On Fri, Jan 11, 2019 at 05:27:30PM -0500, Chuck Lever wrote:
>
>
> > On Jan 11, 2019, at 5:10 PM, Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
> >
> > On Fri, Jan 11, 2019 at 04:54:01PM -0500, Chuck Lever wrote:
> >>> On Jan 11, 2019, at 4:52 PM, Chuck Lever <chuck.lever@xxxxxxxxxx> wrote:
> >>>> So, I think we need your patch plus something like this.
> >>>>
> >>>> Chuck, maybe you could help me with the "XXX: Chuck:" parts?
> >>>
> >>> I haven't been following. Why do you think those are necessary?
> >
> > I'm worried something like this could happen:
> >
> > CPU 1 CPU 2
> > ----- -----
> >
> > set XPT_DATA dec xpt_nr_rqsts
> >
> > svc_xprt_enqueue svc_xprt_enqueue
> >
> > And both decide nothing should be done if neither sees the change that
> > the other made.
> >
> > Maybe I'm still missing some reason that couldn't happen.
> >
> > Even if it can happen, it's an unlikely race that will likely be fixed
> > when another event comes along a little later, which would explain why
> > we've never seen any reports.
> >
> >>> We've had set_bit and atomic_{inc,dec} in this code for ages,
> >>> and I've never noticed a problem.
> >>>
> >>> Rather than adding another CPU pipeline bubble in the RDMA code,
> >>> though, could you simply move the set_bit() call site inside the
> >>> critical sections?
> >>
> >> er, inside the preceding critical section. Just reverse the order
> >> of the spin_unlock and the set_bit.
> >
> > That'd do it, thanks!
>
> I can try that here and see if it results in a performance regression.
Thanks, I've got a version with a typo fixed at
git://linux-nfs.org/~bfields/linux.git nfsd-next
--b.
