[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [rddp] Freeing QN=1 State



 

> -----Original Message-----
> From: rddp-bounces at ietf.org [mailto:rddp-bounces at ietf.org] On 
> Behalf Of Michael C. Cambria
> Sent: Friday, December 10, 2004 9:50 AM
> To: rddp at ietf.org
> Subject: [rddp] Freeing QN=1 State
> 
> 
> When does draft-ietf-rddp-rdmap-02.txt consider an inbound 
> RDMA Read Request Message finished such that the state 
> associated with QN=1 (the
> IRRQ) for the RDMA Read Request can be freed?
> 
> Put another way, if IRRQ is 1 (for simplicity), when is it 
> legal for a new RDMA Read Request to be received?  Is it:
> 
> a) Once the RDMA Read Response Message has been given to DDP?
> b) Once the RDMA Read Response Message has been "acked" by 
> the data sink?
> c) up to the implementation?
> d) specified somewhere that I just missed?
> 
The other side must not consider its RDMA Read Request complete
until it receives the RDMA Read Response. When the Data Source
considers it complete is a matter of implementation.

But conceptually, you should imagine that the RDMA Read engine
is simply an alternate ULP:

- it pre-posts receive buffers to accept RDMA Read Requests. It
  posts enough buffers for the number of credits it has been told
  were issued.
- it posts send requests to transmit RDMA Read Responses.

So basically, it has completed an RDMA Read Request when its
transmit requests complete. Just as with RDMA Write Requests
Submitted by the main ULP, completion is implementation dependent.
It indicates to the ULP that the source buffers are no longer
required. That may be because it has been LLP acked, or it may
be because the LLP has the buffers fully cached on its own.

Of course ultimately, its when the "RDMA Read Engine" has freed
up enough resources to repost its receive request. That depends
on how it tracks pending requests. If the other side does not 
post excess RDMA Read requests it will never have to know the
details of how the peer RDMA Read Engine is implemented.


--
Caitlin Bestler
Director Software Architecture
Siliquent Technologies
caitlinb at siliquent.com
 

_______________________________________________
rddp mailing list
rddp at ietf.org
https://www1.ietf.org/mailman/listinfo/rddp