Messages in this thread |  | | Date | Thu, 7 Mar 2024 11:02:34 +0800 | | Subject | Re: [PATCH v1 1/3] arm64: topology: Add arch_freq_get_on_cpu() support | | From | "lihuisong (C)" <> |
| |
Hi Vanshidhar,
在 2024/3/1 0:25, Vanshidhar Konda 写道: > AMU counters are used by the Frequency Invariance Engine (FIE) to > estimate the CPU utilization during each tick. The delta of the AMU > counters between two ticks can also be used to estimate the average CPU > frequency of each core over the tick duration. Measure the AMU counters > during tick, compute the delta and store it. When the frequency of the > core is queried, use the stored delta to determine the frequency. > > arch_freq_get_on_cpu() is used on x86 systems to estimate the frequency > of each CPU. It can be wired up on arm64 for the same functionality. > > Signed-off-by: Vanshidhar Konda <vanshikonda@os.amperecomputing.com> > --- > arch/arm64/kernel/topology.c | 114 +++++++++++++++++++++++++++++------ > 1 file changed, 96 insertions(+), 18 deletions(-) > > diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c > index 1a2c72f3e7f8..db8d14525cf4 100644 > --- a/arch/arm64/kernel/topology.c > +++ b/arch/arm64/kernel/topology.c > @@ -17,6 +17,8 @@ > #include <linux/cpufreq.h> > #include <linux/init.h> > #include <linux/percpu.h> > +#include <linux/sched/isolation.h> > +#include <linux/seqlock_types.h> > > #include <asm/cpu.h> > #include <asm/cputype.h> > @@ -82,20 +84,54 @@ int __init parse_acpi_topology(void) > #undef pr_fmt > #define pr_fmt(fmt) "AMU: " fmt > > +struct amu_counters { > + seqcount_t seq; > + unsigned long last_update; > + u64 core_cnt; > + u64 const_cnt; > + u64 delta_core_cnt; > + u64 delta_const_cnt; > +}; > + > /* > * Ensure that amu_scale_freq_tick() will return SCHED_CAPACITY_SCALE until > * the CPU capacity and its associated frequency have been correctly > * initialized. > */ > -static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale) = 1UL << (2 * SCHED_CAPACITY_SHIFT); > -static DEFINE_PER_CPU(u64, arch_const_cycles_prev); > -static DEFINE_PER_CPU(u64, arch_core_cycles_prev); > +static DEFINE_PER_CPU_READ_MOSTLY(unsigned long, arch_max_freq_scale) = > + 1UL << (2 * SCHED_CAPACITY_SHIFT); > +static DEFINE_PER_CPU_SHARED_ALIGNED(struct amu_counters, cpu_samples) = { > + .seq = SEQCNT_ZERO(cpu_samples.seq) > +}; > static cpumask_var_t amu_fie_cpus; > > void update_freq_counters_refs(void) > { > - this_cpu_write(arch_core_cycles_prev, read_corecnt()); > - this_cpu_write(arch_const_cycles_prev, read_constcnt()); > + struct amu_counters *cpu_sample = this_cpu_ptr(&cpu_samples); > + u64 core_cnt, const_cnt, delta_core_cnt, delta_const_cnt; > + > + const_cnt = read_constcnt(); > + core_cnt = read_corecnt(); > + > + if (unlikely(core_cnt < cpu_sample->core_cnt) || > + unlikely(const_cnt < cpu_sample->const_cnt)) {
The two counter register might be wrap around. So here is not fully true, right?
If we don't consider this case, below warning should be removed.
> + WARN(1, "AMU counter values should be monotonic.\n"); > + cpu_sample->delta_const_cnt = 0; > + cpu_sample->delta_core_cnt = 0; > + return; > + } > + > + delta_core_cnt = core_cnt - cpu_sample->core_cnt; > + delta_const_cnt = const_cnt - cpu_sample->const_cnt; > + > + cpu_sample->core_cnt = core_cnt; > + cpu_sample->const_cnt = const_cnt; > + > + raw_write_seqcount_begin(&cpu_sample->seq); > + cpu_sample->last_update = jiffies; > + cpu_sample->delta_const_cnt = delta_const_cnt; > + cpu_sample->delta_core_cnt = delta_core_cnt; > + raw_write_seqcount_end(&cpu_sample->seq); > } <...>
|  |