[GIT BISECT] first bad commit: 1f36f774 Switch !O_CREAT case to use of do_last()

[Boaz Harrosh <bharrosh@panasas.com>]

1f36f774b22a0ceb7dd33eca626746c81a97b6a5 is the first bad commit
commit 1f36f774b22a0ceb7dd33eca626746c81a97b6a5
Author: Al Viro <viro@zeniv.linux.org.uk>
Date:   Sat Dec 26 10:56:19 2009 -0500

    Switch !O_CREAT case to use of do_last()
    
    ... and now we have all intents crap well localized
    
    Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

:040000 040000 53c1effc5b22746bb83cdbc6a419bf898067882d 9606f7275fde188a056fd174aa1d622ac0630893 M      fs

I have a file corruption regression  bisected to above. (The one step before def4af30cf945a3735ffca865788ea84b30b25d9
was fine)

The test:

- I have an exofs filesystem mounted on /mnt/exofs
- []$ cd /mnt/exofs/some_linux_git; git status;
  All is fine
- []$ mount -t nfs4 -o minorversion=0 localhost:/ /mnt/nfs
  (Where etc/exports will export /mnt/exofs via nfs4.1)
- []$ cd /mnt/nfs/some_linux_git; git status;
  This will fail and will corrupt the .git/index file. Sometimes the file would be
  too short, and sometimes the file will become a directory (Yes really)

Simple revert of this patch over 2.6.34-rc2 will not work so I've not checked it.
A single threaded read like: "diff -r /mnt/nfs/some_linux_git /mnt/nfs/some_linux_git.back"
or even "diff -r /mnt/exofs/some_linux_git /mnt/nfs/some_linux_git" works fine.
A simple dd write test works fine as well.

So it looks like a race that only something like git can exercise, with multi threaded
reads and writes, through nfs/nfsd.

I know Benny also had failures with files export ontop of 2.6.34-rc2
(I looked at the patch, it is totally out of my league, please help)

I will try to reproduce this with a more conventional FS, or any thing else
I should test?

Thanks
Boaz