Re: [PATCH v2 02/18] KVM: x86: Use __free(put_cpufreq_policy) for policy reference

From
Sean Christopherson <seanjc@google.com>
Date
2025-08-28 17:15:19
Also in
dri-devel, imx, intel-gfx, kvm, linux-acpi, linux-mediatek, linux-omap, linux-pm, linux-samsung-soc, linux-tegra, lkml
Reply to
Zihuan Zhang, 2025-08-28
On Thu, Aug 28, 2025, Zihuan Zhang wrote:
quoted
Hmm, this is technically buggy.  __free() won't invoke put_cpufreq_policy() until
policy goes out of scope, and so using __free() means the code is effectively:

		if (IS_ENABLED(CONFIG_CPU_FREQ)) {
			struct cpufreq_policy *policy;
			int cpu;

			cpu = get_cpu();
			policy = cpufreq_cpu_get(cpu);
			if (policy && policy->cpuinfo.max_freq)
				max_tsc_khz = policy->cpuinfo.max_freq;
			put_cpu();

			if (policy)
				cpufreq_cpu_put(policy);
		}
...
Yes, this will indeed change the execution order.
Can you accept that? 
No, because it's buggy.
Personally, I don’t think it’s ideal either.

		if (IS_ENABLED(CONFIG_CPU_FREQ)) {
 			int cpu;
			cpu = get_cpu();
			{
				struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
				if (policy && policy->cpuinfo.max_freq)
					max_tsc_khz = policy->cpuinfo.max_freq;
			}
			put_cpu();

                                            }

Other places may also have the same issue,

maybe we should consider introducing a macro to handle this properly,
so that initialization and cleanup are well defined without changing
the existing order unexpected.

like this:

#define WITH_CPUFREQ_POLICY(cpu) {\

for(struct cpufreq_policy *policy __free(put_cpufreq_policy) =  \
			cpufreq_cpu_get(cpu);			\
			policy;)

Then Use it:

		if (IS_ENABLED(CONFIG_CPU_FREQ)) {
 			int cpu;
			cpu = get_cpu();
			WITH_CPUFREQ_POLICY(cpu){
				if (policy->cpuinfo.max_freq)
					max_tsc_khz = policy->cpuinfo.max_freq;
			}
			put_cpu();
This all feels very forced, in the sense that we have a shiny new tool and are
trying to use it everywhere without thinking critically about whether or not
doing so is actually an improvement.

At a glance, this is literally the only instance in the entire kernel where the
CPU to use is grabbed immediately before the policy.
 
  $ git grep -B 20 cpufreq_cpu_get | grep -e get_cpu -e smp_processor_id
  arch/x86/kvm/x86.c-			cpu = get_cpu();
  drivers/cpufreq/cppc_cpufreq.c-static int cppc_get_cpu_power(struct device *cpu_dev,
  drivers/cpufreq/cppc_cpufreq.c-static int cppc_get_cpu_cost(struct device *cpu_dev, unsigned long KHz,
  drivers/cpufreq/mediatek-cpufreq-hw.c-mtk_cpufreq_get_cpu_power(struct device *cpu_dev, unsigned long *uW,

Probably because KVM's usage is rather bizarre and honestly kind of dumb.  But
KVM has had this behavior for 15+ years, so as weird as it is, I'm not inclined
to change it without a really, really strong reason to do so, e.g. to iterate
over all CPUs or something.

So given that this is the only intance of the problem patter, I think it makes
sense to leave KVM as-is, and not spend a bunch of time trying to figure out how
to make KVM's usage play nice with __free().
Thread (39 messages) 39 messages, 6 authors, last reply 2025-08-29
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help