Subject: Re: LK_RETRY and coda_lookup
To: Bill Studenmund <wrstuden@netbsd.org>
From: Greg Troxel <gdt@ir.bbn.com>
List: tech-kern
Date: 03/15/2006 16:44:49
Bill Studenmund <wrstuden@netbsd.org> writes:

> On Tue, Feb 28, 2006 at 12:38:30PM -0500, Greg Troxel wrote:
> > In src/sys/coda/coda_vnops.c 1.46, I added LK_RETRY to vn_lock calls.
> > This seems correct because vn_lock will (absent LK_RETRY) fail if the
> > vnode is already locked (presumably by another thread), but it will
> > first sleep until a wakeup happens.
> 
> Sorry for the delay, but I think tis is wrong.

no problem on delay (so leaving extra quoting).

> > 1) Am I correct that some other process could have VXLOCK on the
> >    vnode, that this is really a normal condition, and thus that
> >    LK_RETRY is really in order?  (I would think LK_RETRY should only
> >    be omitted as part of some deadlock avoidance procedure where the
> >    calling code is doing retries, or something like that.)
> 
> VXLOCK should only be used when we are recycling the vnode - when we are 
> blowing it away. It is set when we are changing the type of the vnode, in 
> part as all of the locking originally was done by the fs.

So vn_lock sleeps if VXLOCK is set (which should be rare).  If VXLOCK
is not set, it just calls VOP_LOCK.  coda_lock just does

    return (lockmgr(&vp->v_lock, ap->a_flags, &vp->v_interlock));

In lots of code, I see LK_RETRY (e.g. egrep LK_RETRY sys/vfs_vnops.c).
So I don't understand when that flag should and should not be used -
and why if it's wrong in coda vnops it's ok in cfs_vnops.c.
It seems VOP_LOCK is supposed to sleep until succesful rather than
returning EBUSY, and coda_lock uses lockmgr, so that should be ok.

I wonder if something in coda causes VXLOCK that shouldn't.

If I shouldn't use LK_RETRY, should a vnop that fails to get a lock
return EIO or some such?  It seems like it should not be a panic.  Or
should it, and failing to get a lock on a vnode on which we *should*
have a ref is a sign of incorrect behavior.

> > 2) Why does vn_lock sleep if it's going to fail?  Is this for
> >    caller-managed retry loops?
> 
> To wait for the cleanup to finish, I expect.

thanks, makes sense.

> > 3) For clarity, it seems that LK_INTERLOCK should be cleared after
> >    ltsleep, and separately after VOP_LOCK.  I realize this is a case
> >    of CSE but it makes it harder to realize that the rules are being
> >    followed.
> 
> If I understand the code right, I think it's a style thing. We go through 
> one or the other side of an if statement, and we clear it after the if.

I think it's a style thing - I was making the point that the current
style makes it hard to understand how the rules are being followed.
It sounds like you think the code is right, and I'm not trying to
change it, but I appreciate the feedback since I am gradually coming
to understand vfs locking.


With my recent change to coda_venus (1.20) and LK_RETRY, I can run 20
of pax tarring up 350MB of data in coda to /dev/null in parallel.

getcwd is failing within /coda/realm.  The userland/venus fid to inode
mapping is out of sync with the kernel, and that's on my list to fix,
and may resolve the getcwd.

As alwyas, thanks for the locking advice.

-- 
        Greg Troxel <gdt@ir.bbn.com>