diff mbox series

[RFC,v2,02/35] drivers: base: Use present CPUs in GENERIC_CPU_DEVICES

Message ID 20230913163823.7880-3-james.morse@arm.com (mailing list archive)
State Handled Elsewhere
Headers show
Series ACPI/arm64: add support for virtual cpuhotplug | expand

Checks

Context Check Description
conchuod/cover_letter success Series has a cover letter
conchuod/tree_selection success Guessed tree name to be for-next at HEAD 0bb80ecc33a8
conchuod/fixes_present success Fixes tag not required for -next series
conchuod/maintainers_pattern success MAINTAINERS pattern errors before the patch: 5 and now 5
conchuod/verify_signedoff success Signed-off-by tag matches author and committer
conchuod/kdoc success Errors and warnings before: 0 this patch: 0
conchuod/build_rv64_clang_allmodconfig success Errors and warnings before: 9 this patch: 9
conchuod/module_param success Was 0 now: 0
conchuod/build_rv64_gcc_allmodconfig success Errors and warnings before: 9 this patch: 9
conchuod/build_rv32_defconfig success Build OK
conchuod/dtb_warn_rv64 success Errors and warnings before: 25 this patch: 25
conchuod/header_inline success No static functions without inline keyword in header files
conchuod/checkpatch success total: 0 errors, 0 warnings, 0 checks, 8 lines checked
conchuod/build_rv64_nommu_k210_defconfig success Build OK
conchuod/verify_fixes success No Fixes tag
conchuod/build_rv64_nommu_virt_defconfig success Build OK

Commit Message

James Morse Sept. 13, 2023, 4:37 p.m. UTC
Three of the five ACPI architectures create sysfs entries using
register_cpu() for present CPUs, whereas arm64, riscv and all
GENERIC_CPU_DEVICES do this for possible CPUs.

Registering a CPU is what causes them to show up in sysfs.

It makes very little sense to register all possible CPUs. Registering
a CPU is what triggers the udev notifications allowing user-space to
react to newly added CPUs.

To allow all five ACPI architectures to use GENERIC_CPU_DEVICES, change
it to use for_each_present_cpu(). Making the ACPI architectures use
GENERIC_CPU_DEVICES is a pre-requisite step to centralise their
cpu_register() logic, before moving it into the ACPI processor driver.
When ACPI is disabled this work would be done by
cpu_dev_register_generic().

Of the ACPI architectures that register possible CPUs, arm64 and riscv
do not support making possible CPUs present as they use the weak 'always
fails' version of arch_register_cpu().

Only two of the eight architectures that use GENERIC_CPU_DEVICES have a
distinction between present and possible CPUs.

The following architectures use GENERIC_CPU_DEVICES but are not SMP,
so possible == present:
 * m68k
 * microblaze
 * nios2

The following architectures use GENERIC_CPU_DEVICES and consider
possible == present:
 * csky: setup_smp()
 * parisc: smp_prepare_boot_cpu() marks the boot cpu as present,
   processor_probe() sets possible for all CPUs and present for all CPUs
   except the boot cpu.

um appears to be a subarchitecture of x86.

The remaining architecture using GENERIC_CPU_DEVICES are:
 * openrisc and hexagon:
   where smp_init_cpus() makes all CPUs < NR_CPUS possible,
   whereas smp_prepare_cpus() only makes CPUs < setup_max_cpus present.

After this change, openrisc and hexagon systems that use the max_cpus
command line argument would not see the other CPUs present in sysfs.
This should not be a problem as these CPUs can't bre brought online as
_cpu_up() checks cpu_present().

After this change, only CPUs which are present appear in sysfs.

Signed-off-by: James Morse <james.morse@arm.com>
---
 drivers/base/cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Russell King (Oracle) Sept. 14, 2023, 8:20 a.m. UTC | #1
On Wed, Sep 13, 2023 at 04:37:50PM +0000, James Morse wrote:
> Three of the five ACPI architectures create sysfs entries using
> register_cpu() for present CPUs, whereas arm64, riscv and all
> GENERIC_CPU_DEVICES do this for possible CPUs.
> 
> Registering a CPU is what causes them to show up in sysfs.
> 
> It makes very little sense to register all possible CPUs. Registering
> a CPU is what triggers the udev notifications allowing user-space to
> react to newly added CPUs.
> 
> To allow all five ACPI architectures to use GENERIC_CPU_DEVICES, change
> it to use for_each_present_cpu(). Making the ACPI architectures use
> GENERIC_CPU_DEVICES is a pre-requisite step to centralise their
> cpu_register() logic, before moving it into the ACPI processor driver.
> When ACPI is disabled this work would be done by
> cpu_dev_register_generic().
> 
> Of the ACPI architectures that register possible CPUs, arm64 and riscv
> do not support making possible CPUs present as they use the weak 'always
> fails' version of arch_register_cpu().
> 
> Only two of the eight architectures that use GENERIC_CPU_DEVICES have a
> distinction between present and possible CPUs.
> 
> The following architectures use GENERIC_CPU_DEVICES but are not SMP,
> so possible == present:
>  * m68k
>  * microblaze
>  * nios2
> 
> The following architectures use GENERIC_CPU_DEVICES and consider
> possible == present:
>  * csky: setup_smp()
>  * parisc: smp_prepare_boot_cpu() marks the boot cpu as present,
>    processor_probe() sets possible for all CPUs and present for all CPUs
>    except the boot cpu.

