Libor Michalek wrote:
I see. This appears to come from a difference between the event reporting
model used by the kernel CM versus the usermode CM (callback versus
calldown).
Do you block the destroy on a lock while a callback for that cm_id
is active? I wouldn't say that the difference is attributed to callback
vs. calldown, in both cases it's a matter of serializing the destroy
with the event.
Yes - the destroy call in the kernel blocks while there's a callback in
progress. After destroy returns, the CM guarantees that no additional
callbacks will be received by the user.
The blocking destroy call was the reason for letting the user destroy the
cm_id by returning a non-zero value from the callback.
Maybe there's a way to assist the user here. Can we report a
destruction event, or require a second call to indicate that an event has
been processed?
A destruction event could work, but with some limits which might make
it impracticle. The user would have to be really carefull not to do
_anything_ with the object after calling destroy, and only cleanup in
the same thread that is used to get the destroy completion event. The
destroy completion event could be retreived and processed before the
original destroy call returns. Also, the user would need to make sure
that they are getting events in a _single_ thread, since multiple event
get threads could pose the same problem as before.
I agree that the destroy event could occur before destroy returns, so the
user would need to be careful there. Arlin mentioned that there's a put
event call that needs to be invoked after getting an event. If so, then the
CM can track the number of outstanding events that are in process. It could
then either delay the destroy call while an event is outstanding, or delay
reporting the destroy event until all events have been processed. This
should handle the multi-thread issues.
It may also make sense to have the uCM serialize all events to a single
cm_id anyway. I ended up doing this in the kernel, which simplified the
application's event handling. Otherwise, the events can end up being
processed out of order. E.g. a REJ is reported, followed by a MRA to a REQ.
We could build the serialization table for the API consumer, have
all cm_id calls and events go through a level of indirection in a
table locked against multiple threads. This was the way we ended up
doing it in our old code for the userCM that we used for uDAPL. I
had left this out since it seems reasonable that not all apps would
want/need this guarantee from the API, and that they could implement
it themselves if they did want it... I could be wrong.
I need to spend more time looking at the uCM API/implementation to see if
there's a way to help protect against reporting/processing events.
- Sean
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
