Implement trace_power_frequency for all cpufreq drivers
Subsystems (5)
Is such perf stuff supposed to go through x86-tip? Can someone pick these two up, please.
Thanks,
Thomas
PATCH 1/2 acpi-cpufreq: Fix CPU_ANY CPUFREQ_{PRE,POST}CHANGE notification
This has been found be review.
Currently PRE/POSTCHANGE events seem to only be thrown for one core if the BIOS tells us to run in CPU_ANY mode (switching one/any of the depending cores is enough).
--- arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c +2 -2
@@ -391,7 +391,7 @@ static int acpi_cpufreq_target(struct cpufreq_policy *policy, freqs.old = perf->states[perf->state].core_frequency *1000 ; freqs.new = data->freq_table[next_state].frequency; -for_each_cpu (i, cmd.mask) { +for_each_cpu (i, policy->cpus) { freqs.cpu = i;cpufreq_notify_transition (&freqs,CPUFREQ_PRECHANGE ); }
@@ -407,7 +407,7 @@ static int acpi_cpufreq_target(struct cpufreq_policy *policy, } } -for_each_cpu (i, cmd.mask) { +for_each_cpu (i, policy->cpus) { freqs.cpu = i;cpufreq_notify_transition (&freqs,CPUFREQ_POSTCHANGE ); }
Signature · 1 line
1.6.3
Re: [PATCH 1/2] acpi-cpufreq: Fix CPU_ANY CPUFREQ_{PRE,POST}CHANGE notification
Quoted 44 lines · Thomas Renninger
This has been found be review.
Currently PRE/POSTCHANGE events seem to only be thrown for one core if the BIOS tells us to run inCPU_ANYmode (switching one/any of the depending cores is enough).Signed-off-by Thomas Renninger <[email protected]>---
arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c | 4 ++--1 files changed, 2 insertions(+), 2 deletions(-)arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c +2
-3@@ -391,7 +391,7 @@ static int acpi_cpufreq_target(struct cpufreq_policy *policy, freqs.old = perf->states[perf->state].core_frequency * 1000; freqs.new = data->freq_table[next_state].frequency; - for_each_cpu(i, cmd.mask) { + for_each_cpu(i, policy->cpus) { freqs.cpu = i; cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE); }@@ -407,7 +407,7 @@ static int acpi_cpufreq_target(struct cpufreq_policy *policy, } } - for_each_cpu(i, cmd.mask) { + for_each_cpu(i, policy->cpus) { freqs.cpu = i; cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE); } --1.6.3
PATCH 2/2 x86 cpufreq: Make trace_power_frequency cpufreq driver independent
In reply to Thomas Renningerand fix the broken case if a core's frequency depends on others.
trace_power_frequency was only implemented in a rather ungeneric way in acpi-cpufreq driver's target() function only.
-> Move the call to trace_power_frequency to
cpufreq.c:cpufreq_notify_transition() where CPUFREQ_POSTCHANGE notifier is triggered.
This will support power frequency tracing by all cpufreq drivers
trace_power_frequency did not trace frequency changes correctly when the userspace governor was used or when CPU cores' frequency depend on each other.
-> Moving this into the CPUFREQ_POSTCHANGE notifier and pass the cpu
which gets switched automatically fixes this.
Robert Schoene provided some important fixes on top of my initial quick shot version which are integrated in this patch:
- Forgot some changes in power_end trace (TP_printk/variable names)
- Variable dummy in power_end must now be cpu_id
- Use static 64 bit variable instead of unsigned int for cpu_id
--- arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c | 3 --- arch/x86/kernel/process.c | 8 ++++---- drivers/cpufreq/cpufreq.c | 5 +++++ drivers/cpuidle/cpuidle.c | 2 +- include/trace/events/power.h | 27 +++++++++++++++------------ tools/perf/builtin-timechart.c | 11 ++++++----- 6 files changed, 31 insertions(+), 25 deletions(-)
arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c +0 -3
@@ -34,7 +34,6 @@#include <linux/compiler.h> #include <linux/dmi.h> #include <linux/slab.h> -#include <trace/events/power.h> #include <linux/acpi.h> #include <linux/io.h>
@@ -364,8 +363,6 @@ static int acpi_cpufreq_target(struct cpufreq_policy *policy, } } -trace_power_frequency (POWER_PSTATE , data->freq_table[next_state].frequency); -switch (data->cpu_feature) {case SYSTEM_INTEL_MSR_CAPABLE : cmd.type =SYSTEM_INTEL_MSR_CAPABLE ;