However, init/main.c::start_kernel() calls boot_cpu_init() which sets
the boot CPU in the online, active, present and possible masks. So,
_every_ architecture gets the boot CPU in all these masks no matter
what.

Only of something then clears the boot CPU from these masks (which
would be silly) would the boot CPU not be in all of these masks.
Jonathan Cameron Sept. 14, 2023, 10:56 a.m. UTC | #2
On Thu, 14 Sep 2023 09:20:54 +0100
"Russell King (Oracle)" <linux@armlinux.org.uk> wrote:

> On Wed, Sep 13, 2023 at 04:37:50PM +0000, James Morse wrote:
> > Three of the five ACPI architectures create sysfs entries using
> > register_cpu() for present CPUs, whereas arm64, riscv and all
> > GENERIC_CPU_DEVICES do this for possible CPUs.
> > 
> > Registering a CPU is what causes them to show up in sysfs.
> > 
> > It makes very little sense to register all possible CPUs. Registering
> > a CPU is what triggers the udev notifications allowing user-space to
> > react to newly added CPUs.
> > 
> > To allow all five ACPI architectures to use GENERIC_CPU_DEVICES, change
> > it to use for_each_present_cpu(). Making the ACPI architectures use
> > GENERIC_CPU_DEVICES is a pre-requisite step to centralise their
> > cpu_register() logic, before moving it into the ACPI processor driver.
> > When ACPI is disabled this work would be done by
> > cpu_dev_register_generic().
> > 
> > Of the ACPI architectures that register possible CPUs, arm64 and riscv
> > do not support making possible CPUs present as they use the weak 'always
> > fails' version of arch_register_cpu().
> > 
> > Only two of the eight architectures that use GENERIC_CPU_DEVICES have a
> > distinction between present and possible CPUs.
> > 
> > The following architectures use GENERIC_CPU_DEVICES but are not SMP,
> > so possible == present:
> >  * m68k
> >  * microblaze
> >  * nios2
> > 
> > The following architectures use GENERIC_CPU_DEVICES and consider
> > possible == present:
> >  * csky: setup_smp()
> >  * parisc: smp_prepare_boot_cpu() marks the boot cpu as present,
> >    processor_probe() sets possible for all CPUs and present for all CPUs
> >    except the boot cpu.  
> 
> However, init/main.c::start_kernel() calls boot_cpu_init() which sets
> the boot CPU in the online, active, present and possible masks. So,
> _every_ architecture gets the boot CPU in all these masks no matter
> what.
> 
> Only of something then clears the boot CPU from these masks (which
> would be silly) would the boot CPU not be in all of these masks.
Hi Russel,

Upshot is that the code in parisc smp_prepare_boot_cpu() can be dropped?
Seems like another useful simplification to add to front of this series.
The function will end up with just a print then. 
Seems there are lots of other empty implementations of smp_prepare_boot_cpu()
maybe worth making that optional whilst here and dropping all the empty ones?

There seem to be some other architectures setting at least some of the cpu masks
that could perhaps be tidied up a little via same logic?

Jonathan

>
Jonathan Cameron Sept. 14, 2023, 10:56 a.m. UTC | #3
> Signed-off-by: James Morse <james.morse@arm.com>
Seems sensible and well reasoned cleanup to me.
Technically an ABI change, but would be seriously odd if any real code
relied on the current pointless behavior on the few architectures where
is changing.

