Re: [PATCH] RDMA/rxe: Fix a race condition related to the QP error state

[Leon Romanovsky <leon@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1447 bytes --]

On Thu, Jan 11, 2018 at 07:07:06PM +0000, Bart Van Assche wrote:
> On Thu, 2018-01-11 at 21:00 +0200, Leon Romanovsky wrote:
> > On Thu, Jan 11, 2018 at 04:02:33PM +0000, Bart Van Assche wrote:
> > > On Thu, 2018-01-11 at 08:22 +0200, Leon Romanovsky wrote:
> > > > The proposed patch definitely decreases the chance of races, but it is not fixing them.
> > > > There is a chance to have change in qp state immediately after your "if ..." check.
> > >
> > > Hello Leon,
> > >
> > > Please have a look at rxe_qp_error() and you will see that the patch I posted
> > > is a proper fix. In the scenario you described rxe_qp_error() will trigger a
> > > run of rxe_completer().
> >
> > Bart,
> >
> > What am I missing?
> >
> > CPU1							CPU2
> > 							if (unlikely....
> > 						<---
> > /* move the qp to the error state */
> > void rxe_qp_error(struct rxe_qp *qp)
> > {
> >         qp->req.state = QP_STATE_ERROR;
> >         qp->resp.state = QP_STATE_ERROR;
> >         qp->attr.qp_state = IB_QPS_ERR;
> > 						--->
> > 							rxe_run_task(&qp->req.task, must_sched);
> >
> >
> >
> > It is more or less the same as without "if (unlikely..."
>
> Hello Leon,
>
> In the above the part of rxe_qp_error() that I was referring to in my e-mail
> is missing:
>
> 	if (qp_type(qp) == IB_QPT_RC)
> 		rxe_run_task(&qp->comp.task, 1);


But it is exactly where race exists, as long QP isn't protected, it can
switch CPUs and create race.

Thanks

>
> Bart.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]