interesting times, etc. > >And of course if these "failures" are just symptomatic of metadata >journaling and power failures, and you think they are all explainable >by natural causes, that's fine too. > >I don't read this list, so cc: me if you want. > > >