arch/x86/kernel/process.c +4 -4
@@ -371,7 +371,7 @@ static inline int hlt_use_halt(void)void default_idle (void ) {if (hlt_use_halt ()) { -trace_power_start (POWER_CSTATE ,1 ); +trace_power_start (POWER_CSTATE ,1 ,smp_processor_id ());current_thread_info ()->status &= ~TS_POLLING ;/* * TS_POLLING-cleared state must be visible before we
@@ -441,7 +441,7 @@ EXPORT_SYMBOL_GPL(cpu_idle_wait);*/ void mwait_idle_with_hints (unsigned long ax,unsigned long cx) { -trace_power_start (POWER_CSTATE , (ax>>4 )+1 ); +trace_power_start (POWER_CSTATE , (ax>>4 )+1 ,smp_processor_id ());if (!need_resched ()) {if (cpu_has (¤t_cpu_data,X86_FEATURE_CLFLUSH_MONITOR ))clflush ((void *)¤t_thread_info ()->flags);
@@ -457,7 +457,7 @@ void mwait_idle_with_hints(unsigned long ax, unsigned long cx)static void mwait_idle (void ) {if (!need_resched ()) { -trace_power_start (POWER_CSTATE ,1 ); +trace_power_start (POWER_CSTATE ,1 ,smp_processor_id ());if (cpu_has (¤t_cpu_data,X86_FEATURE_CLFLUSH_MONITOR ))clflush ((void *)¤t_thread_info ()->flags);
@@ -478,7 +478,7 @@ static void mwait_idle(void)*/ static void poll_idle (void ) { -trace_power_start (POWER_CSTATE ,0 ); +trace_power_start (POWER_CSTATE ,0 ,smp_processor_id ());local_irq_enable ();while (!need_resched ())cpu_relax ();
drivers/cpufreq/cpufreq.c +5 -0
@@ -29,6 +29,8 @@#include <linux/completion.h> #include <linux/mutex.h> +#include <trace/events/power.h> + #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_CORE, \ "cpufreq-core", msg)
@@ -354,6 +356,9 @@ void cpufreq_notify_transition(struct cpufreq_freqs *freqs, unsigned int state)case CPUFREQ_POSTCHANGE :adjust_jiffies (CPUFREQ_POSTCHANGE , freqs); +dprintk ("FREQ: %lu - CPU: %lu" , (unsigned long )freqs->new, + (unsigned long )freqs->cpu); +trace_power_frequency (POWER_PSTATE , freqs->new, freqs->cpu);srcu_notifier_call_chain (&cpufreq_transition_notifier_list,CPUFREQ_POSTCHANGE , freqs);if (likely (policy) &&likely (policy->cpu == freqs->cpu))
drivers/cpuidle/cpuidle.c +1 -1
@@ -95,7 +95,7 @@ static void cpuidle_idle_call(void)/* give the governor an opportunity to reflect on the outcome */ if (cpuidle_curr_governor->reflect) cpuidle_curr_governor->reflect (dev); -trace_power_end (0 ); +trace_power_end (smp_processor_id ()); }/**
include/trace/events/power.h +15 -12
@@ -18,52 +18,55 @@ enum {DECLARE_EVENT_CLASS (power, - TP_PROTO(unsigned int type, unsigned int state), + TP_PROTO(unsigned int type, unsigned int state, unsigned int cpu_id), -TP_ARGS (type, state), +TP_ARGS (type, state, cpu_id),TP_STRUCT__entry ( __field( u64, type ) __field( u64, state ) +__field ( u64, cpu_id ) ),TP_fast_assign ( __entry->type = type; __entry->state = state; + __entry->cpu_id = cpu_id; ), -TP_printk ("type=%lu state=%lu" , (unsigned long )__entry->type, (unsigned long )__entry->state) +TP_printk ("type=%lu state=%lu cpu_id=%lu" , (unsigned long )__entry->type, + (unsigned long )__entry->state, (unsigned long )__entry->cpu_id) );DEFINE_EVENT (power, power_start, - TP_PROTO(unsigned int type, unsigned int state), + TP_PROTO(unsigned int type, unsigned int state, unsigned int cpu_id), -TP_ARGS (type, state) +TP_ARGS (type, state, cpu_id) );DEFINE_EVENT (power, power_frequency, - TP_PROTO(unsigned int type, unsigned int state), + TP_PROTO(unsigned int type, unsigned int state, unsigned int cpu_id), -TP_ARGS (type, state) +TP_ARGS (type, state, cpu_id) );TRACE_EVENT (power_end, - TP_PROTO(int dummy), + TP_PROTO(unsigned int cpu_id), -TP_ARGS (dummy), +TP_ARGS (cpu_id),TP_STRUCT__entry ( -__field ( u64, dummy ) +__field ( u64, cpu_id ) ),TP_fast_assign ( - __entry->dummy = 0xffff; + __entry->cpu_id = cpu_id; ), -TP_printk ("dummy=%lu" , (unsigned long )__entry->dummy) +TP_printk ("cpu_id=%lu" , (unsigned long )__entry->cpu_id) );
tools/perf/builtin-timechart.c +6 -5
