On Tue, 20 Feb 2024 15:13:58 +0000
"Russell King (Oracle)" <linux@xxxxxxxxxxxxxxx> wrote:
> On Tue, Feb 20, 2024 at 11:27:15AM +0000, Russell King (Oracle) wrote:
> > On Thu, Feb 15, 2024 at 08:22:29PM +0100, Rafael J. Wysocki wrote:
> > > On Wed, Jan 31, 2024 at 5:50 PM Russell King <rmk+kernel@xxxxxxxxxxxxxxx> wrote:
> > > > diff --git a/drivers/acpi/acpi_processor.c b/drivers/acpi/acpi_processor.c
> > > > index cf7c1cca69dd..a68c475cdea5 100644
> > > > --- a/drivers/acpi/acpi_processor.c
> > > > +++ b/drivers/acpi/acpi_processor.c
> > > > @@ -314,6 +314,18 @@ static int acpi_processor_get_info(struct acpi_device *device)
> > > > cpufreq_add_device("acpi-cpufreq");
> > > > }
> > > >
> > > > + /*
> > > > + * Register CPUs that are present. get_cpu_device() is used to skip
> > > > + * duplicate CPU descriptions from firmware.
> > > > + */
> > > > + if (!invalid_logical_cpuid(pr->id) && cpu_present(pr->id) &&
> > > > + !get_cpu_device(pr->id)) {
> > > > + int ret = arch_register_cpu(pr->id);
> > > > +
> > > > + if (ret)
> > > > + return ret;
> > > > + }
> > > > +
> > > > /*
> > > > * Extra Processor objects may be enumerated on MP systems with
> > > > * less than the max # of CPUs. They should be ignored _iff
> > >
> > > This is interesting, because right below there is the following code:
> > >
> > > if (invalid_logical_cpuid(pr->id) || !cpu_present(pr->id)) {
> > > int ret = acpi_processor_hotadd_init(pr);
> > >
> > > if (ret)
> > > return ret;
> > > }
> > >
> > > and acpi_processor_hotadd_init() essentially calls arch_register_cpu()
> > > with some extra things around it (more about that below).
> > >
> > > I do realize that acpi_processor_hotadd_init() is defined under
> > > CONFIG_ACPI_HOTPLUG_CPU, so for the sake of the argument let's
> > > consider an architecture where CONFIG_ACPI_HOTPLUG_CPU is set.
> > >
> > > So why are the two conditionals that almost contradict each other both
> > > needed? It looks like the new code could be combined with
> > > acpi_processor_hotadd_init() to do the right thing in all cases.
> > >
> > > Now, acpi_processor_hotadd_init() does some extra things that look
> > > like they should be done by the new code too.
> > >
> > > 1. It checks invalid_phys_cpuid() which appears to be a good idea to me.
> > >
> > > 2. It uses locking around arch_register_cpu() which doesn't seem
> > > unreasonable either.
> > >
> > > 3. It calls acpi_map_cpu() and I'm not sure why this is not done by
> > > the new code.
> > >
> > > The only thing that can be dropped from it is the _STA check AFAICS,
> > > because acpi_processor_add() won't even be called if the CPU is not
> > > present (and not enabled after the first patch).
> > >
> > > So why does the code not do 1 - 3 above?
> >
> > Honestly, I'm out of my depth with this and can't answer your
> > questions - and I really don't want to try fiddling with this code
> > because it's just too icky (even in its current form in mainline)
> > to be understandable to anyone who hasn't gained a detailed knowledge
> > of this code.
> >
> > It's going to require a lot of analysis - how acpi_map_cpuid() behaves
> > in all circumstances, what this means for invalid_logical_cpuid() and
> > invalid_phys_cpuid(), what paths will be taken in each case. This code
> > is already just too hairy for someone who isn't an experienced ACPI
> > hacker to be able to follow and I don't see an obvious way to make it
> > more readable.
> >
> > James' additions make it even more complex and less readable.
>
> As an illustration of the problems I'm having here, I was just writing
> a reply to this with a suggestion of transforming this code ultimately
> to:
>
> if (!get_cpu_device(pr->id)) {
> int ret;
>
> if (!invalid_logical_cpuid(pr->id) && cpu_present(pr->id))
> ret = acpi_processor_make_enabled(pr);
> else
> ret = acpi_processor_make_present(pr);
>
> if (ret)
> return ret;
> }
>
> (acpi_processor_make_present() would be acpi_processor_hotadd_init()
> and acpi_processor_make_enabled() would be arch_register_cpu() at this
> point.)
>
> Then I realised that's a bad idea - because we really need to check
> that pr->id is valid before calling get_cpu_device() on it, so this
> won't work. That leaves us with:
>
> int ret;
>
> if (invalid_logical_cpuid(pr->id) || !cpu_present(pr->id)) {
> /* x86 et.al. path */
> ret = acpi_processor_make_present(pr);
> } else if (!get_cpu_device(pr->id)) {
> /* Arm64 path */
> ret = acpi_processor_make_enabled(pr);
> } else {
> ret = 0;
> }
>
> if (ret)
> return ret;
>
> Now, the next transformation would be to move !get_cpu_device(pr->id)
> into acpi_processor_make_enabled() which would eliminate one of those
> if() legs.
>
> Now, if we want to somehow make the call to arch_regster_cpu() common
> in these two paths, the next question is what are the _precise_
> semantics of acpi_map_cpu(), particularly with respect to it
> modifying pr->id. Is it guaranteed to always give the same result
> for the same processor described in ACPI? What acpi_map_cpu() anyway,
> I can find no documentation for it.
>
> Then there's the question whether calling acpi_unmap_cpu() should be
> done on the failure path if arch_register_cpu() fails, which is done
> for the x86 path but not the Arm64 path. Should it be done for the
> Arm64 path? I've no idea, but as Arm64 doesn't implement either of
> these two functions, I guess they could be stubbed out and thus be
> no-ops - but then we open a hole where if pr->id is invalid, we
> end up passing that invalid value to arch_register_cpu() which I'm
> quite sure will explode with a negative CPU number.
>
> So, to my mind, what you're effectively asking for is a total rewrite
> of all the code in and called by acpi_processor_get_info()... and that
> is not something I am willing to do (because it's too far outside of
> my knowledge area.)
>
> As I said in my reply to patch 1, I think your comments on patch 2
> make Arm64 vcpu hotplug unachievable in a reasonable time frame, and
> certainly outside the bounds of what I can do to progress this.
>
> So, at this point I'm going to stand down from further participation
> with this patch set as I believe I've reached the limit of what I can
> do to progress it.
>
Thanks for your hard work on this Russell - we have moved forwards.
Short of anyone else stepping up I'll pick this up with
the help of some my colleagues. As such I'm keen on getting patch
1 upstream ASAP so that we can exclude the need for some of the
other workarounds from earlier versions of this series (the ones
dropped before now).
We will need a little time to get up to speed on the current status
and discussion points Russell raises above.
Jonathan
