On Tue, 2017-08-01 at 17:10 +0300, Leon Romanovsky wrote:
> On Tue, Aug 01, 2017 at 08:38:32AM -0500, Chien Tin Tung wrote:
> > On Tue, Aug 01, 2017 at 03:05:04PM +0300, Leon Romanovsky wrote:
> > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > >
> > > The commit cea05eadded0 ("IB/core: Add flow control to the
> > > portmapper netlink calls")
> > > changed netlink to be blocked for all RDMA clients. This
> > > workaround
> > > worked perfectly for portmapper, but is not correct for the whole
> > > NETLINK_RDMA family.
> >
> > Leon,
> >
> > I've already told you I'm opposed to the revert. There is a patch
> > that will work
> > with your usage of ibnl_unicast() but you chose to abandon that
> > discussion on June 29, 2017
> > (RDMA/core: Add wait/retry version of ibnl_unicast). Either you
> > step up with good sound
> > technical evidence to convince me that patch won't work for you or
> > you stop trying to
> > break existing functionality with this revert.
>
> Chien,
>
> You never explained us what exactly your original patch fixed and why
> it
> should be fixed in kernel and not in user space. I saw that our
> discussion wasn't useful and brought bad blood instead of good will,
> so I stepped out in waited for RDMA maintainer (Doug) step in.
>
> I never heard anything from Doug on the matter, so proceeded and
> resubmitted it.
What makes you think *I* want to be in the middle of a school yard
squabble like this?
Regardless, here's what I think:
1) The blocking/timeout functionality that Chien is using is in the
netlink core, not something he grafted onto netlink himself. Since
it's a core netlink feature, I see no reason he shouldn't be allowed to
use it.
2) We don't break API for user space programs. Reverting this could
have a significant deleterious influence on programs already in the
wild. While those programs should resync on a failure, clearly,
failure is much more frequent without this patch than with it, and as a
resync can be expensive, it is a judgment call whether it is preferable
to resync more frequently because of dropped packets, or utilize a
backlog queue to reduce the frequency of resyncs. Clearly, Intel is of
the position that the backlog queue is worth avoiding the resyncs.
3) Chien submitted a patch that would allow just the iwpmd to get the
behavior they find preferable, while allowing Leon's new code to get
the behavior Leon wants.
Given that #1 and #2 are true and that #3 provides a reasonable
solution, the revert is rejected.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: B826A3330E572FDD
Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
--
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