FWIW review is really of your analysis rather than the change :)

Seems like there may be some additional cleanup that makes sense from
Russell's analysis, but that's perhaps a parallel job.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>


> ---
>  drivers/base/cpu.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
> index 9ea22e165acd..34b48f660b6b 100644
> --- a/drivers/base/cpu.c
> +++ b/drivers/base/cpu.c
> @@ -533,7 +533,7 @@ static void __init cpu_dev_register_generic(void)
>  #ifdef CONFIG_GENERIC_CPU_DEVICES
>  	int i;
>  
> -	for_each_possible_cpu(i) {
> +	for_each_present_cpu(i) {
>  		if (register_cpu(&per_cpu(cpu_devices, i), i))
>  			panic("Failed to register CPU device");
>  	}
Russell King (Oracle) Sept. 14, 2023, 11:11 a.m. UTC | #4
On Thu, Sep 14, 2023 at 11:56:13AM +0100, Jonathan Cameron wrote:
> On Thu, 14 Sep 2023 09:20:54 +0100
> "Russell King (Oracle)" <linux@armlinux.org.uk> wrote:
> 
> > On Wed, Sep 13, 2023 at 04:37:50PM +0000, James Morse wrote:
> > > Three of the five ACPI architectures create sysfs entries using
> > > register_cpu() for present CPUs, whereas arm64, riscv and all
> > > GENERIC_CPU_DEVICES do this for possible CPUs.
> > > 
> > > Registering a CPU is what causes them to show up in sysfs.
> > > 
> > > It makes very little sense to register all possible CPUs. Registering
> > > a CPU is what triggers the udev notifications allowing user-space to
> > > react to newly added CPUs.
> > > 
> > > To allow all five ACPI architectures to use GENERIC_CPU_DEVICES, change
> > > it to use for_each_present_cpu(). Making the ACPI architectures use
> > > GENERIC_CPU_DEVICES is a pre-requisite step to centralise their
> > > cpu_register() logic, before moving it into the ACPI processor driver.
> > > When ACPI is disabled this work would be done by
> > > cpu_dev_register_generic().
> > > 
> > > Of the ACPI architectures that register possible CPUs, arm64 and riscv
> > > do not support making possible CPUs present as they use the weak 'always
> > > fails' version of arch_register_cpu().
> > > 
> > > Only two of the eight architectures that use GENERIC_CPU_DEVICES have a
> > > distinction between present and possible CPUs.
> > > 
> > > The following architectures use GENERIC_CPU_DEVICES but are not SMP,
> > > so possible == present:
> > >  * m68k
> > >  * microblaze
> > >  * nios2
> > > 
> > > The following architectures use GENERIC_CPU_DEVICES and consider
> > > possible == present:
> > >  * csky: setup_smp()
> > >  * parisc: smp_prepare_boot_cpu() marks the boot cpu as present,
> > >    processor_probe() sets possible for all CPUs and present for all CPUs
> > >    except the boot cpu.  
> > 
> > However, init/main.c::start_kernel() calls boot_cpu_init() which sets
> > the boot CPU in the online, active, present and possible masks. So,
> > _every_ architecture gets the boot CPU in all these masks no matter
> > what.
> > 
> > Only of something then clears the boot CPU from these masks (which
> > would be silly) would the boot CPU not be in all of these masks.
> Hi Russel,
> 
> Upshot is that the code in parisc smp_prepare_boot_cpu() can be dropped?
> Seems like another useful simplification to add to front of this series.

Yes - but I personally (and probably others) would like to see progress
made towards getting at least some of the changes in this series merged,
rather than seeing this series hang around longer and grow. Nothing in
this series touches any architecture's smp_prepare_boot_cpu(), so such
a change would not interfere with this series.

Therefore, I suggest that removing those two set_cpu_*() calls in
smp_prepare_boot_cpu() is something that could happen irrespective of
anything in this series, and I would encourage PA-RISC folk to do that
anway.

The same is true of Loongarch, mips, sh, and sparc32, and they can
independently sort this.

> Seems there are lots of other empty implementations of smp_prepare_boot_cpu()
> maybe worth making that optional whilst here and dropping all the empty ones?

Yes, and again, this could be a series separate from this one. If one
arch wants to add the empty weak version of smp_prepare_boot_cpu(),
then it would be a matter of others deleting their empty implementation
(possibly after first having cleaned up the unnecessary set_cpu_*()
calls).

In any case, I would expect that patches doing any of the above would
end up being cherry-picked from a series by arch maintainers, so at
least to me it makes zero sense to include it with this already large
series, and would make the management of this series more complex.
diff mbox series

Patch

diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
index 9ea22e165acd..34b48f660b6b 100644
--- a/drivers/base/cpu.c
+++ b/drivers/base/cpu.c
@@ -533,7 +533,7 @@  static void __init cpu_dev_register_generic(void)
 #ifdef CONFIG_GENERIC_CPU_DEVICES
 	int i;
 
-	for_each_possible_cpu(i) {
+	for_each_present_cpu(i) {
 		if (register_cpu(&per_cpu(cpu_devices, i), i))
 			panic("Failed to register CPU device");
 	}