XFS and linux-image-2.6.15-1-686[-smp]

[Robert Edmonds <edmonds42@bellsouth.net>]

Hi, I've been using the backported 2.6.15 -686 flavor kernel package,
and most recently the -686-smp flavor (version 2.6.15-5bpo1).

One of our users has discovered a use case that triggers some nasty XFS
error messages that look like this:

  XFS: possible memory allocation deadlock in kmem_alloc (mode:0x250)

And now, more seriously, processes have been going into uninterruptible
sleep when accessing certain files on this XFS filesystem (a reboot
clears this up).  I now see these kinds of messages in the logs:

  kernel: printk: 13 messages suppressed.
  kernel: allocation failed: out of vmalloc space - use vmalloc=<size>
    to increase size.
  last message repeated 6 times
  kernel: allocation failed: out of vmalloc space - use vmalloc=<size>
    to increase size.
  last message repeated 6 times

Here's what the use case looks like in pseudo code:

    do (a few billion times)
        if !opendir("my/new/directory")
            system("mkdir -p my/new/directory")
        else
            closedir()

        for rec in a few records:
            open("a/new/directory/" + rec.name)
            print(rec.data)
            close()

Note that it's not really creating a billion new directories.  There's
only 80 subdirectories and about 20,000 files in those directories, so
the same directories are being tested and the same files are being
repeatedly opened, written to, and closed.  Very inefficient, but it
shouldn't be fscking up the kernel.

I suspect the solution here is to backup and restore to a non-SGI
filesystem.

-- 
Robert Edmonds