> >The rdma provider must ensure that event upcalls are serialized per
> >object.
> >Is this being done somewhere? See "Callbacks" in
> >Documentation/infiniband/core_locking.txt.
> >
>
> That says, if multiple QP's etc. contribute to a CQ, only one
> CQ event handler is allowed to run? That's not there, I would have
> to make a lock around...
Correct. Basically, for a given CQ object, the driver can only have a single outstanding call to the ULP's event handler for that CQ at any point in time.
...
> >> + /*
> >> + * Try to acquire QP state lock. Must be non-blocking
> >> + * to accommodate kernel clients needs.
> >> + */
> >> + if (!down_read_trylock(&qp->state_lock)) {
> >> + *bad_wr = wr;
> >> + return -ENOTCONN;
> >> + }
> >> +
> >
> >Under what conditions does down_read_trylock() return 0? This seems
> >like a
> >kernel ULP that has multiple threads posting to the sq might get an
> >error
> >due to lock contention?
>
> This lock is only taken if the QP is to be transitioned into
> another state. If we are here, we assume to be in RTS state.
> If the lock is not available, someone else just moves the QP out
> of RTS and we should not do much here.
>
Ah, I see.
> We do trylock() since we cannot sleep in some of our contexts...
>
Right.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
