Re: [PATCH] nfs: don't queue synchronous NFSv4 close rpc_release to nfsiod

[Trond Myklebust <Trond.Myklebust@netapp.com>]

On Fri, 2011-02-18 at 15:04 -0500, Jeff Layton wrote: 
> On Thu, 17 Feb 2011 16:37:41 -0500
> Jeff Layton <jlayton@redhat.com> wrote:
> 
> > On Thu, 17 Feb 2011 11:47:24 -0800
> > Trond Myklebust <Trond.Myklebust@netapp.com> wrote:
> > 
> > > On Thu, 2011-02-17 at 10:10 -0500, Jeff Layton wrote: 
> > > > On Thu, 17 Feb 2011 08:40:52 -0500
> > > > Jeff Layton <jlayton@redhat.com> wrote:
> > > > 
> > > > > 
> > > > > Looks like it finally failed on the 39th pass:
> > > > > 
> > > > > second check for lost reply on non-idempotent requests
> > > > > testing 50 idempotencies in directory "testdir"
> > > > > rmdir 1: Directory not empty
> > > > > special tests failed
> > > > > 
> > > > > When I look in the directory (several hours after it failed), the
> > > > > silly-renamed file is still there:
> > > > > 
> > > > > -rw---x--x. 1 root root   30 Feb 16 15:04 .nfs000000000000002d00000090
> > > > > 
> > > > > ...so I'm not sure what exactly is wrong yet, but it looks like the
> > > > > silly delete just never happened. Maybe there's a dentry refcount leak
> > > > > of some sort? There are no queued RPC's.
> > > > > 
> > > > > I'll keep looking at it but if you have ideas as to what it could be,
> > > > > let me know.
> > > > > 
> > > > 
> > > > I walked down the directory tree in crash on the live kernel and found
> > > > the dentry. The d_count is 0x0, so I'm not clear on why it didn't get
> > > > cleaned up:
> > > > 
> > > > crash> struct dentry.d_flags,d_count,d_name 0xffff880017d46a80
> > > >   d_flags = 0xc000, 
> > > >   d_count = 0x0, 
> > > >   d_name = {
> > > >     hash = 0xe08ab5c8, 
> > > >     len = 0x1c, 
> > > >     name = 0xffff880017d46ab8 ".nfs000000000000002d00000090"
> > > >   }, 
> > > > 
> > > > 
> > > > The d_flags are:
> > > > 
> > > > #define DCACHE_OP_REVALIDATE    0x4000
> > > > #define DCACHE_OP_DELETE        0x8000
> > > > 
> > > > ...very odd. I'd have expected to see this one set too:
> > > > 
> > > > #define DCACHE_NFSFS_RENAMED  0x0002
> > > > 
> > > > I suppose the async sillyrename call could have failed and we ended up
> > > > calling nfs_cancel_async_unlink? I'll stick in some printk's around
> > > > that area and see if I can figure out what's going on...
> > > > 
> > > 
> > > Perhaps I missed a call site that needs an rpc_put_task_async()?
> > > 
> > 
> > I'm not sure that would explain what we're seeing here, unless I'm just
> > missing something. I do know that nfs_cancel_async_unlink was never
> > called in my latest test run, so that does not seem to be it.
> > 
> > Another anomaly -- d_fsdata is NULL. How do we get into that state with
> > a dentry that has been silly-renamed? FWIW, here's the entire dentry
> > struct with my latest reproducer. Let me know if anything stands out to
> > you...
> > 
> 
> My current thinking is that the problems I'm seeing here are not
> directly related to the bug we're trying to fix. It looks like
> sillyrename is still just plain broken a'la your email entitled:
> 
>     2.6.38-rc2... NFS sillyrename is broken...
> 
> ...I'll plan to keep an eye out for the fix from Nick, and re-test this
> when it hits.
> 
> Thanks,

Oh, I'm sorry. I wasn't aware you were testing on 2.6.38-rc...

Yes, sillyrename is still broken there due to the NULL parent dentry
being passed to the d_iput() callback. Nick promised us a patch, but
hasn't delivered yet.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com