Re: [RFC PATCH]autofs4: hang and proposed fix

[Ian Kent <raven@themaw.net>]

On Sun, 4 Dec 2005, Ian Kent wrote:

> On Sun, 4 Dec 2005, Ian Kent wrote:
> 
> > On Fri, 2 Dec 2005, Will Taber wrote:
> > 
> > > > > 
> > > > > Well, I think we've determined that the reported problem doesn't happen
> > > > > with any in-tree callers.  The question, then, is do you want to fix the
> > > > > locking problem?  Two approaches were presented in this thread.  I don't
> > > > > really like the idea of the hack used by devfs, since it relies on
> > > > > implicit
> > > > > semantics.  I haven't given much thought to the second approach, though
> > > > > (are we sure it can be made to work?).  It may require a good deal of
> > > > > effort, but if it makes things work properly, it's worth considering.  I'm
> > > > > just not sure where it sits in the list of priorities, as I know you've
> > > > > got
> > > > > a lot on your plate, Ian.
> > > > 
> > > > 
> > > > It appears to me that the unhashed directory approach proposed by Will does
> > > > not account for directories that exist but don't have current mounts.
> > > > 
> > > > I will re-read the posts, I expect I missed something, and give it more
> > > > thought.
> > > > 
> > > It doesn't consider that case.  You had mentioned it but I had forgotten.
> > > 
> > 
> > OK so I decided to give Wills recommendation a bit of a run and I've come 
> > up with a first cut patch which of course doesn't work.
> > 
> > The approach is to force all callbacks to go through lookup instead of 
> > some through revalidate as well. The patch basically posts the dentry to a 
> > pending list and unhashs it, then picks it up from the list in the lookup 
> > and rehash it. Should be fairly simple really but I'm doing something 
> > obviously wrong somewhere.
> > 
> > I'm seeing slab corruption and I really can't see why this should be the 
> > case. Anyone got any ideas. The patch is against 2.6.15-rc1 but the kernel 
> > I'm compliling against is a RedHat patched 2.6.11 (Aurora).
> > 

Think I failed to return a dgot dentry from lookup. Ooops.

> 
> Interestingly it seems to function OK on my Intel FC3 box?

Pure chance I think.

> 
> And the patch only deals with the revalidate and lookup logic, the readdir 
> stuff will need to be reworked as well.

I'll push on with this.

Ian