@@ -306,8 +306,9 @@ struct trace_entry {struct power_entry {struct trace_entry te; -s64 type; -s64 value; +u64 type; +u64 value; +u64 cpu_id; };#define TASK_COMM_LEN 16
@@ -504,13 +505,13 @@ static int process_sample_event(event_t *event, struct perf_session *session)return 0 ;if (strcmp (event_str,"power:power_start" ) ==0 ) -c_state_start (data.cpu, data.time, pe->value); +c_state_start (pe->cpu_id, data.time, pe->value);if (strcmp (event_str,"power:power_end" ) ==0 ) -c_state_end (data.cpu, data.time); +c_state_end (pe->cpu_id, data.time);if (strcmp (event_str,"power:power_frequency" ) ==0 ) -p_state_change (data.cpu, data.time, pe->value); +p_state_change (pe->cpu_id, data.time, pe->value);if (strcmp (event_str,"sched:sched_wakeup" ) ==0 )sched_wakeup (data.cpu, data.time, data.pid, te);
Signature · 1 line
1.6.3
Re: [PATCH 2/2] x86 cpufreq: Make trace_power_frequency cpufreq driver independent
Review
Quoted 6 lines · Thomas Renninger
(quoted diff) +6
-0+ trace_power_start(POWER_CSTATE, 1, smp_processor_id()); + trace_power_start(POWER_CSTATE, (ax>>4)+1, smp_processor_id()); + trace_power_start(POWER_CSTATE, 1, smp_processor_id()); + trace_power_start(POWER_CSTATE, 0, smp_processor_id()); + trace_power_frequency(POWER_PSTATE, freqs->new, freqs->cpu); + trace_power_end(smp_processor_id());
Extending power tracing to all cpufreq modules is obviously a good thing.
But why is trace_power_start() adding a CPU ID argument? CPU ids are already available and can be sampled via PERF_SAMPLE_CPU if needed. AFAICS only power_frequency needs a new 'target_cpu_id' field.
tools/perf/builtin-timechart.c
Quoted 4 lines
tools/perf/builtin-timechart.c +0
-0
Timechart is maintained by Arjan so we need an ack from him as well. I've seen some back and forth in the discussions - what's the technical resolution of that?
(Also, there's some whitespace noise in the patch.)
Thanks,
Ingo
Re: [PATCH 2/2] x86 cpufreq: Make trace_power_frequency cpufreq driver independent
Quoted 12 lines · On Friday 30 April 2010 11:08:14 Ingo Molnar wrote:
Review
Quoted 6 lines · Thomas Renninger
(quoted diff) +6
-0+ trace_power_start(POWER_CSTATE, 1, smp_processor_id()); + trace_power_start(POWER_CSTATE, (ax>>4)+1, smp_processor_id()); + trace_power_start(POWER_CSTATE, 1, smp_processor_id()); + trace_power_start(POWER_CSTATE, 0, smp_processor_id()); + trace_power_frequency(POWER_PSTATE, freqs->new, freqs->cpu); + trace_power_end(smp_processor_id());Extending power tracing to all cpufreq modules is obviously a good thing.
Quoted 3 lines
But why is
trace_power_start()adding a CPU ID argument? CPU ids are already available and can be sampled viaPERF_SAMPLE_CPUif needed. AFAICS onlypower_frequencyneeds a new 'target_cpu_id' field.
Currently the C-states get triggered on the machine you are running on. But you can have C-state dependencies (ACPI spec at least defines this the same way as done for P-states, not sure whether this exists in reality already, Linux does not evaluate them yet).
The same way we run into trouble with P-states (SW_ALL, SW_ANY, HW). Not sure how this will/could show up in reality in HW or implementation, but as these dependencies are already defined in spec, it sounds like a good idea to pass the CPU through the POWER_CSTATE events as well.
Compare with chapter 8.4.2.2 _CSD (C-State Dependency) of an ACPI spec 3.0 or newer.
There are also other architectures starting to use processor sleep states.
So this is not actually used (smp_processor_id() is the same as already tracked by trace event internally), but to make the interface more robust for the future. Changes in the trace_power layout later may hurt.
Quoted 10 lines
tools/perf/builtin-timechart.c
Quoted 4 lines
tools/perf/builtin-timechart.c +0
-0Timechart is maintained by Arjan so we need an ack from him as well. I've seen some back and forth in the discussions - what's the technical resolution of that?
(Also, there's some whitespace noise in the patch.)
Oops. Tell me if you like to push it and I send a checkpatch cleaned up version. Sorry about that.
Thomas