svn commit: r330451 - in stable/11/sys: dev/iwm dev/otus dev/usb/wlan net80211

[Warner Losh]

On Fri, Mar 9, 2018 at 9:09 AM, Rodney W. Grimes <
freebsd at pdx.rh.cn85.dnsmgr.net> wrote:

> > On Fri, Mar 9, 2018 at 9:28 AM, Rodney W. Grimes
> > <freebsd at pdx.rh.cn85.dnsmgr.net> wrote:
> > >> On Wed, Mar 07, 2018 at 02:53:49PM -0800, Eitan Adler wrote:
> > >> > On 7 March 2018 at 09:37, John Baldwin <jhb at freebsd.org> wrote:
> > >> > > ...
> > >> > > I suspect many of these changes for iwm, etc. are all intertwined
> > >> > > so I'm not sure if you can leave out individual ones.
> > >> >
> > >> > Possibly. I do have iwm working on my laptop though. I also know of
> > >> > one open PR assigned to me w.r.t. a model I don't own. I'll be
> > >> > addressing it some time this week.
> > >>
> > >> I often have mixed feelings when I see lots of similar changes (i.e.
> > >> that make up for better hardware support, esp. on a laptop) MFCed.
> > >> I'd rather see laptop users run -CURRENT and leave -STABLE branches
> > >> for very conservative (server?) users who can't/don't want to afford
> > >> the risks of running -CURRENT or require ABI stability in a really
> > >> long run, rather than binge-merging things. :-)
> > >>
> > >> By default it should be -CURRENT all over; it's a very good thing
> > >> that we as a Project ourselves are doing this as part of our own
> > >> dogfood eating strategy.
> > >
> > > As a data point just last night a person came into #freebsd irc
> > > channel with a none working wireless nic on a desktop, he was
> > > attemtping to use a Realteak RTL8192EU on 11.1-RELEASE.
> > >
> > > Someone had already told him to try urtwn driver, which in
> > > -current is the right driver and supports this device.
> > >
> > > It did not work for him.  This is about when I came in to the
> > > discussion, and helped to confirm that -current did infact
> > > have support for this device, and that this supported had
> > > existed in -current for 16 months, and that this support
> > > would not be a simple grab a couple driver files and build
> > > it on 11.1.
> > >
> > > The commit to add this support involves 46 files, which 18
> > > of are new files.
> > >
> > > My feelings are if this driver has been in -current for
> > > 16 months why is it NOT in -stable yet?  I know part of
> > > the answer "its not a simple merge".  But as a project
> > > this is egg on our face.  We can merge a new very complicated
> > > change to a our boot code, within a few months (thank you
> > > kevans for that work), but we can not merge back a
> > > device driver in 16?
> > >
> >
> > I had the same questions- this exact same person had hopped over to
> > #freebsd-wifi and I had walked through this same process, identifying
> > it as "not MFC-able" at this point because so many commits having been
> > left un-merged prior to it. I've already recently gone through the fun
> > of catching up on one and a half years worth of unmerged work in
> > stand, I'm not really prepared to do it again quite yet.
>
> But but but.. you did it so well the first time!!! :-)
> I can fully appreciate that you do not desire to do another
> massive merge.
>
> > It felt pretty bad having to tell him that his only option here was to
> > either hop on -CURRENT + rtwn(4) or grab a stable/11 supported NIC-
> > especially since stable/11 is still supposed to be supported for
> > another three (3!) years, and I had to leave before I could help him
> > walk through getting it setup on -CURRENT properly.
>
> When I left the #freebsd channel he was downloading, and others
> there are capable of helping him get up and running.
>
> One thing that did come up while discussing some of the issues
> with merging head to stable was it might be usefull if we add
> yet another marker line:
>         Changes ABI/KABI:       yes
> to the commit messages, at least if marked you can be pretty
> sure that merging is going to involve additional work, lack
> of this mark would not guarantee you didnt have an ABI issue
> as they are easy to overlook and you would still need to
> look for those types of problems.
>

Given our past experience, it is often the case that people make changes
and aren't aware it's an ABI change, or forget that it is an ABI change
once testing is done. Or not realize a change in place X affects Y that
affects Z that makes it an ABI change. We've had several changes to CAM
that we had to redo because people didn't realize they were breaking
something. And at least one of them I approved knowing what changed, but
not groking the implications of the change.


> I have been told by someone there are some tools that can
> mechanically look for ABI changes.
>

This is the only way forward that may work, but it's much harder in the
kernel where the interfaces aren't well enumerated.

But we have to look at why so much is desirable to MFC. It's been a long
time since a major release and a lot has gone into 12 on the assumption the
release would be maybe 30 months from 11.0. The current schedule doesn't
have it going out for quite some time. Another way forward is to pull in
the 12.0R release. If there were an impending major release, it wouldn't be
so bad. However, the agreed to release schedule a few years ago is
slipping, which is putting pressure to MFC things, which is why we're
talking about this. 11.0 was released in October 2016. That means we should
be entering a code slush this summer for 12.0 towards the end of the year,
but so far the best guess is a vague hand wave of sometime in early 2019,
maybe. That's what's creating the pressure to MFC: we have at least one
more year of 11 being the latest release.

Warner