wireshark's RPC-over-RDMA dissector

wireshark's RPC-over-RDMA dissector

[Chuck Lever]

Hi-

Thanks to Yan Berman, for a couple of years now we've had a basic
RPC-over-RDMA dissector in wireshark that can be used with ibdump
captures. There have been some bugs noted, but no-one has had the
cycles to dig in and address.

Recently Tom Haynes helped me set up my own wireshark build so I
could help address some of the known issues.

http://git.linux-nfs.org/?p=cel/wireshark.git;a=shortlog;h=refs/heads/rpc-rdma-fixes

Posting here for review before I take the next steps to push these
to the wireshark community. Constructive critique and other
suggestions are welcome.

The fixes so far focus on dissecting transport headers correctly.
There continue to be significant open issues with the dissector:
	• There does not appear to be any support for dissecting
	  RPC-over-RDMA on iWARP or RoCE
	• The NFS dissector does not handle portions of the XDR
	  stream that were transmitted via RDMA Read/Write
	• RPC messages conveyed via RDMA_NOMSG are not recognized
	  or dissected
	• There is no association between RDMA Reads and Writes
	  and the RPC-over-RDMA message they go with
	• A CREQ / CREP pair are needed to identify which QP
	  numbers are used for RPC-over-RDMA traffic
	• With TCP, the dissector fully outdents the RPC and NFS
	  dissection results; but with RDMA, the dissector places
	  the results in the tree under the Infiniband header
	• Not enough error detection in the dissector

--
Chuck Lever

Re: wireshark's RPC-over-RDMA dissector

[Parav Pandit]

Hi Chuck,

Just FYI.
I have committed few fixes in wireshark trunk for RoCE and IB
dissectors which will in general benefit other ULPs as well (primary
for statefulness of ULPs) last month.

I have few patches pending in my sandbox in area of RoCE and for other
ULP that I will push in coming days, likely next week.
I am currently actively testing them and have made steady progress so far.

I will try to find sometime to review them next week.

Regards,
Parav Pandit


On Wed, Nov 9, 2016 at 9:35 PM, Chuck Lever <chuck.lever@oracle.com> wrote:
> Hi-
>
> Thanks to Yan Berman, for a couple of years now we've had a basic
> RPC-over-RDMA dissector in wireshark that can be used with ibdump
> captures. There have been some bugs noted, but no-one has had the
> cycles to dig in and address.
>
> Recently Tom Haynes helped me set up my own wireshark build so I
> could help address some of the known issues.
>
> http://git.linux-nfs.org/?p=3Dcel/wireshark.git;a=3Dshortlog;h=3Drefs/hea=
ds/rpc-rdma-fixes
>
> Posting here for review before I take the next steps to push these
> to the wireshark community. Constructive critique and other
> suggestions are welcome.
>
> The fixes so far focus on dissecting transport headers correctly.
> There continue to be significant open issues with the dissector:
>         =E2=80=A2 There does not appear to be any support for dissecting
>           RPC-over-RDMA on iWARP or RoCE
>         =E2=80=A2 The NFS dissector does not handle portions of the XDR
>           stream that were transmitted via RDMA Read/Write
>         =E2=80=A2 RPC messages conveyed via RDMA_NOMSG are not recognized
>           or dissected
>         =E2=80=A2 There is no association between RDMA Reads and Writes
>           and the RPC-over-RDMA message they go with
>         =E2=80=A2 A CREQ / CREP pair are needed to identify which QP
>           numbers are used for RPC-over-RDMA traffic
>         =E2=80=A2 With TCP, the dissector fully outdents the RPC and NFS
>           dissection results; but with RDMA, the dissector places
>           the results in the tree under the Infiniband header
>         =E2=80=A2 Not enough error detection in the dissector
>
> --
> Chuck Lever
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: wireshark's RPC-over-RDMA dissector

[Chuck Lever]

> On Nov 9, 2016, at 11:41 AM, Parav Pandit <pandit.parav@gmail.com> wrote:
> 
> Hi Chuck,
> 
> Just FYI.
> I have committed few fixes in wireshark trunk for RoCE and IB
> dissectors which will in general benefit other ULPs as well (primary
> for statefulness of ULPs) last month.
> 
> I have few patches pending in my sandbox in area of RoCE and for other
> ULP that I will push in coming days, likely next week.
> I am currently actively testing them and have made steady progress so far.
> 
> I will try to find sometime to review them next week.
> 
> Regards,
> Parav Pandit

Stateful dissection might be helpful for associating RPC-over-RDMA
headers with RDMA Reads and Writes on the same QP. Thanks Parav,
I'll watch for your work.


> On Wed, Nov 9, 2016 at 9:35 PM, Chuck Lever <chuck.lever@oracle.com> wrote:
>> Hi-
>> 
>> Thanks to Yan Berman, for a couple of years now we've had a basic
>> RPC-over-RDMA dissector in wireshark that can be used with ibdump
>> captures. There have been some bugs noted, but no-one has had the
>> cycles to dig in and address.
>> 
>> Recently Tom Haynes helped me set up my own wireshark build so I
>> could help address some of the known issues.
>> 
>> http://git.linux-nfs.org/?p=cel/wireshark.git;a=shortlog;h=refs/heads/rpc-rdma-fixes
>> 
>> Posting here for review before I take the next steps to push these
>> to the wireshark community. Constructive critique and other
>> suggestions are welcome.
>> 
>> The fixes so far focus on dissecting transport headers correctly.
>> There continue to be significant open issues with the dissector:
>>        • There does not appear to be any support for dissecting
>>          RPC-over-RDMA on iWARP or RoCE
>>        • The NFS dissector does not handle portions of the XDR
>>          stream that were transmitted via RDMA Read/Write
>>        • RPC messages conveyed via RDMA_NOMSG are not recognized
>>          or dissected
>>        • There is no association between RDMA Reads and Writes
>>          and the RPC-over-RDMA message they go with
>>        • A CREQ / CREP pair are needed to identify which QP
>>          numbers are used for RPC-over-RDMA traffic
>>        • With TCP, the dissector fully outdents the RPC and NFS
>>          dissection results; but with RDMA, the dissector places
>>          the results in the tree under the Infiniband header
>>        • Not enough error detection in the dissector
>> 
>> --
>> Chuck Lever


--
Chuck Lever