Re: [resend][bug] low-probability console lockups since 5.19

[Conor Dooley]

On Thu, Sep 29, 2022 at 12:12:39PM +0200, Petr Mladek wrote:
> On Thu 2022-09-29 10:29:05, Conor Dooley wrote:
> > On Thu, Sep 29, 2022 at 11:06:01AM +0200, Thorsten Leemhuis wrote:
> > > Hi Conor
> > > 
> > > On 28.09.22 18:55, Conor Dooley wrote:
> > > > On Fri, Sep 23, 2022 at 05:24:17PM +0100, Conor Dooley wrote:
> > > >>
> > > >> Been bisecting a bug that is causing a boot failure in my CI & have
> > > >> ended up here.. The bug in question is a low(ish) probability lock up
> > > >> of the serial console, I would estimate about 1-in-5 chance on the
> > > >> boards I could actually trigger it on which it has taken me so long
> > > >> to realise that this was an actual problem. Thinking back on it, there
> > > >> were other failures that I would retroactively attribute to this
> > > >> problem too, but I had earlycon disabled
> > > 
> > > There is one thing I wonder when skimming this thread: was there maybe
> > > some other change somewhere in the kernel between the introduction and
> > > the revert of the printk console kthreads patches that is the real
> > > culprit here that makes existing, older races easier to hit? But I guess
> > > in the end that would be very hard to find and it's easier to fix the
> > > problem in the console driver... :-/
> > 
> > Entirely possible that something arrived in the middle, yeah. I've done
> > 100s of reboots on that interim section, albeit with the threaded
> > printers enabled, as I restarted the bisection several times & never hit
> > this failure then.
> 
> Interesting. I wonder if the used console was fixed during the window
> when the kthreads were enabled.

I will, possibly tonight but probably not, run the bisection again with
the threaded printer merge reverted. Hopefully it is not filled with
conflicts if I go that way...

> 
> > I don't know anything about console/printk/serial drivers unfortunately
> > so I will almost certainly not be able to find the problem by
> > inspection. I'd rather submit patches than send reports, but I really
> > really need some help here. I looked at the two patterns Petr suggested,
> > but the former I am not sure applies since the issue is present even
> > when earlycon is disabled & the latter appears (to my untrained eye) to
> > be accounted for in the 8250 driver.
> 
> The problem with the missing port->lock is visible only when the
> early console is enabled. But It is really hard to hit without
> the kthreads.

Right, so sounds like that can be excluded since my CI was hitting it
with earlycon disabled. I'll triple check this, possibly later today.

> 
> The problem with enabled IRQs was visible only with kthreads. The
> original code called console->write() callback already with IRQs
> disabled.
> 
> The kthreads called console->write() callback with IRQs enabled.
> It made sense. They need to be disabled only when really needed
> and the tested drivers did this correctly.

And that sounds like it can be also excluded, since my issue started
post-revert. Unless there's still some kthreads code in there that was
not reverted?

Thanks,
Conor.