[PATCH v2 0/2] Fix console probe delay when stdout-path isn't set

[Andy Shevchenko]

On Mon, Sep 26, 2022 at 01:25:05PM -0500, Rob Herring wrote:
> On Mon, Sep 19, 2022 at 5:56 PM Olof Johansson <olof at lixom.net> wrote:
> >
> > On Mon, Sep 19, 2022 at 1:44 AM Greg Kroah-Hartman
> > <gregkh at linuxfoundation.org> wrote:
> > >
> > > On Sun, Sep 18, 2022 at 08:44:27PM -0700, Olof Johansson wrote:
> > > > On Tue, Aug 23, 2022 at 8:37 AM Greg Kroah-Hartman
> > > > <gregkh at linuxfoundation.org> wrote:
> > > > >
> > > > > On Thu, Jun 30, 2022 at 06:26:38PM -0700, Saravana Kannan wrote:
> > > > > > These patches are on top of driver-core-next.
> > > > > >
> > > > > > Even if stdout-path isn't set in DT, this patch should take console
> > > > > > probe times back to how they were before the deferred_probe_timeout
> > > > > > clean up series[1].
> > > > >
> > > > > Now dropped from my queue due to lack of a response to other reviewer's
> > > > > questions.
> > > >
> > > > What happened to this patch? I have a 10 second timeout on console
> > > > probe on my SiFive Unmatched, and I don't see this flag being set for
> > > > the serial driver. In fact, I don't see it anywhere in-tree. I can't
> > > > seem to locate another patchset from Saravana around this though, so
> > > > I'm not sure where to look for a missing piece for the sifive serial
> > > > driver.
> > > >
> > > > This is the second boot time regression (this one not fatal, unlike
> > > > the Layerscape PCIe one) from the fw_devlink patchset.
> > > >
> > > > Greg, can you revert the whole set for 6.0, please? It's obviously
> > > > nowhere near tested enough to go in and I expect we'll see a bunch of
> > > > -stable fixups due to this if we let it remain in.
> > >
> > > What exactly is "the whole set"?  I have the default option fix queued
> > > up and will send that to Linus later this week (am traveling back from
> > > Plumbers still), but have not heard any problems about any other issues
> > > at all other than your report.
> >
> > I stand corrected in this case, the issue on the Hifive Unmatched was
> > a regression due to a PWM clock change -- I just sent a patch for that
> > (serial driver fix).
> >
> > So it seems like as long as the fw_devlink.strict=1 patch is reverted,
> > things are back to a working state here.
> >
> > I still struggle with how the fw_devlink patchset is expected to work
> > though, since DT is expected to describe the hardware configuration,
> > and it has no knowledge of whether there are drivers that will be
> > bound to any referenced supplier devnodes. It's not going to work well
> > to assume that they will always be bound, and to add 10 second
> > timeouts for those cases isn't a good solution. Seems like the number
> > of special cases will keep adding up.
> 
> Since the introduction of deferred probe, the kernel has always
> assumed if there is a device described, then there is or will be a
> driver for it. The result is you can't use new DTs (if they add
> providers) with older kernels.
> 
> We've ended up with a timeout because no one has come up with a better
> way to handle it. What the kernel needs is userspace saying "I'm done
> loading modules", but it's debatable whether that's a good solution
> too.

In my opinion the deferred probe is a big hack and that is the root
cause of the issues we have here and there. It has to be redesigned
to be mathematically robust. It was an attempt by Andrzej Hajda to
solve this [1].

[1]: https://events19.linuxfoundation.org/wp-content/uploads/2017/12/Deferred-Problem-Issues-With-Complex-Dependencies-Between-Devices-in-Linux-Kernel-Andrzej-Hajda-Samsung.pdf

-- 
With Best Regards,
Andy Shevchenko