[Ocfs2-users] git checkout on an ocfs2 filesystem

[Joel Becker]

On Tue, Sep 01, 2009 at 09:32:14AM +0800, Tao Ma wrote:
> >5441  open("t/t6015-rev-list-show-all-parents.sh", O_WRONLY|O_CREAT|O_EXCL|O_LARGEFILE, 0777) = ? ERESTARTSYS (To be restarted)
> >5441  --- SIGALRM (Alarm clock) @ 0 (0) ---
> >5441  sigreturn()                       = ? (mask now [])
> >5441  open("t/t6015-rev-list-show-all-parents.sh", O_WRONLY|O_CREAT|O_EXCL|O_LARGEFILE, 0777) = -1 EEXIST (File exists)
> >5441  write(2, "error: git-checkout-index: unabl"..., 100) = 100
> >
> >How on earth is userspace seeing ERESTARTSYS?  Did someone forget to
> >-ive it?
> The ERESTARTSYS may happen when we get interrupted from ocfs2_cluster_lock.
> I met with it when I rm -rf a very large dir and use "ctrl+c" to
> stop it when I tested bug 1162.

	Yeah, I got there.  In the git case, we do ocfs2_add_entry(),
but then a signal interrupts ocfs2_dentry_lock().  So ERESTARTSYS is
returned, but the file has been created.  When entry.S goes to retry the
open(O_EXCL), it gets EEXIST.
	I've added code to block signals around that dentry lock call,
but there's another place that the git code is triggering.  I'm hunting
that down.
	In general, we can't return ERESTARTSYS once we've done
something that isn't idempotent.  I think we need to audit our code a
bit.

Joel

-- 

Life's Little Instruction Book #173

	"Be kinder than necessary."

Joel Becker
Principal Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127