IP More Fragements bit problem.

IP More Fragements bit problem.

[Vincent Roqueta]

Hello all,

I am working on NFSv4 and I found an IP problem while doing interoperability 
testing with AIX.

The More Fragements (MF) bits is set on the last packet sent in a list of 
packets that sould be reassembled
 
For example :

If we get the A, B, C, D packets that may be reassembled into ABCD we sould 
have 

A + MF
B + MF
C + MF
D

On kernel 2.6.10rc2 and laters (last tested is 2.6.11-rc3) we have: 

A + MF
B + MF
C + MF
D + MF

That cause AIX waiting for the last packet with out sending ack
So Linux suppose  AIX didn't receive all, and re send all packets.
So AIX is receiving that: 
A + MF
B + MF
C + MF
D + MF
*DELAY *
A + MF
B + MF
C + MF
D + MF
*DELAY*
A + MF
B + MF
C + MF
D + MF
*DELAY*
...

After a while AIX destroy first fragments because of the IP fragements life 
time. Trond Myklebust said me you can do anything for that?

Cher,
Vincent ROQUETA

Re: IP More Fragements bit problem.

[Andi Kleen]

Vincent Roqueta <vincent.roqueta@ext.bull.net> writes:
> D + MF
> *DELAY*
> A + MF
> B + MF
> C + MF
> D + MF
> *DELAY*
> ...
>
> After a while AIX destroy first fragments because of the IP fragements life 
> time. Trond Myklebust said me you can do anything for that?

Are you sure? I tested 2.6.10rc3 and it works correctly for 
me with ping. The algorithm in ip_fragment() looks good too
from visual inspection.

And ping uses the same code to fragment as NFS sunrpc
over UDP.

19:15:24.100934 averell > trent: (frag 64564:1480@1480+)
19:15:24.100938 averell > trent: (frag 64564:1480@2960+)
19:15:24.100943 averell > trent: (frag 64564:1480@4440+)
19:15:24.100947 averell > trent: (frag 64564:1480@5920+)
19:15:24.100951 averell > trent: (frag 64564:1480@7400+)
19:15:24.100957 averell > trent: (frag 64564:1128@8880)     <--- No MF.

Also why are you testing NFSv4 over UDP anyways? I thought
everybody was finally running it over TCP now.

-Andi

Re: IP More Fragements bit problem.

[Vincent Roqueta]

Le vendredi 11 Février 2005 19:24, vous avez écrit :
> Vincent Roqueta <vincent.roqueta@ext.bull.net> writes:
> > D + MF
> > *DELAY*
> > A + MF
> > B + MF
> > C + MF
> > D + MF
> > *DELAY*
> > ...
> >
> > After a while AIX destroy first fragments because of the IP fragements
> > life time. Trond Myklebust said me you can do anything for that?
>
> Are you sure? I tested 2.6.10rc3 and it works correctly for
> me with ping. The algorithm in ip_fragment() looks good too
> from visual inspection.
Yes, I am sure testing with NFSv3 over UDP.
I can send you traces seen by AIX.

> And ping uses the same code to fragment as NFS sunrpc
> over UDP.
Ok.
> 19:15:24.100934 averell > trent: (frag 64564:1480@1480+)
> 19:15:24.100938 averell > trent: (frag 64564:1480@2960+)
> 19:15:24.100943 averell > trent: (frag 64564:1480@4440+)
> 19:15:24.100947 averell > trent: (frag 64564:1480@5920+)
> 19:15:24.100951 averell > trent: (frag 64564:1480@7400+)
> 19:15:24.100957 averell > trent: (frag 64564:1128@8880)     <--- No MF.
I don't have that :(

> Also why are you testing NFSv4 over UDP anyways? I thought
> everybody was finally running it over TCP now.
I am not testing NFSv4 over UDP, I am doing NFSv3 tests over UDP to compare 
performances between NFSv4 and NFSv3, and some interoperablity testing on NFS 
(v3 and v4). That the last tests that does not work fine.

Vincent

Re: IP More Fragements bit problem.

[Nivedita Singhvi]

Vincent Roqueta wrote:

> On kernel 2.6.10rc2 and laters (last tested is 2.6.11-rc3) we have: 
> 
> A + MF
> B + MF
> C + MF
> D + MF
> 
> That cause AIX waiting for the last packet with out sending ack
> So Linux suppose  AIX didn't receive all, and re send all packets.
> So AIX is receiving that: 
> A + MF
> B + MF
> C + MF
> D + MF
> *DELAY *

Silly question, but are you absolutely sure that the original wasn't
actually ABCDE?

thanks,
Nivedita

Re: IP More Fragements bit problem.

[Vincent Roqueta]

Le vendredi 11 Février 2005 19:44, vous avez écrit :
> Vincent Roqueta wrote:
> > On kernel 2.6.10rc2 and laters (last tested is 2.6.11-rc3) we have:
> >
> > A + MF
> > B + MF
> > C + MF
> > D + MF
> >
> > That cause AIX waiting for the last packet with out sending ack
> > So Linux suppose  AIX didn't receive all, and re send all packets.
> > So AIX is receiving that:
> > A + MF
> > B + MF
> > C + MF
> > D + MF
> > *DELAY *
>
> Silly question, but are you absolutely sure that the original wasn't
> actually ABCDE?
No, we are not sure. It is possible we lack the last fragement witch is not 
marked with the MF bit.

Vincent ROQUETA

Re: IP More Fragements bit problem.

[Vincent Roqueta]

Hello,
> >
> > Silly question, but are you absolutely sure that the original wasn't
> > actually ABCDE?
>
> No, we are not sure. It is possible we lack the last fragement witch is not
> marked with the MF bit.

You are right. All fragments are not sent !!! And so the last fragment 
received is not the last fragment sent by the client. The MF is set and that 
is correct.

Sorry :/


Vincent