[openib-general] (no subject)

[mark kowalski]

Thanks Caitlin,
     This answer helps explain what I'm seeing.


>From a strict DAT specification point of view:

>After a dat_ep_reset the DAT Provider MAY rotate the underlying
>transport resources (QP) that the EP is associated with in order
>to avoid time-wait states. But I haven't seen any implementations
>actually do that. Note that I stated "rotate", not "release",
>it is not acceptable to suddenly find that you no longer have
>the required resources when the Consumer attempts to establish
>a new connection with the same EP.

>Generally a QP is associated with an EP from beginning to end.
>The only way to be certain that the resource is freed is to
>free it, and you really should do a reset first. In no cases
>can disconnecting alone be expected to release transport
>layer resources, it merely changes their state