From 9590b7ba3fefdfe0c7741f5e2f61faf2ffcea19c Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Mon, 6 Jul 2009 22:01:31 +1000 Subject: [PATCH 01/31] perf_counter tools: Rename cache events to remove $ The cache events contain '$' which will hit shell variable expansion. To avoid confusion change this to 'cache', ie L1-d$-loads becomes L1-dcache-loads. Signed-off-by: Anton Blanchard Cc: Roland Dreier Cc: Jaswinder Singh Rajput Cc: Peter Zijlstra LKML-Reference: <20090706120131.GB4391@kryten> Signed-off-by: Ingo Molnar --- tools/perf/util/parse-events.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 5184959e061..518a33affe1 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -71,8 +71,8 @@ static char *sw_event_names[] = { #define MAX_ALIASES 8 static char *hw_cache[][MAX_ALIASES] = { - { "L1-d$", "l1-d", "l1d", "L1-data", }, - { "L1-i$", "l1-i", "l1i", "L1-instruction", }, + { "L1-dcache", "l1-d", "l1d", "L1-data", }, + { "L1-icache", "l1-i", "l1i", "L1-instruction", }, { "LLC", "L2" }, { "dTLB", "d-tlb", "Data-TLB", }, { "iTLB", "i-tlb", "Instruction-TLB", }, From 11d1578f9454159c43499d1d8fe8a7d728c176a3 Mon Sep 17 00:00:00 2001 From: Vince Weaver Date: Wed, 8 Jul 2009 17:46:14 -0400 Subject: [PATCH 02/31] perf_counter: Add P6 PMU support Add basic P6 PMU support. The P6 uses the EVNTSEL0 EN bit to enable/disable both its counters. We use this for the global enable/disable, and clear all config bits (except EN) to disable individual counters. Actual ia32 hardware doesn't support lfence, so use a locked op without side-effect to implement a full barrier. perf stat and perf record seem to function correctly. [a.p.zijlstra@chello.nl: cleanups and complete the enable/disable code] Signed-off-by: Vince Weaver Signed-off-by: Peter Zijlstra LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_counter.c | 227 +++++++++++++++++++++++++++-- tools/perf/perf.h | 8 +- 2 files changed, 220 insertions(+), 15 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index 36c3dc7b899..1910f39ff19 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -65,6 +65,44 @@ static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = { .enabled = 1, }; +/* + * Not sure about some of these + */ +static const u64 p6_perfmon_event_map[] = +{ + [PERF_COUNT_HW_CPU_CYCLES] = 0x0079, + [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, + [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0000, + [PERF_COUNT_HW_CACHE_MISSES] = 0x0000, + [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, + [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, + [PERF_COUNT_HW_BUS_CYCLES] = 0x0062, +}; + +static u64 p6_pmu_event_map(int event) +{ + return p6_perfmon_event_map[event]; +} + +static u64 p6_pmu_raw_event(u64 event) +{ +#define P6_EVNTSEL_EVENT_MASK 0x000000FFULL +#define P6_EVNTSEL_UNIT_MASK 0x0000FF00ULL +#define P6_EVNTSEL_EDGE_MASK 0x00040000ULL +#define P6_EVNTSEL_INV_MASK 0x00800000ULL +#define P6_EVNTSEL_COUNTER_MASK 0xFF000000ULL + +#define P6_EVNTSEL_MASK \ + (P6_EVNTSEL_EVENT_MASK | \ + P6_EVNTSEL_UNIT_MASK | \ + P6_EVNTSEL_EDGE_MASK | \ + P6_EVNTSEL_INV_MASK | \ + P6_EVNTSEL_COUNTER_MASK) + + return event & P6_EVNTSEL_MASK; +} + + /* * Intel PerfMon v3. Used on Core2 and later. */ @@ -726,6 +764,23 @@ static int __hw_perf_counter_init(struct perf_counter *counter) return 0; } +static void p6_pmu_disable_all(void) +{ + struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); + unsigned long val; + + if (!cpuc->enabled) + return; + + cpuc->enabled = 0; + barrier(); + + /* p6 only has one enable register */ + rdmsrl(MSR_P6_EVNTSEL0, val); + val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE; + wrmsrl(MSR_P6_EVNTSEL0, val); +} + static void intel_pmu_disable_all(void) { wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); @@ -767,6 +822,23 @@ void hw_perf_disable(void) return x86_pmu.disable_all(); } +static void p6_pmu_enable_all(void) +{ + struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); + unsigned long val; + + if (cpuc->enabled) + return; + + cpuc->enabled = 1; + barrier(); + + /* p6 only has one enable register */ + rdmsrl(MSR_P6_EVNTSEL0, val); + val |= ARCH_PERFMON_EVENTSEL0_ENABLE; + wrmsrl(MSR_P6_EVNTSEL0, val); +} + static void intel_pmu_enable_all(void) { wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); @@ -819,16 +891,13 @@ static inline void intel_pmu_ack_status(u64 ack) static inline void x86_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) { - int err; - err = checking_wrmsrl(hwc->config_base + idx, + (void)checking_wrmsrl(hwc->config_base + idx, hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE); } static inline void x86_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) { - int err; - err = checking_wrmsrl(hwc->config_base + idx, - hwc->config); + (void)checking_wrmsrl(hwc->config_base + idx, hwc->config); } static inline void @@ -836,13 +905,24 @@ intel_pmu_disable_fixed(struct hw_perf_counter *hwc, int __idx) { int idx = __idx - X86_PMC_IDX_FIXED; u64 ctrl_val, mask; - int err; mask = 0xfULL << (idx * 4); rdmsrl(hwc->config_base, ctrl_val); ctrl_val &= ~mask; - err = checking_wrmsrl(hwc->config_base, ctrl_val); + (void)checking_wrmsrl(hwc->config_base, ctrl_val); +} + +static inline void +p6_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) +{ + struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); + unsigned long val = ARCH_PERFMON_EVENTSEL0_ENABLE; + + if (!cpuc->enabled) + val = 0; + + (void)checking_wrmsrl(hwc->config_base + idx, val); } static inline void @@ -943,6 +1023,17 @@ intel_pmu_enable_fixed(struct hw_perf_counter *hwc, int __idx) err = checking_wrmsrl(hwc->config_base, ctrl_val); } +static void p6_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) +{ + struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); + + if (cpuc->enabled) + x86_pmu_enable_counter(hwc, idx); + else + x86_pmu_disable_counter(hwc, idx); +} + + static void intel_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) { if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { @@ -1176,6 +1267,49 @@ static void intel_pmu_reset(void) local_irq_restore(flags); } +static int p6_pmu_handle_irq(struct pt_regs *regs) +{ + struct perf_sample_data data; + struct cpu_hw_counters *cpuc; + struct perf_counter *counter; + struct hw_perf_counter *hwc; + int idx, handled = 0; + u64 val; + + data.regs = regs; + data.addr = 0; + + cpuc = &__get_cpu_var(cpu_hw_counters); + + for (idx = 0; idx < x86_pmu.num_counters; idx++) { + if (!test_bit(idx, cpuc->active_mask)) + continue; + + counter = cpuc->counters[idx]; + hwc = &counter->hw; + + val = x86_perf_counter_update(counter, hwc, idx); + if (val & (1ULL << (x86_pmu.counter_bits - 1))) + continue; + + /* + * counter overflow + */ + handled = 1; + data.period = counter->hw.last_period; + + if (!x86_perf_counter_set_period(counter, hwc, idx)) + continue; + + if (perf_counter_overflow(counter, 1, &data)) + p6_pmu_disable_counter(hwc, idx); + } + + if (handled) + inc_irq_stat(apic_perf_irqs); + + return handled; +} /* * This handler is triggered by the local APIC, so the APIC IRQ handling @@ -1185,14 +1319,13 @@ static int intel_pmu_handle_irq(struct pt_regs *regs) { struct perf_sample_data data; struct cpu_hw_counters *cpuc; - int bit, cpu, loops; + int bit, loops; u64 ack, status; data.regs = regs; data.addr = 0; - cpu = smp_processor_id(); - cpuc = &per_cpu(cpu_hw_counters, cpu); + cpuc = &__get_cpu_var(cpu_hw_counters); perf_disable(); status = intel_pmu_get_status(); @@ -1249,14 +1382,13 @@ static int amd_pmu_handle_irq(struct pt_regs *regs) struct cpu_hw_counters *cpuc; struct perf_counter *counter; struct hw_perf_counter *hwc; - int cpu, idx, handled = 0; + int idx, handled = 0; u64 val; data.regs = regs; data.addr = 0; - cpu = smp_processor_id(); - cpuc = &per_cpu(cpu_hw_counters, cpu); + cpuc = &__get_cpu_var(cpu_hw_counters); for (idx = 0; idx < x86_pmu.num_counters; idx++) { if (!test_bit(idx, cpuc->active_mask)) @@ -1353,6 +1485,32 @@ static __read_mostly struct notifier_block perf_counter_nmi_notifier = { .priority = 1 }; +static struct x86_pmu p6_pmu = { + .name = "p6", + .handle_irq = p6_pmu_handle_irq, + .disable_all = p6_pmu_disable_all, + .enable_all = p6_pmu_enable_all, + .enable = p6_pmu_enable_counter, + .disable = p6_pmu_disable_counter, + .eventsel = MSR_P6_EVNTSEL0, + .perfctr = MSR_P6_PERFCTR0, + .event_map = p6_pmu_event_map, + .raw_event = p6_pmu_raw_event, + .max_events = ARRAY_SIZE(p6_perfmon_event_map), + .max_period = (1ULL << 31) - 1, + .version = 0, + .num_counters = 2, + /* + * Counters have 40 bits implemented. However they are designed such + * that bits [32-39] are sign extensions of bit 31. As such the + * effective width of a counter for P6-like PMU is 32 bits only. + * + * See IA-32 Intel Architecture Software developer manual Vol 3B + */ + .counter_bits = 32, + .counter_mask = (1ULL << 32) - 1, +}; + static struct x86_pmu intel_pmu = { .name = "Intel", .handle_irq = intel_pmu_handle_irq, @@ -1392,6 +1550,41 @@ static struct x86_pmu amd_pmu = { .max_period = (1ULL << 47) - 1, }; +static int p6_pmu_init(void) +{ + int high, low; + + switch (boot_cpu_data.x86_model) { + case 1: + case 3: /* Pentium Pro */ + case 5: + case 6: /* Pentium II */ + case 7: + case 8: + case 11: /* Pentium III */ + break; + case 9: + case 13: + /* for Pentium M, we need to check if PMU exist */ + rdmsr(MSR_IA32_MISC_ENABLE, low, high); + if (low & MSR_IA32_MISC_ENABLE_EMON) + break; + default: + pr_cont("unsupported p6 CPU model %d ", + boot_cpu_data.x86_model); + return -ENODEV; + } + + if (!cpu_has_apic) { + pr_info("no Local APIC, try rebooting with lapic"); + return -ENODEV; + } + + x86_pmu = p6_pmu; + + return 0; +} + static int intel_pmu_init(void) { union cpuid10_edx edx; @@ -1400,8 +1593,14 @@ static int intel_pmu_init(void) unsigned int ebx; int version; - if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) + if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { + /* check for P6 processor family */ + if (boot_cpu_data.x86 == 6) { + return p6_pmu_init(); + } else { return -ENODEV; + } + } /* * Check whether the Architectural PerfMon supports diff --git a/tools/perf/perf.h b/tools/perf/perf.h index 27887c91643..53bb9550def 100644 --- a/tools/perf/perf.h +++ b/tools/perf/perf.h @@ -1,7 +1,13 @@ #ifndef _PERF_PERF_H #define _PERF_PERF_H -#if defined(__x86_64__) || defined(__i386__) +#if defined(__i386__) +#include "../../arch/x86/include/asm/unistd.h" +#define rmb() asm volatile("lock; addl $0,0(%%esp)" ::: "memory") +#define cpu_relax() asm volatile("rep; nop" ::: "memory"); +#endif + +#if defined(__x86_64__) #include "../../arch/x86/include/asm/unistd.h" #define rmb() asm volatile("lfence" ::: "memory") #define cpu_relax() asm volatile("rep; nop" ::: "memory"); From 9c74fb50867e8fb5f3be3be06716492c0f79309e Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 8 Jul 2009 10:21:41 +0200 Subject: [PATCH 03/31] perf_counter: Fix up P6 PMU details The P6 doesn't seem to support cache ref/hit/miss counts, so we extend the generic hardware event codes to have 0 and -1 mean the same thing as for the generic cache events. Furthermore, it turns out the 0 event does not count (that is, its reported that on PPro it actually does count something), therefore use a event configuration that's specified not to count to disable the counters. Signed-off-by: Peter Zijlstra LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_counter.c | 28 +++++++++++++++++++++++----- 1 file changed, 23 insertions(+), 5 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index 1910f39ff19..c7cc6eac14e 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -84,6 +84,14 @@ static u64 p6_pmu_event_map(int event) return p6_perfmon_event_map[event]; } +/* + * Counter setting that is specified not to count anything. + * We use this to effectively disable a counter. + * + * L2_RQSTS with 0 MESI unit mask. + */ +#define P6_NOP_COUNTER 0x0000002EULL + static u64 p6_pmu_raw_event(u64 event) { #define P6_EVNTSEL_EVENT_MASK 0x000000FFULL @@ -704,6 +712,7 @@ static int __hw_perf_counter_init(struct perf_counter *counter) { struct perf_counter_attr *attr = &counter->attr; struct hw_perf_counter *hwc = &counter->hw; + u64 config; int err; if (!x86_pmu_initialized()) @@ -756,10 +765,19 @@ static int __hw_perf_counter_init(struct perf_counter *counter) if (attr->config >= x86_pmu.max_events) return -EINVAL; + /* * The generic map: */ - hwc->config |= x86_pmu.event_map(attr->config); + config = x86_pmu.event_map(attr->config); + + if (config == 0) + return -ENOENT; + + if (config == -1LL) + return -EINVAL; + + hwc->config |= config; return 0; } @@ -767,7 +785,7 @@ static int __hw_perf_counter_init(struct perf_counter *counter) static void p6_pmu_disable_all(void) { struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); - unsigned long val; + u64 val; if (!cpuc->enabled) return; @@ -917,10 +935,10 @@ static inline void p6_pmu_disable_counter(struct hw_perf_counter *hwc, int idx) { struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); - unsigned long val = ARCH_PERFMON_EVENTSEL0_ENABLE; + u64 val = P6_NOP_COUNTER; - if (!cpuc->enabled) - val = 0; + if (cpuc->enabled) + val |= ARCH_PERFMON_EVENTSEL0_ENABLE; (void)checking_wrmsrl(hwc->config_base + idx, val); } From 984b838ce69c063a91b87550598ab7f3439dd94a Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Fri, 10 Jul 2009 09:59:56 +0200 Subject: [PATCH 04/31] perf_counter: Clean up global vs counter enable Ingo noticed that both AMD and P6 call x86_pmu_disable_counter() on *_pmu_enable_counter(). This is because we rely on the side effect of that call to program the event config but not touch the EN bit. We change that for AMD by having enable_all() simply write the full config in, and for P6 by explicitly coding it. Signed-off-by: Peter Zijlstra LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_counter.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index c7cc6eac14e..bed1c4c2f25 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -874,13 +874,13 @@ static void amd_pmu_enable_all(void) barrier(); for (idx = 0; idx < x86_pmu.num_counters; idx++) { + struct perf_counter *counter = cpuc->counters[idx]; u64 val; if (!test_bit(idx, cpuc->active_mask)) continue; - rdmsrl(MSR_K7_EVNTSEL0 + idx, val); - if (val & ARCH_PERFMON_EVENTSEL0_ENABLE) - continue; + + val = counter->hw.config; val |= ARCH_PERFMON_EVENTSEL0_ENABLE; wrmsrl(MSR_K7_EVNTSEL0 + idx, val); } @@ -1044,11 +1044,13 @@ intel_pmu_enable_fixed(struct hw_perf_counter *hwc, int __idx) static void p6_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) { struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters); + u64 val; + val = hwc->config; if (cpuc->enabled) - x86_pmu_enable_counter(hwc, idx); - else - x86_pmu_disable_counter(hwc, idx); + val |= ARCH_PERFMON_EVENTSEL0_ENABLE; + + (void)checking_wrmsrl(hwc->config_base + idx, val); } @@ -1068,8 +1070,6 @@ static void amd_pmu_enable_counter(struct hw_perf_counter *hwc, int idx) if (cpuc->enabled) x86_pmu_enable_counter(hwc, idx); - else - x86_pmu_disable_counter(hwc, idx); } static int From 71a851b4d2a815adcfac09c1adda7ef6811fde66 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Fri, 10 Jul 2009 09:06:56 +0200 Subject: [PATCH 05/31] perf_counter: Stop open coding unclone_ctx Instead of open coding the unclone context thingy, put it in a common function. Signed-off-by: Peter Zijlstra LKML-Reference: Signed-off-by: Ingo Molnar --- kernel/perf_counter.c | 36 +++++++++++++++++------------------- 1 file changed, 17 insertions(+), 19 deletions(-) diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index d55a50da234..8bf997d86bf 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -146,6 +146,14 @@ static void put_ctx(struct perf_counter_context *ctx) } } +static void unclone_ctx(struct perf_counter_context *ctx) +{ + if (ctx->parent_ctx) { + put_ctx(ctx->parent_ctx); + ctx->parent_ctx = NULL; + } +} + /* * Get the perf_counter_context for a task and lock it. * This has to cope with with the fact that until it is locked, @@ -1463,10 +1471,8 @@ static void perf_counter_enable_on_exec(struct task_struct *task) /* * Unclone this context if we enabled any counter. */ - if (enabled && ctx->parent_ctx) { - put_ctx(ctx->parent_ctx); - ctx->parent_ctx = NULL; - } + if (enabled) + unclone_ctx(ctx); spin_unlock(&ctx->lock); @@ -1526,7 +1532,6 @@ __perf_counter_init_context(struct perf_counter_context *ctx, static struct perf_counter_context *find_get_context(pid_t pid, int cpu) { - struct perf_counter_context *parent_ctx; struct perf_counter_context *ctx; struct perf_cpu_context *cpuctx; struct task_struct *task; @@ -1586,11 +1591,7 @@ static struct perf_counter_context *find_get_context(pid_t pid, int cpu) retry: ctx = perf_lock_task_context(task, &flags); if (ctx) { - parent_ctx = ctx->parent_ctx; - if (parent_ctx) { - put_ctx(parent_ctx); - ctx->parent_ctx = NULL; /* no longer a clone */ - } + unclone_ctx(ctx); spin_unlock_irqrestore(&ctx->lock, flags); } @@ -4255,15 +4256,12 @@ void perf_counter_exit_task(struct task_struct *child) */ spin_lock(&child_ctx->lock); child->perf_counter_ctxp = NULL; - if (child_ctx->parent_ctx) { - /* - * This context is a clone; unclone it so it can't get - * swapped to another process while we're removing all - * the counters from it. - */ - put_ctx(child_ctx->parent_ctx); - child_ctx->parent_ctx = NULL; - } + /* + * If this context is a clone; unclone it so it can't get + * swapped to another process while we're removing all + * the counters from it. + */ + unclone_ctx(child_ctx); spin_unlock(&child_ctx->lock); local_irq_restore(flags); From 52d422de22b26d96bbdfbc605eb31c2994df6d0b Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Fri, 10 Jul 2009 22:47:28 -0300 Subject: [PATCH 06/31] perf report: Adjust column width to the values sampled Auto-adjust column width of perf report output to the longest occuring string length. Example: [acme@doppio pahole]$ perf report --sort comm,dso,symbol | head -13 12.79% pahole /usr/lib64/libdw-0.141.so [.] __libdw_find_attr 8.90% pahole /lib64/libc-2.10.1.so [.] _int_malloc 8.68% pahole /usr/lib64/libdw-0.141.so [.] __libdw_form_val_len 8.15% pahole /lib64/libc-2.10.1.so [.] __GI_strcmp 6.80% pahole /lib64/libc-2.10.1.so [.] __tsearch 5.54% pahole ./build/libdwarves.so.1.0.0 [.] tag__recode_dwarf_type [acme@doppio pahole]$ [acme@doppio pahole]$ perf report --sort comm,dso,symbol -d /lib64/libc-2.10.1.so | head -10 21.92% pahole /lib64/libc-2.10.1.so [.] _int_malloc 20.08% pahole /lib64/libc-2.10.1.so [.] __GI_strcmp 16.75% pahole /lib64/libc-2.10.1.so [.] __tsearch [acme@doppio pahole]$ Also add these extra options to control the new behaviour: -w, --field-width Force each column width to the provided list, for large terminal readability. -t, --field-separator: Use a special separator character and don't pad with spaces, replacing all occurances of this separator in symbol names (and other output) with a '.' character, that thus it's the only non valid separator. Signed-off-by: Arnaldo Carvalho de Melo Cc: Frederic Weisbecker Cc: Mike Galbraith Cc: Peter Zijlstra Cc: Paul Mackerras LKML-Reference: <20090711014728.GH3452@ghostprotocols.net> Signed-off-by: Ingo Molnar --- tools/perf/Documentation/perf-report.txt | 12 ++ tools/perf/builtin-report.c | 174 ++++++++++++++++++----- tools/perf/util/include/linux/kernel.h | 8 ++ tools/perf/util/symbol.c | 1 + tools/perf/util/symbol.h | 1 + 5 files changed, 164 insertions(+), 32 deletions(-) diff --git a/tools/perf/Documentation/perf-report.txt b/tools/perf/Documentation/perf-report.txt index 8aa3f8c8870..05774dfbd14 100644 --- a/tools/perf/Documentation/perf-report.txt +++ b/tools/perf/Documentation/perf-report.txt @@ -33,6 +33,18 @@ OPTIONS Only consider these symbols. CSV that understands file://filename entries. +-w:: +--field-width=:: + Force each column width to the provided list, for large terminal + readability. + +-t:: +--field-separator=:: + + Use a special separator character and don't pad with spaces, replacing + all occurances of this separator in symbol names (and other output) + with a '.' character, that thus it's the only non valid separator. + SEE ALSO -------- linkperf:perf-stat[1] diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 4e5cc266311..740da43f313 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -33,8 +33,10 @@ static char *vmlinux = NULL; static char default_sort_order[] = "comm,dso"; static char *sort_order = default_sort_order; -static char *dso_list_str, *comm_list_str, *sym_list_str; +static char *dso_list_str, *comm_list_str, *sym_list_str, + *col_width_list_str; static struct strlist *dso_list, *comm_list, *sym_list; +static char *field_sep; static int input; static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV; @@ -129,6 +131,33 @@ typedef union event_union { struct read_event read; } event_t; +static int repsep_fprintf(FILE *fp, const char *fmt, ...) +{ + int n; + va_list ap; + + va_start(ap, fmt); + if (!field_sep) + n = vfprintf(fp, fmt, ap); + else { + char *bf = NULL; + n = vasprintf(&bf, fmt, ap); + if (n > 0) { + char *sep = bf; + while (1) { + sep = strchr(sep, *field_sep); + if (sep == NULL) + break; + *sep = '.'; + } + } + fputs(bf, fp); + free(bf); + } + va_end(ap); + return n; +} + static LIST_HEAD(dsos); static struct dso *kernel_dso; static struct dso *vdso; @@ -360,12 +389,28 @@ static struct thread *thread__new(pid_t pid) return self; } +static unsigned int dsos__col_width, + comms__col_width, + threads__col_width; + static int thread__set_comm(struct thread *self, const char *comm) { if (self->comm) free(self->comm); self->comm = strdup(comm); - return self->comm ? 0 : -ENOMEM; + if (!self->comm) + return -ENOMEM; + + if (!col_width_list_str && !field_sep && + (!comm_list || strlist__has_entry(comm_list, comm))) { + unsigned int slen = strlen(comm); + if (slen > comms__col_width) { + comms__col_width = slen; + threads__col_width = slen + 6; + } + } + + return 0; } static size_t thread__fprintf(struct thread *self, FILE *fp) @@ -536,7 +581,8 @@ struct sort_entry { int64_t (*cmp)(struct hist_entry *, struct hist_entry *); int64_t (*collapse)(struct hist_entry *, struct hist_entry *); - size_t (*print)(FILE *fp, struct hist_entry *); + size_t (*print)(FILE *fp, struct hist_entry *, unsigned int width); + unsigned int *width; }; static int64_t cmp_null(void *l, void *r) @@ -558,15 +604,17 @@ sort__thread_cmp(struct hist_entry *left, struct hist_entry *right) } static size_t -sort__thread_print(FILE *fp, struct hist_entry *self) +sort__thread_print(FILE *fp, struct hist_entry *self, unsigned int width) { - return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid); + return repsep_fprintf(fp, "%*s:%5d", width - 6, + self->thread->comm ?: "", self->thread->pid); } static struct sort_entry sort_thread = { - .header = " Command: Pid", + .header = "Command: Pid", .cmp = sort__thread_cmp, .print = sort__thread_print, + .width = &threads__col_width, }; /* --sort comm */ @@ -590,16 +638,17 @@ sort__comm_collapse(struct hist_entry *left, struct hist_entry *right) } static size_t -sort__comm_print(FILE *fp, struct hist_entry *self) +sort__comm_print(FILE *fp, struct hist_entry *self, unsigned int width) { - return fprintf(fp, "%16s", self->thread->comm); + return repsep_fprintf(fp, "%*s", width, self->thread->comm); } static struct sort_entry sort_comm = { - .header = " Command", + .header = "Command", .cmp = sort__comm_cmp, .collapse = sort__comm_collapse, .print = sort__comm_print, + .width = &comms__col_width, }; /* --sort dso */ @@ -617,18 +666,19 @@ sort__dso_cmp(struct hist_entry *left, struct hist_entry *right) } static size_t -sort__dso_print(FILE *fp, struct hist_entry *self) +sort__dso_print(FILE *fp, struct hist_entry *self, unsigned int width) { if (self->dso) - return fprintf(fp, "%-25s", self->dso->name); + return repsep_fprintf(fp, "%-*s", width, self->dso->name); - return fprintf(fp, "%016llx ", (u64)self->ip); + return repsep_fprintf(fp, "%*llx", width, (u64)self->ip); } static struct sort_entry sort_dso = { - .header = "Shared Object ", + .header = "Shared Object", .cmp = sort__dso_cmp, .print = sort__dso_print, + .width = &dsos__col_width, }; /* --sort symbol */ @@ -648,22 +698,23 @@ sort__sym_cmp(struct hist_entry *left, struct hist_entry *right) } static size_t -sort__sym_print(FILE *fp, struct hist_entry *self) +sort__sym_print(FILE *fp, struct hist_entry *self, unsigned int width __used) { size_t ret = 0; if (verbose) - ret += fprintf(fp, "%#018llx ", (u64)self->ip); + ret += repsep_fprintf(fp, "%#018llx ", (u64)self->ip); if (self->sym) { - ret += fprintf(fp, "[%c] %s", + ret += repsep_fprintf(fp, "[%c] %s", self->dso == kernel_dso ? 'k' : self->dso == hypervisor_dso ? 