[openib-general] matching an RMPP NACK to send or receive?
Sean Hefty
mshefty at ichips.intel.com
Fri Jul 15 12:17:50 PDT 2005
Hal Rosenstock wrote:
>>I'm returning IB_WC_REM_ABORT_ERR and placing the actual RMPP status code in
>>the vendor_err field.
>
> What vendor errors are being defined ? I realize that is the only way to
> do this now but aren't they really standard rather than vendor errors ?
No vendor errors are being defined. I'm simply returning the received RMPP
Status in that field since it's not used in this case. We can rename
vendor_err to something else.
>>I believe that the following status values are possible for either a send or
>>receive:
>>
>>resources exhausted
>>bad RMPP type
>>illegal status
>>unspecified and class specific
>
> What's unspecified ? Is it reserved status in ABORT ?
Unspecified is status value 127 - "an error .. not covered by any other
status code". This is handled, but only aborts a send request if received.
> There are some others too. Does the OpenIB RMPP not generate them ? Will
> it pass them back up if it received an abort with these other statuses ?
I think that the other status codes occur only in response to a send OR to a
receive. So it's not that they're not generated, I just have a better idea
on how to handle them.
Total time too long - assuming sent by receiver to sender. This is handled.
Inconsistent last and payload length - sent by receiver to sender. This is
not transmitted currently, but should work fine if received. (I need to add
state information to transmit this.)
Inconsistent first and segment number - transmitted by receiver to sender.
This is handled.
Bad RMPP type - transmitted either way. This is checked for all messages,
but will only abort a send request if received.
New window last too small - transmitted by sender to receiver. This is handled.
Segment number too big - transmitted by sender to receiver. This is handled.
Illegal status - transmitted either way. This is checked for all messages,
but will only abort a send request if received.
Unsupported version - transmitted by receiver to sender. This is handled.
Too many retries - assuming sent by sender to receiver. This is not
handled. (I need to add a new call from the MAD code into the RMPP layer to
support this.)
- Sean
More information about the general
mailing list