SRP Logout Ack? (was RE: comments on T10/01-150r0)
cris.simpson at intel.com
Thu May 10 15:17:21 PDT 2001
* From the T10 Reflector (t10 at t10.org), posted by:
* "Simpson, Cris" <cris.simpson at intel.com>
>> The group discussed whether a CM disconnect can destroy the
>> queue pair before the application has processed all outstanding
>> IOs on that queue pair. If so, a Logout Acknowledge IU would
>> be needed to ensure clean disconnections.
>> Cris Simpson will research the issue.
Once my brain got back to PDT, it didn't need much. Once the
Completion Queue Entry (CQE) is on the Completion Queue, it's
no longer dependent on the Rcv Queue, which may be blown away at
will. The data will have been delivered into memory, and the
completion status will be in the CQE. The sender will know
that the SRP_LOGOUT has been delivered when it receives the
No change needed.
> I don't see how this is an issue. A queue pair exists until
> the consumer (i.e. the application) issues a Destroy Queue Pair
> verb (v1 page 375, 446 and 469). How the CM tells the consumer
> that a connection has disappeared is on of the "internal interfaces"
> that "are outside the scope of the Infiniband Architecture"
> (v1 page 516). But the CM can't delete queue pairs or any
> other resources out from under the consumer with first
> synchronizing with it.
Other models are possible. Remember that each OS defines its own
resource management model and API.
Cris Simpson 503.712.4333
Intel EPG/FCD Architecture Hillsboro, OR
PGP Fingerprint:0DA0 418E A27B 0B76 5F02 3DD4 0546 6D13 F88A 1E60
* For T10 Reflector information, send a message with
* 'info t10' (no quotes) in the message body to majordomo at t10.org
More information about the T10