'h' : '.', self->sym->name); if (self->sym->module) - ret += fprintf(fp, "\t[%s]", self->sym->module->name); + ret += repsep_fprintf(fp, "\t[%s]", + self->sym->module->name); } else { - ret += fprintf(fp, "%#016llx", (u64)self->ip); + ret += repsep_fprintf(fp, "%#016llx", (u64)self->ip); } return ret; @@ -690,19 +741,19 @@ sort__parent_cmp(struct hist_entry *left, struct hist_entry *right) } static size_t -sort__parent_print(FILE *fp, struct hist_entry *self) +sort__parent_print(FILE *fp, struct hist_entry *self, unsigned int width) { - size_t ret = 0; - - ret += fprintf(fp, "%-20s", self->parent ? self->parent->name : "[other]"); - - return ret; + return repsep_fprintf(fp, "%-*s", width, + self->parent ? self->parent->name : "[other]"); } +static unsigned int parent_symbol__col_width; + static struct sort_entry sort_parent = { - .header = "Parent symbol ", + .header = "Parent symbol", .cmp = sort__parent_cmp, .print = sort__parent_print, + .width = &parent_symbol__col_width, }; static int sort__need_collapse = 0; @@ -967,17 +1018,18 @@ hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples) return 0; if (total_samples) - ret = percent_color_fprintf(fp, " %6.2f%%", - (self->count * 100.0) / total_samples); + ret = percent_color_fprintf(fp, + field_sep ? "%.2f" : " %6.2f%%", + (self->count * 100.0) / total_samples); else - ret = fprintf(fp, "%12Ld ", self->count); + ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count); list_for_each_entry(se, &hist_entry__sort_list, list) { if (exclude_other && (se == &sort_parent)) continue; - fprintf(fp, " "); - ret += se->print(fp, self); + fprintf(fp, "%s", field_sep ?: " "); + ret += se->print(fp, self, se->width ? *se->width : 0); } ret += fprintf(fp, "\n"); @@ -992,6 +1044,18 @@ hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples) * */ +static void dso__calc_col_width(struct dso *self) +{ + if (!col_width_list_str && !field_sep && + (!dso_list || strlist__has_entry(dso_list, self->name))) { + unsigned int slen = strlen(self->name); + if (slen > dsos__col_width) + dsos__col_width = slen; + } + + self->slen_calculated = 1; +} + static struct symbol * resolve_symbol(struct thread *thread, struct map **mapp, struct dso **dsop, u64 *ipp) @@ -1011,6 +1075,14 @@ resolve_symbol(struct thread *thread, struct map **mapp, map = thread__find_map(thread, ip); if (map != NULL) { + /* + * We have to do this here as we may have a dso + * with no symbol hit that has a name longer than + * the ones with symbols sampled. + */ + if (!map->dso->slen_calculated) + dso__calc_col_width(map->dso); + if (mapp) *mapp = map; got_map: @@ -1282,6 +1354,8 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) struct sort_entry *se; struct rb_node *nd; size_t ret = 0; + unsigned int width; + char *col_width = col_width_list_str; fprintf(fp, "\n"); fprintf(fp, "#\n"); @@ -1292,10 +1366,29 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) list_for_each_entry(se, &hist_entry__sort_list, list) { if (exclude_other && (se == &sort_parent)) continue; - fprintf(fp, " %s", se->header); + if (field_sep) { + fprintf(fp, "%c%s", *field_sep, se->header); + continue; + } + width = strlen(se->header); + if (se->width) { + if (col_width_list_str) { + if (col_width) { + *se->width = atoi(col_width); + col_width = strchr(col_width, ','); + if (col_width) + ++col_width; + } + } + width = *se->width = max(*se->width, width); + } + fprintf(fp, " %*s", width, se->header); } fprintf(fp, "\n"); + if (field_sep) + goto print_entries; + fprintf(fp, "# ........"); list_for_each_entry(se, &hist_entry__sort_list, list) { unsigned int i; @@ -1304,13 +1397,18 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) continue; fprintf(fp, " "); - for (i = 0; i < strlen(se->header); i++) + if (se->width) + width = *se->width; + else + width = strlen(se->header); + for (i = 0; i < width; i++) fprintf(fp, "."); } fprintf(fp, "\n"); fprintf(fp, "#\n"); +print_entries: for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) { pos = rb_entry(nd, struct hist_entry, rb_node); ret += hist_entry__fprintf(fp, pos, total_samples); @@ -1900,6 +1998,12 @@ static const struct option options[] = { "only consider symbols in these comms"), OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]", "only consider these symbols"), + OPT_STRING('w', "column-widths", &col_width_list_str, + "width[,width...]", + "don't try to adjust column width, use these fixed values"), + OPT_STRING('t', "field-separator", &field_sep, "separator", + "separator for columns, no spaces will be added between " + "columns '.' is reserved."), OPT_END() }; @@ -1956,6 +2060,12 @@ int cmd_report(int argc, const char **argv, const char *prefix __used) setup_list(&comm_list, comm_list_str, "comm"); setup_list(&sym_list, sym_list_str, "symbol"); + if (field_sep && *field_sep == '.') { + fputs("'.' is the only non valid --field-separator argument\n", + stderr); + exit(129); + } + setup_pager(); return __cmd_report(); diff --git a/tools/perf/util/include/linux/kernel.h b/tools/perf/util/include/linux/kernel.h index 99c1b3d1edd..a6b87390cb5 100644 --- a/tools/perf/util/include/linux/kernel.h +++ b/tools/perf/util/include/linux/kernel.h @@ -18,4 +18,12 @@ (type *)((char *)__mptr - offsetof(type, member)); }) #endif +#ifndef max +#define max(x, y) ({ \ + typeof(x) _max1 = (x); \ + typeof(y) _max2 = (y); \ + (void) (&_max1 == &_max2); \ + _max1 > _max2 ? _max1 : _max2; }) +#endif + #endif diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c index 4683b67b5ee..8efe7e41109 100644 --- a/tools/perf/util/symbol.c +++ b/tools/perf/util/symbol.c @@ -65,6 +65,7 @@ struct dso *dso__new(const char *name, unsigned int sym_priv_size) self->syms = RB_ROOT; self->sym_priv_size = sym_priv_size; self->find_symbol = dso__find_symbol; + self->slen_calculated = 0; } return self; diff --git a/tools/perf/util/symbol.h b/tools/perf/util/symbol.h index 7918cffb23c..2f92b21c712 100644 --- a/tools/perf/util/symbol.h +++ b/tools/perf/util/symbol.h @@ -25,6 +25,7 @@ struct dso { struct symbol *(*find_symbol)(struct dso *, u64 ip); unsigned int sym_priv_size; unsigned char adjust_symbols; + unsigned char slen_calculated; char name[0]; }; From 60c1baf1248e00d423604f018c8af1cf750ad885 Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Sat, 11 Jul 2009 12:18:33 -0300 Subject: [PATCH 07/31] perf report: Tidy up reporting of symbols not found Always printing the level info about if it is in the kernel, hypervisor or userspace as that is in the hist_entry. Signed-off-by: Arnaldo Carvalho de Melo LKML-Reference: <1247325517-12272-1-git-send-email-acme@redhat.com> Signed-off-by: Ingo Molnar --- tools/perf/builtin-report.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 740da43f313..617f4cb7f16 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -705,10 +705,9 @@ sort__sym_print(FILE *fp, struct hist_entry *self, unsigned int width __used) if (verbose) ret += repsep_fprintf(fp, "%#018llx ", (u64)self->ip); + ret += repsep_fprintf(fp, "[%c] ", self->level); if (self->sym) { - ret += repsep_fprintf(fp, "[%c] %s", - self->dso == kernel_dso ? 'k' : - self->dso == hypervisor_dso ? 'h' : '.', self->sym->name); + ret += repsep_fprintf(fp, "%s", self->sym->name); if (self->sym->module) ret += repsep_fprintf(fp, "\t[%s]", From 27d0fd410c3cee00ece2e55f4354a7a9ec1a6a6a Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Sat, 11 Jul 2009 12:18:34 -0300 Subject: [PATCH 08/31] strlist: Introduce strlist__entry and strlist__nr_entries methods The strlist__entry method allows accessing strlists like an array, will be used in the 'perf report' to access the first entry. We now keep the nr_entries so that we can check if we have just one entry, will be used in 'perf report' to improve the output by showing just at the top when we have just, say, one DSO. While at it use nr_entries to optimize strlist__is_empty by not using the far more costly rb_first based implementation. Signed-off-by: Arnaldo Carvalho de Melo LKML-Reference: <1247325517-12272-2-git-send-email-acme@redhat.com> Signed-off-by: Ingo Molnar --- tools/perf/util/strlist.c | 20 ++++++++++++++++++-- tools/perf/util/strlist.h | 11 +++++++++-- 2 files changed, 27 insertions(+), 4 deletions(-) diff --git a/tools/perf/util/strlist.c b/tools/perf/util/strlist.c index 025a78edfff..7ad38171dc2 100644 --- a/tools/perf/util/strlist.c +++ b/tools/perf/util/strlist.c @@ -64,6 +64,7 @@ int strlist__add(struct strlist *self, const char *new_entry) rb_link_node(&sn->rb_node, parent, p); rb_insert_color(&sn->rb_node, &self->entries); + ++self->nr_entries; return 0; } @@ -155,8 +156,9 @@ struct strlist *strlist__new(bool dupstr, const char *slist) struct strlist *self = malloc(sizeof(*self)); if (self != NULL) { - self->entries = RB_ROOT; - self->dupstr = dupstr; + self->entries = RB_ROOT; + self->dupstr = dupstr; + self->nr_entries = 0; if (slist && strlist__parse_list(self, slist) != 0) goto out_error; } @@ -182,3 +184,17 @@ void strlist__delete(struct strlist *self) free(self); } } + +struct str_node *strlist__entry(const struct strlist *self, unsigned int idx) +{ + struct rb_node *nd; + + for (nd = rb_first(&self->entries); nd; nd = rb_next(nd)) { + struct str_node *pos = rb_entry(nd, struct str_node, rb_node); + + if (!idx--) + return pos; + } + + return NULL; +} diff --git a/tools/perf/util/strlist.h b/tools/perf/util/strlist.h index 2fdcfee8758..921818e44a5 100644 --- a/tools/perf/util/strlist.h +++ b/tools/perf/util/strlist.h @@ -11,7 +11,8 @@ struct str_node { struct strlist { struct rb_root entries; - bool dupstr; + unsigned int nr_entries; + bool dupstr; }; struct strlist *strlist__new(bool dupstr, const char *slist); @@ -21,11 +22,17 @@ void strlist__remove(struct strlist *self, struct str_node *sn); int strlist__load(struct strlist *self, const char *filename); int strlist__add(struct strlist *self, const char *str); +struct str_node *strlist__entry(const struct strlist *self, unsigned int idx); bool strlist__has_entry(struct strlist *self, const char *entry); static inline bool strlist__empty(const struct strlist *self) { - return rb_first(&self->entries) == NULL; + return self->nr_entries == 0; +} + +static inline unsigned int strlist__nr_entries(const struct strlist *self) +{ + return self->nr_entries; } int strlist__parse_list(struct strlist *self, const char *s); From 021191b35cdfb1b5ee6e78ed5ae010114a40902c Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Sat, 11 Jul 2009 12:18:35 -0300 Subject: [PATCH 09/31] perf report: Make the output more compact When we filter by column content we may end up with a column that has the same value for all the lines. So remove that column and tell its unique value on the top, as a comment. Example: [acme@doppio pahole]$ perf report --sort comm,dso,symbol -d ./build/libdwarves.so.1.0.0 -C pahole | head -15 # dso: ./build/libdwarves.so.1.0.0 # comm: pahole # Samples: 58409 # # Overhead Symbol # ........ ...... # 20.93% [.] tag__recode_dwarf_type 14.94% [.] namespace__recode_dwarf_types 10.38% [.] cu__table_add_tag 6.69% [.] __die__process_tag 5.05% [.] die__process_function 4.70% [.] list__for_all_tags 3.68% [.] tag__init 3.48% [.] die__create_new_parameter [acme@doppio pahole]$ Signed-off-by: Arnaldo Carvalho de Melo LKML-Reference: <1247325517-12272-3-git-send-email-acme@redhat.com> Signed-off-by: Ingo Molnar --- tools/perf/builtin-report.c | 36 +++++++++++++++++++++--------------- 1 file changed, 21 insertions(+), 15 deletions(-) diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 617f4cb7f16..f3422121d85 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -583,6 +583,7 @@ struct sort_entry { int64_t (*collapse)(struct hist_entry *, struct hist_entry *); size_t (*print)(FILE *fp, struct hist_entry *, unsigned int width); unsigned int *width; + bool elide; }; static int64_t cmp_null(void *l, void *r) @@ -1024,7 +1025,7 @@ hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples) ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count); list_for_each_entry(se, &hist_entry__sort_list, list) { - if (exclude_other && (se == &sort_parent)) + if (se->elide) continue; fprintf(fp, "%s", field_sep ?: " "); @@ -1079,7 +1080,7 @@ resolve_symbol(struct thread *thread, struct map **mapp, * with no symbol hit that has a name longer than * the ones with symbols sampled. */ - if (!map->dso->slen_calculated) + if (!sort_dso.elide && !map->dso->slen_calculated) dso__calc_col_width(map->dso); if (mapp) @@ -1356,14 +1357,12 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) unsigned int width; char *col_width = col_width_list_str; - fprintf(fp, "\n"); - fprintf(fp, "#\n"); - fprintf(fp, "# (%Ld samples)\n", (u64)total_samples); + fprintf(fp, "# Samples: %Ld\n", (u64)total_samples); fprintf(fp, "#\n"); fprintf(fp, "# Overhead"); list_for_each_entry(se, &hist_entry__sort_list, list) { - if (exclude_other && (se == &sort_parent)) + if (se->elide) continue; if (field_sep) { fprintf(fp, "%c%s", *field_sep, se->header); @@ -1392,7 +1391,7 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) list_for_each_entry(se, &hist_entry__sort_list, list) { unsigned int i; - if (exclude_other && (se == &sort_parent)) + if (se->elide) continue; fprintf(fp, " "); @@ -2022,7 +2021,8 @@ static void setup_sorting(void) } static void setup_list(struct strlist **list, const char *list_str, - const char *list_name) + struct sort_entry *se, const char *list_name, + FILE *fp) { if (list_str) { *list = strlist__new(true, list_str); @@ -2031,6 +2031,11 @@ static void setup_list(struct strlist **list, const char *list_str, list_name); exit(129); } + if (strlist__nr_entries(*list) == 1) { + fprintf(fp, "# %s: %s\n", list_name, + strlist__entry(*list, 0)->s); + se->elide = true; + } } } @@ -2044,9 +2049,10 @@ int cmd_report(int argc, const char **argv, const char *prefix __used) setup_sorting(); - if (parent_pattern != default_parent_pattern) + if (parent_pattern != default_parent_pattern) { sort_dimension__add("parent"); - else + sort_parent.elide = 1; + } else exclude_other = 0; /* @@ -2055,9 +2061,11 @@ int cmd_report(int argc, const char **argv, const char *prefix __used) if (argc) usage_with_options(report_usage, options); - setup_list(&dso_list, dso_list_str, "dso"); - setup_list(&comm_list, comm_list_str, "comm"); - setup_list(&sym_list, sym_list_str, "symbol"); + setup_pager(); + + setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout); + setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout); + setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout); if (field_sep && *field_sep == '.') { fputs("'.' is the only non valid --field-separator argument\n", @@ -2065,7 +2073,5 @@ int cmd_report(int argc, const char **argv, const char *prefix __used) exit(129); } - setup_pager(); - return __cmd_report(); } From a25e46c46311316cd1b3f27f8bb036df1693c032 Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Sat, 11 Jul 2009 12:18:36 -0300 Subject: [PATCH 10/31] perf_counter tools: PLT info is stripped in -debuginfo packages So we need to get the richer .symtab from the debuginfo packages but the PLT info from the original DSO where we have just the leaner .dynsym symtab. Example: | [acme@doppio pahole]$ perf report --sort comm,dso,symbol > before | [acme@doppio pahole]$ perf report --sort comm,dso,symbol > after | [acme@doppio pahole]$ diff -U1 before after | --- before 2009-07-11 11:04:22.688595741 -0300 | +++ after 2009-07-11 11:04:33.380595676 -0300 | @@ -80,3 +80,2 @@ | 0.07% pahole ./build/pahole [.] pahole_stealer | - 0.06% pahole /usr/lib64/libdw-0.141.so [.] 0x00000000007140 | 0.06% pahole /usr/lib64/libdw-0.141.so [.] __libdw_getabbrev | @@ -91,2 +90,3 @@ | 0.06% pahole [kernel] [k] free_hot_cold_page | + 0.06% pahole /usr/lib64/libdw-0.141.so [.] tfind@plt | 0.05% pahole ./build/libdwarves.so.1.0.0 [.] ftype__add_parameter | @@ -242,2 +242,3 @@ | 0.01% pahole [kernel] [k] account_group_user_time | + 0.01% pahole /usr/lib64/libdw-0.141.so [.] strlen@plt | 0.01% pahole ./build/pahole [.] strcmp@plt | [acme@doppio pahole]$ Signed-off-by: Arnaldo Carvalho de Melo LKML-Reference: <1247325517-12272-4-git-send-email-acme@redhat.com> Signed-off-by: Ingo Molnar --- tools/perf/util/symbol.c | 118 ++++++++++++++++++++++----------------- 1 file changed, 66 insertions(+), 52 deletions(-) diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c index 8efe7e41109..f40266b4845 100644 --- a/tools/perf/util/symbol.c +++ b/tools/perf/util/symbol.c @@ -374,36 +374,61 @@ static Elf_Scn *elf_section_by_name(Elf *elf, GElf_Ehdr *ep, idx < nr_entries; \ ++idx, pos = gelf_getrela(reldata, idx, &pos_mem)) -static int dso__synthesize_plt_symbols(struct dso *self, Elf *elf, - GElf_Ehdr *ehdr, Elf_Scn *scn_dynsym, - GElf_Shdr *shdr_dynsym, - size_t dynsym_idx, int verbose) +/* + * We need to check if we have a .dynsym, so that we can handle the + * .plt, synthesizing its symbols, that aren't on the symtabs (be it + * .dynsym or .symtab). + * And always look at the original dso, not at debuginfo packages, that + * have the PLT data stripped out (shdr_rel_plt.sh_type == SHT_NOBITS). + */ +static int dso__synthesize_plt_symbols(struct dso *self, int verbose) { uint32_t nr_rel_entries, idx; GElf_Sym