[Skiboot] [PATCH] phb4: Avoid complete reset without finishing slot polling

[Russell Currey]

On Thu, 2017-09-28 at 14:53 +1000, Stewart Smith wrote:
> Russell Currey <ruscur at russell.cc> writes:
> > On Fri, 2017-09-15 at 17:34 +1000, Stewart Smith wrote:
> > > Russell Currey <ruscur at russell.cc> writes:
> > > > On Fri, 2017-09-15 at 17:02 +1000, Stewart Smith wrote:
> > > > > Russell Currey <ruscur at russell.cc> writes:
> > > > > > In rare circumstances, a complete reset can be requested on a slot
> > > > > > while
> > > > > > it is still going through the link polling process.  This results in
> > > > > > the
> > > > > > link never coming up.  In complete reset, check to see if the link
> > > > > > needs
> > > > > > to complete its state transitions, and if so poll to completion.
> > > > > 
> > > > > I was having a bit of a problem with this on my (DD1) zz. I naturally
> > > > > saved all the logs and spent a whole bunch of time debugging. No,
> > > > > wait, that would have been all useful or something.
> > > > > 
> > > > > Instead, I went "err... stopped in PCI in kernel, let's randomly pull
> > > > > this
> > > > > patch out to see if it fixes it"... and then it booted.
> > > > > 
> > > > > Any ideas? Want to borrow the system to tell me I just saw something
> > > > > random and rather that your code is perfect?
> > > > > 
> > > > 
> > > > Sounds good.  This patch has been tested quite a bit and seems to be
> > > > doing
> > > > more
> > > > fixing than breaking, maybe there's some weird case on DD1
> > > 
> > > or zz. or my zz. Or the random old FSP build that's on my zz. All the
> > > things that could be wrong with prerelease hardware could well be wrong
> > > here.
> > > 
> > 
> > Not sure if it's your issue or not, but there are some known problems with
> > this
> > patch so hold off on merging, I'll respin once I figure it out.
> > 
> > Or merge it anyway.  The overall level of broken probably won't change much.
> 
> Any further thoughts?
> 
I haven't had time to look back into it yet.  It still needs to be reworked.