Ubuntu
linux-source-2.6.15 package

  1. Bug #37435
  2. Comment #4

Comment 4 for bug 37435

Revision history for this message
Gary Coady (garycoady) wrote : Re: maintainer of /var/lib/dpkg/alternatives/ doesn't do updates safely

There's a mail from Chris Wedgwood (who seems to be one of the main XFS developers) in 2004 on the subject at http://marc.theaimsgroup.com/?l=linux-xfs&m=109030382432326&w=2

----- quote ------
sounds like:

  old file foo on disk, all safe

  new file bar is written metadata on disk, file
      data in ram

  [*]

  rename bar to foo old file unlinked, new
      file in place but data
      not flushed yet

now, if there was an fsync at [*] it would work just fine
---- end quote ------

I had a look at the POSIX semantics for rename() at http://www.opengroup.org/onlinepubs/009695399/functions/rename.html and there is no reference to the referenced file having been written to persistent storage. XFS only claims to follow POSIX semantics.

I used to run XFS on my desktop a few years ago, and had quite a few instances of data corruption - its delayed allocation DOES seem to make problems like this more likely to occur than with other systems (I saw lots of NULs in corrupted files after a crash).