Michael> Let me rephrase: after post send/receive to QP in error
Michael> state in mthca, a completion with error might never get
Michael> generated.
Won't a flush error be generated for every request posted to a QP in
the error state?
Michael> To fix mthca, we'd need to change QP state on completion
Michael> with error and on modify to error, and add actual code
Michael> where it now says
Michael> /* XXX check that state is OK to post send */ /* XXX
Michael> check that state is OK to post receive */
Michael> I guess the reason we never fixed this was because it did
Michael> not seem to actually hurt any real ULPs, and testing QP
Michael> state will affect fast path performance.
Michael> However, IB spec is quite explicit on this point, and
Michael> fixing a low level drivers seems a better approach than
Michael> adding work-arounds in ULPs.
Michael> Roland, what do you think?
Yes, that was something I just never got around to implementing. Of
course since transition to error state may be done asynchronously by
hardware, we still have the case where the consumer tries to post a
work request to a QP in error but the low-level driver still thinks
the QP is in RTS.
- R.
In-Reply-To: <[EMAIL PROTECTED]> (Michael S. Tsirkin's message of "Tue, 26 Sep
2006 17:24:00 +0300")
User-Agent: Gnus/5.1007 (Gnus v5.10.7) XEmacs/21.4.18 (linux)
Date: Tue, 26 Sep 2006 08:11:54 -0700
Message-ID: <[EMAIL PROTECTED]>
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