sym; u64 plt_offset; GElf_Shdr shdr_plt; struct symbol *f; - GElf_Shdr shdr_rel_plt; + GElf_Shdr shdr_rel_plt, shdr_dynsym; Elf_Data *reldata, *syms, *symstrs; - Elf_Scn *scn_plt_rel, *scn_symstrs; + Elf_Scn *scn_plt_rel, *scn_symstrs, *scn_dynsym; + size_t dynsym_idx; + GElf_Ehdr ehdr; char sympltname[1024]; - int nr = 0, symidx; + Elf *elf; + int nr = 0, symidx, fd, err = 0; - scn_plt_rel = elf_section_by_name(elf, ehdr, &shdr_rel_plt, + fd = open(self->name, O_RDONLY); + if (fd < 0) + goto out; + + elf = elf_begin(fd, ELF_C_READ_MMAP, NULL); + if (elf == NULL) + goto out_close; + + if (gelf_getehdr(elf, &ehdr) == NULL) + goto out_elf_end; + + scn_dynsym = elf_section_by_name(elf, &ehdr, &shdr_dynsym, + ".dynsym", &dynsym_idx); + if (scn_dynsym == NULL) + goto out_elf_end; + + scn_plt_rel = elf_section_by_name(elf, &ehdr, &shdr_rel_plt, ".rela.plt", NULL); if (scn_plt_rel == NULL) { - scn_plt_rel = elf_section_by_name(elf, ehdr, &shdr_rel_plt, + scn_plt_rel = elf_section_by_name(elf, &ehdr, &shdr_rel_plt, ".rel.plt", NULL); if (scn_plt_rel == NULL) - return 0; + goto out_elf_end; } - if (shdr_rel_plt.sh_link != dynsym_idx) - return 0; + err = -1; - if (elf_section_by_name(elf, ehdr, &shdr_plt, ".plt", NULL) == NULL) - return 0; + if (shdr_rel_plt.sh_link != dynsym_idx) + goto out_elf_end; + + if (elf_section_by_name(elf, &ehdr, &shdr_plt, ".plt", NULL) == NULL) + goto out_elf_end; /* * Fetch the relocation section to find the indexes to the GOT @@ -411,19 +436,19 @@ static int dso__synthesize_plt_symbols(struct dso *self, Elf *elf, */ reldata = elf_getdata(scn_plt_rel, NULL); if (reldata == NULL) - return -1; + goto out_elf_end; syms = elf_getdata(scn_dynsym, NULL); if (syms == NULL) - return -1; + goto out_elf_end; - scn_symstrs = elf_getscn(elf, shdr_dynsym->sh_link); + scn_symstrs = elf_getscn(elf, shdr_dynsym.sh_link); if (scn_symstrs == NULL) - return -1; + goto out_elf_end; symstrs = elf_getdata(scn_symstrs, NULL); if (symstrs == NULL) - return -1; + goto out_elf_end; nr_rel_entries = shdr_rel_plt.sh_size / shdr_rel_plt.sh_entsize; plt_offset = shdr_plt.sh_offset; @@ -442,7 +467,7 @@ static int dso__synthesize_plt_symbols(struct dso *self, Elf *elf, f = symbol__new(plt_offset, shdr_plt.sh_entsize, sympltname, self->sym_priv_size, 0, verbose); if (!f) - return -1; + goto out_elf_end; dso__insert_symbol(self, f); ++nr; @@ -460,19 +485,25 @@ static int dso__synthesize_plt_symbols(struct dso *self, Elf *elf, f = symbol__new(plt_offset, shdr_plt.sh_entsize, sympltname, self->sym_priv_size, 0, verbose); if (!f) - return -1; + goto out_elf_end; dso__insert_symbol(self, f); ++nr; } - } else { - /* - * TODO: There are still one more shdr_rel_plt.sh_type - * I have to investigate, but probably should be ignored. - */ } - return nr; + err = 0; +out_elf_end: + elf_end(elf); +out_close: + close(fd); + + if (err == 0) + return nr; +out: + fprintf(stderr, "%s: problems reading %s PLT info.\n", + __func__, self->name); + return 0; } static int dso__load_sym(struct dso *self, int fd, const char *name, @@ -486,9 +517,8 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, GElf_Shdr shdr; Elf_Data *syms; GElf_Sym sym; - Elf_Scn *sec, *sec_dynsym, *sec_strndx; + Elf_Scn *sec, *sec_strndx; Elf *elf; - size_t dynsym_idx; int nr = 0; elf = elf_begin(fd, ELF_C_READ_MMAP, NULL); @@ -505,32 +535,11 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, goto out_elf_end; } - /* - * We need to check if we have a .dynsym, so that we can handle the - * .plt, synthesizing its symbols, that aren't on the symtabs (be it - * .dynsym or .symtab) - */ - sec_dynsym = elf_section_by_name(elf, &ehdr, &shdr, - ".dynsym", &dynsym_idx); - if (sec_dynsym != NULL) { - nr = dso__synthesize_plt_symbols(self, elf, &ehdr, - sec_dynsym, &shdr, - dynsym_idx, verbose); - if (nr < 0) - goto out_elf_end; - } - - /* - * But if we have a full .symtab (that is a superset of .dynsym) we - * should add the symbols not in the .dynsyn - */ sec = elf_section_by_name(elf, &ehdr, &shdr, ".symtab", NULL); if (sec == NULL) { - if (sec_dynsym == NULL) + sec = elf_section_by_name(elf, &ehdr, &shdr, ".dynsym", NULL); + if (sec == NULL) goto out_elf_end; - - sec = sec_dynsym; - gelf_getshdr(sec, &shdr); } syms = elf_getdata(sec, NULL); @@ -669,6 +678,11 @@ more: if (!ret) goto more; + if (ret > 0) { + int nr_plt = dso__synthesize_plt_symbols(self, verbose); + if (nr_plt > 0) + ret += nr_plt; + } out: free(name); return ret; From e3d7e183dc276df2fcaf02af173a49ad119ba9f9 Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Sat, 11 Jul 2009 12:18:37 -0300 Subject: [PATCH 11/31] perf report: Introduce -n/--show-nr-samples [acme@doppio pahole]$ perf report -ns comm,dso,symbol -d /lib64/libc-2.10.1.so -C pahole | head -17 21.94% 32101 [.] _int_malloc 20.10% 29402 [.] __GI_strcmp 16.77% 24533 [.] __tsearch 12.61% 18450 [.] malloc_consolidate 6.42% 9394 [.] _int_free 6.28% 9191 [.] __tfind 4.56% 6678 [.] __GI___libc_free 4.46% 6520 [.] _IO_vfprintf_internal 2.59% 3786 [.] __malloc 1.17% 1716 [.] __GI_memcpy [acme@doppio pahole]$ Signed-off-by: Arnaldo Carvalho de Melo LKML-Reference: <1247325517-12272-5-git-send-email-acme@redhat.com> Signed-off-by: Ingo Molnar --- tools/perf/Documentation/perf-report.txt | 3 +++ tools/perf/builtin-report.c | 18 ++++++++++++++++++ 2 files changed, 21 insertions(+) diff --git a/tools/perf/Documentation/perf-report.txt b/tools/perf/Documentation/perf-report.txt index 05774dfbd14..e72e9311078 100644 --- a/tools/perf/Documentation/perf-report.txt +++ b/tools/perf/Documentation/perf-report.txt @@ -24,6 +24,9 @@ OPTIONS --dsos=:: Only consider symbols in these dsos. CSV that understands file://filename entries. +-n +--show-nr-samples + Show the number of samples for each symbol -C:: --comms=:: Only consider symbols in these comms. CSV that understands diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index f3422121d85..430a195b858 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -51,6 +51,7 @@ static int verbose; static int modules; static int full_paths; +static int show_nr_samples; static unsigned long page_size; static unsigned long mmap_window = 32; @@ -1024,6 +1025,13 @@ hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples) else ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count); + if (show_nr_samples) { + if (field_sep) + fprintf(fp, "%c%lld", *field_sep, self->count); + else + fprintf(fp, "%11lld", self->count); + } + list_for_each_entry(se, &hist_entry__sort_list, list) { if (se->elide) continue; @@ -1361,6 +1369,12 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) fprintf(fp, "#\n"); fprintf(fp, "# Overhead"); + if (show_nr_samples) { + if (field_sep) + fprintf(fp, "%cSamples", *field_sep); + else + fputs(" Samples ", fp); + } list_for_each_entry(se, &hist_entry__sort_list, list) { if (se->elide) continue; @@ -1388,6 +1402,8 @@ static size_t output__fprintf(FILE *fp, u64 total_samples) goto print_entries; fprintf(fp, "# ........"); + if (show_nr_samples) + fprintf(fp, " .........."); list_for_each_entry(se, &hist_entry__sort_list, list) { unsigned int i; @@ -1979,6 +1995,8 @@ static const struct option options[] = { OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"), OPT_BOOLEAN('m', "modules", &modules, "load module symbols - WARNING: use only with -k and LIVE kernel"), + OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples, + "Show a column with the number of samples"), OPT_STRING('s', "sort", &sort_order, "key[,key2...]", "sort by key(s): pid, comm, dso, symbol, parent"), OPT_BOOLEAN('P', "full-paths", &full_paths, From f1c6a58121f9846ac665b0fbd3cbab90ce8bcbac Mon Sep 17 00:00:00 2001 From: Daniel Qarras Date: Sun, 12 Jul 2009 04:32:40 -0700 Subject: [PATCH 12/31] perf_counter, x86: Extend perf_counter Pentium M support I've attached a patch to remove the Pentium M special casing of EMON and as noticed at least with my Pentium M the hardware PMU now works: Performance counter stats for '/bin/ls /var/tmp': 1.809988 task-clock-msecs # 0.125 CPUs 1 context-switches # 0.001 M/sec 0 CPU-migrations # 0.000 M/sec 224 page-faults # 0.124 M/sec 1425648 cycles # 787.656 M/sec 912755 instructions # 0.640 IPC Vince suggested that this code was trying to address erratum Y17 in Pentium-M's: http://download.intel.com/support/processors/mobile/pm/sb/25266532.pdf But that erratum (related to IA32_MISC_ENABLES.7) does not affect perfcounters as we dont use this toggle to disable RDPMC and WRMSR/RDMSR access to performance counters. We keep cr4's bit 8 (X86_CR4_PCE) clear so unprivileged RDPMC access is not allowed anyway. Cc: Vince Weaver Cc: Peter Zijlstra Cc: Stephane Eranian LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_counter.