RE: [PATCH] x86/msr: Use num_possible_cpus() for chrdev minor count

[Michael Kelley]

From: Roman 'Hedin' Storozhenko <romeusmeister@xxxxxxxxx> Sent: Wednesday, April 1, 2026 11:56 AM
> 
> The MSR driver currently register their character device
> regions using NR_CPUS as the minor count:
> 
>   __register_chrdev(..., 0, NR_CPUS, ...)
> 
> This dates back to commit 0b962d473af3 ("x86, msr/cpuid: Register
> enough minors for the MSR and CPUID drivers"), which intentionally
> used NR_CPUS as a static upper bound on the maximum CPU index.
> 
> However, NR_CPUS is a compile-time upper bound and may significantly
> exceed the actual number of CPUs supported by the running system
> (i.e. cpu_possible_mask). As a result, the drivers may reserve a much
> larger minor range than can ever be used at runtime.
> 
> Replace NR_CPUS with num_possible_cpus() so that the registered minor
> range reflects the actual system topology instead of a static maximum.

Using num_possible_cpus() introduces the potential for a different
problem. There's no guarantee that the cpu_possible_mask is dense,
so num_possible_cpus() might be smaller than the largest possible
Linux cpu number. Since msr_device_create() uses the Linux cpu number
as the minor device number, that minor number could exceed the size
of the registered minor range.

There's no problem seen today because I believe the possible_cpus
bitmap is currently always dense, but there's the potential in the future
that it is not. See email thread [1].

Instead of num_possible_cpus(), use nr_cpu_ids.

Michael

[1] https://lore.kernel.org/lkml/SN6PR02MB4157210CC36B2593F8572E5ED4692@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

> 
> The driver already creates and destroy devices dynamically based on
> CPU hotplug state, so this change only affects the size of the
> internally reserved minor range and does not alter runtime behavior
> or userspace-visible interfaces.
> 
> Signed-off-by: Roman 'Hedin' Storozhenko <romeusmeister@xxxxxxxxx>
> ---
> This series replaces NR_CPUS with num_possible_cpus() when registering
> character device regions for the MSR and CPUID drivers.
> 
> Historically, commit 0b962d473af3 ("x86, msr/cpuid: Register enough minors
> for the MSR and CPUID drivers") switched to using NR_CPUS as a static upper
> bound on the maximum CPU index. While correct, this may result in reserving
> a significantly larger minor range than is actually usable on a given system.
> 
> This series updates the drivers to use num_possible_cpus(), aligning the
> registered minor range with the system's CPU topology.
> 
> The change does not affect runtime behavior: devices are still created and
> destroyed dynamically based on CPU hotplug state.
> 
> Validation
> ----------
> 
> The change was validated in a virtme-ng environment with:
> 
>   CONFIG_NR_CPUS=64
>   possible CPUs: 0-7 (8 CPUs)
> 
> 1) Internal state (eBPF)
> 
> Using bpftrace on __register_chrdev_region():
> 
>   Before:
>     count = 64
> 
>   After:
>     count = 8
> 
> This confirms that the registered minor range now reflects
> num_possible_cpus() instead of NR_CPUS.
> 
> 2) Userspace-visible behavior
> 
> No change observed:
> 
>   /dev/cpu/*/msr count: unchanged (8)
>   /sys/class/msr entries: unchanged (8)
> 
> 3) Functional correctness
> 
> rdmsr access continues to work as expected.
> 
> 4) CPU hotplug
> 
> Tracing msr_device_create shows identical behavior before and after,
> confirming that device lifecycle remains driven by CPU hotplug events.
> 
> ----------
> Overall, this change reduces over-allocation of minor numbers while
> preserving existing behavior and interfaces.
> ---
>  arch/x86/kernel/msr.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/kernel/msr.c b/arch/x86/kernel/msr.c
> index 4469c784eaa0..7f93f59d3cd9 100644
> --- a/arch/x86/kernel/msr.c
> +++ b/arch/x86/kernel/msr.c
> @@ -263,7 +263,7 @@ static int __init msr_init(void)
>  {
>  	int err;
> 
> -	if (__register_chrdev(MSR_MAJOR, 0, NR_CPUS, "cpu/msr", &msr_fops)) {
> +	if (__register_chrdev(MSR_MAJOR, 0, num_possible_cpus(), "cpu/msr",
> &msr_fops)) {
>  		pr_err("unable to get major %d for msr\n", MSR_MAJOR);
>  		return -EBUSY;
>  	}
> @@ -281,7 +281,7 @@ static int __init msr_init(void)
>  out_class:
>  	class_unregister(&msr_class);
>  out_chrdev:
> -	__unregister_chrdev(MSR_MAJOR, 0, NR_CPUS, "cpu/msr");
> +	__unregister_chrdev(MSR_MAJOR, 0, num_possible_cpus(), "cpu/msr");
>  	return err;
>  }
>  module_init(msr_init);
> @@ -290,7 +290,7 @@ static void __exit msr_exit(void)
>  {
>  	cpuhp_remove_state(cpuhp_msr_state);
>  	class_unregister(&msr_class);
> -	__unregister_chrdev(MSR_MAJOR, 0, NR_CPUS, "cpu/msr");
> +	__unregister_chrdev(MSR_MAJOR, 0, num_possible_cpus(), "cpu/msr");
>  }
>  module_exit(msr_exit)
> 
> 
> ---
> base-commit: 25f9333c7244d9c5a0243bcf045903ba19635d35
> change-id: 20260401-msr_cpu-d88ade162ca0
> 
> Best regards,
> --
> Roman 'Hedin' Storozhenko <romeusmeister@xxxxxxxxx>
>