Re: [PATCH v8 04/16] ACPI: processor: Move checks and availability of acpi_processor earlier

["Rafael J. Wysocki" <rafael@kernel.org>]

On Tue, Apr 30, 2024 at 11:28 AM Jonathan Cameron
<Jonathan.Cameron@huawei.com> wrote:
>
> On Tue, 30 Apr 2024 14:17:24 +1000
> Gavin Shan <gshan@redhat.com> wrote:
>
> > On 4/26/24 23:51, Jonathan Cameron wrote:
> > > Make the per_cpu(processors, cpu) entries available earlier so that
> > > they are available in arch_register_cpu() as ARM64 will need access
> > > to the acpi_handle to distinguish between acpi_processor_add()
> > > and earlier registration attempts (which will fail as _STA cannot
> > > be checked).
> > >
> > > Reorder the remove flow to clear this per_cpu() after
> > > arch_unregister_cpu() has completed, allowing it to be used in
> > > there as well.
> > >
> > > Note that on x86 for the CPU hotplug case, the pr->id prior to
> > > acpi_map_cpu() may be invalid. Thus the per_cpu() structures
> > > must be initialized after that call or after checking the ID
> > > is valid (not hotplug path).
> > >
> > > Signed-off-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
> > >
> > > ---
> > > v8: On buggy bios detection when setting per_cpu structures
> > >      do not carry on.
> > >      Fix up the clearing of per cpu structures to remove unwanted
> > >      side effects and ensure an error code isn't use to reference them.
> > > ---
> > >   drivers/acpi/acpi_processor.c | 79 +++++++++++++++++++++--------------
> > >   1 file changed, 48 insertions(+), 31 deletions(-)
> > >
> > > diff --git a/drivers/acpi/acpi_processor.c b/drivers/acpi/acpi_processor.c
> > > index ba0a6f0ac841..3b180e21f325 100644
> > > --- a/drivers/acpi/acpi_processor.c
> > > +++ b/drivers/acpi/acpi_processor.c
> > > @@ -183,8 +183,38 @@ static void __init acpi_pcc_cpufreq_init(void) {}
> > >   #endif /* CONFIG_X86 */
> > >
> > >   /* Initialization */
> > > +static DEFINE_PER_CPU(void *, processor_device_array);
> > > +
> > > +static bool acpi_processor_set_per_cpu(struct acpi_processor *pr,
> > > +                                  struct acpi_device *device)
> > > +{
> > > +   BUG_ON(pr->id >= nr_cpu_ids);
> >
> > One blank line after BUG_ON() if we need to follow original implementation.
>
> Sure unintentional - I'll put that back.
>
> >
> > > +   /*
> > > +    * Buggy BIOS check.
> > > +    * ACPI id of processors can be reported wrongly by the BIOS.
> > > +    * Don't trust it blindly
> > > +    */
> > > +   if (per_cpu(processor_device_array, pr->id) != NULL &&
> > > +       per_cpu(processor_device_array, pr->id) != device) {
> > > +           dev_warn(&device->dev,
> > > +                    "BIOS reported wrong ACPI id %d for the processor\n",
> > > +                    pr->id);
> > > +           /* Give up, but do not abort the namespace scan. */
> >
> > It depends on how the return value is handled by the caller if the namespace
> > is continued to be scanned. The caller can be acpi_processor_hotadd_init()
> > and acpi_processor_get_info() after this patch is applied. So I think this
> > specific comment need to be moved to the caller.
>
> Good point. This gets messy and was an unintended change.
>
> Previously the options were:
> 1) acpi_processor_get_info() failed for other reasons - this code was never called.
> 2) acpi_processor_get_info() succeeded without acpi_processor_hotadd_init (non hotplug)
>    this code then ran and would paper over the problem doing a bunch of cleanup under err.
> 3) acpi_processor_get_info() succeeded with acpi_processor_hotadd_init called.
>    This code then ran and would paper over the problem doing a bunch of cleanup under err.
>
> We should maintain that or argue cleanly against it.

The return value needs to be propagated to acpi_processor_add() so it
can decide what to do with it.

Now, acpi_processor_add() can only return 1 if the CPU has been
successfully registered and initialized, so it is regarded as
available (but it may not be online to start with).

Returning 0 from it may get messy, because acpi_default_enumeration()
will get called and it will attempt to create a platform device for
the CPU, so in all cases in which the CPU is not regarded as available
when acpi_processor_add() returns, it should return an error code (the
exact value doesn't matter for its caller so long as it is negative).

> This isn't helped the the fact I have no idea which cases we care about for that bios
> bug handling.  Do any of those bios's ever do hotplug?  Guess we have to try and maintain
> whatever protection this was offering.
>
> Also, the original code leaks data in some paths and I have limited idea
> of whether it is intentional or not. So to tidy the issue up that you've identified
> I'll need to try and make that code consistent first.

I agree.

> I suspect the only way to do that is going to be to duplicate the allocations we
> 'want' to leak to deal with the bios bug detection.
>
> For example acpi_processor_get_info() failing leaks pr and pr->throttling.shared_cpu_map
> before this series. After this series we need pr to leak because it's used for the detection
> via processor_device_array.
>
> I'll work through this but it's going to be tricky to tell if we get right.
> Step 1 will be closing the existing leaks and then we will have something
> consistent to build on.

Sounds good to me.