nfs no longer reconnects for udp sockets

Hey guys, someone was having issues with NFS mounts and
I happened to notice that it appears that the
"reconnect if socket went south" semantics I added
a few
years ago were basically disabled by the nfs optimizations
added for "recv side processing".

The problem is as such:

You have an NFS mount on UDP.
Somehow the route goes bad.
The UDP socket is now "broken" as the route will
remain
hosed forever.  This is particularly bad when an interface
flaps and loses its IP address as the UDP socket's route is
then set to nul or loopback or something and never gets
fixed.
Your nfs mount goes dead even if the routing issues is
resolved (interface brought back up).

Please see attached patch.

Easy way to reproduce problem:

mount an nfs filesystem using UDP.
ifconfig interface down
try to access mount
ifconfig interface up
mount should still be dead.

Please review.

-- 
- Alfred Perlstein

_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

The patch looks good to me, but I hope someone will test to
make sure tcp
reconnects still work too.  I will not be able to test any
time soon.

In nfs version 4, udp mounts are no longer supported.
_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

Ah, Very nice, I have been hitting this issue for a while
now but had  
not completely realized what was causing it.
Will have to set aside some time to test this but the patch
looks good  
(it's 50% comments so I can understand what you are doing
anyway ).

"Please don't break this, it's annoying."
Indeed, this bug is quite annoying.

Thanks for your work on this (and, obviously, the work you
did on the  
reconnect support prior).

-Adam


On Mar 14, 2008, at 6:01 AM, Alfred Perlstein wrote:

> Hey guys, someone was having issues with NFS mounts
and
> I happened to notice that it appears that the
> "reconnect if socket went south" semantics I
added a few
> years ago were basically disabled by the nfs
optimizations
> added for "recv side processing".
>
> The problem is as such:
>
> You have an NFS mount on UDP.
> Somehow the route goes bad.
> The UDP socket is now "broken" as the route
will remain
> hosed forever.  This is particularly bad when an
interface
> flaps and loses its IP address as the UDP socket's
route is
> then set to nul or loopback or something and never gets
fixed.
> Your nfs mount goes dead even if the routing issues is
> resolved (interface brought back up).
>
> Please see attached patch.
>
> Easy way to reproduce problem:
>
> mount an nfs filesystem using UDP.
> ifconfig interface down
> try to access mount
> ifconfig interface up
> mount should still be dead.
>
> Please review.
>
> -- 
> - Alfred Perlstein
> < 
>
nfs_reconnect_udp.diff>__________________________________
_____________
> freebsd-fsfreebsd.org mailing list
> 
http://lists.freebsd.org/mailman/listinfo/freebsd-fs
> To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

* Adam Jacob Muller <listsadam.gs> [080316 18:43]
wrote:
> Ah, Very nice, I have been hitting this issue for a
while now but had  
> not completely realized what was causing it.
> Will have to set aside some time to test this but the
patch looks good  
> (it's 50% comments so I can understand what you are
doing anyway ).
> 
> "Please don't break this, it's annoying."
> Indeed, this bug is quite annoying.
> 
> Thanks for your work on this (and, obviously, the work
you did on the  
> reconnect support prior).

Can you give me an ETA on testing?  If it's soon I can hold
off
committing for your feedback and that would help much.

Don't forget you need to use UDP. 

-Alfred

> 
> -Adam
> 
> 
> On Mar 14, 2008, at 6:01 AM, Alfred Perlstein wrote:
> 
> >Hey guys, someone was having issues with NFS mounts
and
> >I happened to notice that it appears that the
> >"reconnect if socket went south"
semantics I added a few
> >years ago were basically disabled by the nfs
optimizations
> >added for "recv side processing".
> >
> >The problem is as such:
> >
> >You have an NFS mount on UDP.
> >Somehow the route goes bad.
> >The UDP socket is now "broken" as the
route will remain
> >hosed forever.  This is particularly bad when an
interface
> >flaps and loses its IP address as the UDP socket's
route is
> >then set to nul or loopback or something and never
gets fixed.
> >Your nfs mount goes dead even if the routing issues
is
> >resolved (interface brought back up).
> >
> >Please see attached patch.
> >
> >Easy way to reproduce problem:
> >
> >mount an nfs filesystem using UDP.
> >ifconfig interface down
> >try to access mount
> >ifconfig interface up
> >mount should still be dead.
> >
> >Please review.
> >
> >-- 
> >- Alfred Perlstein
> >< 
>
>nfs_reconnect_udp.diff>______________________________
_________________
> >freebsd-fsfreebsd.org mailing list
> >
http://lists.freebsd.org/mailman/listinfo/freebsd-fs
> >To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

-- 
- Alfred Perlstein
_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

Alfred Perlstein wrote:
> Hey guys, someone was having issues with NFS mounts
and
> I happened to notice that it appears that the
> "reconnect if socket went south" semantics I
added a few
> years ago were basically disabled by the nfs
optimizations
> added for "recv side processing".
>
> The problem is as such:
>
> You have an NFS mount on UDP.
> Somehow the route goes bad.
> The UDP socket is now "broken" as the route
will remain
> hosed forever.  This is particularly bad when an
interface
> flaps and loses its IP address as the UDP socket's
route is
> then set to nul or loopback or something and never gets
fixed.
> Your nfs mount goes dead even if the routing issues is
> resolved (interface brought back up).
>
> Please see attached patch.
>
> Easy way to reproduce problem:
>
> mount an nfs filesystem using UDP.
> ifconfig interface down
> try to access mount
> ifconfig interface up
> mount should still be dead.
>
> Please review.
>   

This patch doesn't apply against HEAD.  There are also
gratuitous style 
changes.   Looks fine, but please re-spin (and test against
HEAD).

    Sam

_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"