fixing "umount -f" for the NFS client

[Rick Macklem]

Marcelo Araujo wrote:
> 
> 
> On 2013/10/15, at 21:05, Rick Macklem <rmacklem at uoguelph.ca> wrote:
> 
> > araujo wrote:
> >> 
> >> 
> >> 
> >> 
> >> 
> >> 
> >> 2013/9/5 Benjamin Kaduk < kaduk at mit.edu >
> >> 
> >> 
> >> 
> >> On Wed, 4 Sep 2013, Rick Macklem wrote:
> >> 
> >> 
> >> 
> >> 
> >> Benjamin Kaduk wrote:
> >> 
> >> 
> >> 
> >> I think there are spare vfsops fields, so the MFC can be done in
> >> an
> >> ABI-compatible way. The new routine is for optional functionality,
> >> so it
> >> seems fine.
> >> 
> >> There are spares vfs ops in 10/current, but not in stable/9. An
> >> MFC
> >> will
> >> result in a VFS ABI change. (Since 10.0 hasn't been released yet,
> >> I
> >> didn't
> >> use one of the recently added spares.)
> >> 
> >> Oh, right, I was looking at 10/current.
> >> 
> >> Unless there are pressing calls for the feature in the stable
> >> branches, it's probably best to hold off on the MFC, then. OpenAFS
> >> has encountered a few KBI incompatibilities over the years (mostly
> >> in the networking bits, if I remember correctly), and we can deal
> >> in
> >> the future, but not having to is nice.
> >> 
> >> 
> >> 
> >> 
> >> Hello Guys,
> >> 
> >> 
> >> Is it possible to have it on 9-STABLE?
> >> I tried to port the changes of revision 255136 made by rmacklem@
> >> to a
> >> 9.1-RELEASE but the bug is still there.
> >> 
> >> 
> >> Any change to make it works on 9.1, 9.2 or 9-STABLE?
> >> 
> >> The patch attached is based on 9.1-RELEASE.
> > The patch looks ok at a glance. Note that it can take
> > up to 2-3minutes for a forced dismount to complete,
> > depending on where the threads are waiting.
> > 
> > If the mount is still there 5minutes after doing
> > "umount -f", do a "ps axhl" and post the output
> > of that to me. It may be getting stuck somewhere
> > else than where I've seen during testing.
> 
> Hello Rick,
> 
> Thanks by the prompt reply, I'm gonna make more tests tomorrow, and
> give you the output if necessary!
> 
> However, is there any way to improve this time to force the umount?
> 
To be honest, I have enough trouble making it work at all.
I think the long delays happen when the kernel RPC is stuck
somewhere in the TCP code waiting for timers to go off.
(The "umount -f" forces a soclose() and then the RPCs stuck
 in sosend() and friends will eventually fail.)

rick

> 
> 
> > 
> > rick
> > 
> >> 
> >> Best Regards, --
> >> Marcelo Araujo
> >> araujo at FreeBSD.org
>