RE: [PATCH v3 05/11] xprtrdma: Do not wait if ib_post_send() fails

["Steve Wise" <swise@opengridcomputing.com>]

> >>>>>>>> Moving the QP into error state right after with rdma_disconnect
> >>>>>>>> you are not sure that none of the subset of the invalidations
> >>>>>>>> that _were_ posted completed and you get the corresponding MRs
> >>>>>>>> in a bogus state...
> >>>>>>>
> >>>>>>> Moving the QP to error state and then draining the CQs means
> >>>>>>> that all LOCAL_INV WRs that managed to get posted will get
> >>>>>>> completed or flushed. That's already handled today.
> >>>>>>>
> >>>>>>> It's the WRs that didn't get posted that I'm worried about
> >>>>>>> in this patch.
> >>>>>>>
> >>>>>>> Are there RDMA consumers in the kernel that use that third
> >>>>>>> argument to recover when LOCAL_INV WRs cannot be posted?
> >>>>>>
> >>>>>> None :)
> >>>>>>
> >>>>>>>>> I suppose I could reset these MRs instead (that is,
> >>>>>>>>> pass them to ib_dereg_mr).
> >>>>>>>>
> >>>>>>>> Or, just wait for a completion for those that were posted
> >>>>>>>> and then all the MRs are in a consistent state.
> >>>>>>>
> >>>>>>> When a LOCAL_INV completes with IB_WC_SUCCESS, the associated
> >>>>>>> MR is in a known state (ie, invalid).
> >>>>>>>
> >>>>>>> The WRs that flush mean the associated MRs are not in a known
> >>>>>>> state. Sometimes the MR state is different than the hardware
> >>>>>>> state, for example. Trying to do anything with one of these
> >>>>>>> inconsistent MRs results in IB_WC_BIND_MW_ERR until the thing
> >>>>>>> is deregistered.
> >>>>>>
> >>>>>> Correct.
> >>>>>>
> >>>>>
> >>>>> It is legal to invalidate an MR that is not in the valid state.  So you
> >>> don't
> >>>>> have to deregister it, you can assume it is valid and post another LINV
> > WR.
> >>>>
> >>>> I've tried that. Once the MR is inconsistent, even LOCAL_INV
> >>>> does not work.
> >>>>
> >>>
> >>> Maybe IB Verbs don't mandate that invalidating an invalid MR must be
> > allowed?
> >>> (looking at the verbs spec now).
> >>
> >
> > IB Verbs doesn't have specify this requirement.  iW verbs does.  So
transport
> > independent applications cannot rely on it.  So ib_dereg_mr() seems to be
the
> > only thing you can do.
> >
> >> If the MR is truly invalid, then there is no issue, and
> >> the second LOCAL_INV completes successfully.
> >>
> >> The problem is after a flushed LOCAL_INV, the MR state
> >> sometimes does not match the hardware state. The MR is
> >> neither registered or invalid.
> >>
> >
> > There is a difference, at least with iWARP devices, between the MR state:
VALID
> > vs INVALID, and if the MR is allocated or not.
> >
> >> A flushed LOCAL_INV tells you nothing more than that the
> >> LOCAL_INV didn't complete. The MR state at that point is
> >> unknown.
> >>
> >
> > With respect to iWARP and cxgb4: when you allocate a fastreg MR, HW has an
> entry
> > for that MR and it is marked "allocated".  The MR record in HW also has a
state:
> > VALID or INVALID.  While the MR is "allocated" you can post WRs to
invalidate it
> > which changes the state to INVALID, or fast-register memory which makes it
> > VALID.  Regardless of what happens on any given QP, the MR remains
"allocated"
> > until you call ib_dereg_mr().  So at least for cxgb4, you could in fact just
> > post another LINV to get it back to a known state that allows subsequent
> > fast-reg WRs.
> >
> > Perhaps IB devices don't work this way.
> >
> > What error did you get when you tried just doing an LINV after a flush?
> 
> With CX-2 and CX-3, after a flushed LOCAL_INV, trying either
> a FASTREG or LOCAL_INV on that MR can sometimes complete with
> IB_WC_MW_BIND_ERR.


I wonder if you post a FASREG+LINV+LINV if you'd get the same failure?  IE
invalidate the same rkey twice.  Just as an experiment...