Re: linux-next: build failure after merge of the scsi tree

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sat, 2016-12-24 at 12:46 +0100, Thomas Gleixner wrote:
> On Sat, 24 Dec 2016, Stephen Rothwell wrote:
> > On Thu, 22 Dec 2016 16:56:34 -0800 James Bottomley <
> > James.Bottomley@HansenPartnership.com> wrote:
> > > 
> > > On Fri, 2016-12-23 at 11:45 +1100, Stephen Rothwell wrote:
> > > > Hi James,
> > > > 
> > > > After merging the scsi tree, today's linux-next build (x86_64
> > > > allmodconfig) failed like this:
> > > > 
> > > > drivers/scsi/qedi/qedi_main.c: In function 'qedi_init':
> > > > drivers/scsi/qedi/qedi_main.c:2073:2: error: implicit
> > > > declaration of
> > > > function 'register_hotcpu_notifier' [-Werror=implicit-function
> > > > -declaration]
> > > >   register_hotcpu_notifier(&qedi_cpu_notifier);
> > > >   ^
> > > > drivers/scsi/qedi/qedi_main.c: In function 'qedi_cleanup':
> > > > drivers/scsi/qedi/qedi_main.c:2113:2: error: implicit
> > > > declaration of
> > > > function 'unregister_hotcpu_notifier' [-Werror=implicit
> > > > -function
> > > > -declaration]
> > > >   unregister_hotcpu_notifier(&qedi_cpu_notifier);
> > > >   ^
> > > > 
> > > > Caused by commit
> > > > 
> > > >   ace7f46ba5fd ("scsi: qedi: Add QLogic FastLinQ offload iSCSI
> > > > driver
> > > > framework.")
> > > > 
> > > > Interacting with commit
> > > > 
> > > >   8e38db753d95 ("cpu/hotplug: Remove obsolete cpu hotplug
> > > > register/unregister functions")
> > > > 
> > > > from the tip tree.  
> > > 
> > > Well, that's a bit of a problem given that the SCSI tree has a 
> > > pending pull request including this driver.
> > > 
> > > Thomas, can you do another fixup in your tip tree like you did 
> > > for the two BNX2* drivers?
> > 
> > OK, so this is now a problem for the tip tree merge since James 
> > tree has been pulled by Linus.
> 
> Sure, because SSCI people merge broken crap and I can wipe up the 
> mess they create.

Well, the driver was reviewed, but I don't think biting the heads of
the reviewers would help.

> Dammit, SCSI folks knew for a long time that the old interface goes
> away, but just waving crap through and let other people deal with the
> outcome is way simpler.

So this is the problem: I don't think this was actually communicated
this in a meaningful way; I think that's because we don't really have a
functional process for deprecating stuff.  Saying "this is broken or
deprecated" doesn't work because people don't necessarily see it and
even if they do, they often forget.  Putting it in some document
doesn't work either because we can't agree which one (and even if we
could, I bet the reviewers won't always consult it).

So here's the thing: if you want me to notice that a driver is using a
deprecated API, the API must be *marked* as deprecated so that I get a
build warning to investigate.  Absent that, I'm going to assume the
reviewers knew what they were talking about ... and likely a build
warning is the only way they'd know as well.

> And of course that hotplug code in this new driver is broken as hell. 
> It leaks notifiers in cases of errors and is racy against cpu 
> hotplug. The proper thing would be to mark this trainwreck broken and 
> be done with it.
> 
> I'm seriously pissed off as I now have to rebase my stuff and cleanup 
> that sad affair in order to not break bisects completely.

Well, I'm sorry, but it's a timing thing.  I could have removed this
driver from the pull it if I'd got the next failure before it was in
-flight.  I did incubate this final pull tree in next for a week which
was supposed to catch all the issues like this.

James