[ofa-general] [PATCH RFC v3 1/2] RDMA/Core:MEM_MGT_EXTENSIONS support
Steve Wise
swise at opengridcomputing.com
Tue May 27 12:59:16 PDT 2008
Talpey, Thomas wrote:
> At 02:40 PM 5/27/2008, Steve Wise wrote:
>
>> Talpey, Thomas wrote:
>>
>>> At 12:58 PM 5/27/2008, Felix Marti wrote:
>>>
>>>
>>>> RDMA Read with Local Invalidate does not affect the wire. The 'must
>>>> invalidate' state is kept in the RNIC that issues the RDMA Read
>>>> Request...
>>>>
>>>>
>>> Aha, okay that was not clear to me. What information does the RNIC use
>>> to line up the arrival of the RDMA Read response with the "must invalidate"
>>> state?
>>>
>> The rnic already tracks outstanding read requests. It now also will
>> track the local stag to invalidate when the read completes.
>>
>
> Ah - okay, so the stag that actually gets invalidated was provided with
> the RDMA Read request posting, and is not necessarily the stag that
> arrived in the peer's RDMA Read response. That helps.
>
> What happens if the upper layer gives up and invalidates the stag itself,
> and the peer's RDMA Read response arrives later? Nothing bad, I assume,
> and the peer's response is denied?
>
>
It behaves just like any other tagged message arriving and the target
stag is invalid. The connection is torn down via an RDMAP TERMINATE...
>>> Also, how does the RNIC signal whether the invalidation actually
>>> occurred, so the upper layer can defend itself from attack?
>>>
>>>
>>>
>> The stag is guaranteed to be in the invalid state by the time the app
>> reaps the read-inv-local work completion...
>>
>
> Ok, given my correct understanding of the source of the stag above.
>
> Tom.
>
More information about the general
mailing list