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index bed1c4c2f25..7e346d4bc0f 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -1583,10 +1583,8 @@ static int p6_pmu_init(void) break; case 9: case 13: - /* for Pentium M, we need to check if PMU exist */ - rdmsr(MSR_IA32_MISC_ENABLE, low, high); - if (low & MSR_IA32_MISC_ENABLE_EMON) - break; + /* Pentium M */ + break; default: pr_cont("unsupported p6 CPU model %d ", boot_cpu_data.x86_model); From d4d7d0b9545721d3cabb19d15163bbc66b797707 Mon Sep 17 00:00:00 2001 From: Chris Wilson Date: Mon, 6 Jul 2009 09:31:33 +0100 Subject: [PATCH 13/31] perf_counter: Fix the tracepoint channel to perfcounters Fix a missed rename in EVENT_PROFILE support so that it gets built and allows tracepoint tracing from the 'perf' tool. Fix a typo in the (never before built & enabled) portion in perf_counter.c as well, and update that code to the attr.config changes as well. Signed-off-by: Chris Wilson Cc: Ben Gamari Cc: Jason Baron Cc: Frederic Weisbecker Cc: Mike Galbraith Cc: Peter Zijlstra Cc: Paul Mackerras Cc: Steven Rostedt LKML-Reference: <1246869094-21237-1-git-send-email-chris@chris-wilson.co.uk> Signed-off-by: Ingo Molnar --- init/Kconfig | 2 +- kernel/perf_counter.c | 10 +++------- 2 files changed, 4 insertions(+), 8 deletions(-) diff --git a/init/Kconfig b/init/Kconfig index 1ce05a4cb5f..cb2c0927022 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -962,7 +962,7 @@ config PERF_COUNTERS config EVENT_PROFILE bool "Tracepoint profile sources" - depends on PERF_COUNTERS && EVENT_TRACER + depends on PERF_COUNTERS && EVENT_TRACING default y endmenu diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index d55a50da234..c6c38fb7766 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -3671,7 +3671,7 @@ static const struct pmu perf_ops_task_clock = { void perf_tpcounter_event(int event_id) { struct perf_sample_data data = { - .regs = get_irq_regs(); + .regs = get_irq_regs(), .addr = 0, }; @@ -3687,16 +3687,12 @@ extern void ftrace_profile_disable(int); static void tp_perf_counter_destroy(struct perf_counter *counter) { - ftrace_profile_disable(perf_event_id(&counter->attr)); + ftrace_profile_disable(counter->attr.config); } static const struct pmu *tp_perf_counter_init(struct perf_counter *counter) { - int event_id = perf_event_id(&counter->attr); - int ret; - - ret = ftrace_profile_enable(event_id); - if (ret) + if (ftrace_profile_enable(counter->attr.config)) return NULL; counter->destroy = tp_perf_counter_destroy; From 23cdb5d5171d591ec911aada682e09d53c14a810 Mon Sep 17 00:00:00 2001 From: Roel Kluin Date: Mon, 13 Jul 2009 02:25:47 +0200 Subject: [PATCH 14/31] perf_counter tools: Fix index boundary check Keep index within event_type_descriptors[] Signed-off-by: Roel Kluin Cc: Andrew Morton Cc: Peter Zijlstra Cc: Mike Galbraith Cc: Paul Mackerras Cc: Arnaldo Carvalho de Melo Cc: Frederic Weisbecker LKML-Reference: <4A5A7F0B.4070106@gmail.com> Signed-off-by: Ingo Molnar --- tools/perf/util/parse-events.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 518a33affe1..d18c98edd00 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -436,7 +436,7 @@ void print_events(void) for (i = 0; i < ARRAY_SIZE(event_symbols); i++, syms++) { type = syms->type + 1; - if (type > ARRAY_SIZE(event_type_descriptors)) + if (type >= ARRAY_SIZE(event_type_descriptors)) type = 0; if (type != prev_type) From 413ee3b48ab582ffea33e7e140c7a2c5ea657e9a Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Thu, 16 Jul 2009 15:15:52 +0200 Subject: [PATCH 15/31] perf_counter: Make sure we dont leak kernel memory to userspace There are a few places we are leaking tiny amounts of kernel memory to userspace. This happens when writing out strings because we always align the end to 64 bits. To avoid this we should always use an appropriately sized temporary buffer and ensure it is zeroed. Since d_path assembles the string from the end of the buffer backwards, we need to add 64 bits after the buffer to allow for alignment. We also need to copy arch_vma_name to the temporary buffer, because if we use it directly we may end up copying to userspace a number of bytes after the end of the string constant. Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090716104817.273972048@samba.org> Signed-off-by: Ingo Molnar --- kernel/perf_counter.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index c6c38fb7766..f7a8ab9576e 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -2968,8 +2968,10 @@ static void perf_counter_comm_event(struct perf_comm_event *comm_event) struct perf_cpu_context *cpuctx; struct perf_counter_context *ctx; unsigned int size; - char *comm = comm_event->task->comm; + char comm[TASK_COMM_LEN]; + memset(comm, 0, sizeof(comm)); + strncpy(comm, comm_event->task->comm, sizeof(comm)); size = ALIGN(strlen(comm)+1, sizeof(u64)); comm_event->comm = comm; @@ -3088,8 +3090,15 @@ static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event) char *buf = NULL; const char *name; + memset(tmp, 0, sizeof(tmp)); + if (file) { - buf = kzalloc(PATH_MAX, GFP_KERNEL); + /* + * d_path works from the end of the buffer backwards, so we + * need to add enough zero bytes after the string to handle + * the 64bit alignment we do later. + */ + buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); if (!buf) { name = strncpy(tmp, "//enomem", sizeof(tmp)); goto got_name; @@ -3100,9 +3109,11 @@ static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event) goto got_name; } } else { - name = arch_vma_name(mmap_event->vma); - if (name) + if (arch_vma_name(mmap_event->vma)) { + name = strncpy(tmp, arch_vma_name(mmap_event->vma), + sizeof(tmp)); goto got_name; + } if (!vma->vm_mm) { name = strncpy(tmp, "[vdso]", sizeof(tmp)); From 11b5f81e1b0ea0bc84fe32f0a27054e052b2bf84 Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Thu, 16 Jul 2009 15:44:29 +0200 Subject: [PATCH 16/31] perf_counter: Synthesize VDSO mmap event perf record synthesizes mmap events for the running process. Right now it just catches file mappings, but we can check for the vdso symbol and add that too. Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090716104817.517264409@samba.org> Signed-off-by: Ingo Molnar --- tools/perf/builtin-record.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 4ef78a5e6f3..072aaf0369f 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -313,6 +313,10 @@ static void pid_synthesize_mmap_samples(pid_t pid) if (*pbf == 'x') { /* vm_exec */ char *execname = strchr(bf, '/'); + /* Catch VDSO */ + if (execname == NULL) + execname = strstr(bf, "[vdso]"); + if (execname == NULL) continue; From ed900c054b541254f0ce5cedaf75206e29bd614e Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Thu, 16 Jul 2009 15:44:29 +0200 Subject: [PATCH 17/31] perf_counter: Log vfork as a fork event Right now we don't output vfork events. Even though we should always see an exec after a vfork, we may get perfcounter samples between the vfork and exec. These samples can lead to some confusion when parsing perfcounter data. To keep things consistent we should always log a fork event. It will result in a little more log data, but is less confusing to trace parsing tools. Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090716104817.589309391@samba.org> Signed-off-by: Ingo Molnar --- kernel/fork.c | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/kernel/fork.c b/kernel/fork.c index 467746b3f0a..4812d60b29f 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -1408,14 +1408,11 @@ long do_fork(unsigned long clone_flags, if (clone_flags & CLONE_VFORK) { p->vfork_done = &vfork; init_completion(&vfork); - } else if (!(clone_flags & CLONE_VM)) { - /* - * vfork will do an exec which will call - * set_task_comm() - */ - perf_counter_fork(p); } + if (!(clone_flags & CLONE_THREAD)) + perf_counter_fork(p); + audit_finish_fork(p); tracehook_report_clone(regs, clone_flags, nr, p); From 4bba828dd9bb950ad1fe340ef148a5436a10f131 Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Thu, 16 Jul 2009 15:44:29 +0200 Subject: [PATCH 18/31] perf_counter: Add perf record option to log addresses Add the -d or --data option to log event addresses (eg page faults). Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090716104817.697698033@samba.org> Signed-off-by: Ingo Molnar --- tools/perf/builtin-record.c | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 072aaf0369f..68a9be0d151 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -43,6 +43,7 @@ static int call_graph = 0; static int verbose = 0; static int inherit_stat = 0; static int no_samples = 0; +static int sample_address = 0; static long samples; static struct timeval last_read; @@ -405,6 +406,9 @@ static void create_counter(int counter, int cpu, pid_t pid) if (inherit_stat) attr->inherit_stat = 1; + if (sample_address) + attr->sample_type |= PERF_SAMPLE_ADDR; + if (call_graph) attr->sample_type |= PERF_SAMPLE_CALLCHAIN; @@ -649,6 +653,8 @@ static const struct option options[] = { "be more verbose (show counter open errors, etc)"), OPT_BOOLEAN('s', "stat", &inherit_stat, "per thread counts"), + OPT_BOOLEAN('d', "data", &sample_address, + "Sample addresses"), OPT_BOOLEAN('n', "no-samples", &no_samples, "don't sample"), OPT_END() From 1483b19f8f5e8ad0c8816de368b099322dad4db5 Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Thu, 16 Jul 2009 15:44:29 +0200 Subject: [PATCH 19/31] perf_counter: Make call graph option consistent perf record uses -g for logging call graph data but perf report uses -c to print call graph data. Be consistent and use -g everywhere for call graph data. Also update the help text to reflect the current default - fractal,0.5 Signed-off-by: Anton Blanchard Acked-by: Frederic Weisbecker Signed-off-by: Peter Zijlstra LKML-Reference: <20090716104817.803604373@samba.org> Signed-off-by: Ingo Molnar --- tools/perf/builtin-report.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 4e5cc266311..4b980cce705 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -1891,9 +1891,9 @@ static const struct option options[] = { "regex filter to identify parent, see: '--sort parent'"), OPT_BOOLEAN('x', "exclude-other", &exclude_other, "Only display entries with parent-match"), - OPT_CALLBACK_DEFAULT('c', "callchain", NULL, "output_type,min_percent", + OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent", "Display callchains using output_type and min percent threshold. " - "Default: flat,0", &parse_callchain_opt, callchain_default_opt), + "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt), OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]", "only consider symbols in these dsos"), OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]", From 9b7019ae6a542a801a80df6694c66e240e2c3e39 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 22 Jul 2009 16:31:36 +0200 Subject: [PATCH 20/31] perf_counter: Remove unused variables Fix a gcc unused variables warning. Signed-off-by: Peter Zijlstra --- arch/x86/kernel/cpu/perf_counter.c | 2 -- 1 file changed, 2 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index 7e346d4bc0f..a7aa8f90095 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -1570,8 +1570,6 @@ static struct x86_pmu amd_pmu = { static int p6_pmu_init(void) { - int high, low; - switch (boot_cpu_data.x86_model) { case 1: case 3: /* Pentium Pro */ From c9f73a3dd27e03411f18a58c0814d51392d2b17a Mon Sep 17 00:00:00 2001 From: Arjan van de Ven Date: Tue, 21 Jul 2009 00:55:05 -0700 Subject: [PATCH 21/31] perf: Fix stack data leak the "reserved" field was not initialized to zero, resulting in 4 bytes of stack data leaking to userspace.... Signed-off-by: Arjan van de Ven Signed-off-by: Peter Zijlstra --- kernel/perf_counter.c | 1 + 1 file changed, 1 insertion(+) diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index 5c6fae4f43d..ff854fd89a8 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -2666,6 +2666,7 @@ static void perf_counter_output(struct perf_counter *counter, int nmi, header.size += sizeof(cpu_entry); cpu_entry.cpu = raw_smp_processor_id(); + cpu_entry.reserved = 0; } if (sample_type & PERF_SAMPLE_PERIOD) From 573402db02746179b3f95f83a11a787501f52d0a Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 22 Jul 2009 11:13:50 +0200 Subject: [PATCH 22/31] perf_counter: Plug more stack leaks Per example of Arjan's patch, I went through and found a few more. Signed-off-by: Peter Zijlstra --- kernel/perf_counter.c | 23 +++++++++++++++++++++-- 1 file changed, 21 insertions(+), 2 deletions(-) diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index ff854fd89a8..e1d6a3aa133 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -2897,8 +2897,11 @@ void perf_counter_fork(struct task_struct *task) .event = { .header = { .type = PERF_EVENT_FORK, + .misc = 0, .size = sizeof(fork_event.event), }, + /* .pid */ + /* .ppid */ }, }; @@ -3008,8 +3011,16 @@ void perf_counter_comm(struct task_struct *task) comm_event = (struct perf_comm_event){ .task = task, + /* .comm */ + /* .comm_size */ .event = { - .header = { .type = PERF_EVENT_COMM, }, + .header = { + .type = PERF_EVENT_COMM, + .misc = 0, + /* .size */ + }, + /* .pid */ + /* .tid */ }, }; @@ -3160,8 +3171,16 @@ void __perf_counter_mmap(struct vm_area_struct *vma) mmap_event = (struct perf_mmap_event){ .vma = vma, + /* .file_name */ + /* .file_size */ .event = { - .header = { .type = PERF_EVENT_MMAP, }, + .header = { + .type = PERF_EVENT_MMAP, + .misc = 0, + /* .size */ + }, + /* .pid */ + /* .tid */ .start = vma->vm_start, .len = vma->vm_end - vma->vm_start, .pgoff = vma->vm_pgoff, From 7f453c24b95a085fc7bd35d53b33abc4dc5a048b Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Tue, 21 Jul 2009 13:19:40 +0200 Subject: [PATCH 23/31] perf_counter: PERF_SAMPLE_ID and inherited counters Anton noted that for inherited counters the counter-id as provided by PERF_SAMPLE_ID isn't mappable to the id found through PERF_RECORD_ID because each inherited counter gets its own id. His suggestion was to always return the parent counter id, since that is the primary counter id as exposed. However, these inherited counters have a unique identifier so that events like PERF_EVENT_PERIOD and PERF_EVENT_THROTTLE can be specific about which counter gets modified, which is important when trying to normalize the sample streams. This patch removes PERF_EVENT_PERIOD in favour of PERF_SAMPLE_PERIOD, which is more useful anyway, since changing periods became a lot more common than initially thought -- rendering PERF_EVENT_PERIOD the less useful solution (also, PERF_SAMPLE_PERIOD reports the more accurate value, since it reports the value used to trigger the overflow, whereas PERF_EVENT_PERIOD simply reports the requested period changed, which might only take effect on the next cycle). This still leaves us PERF_EVENT_THROTTLE to consider, but since that _should_ be a rare occurrence, and linking it to a primary id is the most useful bit to diagnose the problem, we introduce a PERF_SAMPLE_STREAM_ID, for those few cases where the full reconstruction is important. [Does change the ABI a little, but I see no other way out] Suggested-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <1248095846.15751.8781.camel@twins> --- include/linux/perf_counter.h | 15 ++---- kernel/perf_counter.c | 92 ++++++++++++----------------------- tools/perf/builtin-annotate.c | 24 --------- tools/perf/builtin-report.c | 24 --------- 4 files changed, 35 insertions(+), 120 deletions(-) diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index 5e970c7d3fd..bd15d7a5f5c 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -120,8 +120,9 @@ enum perf_counter_sample_format { PERF_SAMPLE_ID = 1U << 6, PERF_SAMPLE_CPU = 1U << 7, PERF_SAMPLE_PERIOD = 1U << 8, + PERF_SAMPLE_STREAM_ID = 1U << 9, - PERF_SAMPLE_MAX = 1U << 9, /* non-ABI */ + PERF_SAMPLE_MAX = 1U << 10, /* non-ABI */ }; /* @@ -312,16 +313,7 @@ enum perf_event_type { * struct perf_event_header header; * u64 time; * u64 id; - * u64 sample_period; - * }; - */ - PERF_EVENT_PERIOD = 4, - - /* - * struct { - * struct perf_event_header header; - * u64 time; - * u64 id; + * u64 stream_id; * }; */ PERF_EVENT_THROTTLE = 5, @@ -356,6 +348,7 @@ enum perf_event_type { * { u64 time; } && PERF_SAMPLE_TIME * { u64 addr; } && PERF_SAMPLE_ADDR * { u64 id; } && PERF_SAMPLE_ID + * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID * { u32 cpu, res; } && PERF_SAMPLE_CPU * { u64 period; } && PERF_SAMPLE_PERIOD * diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index e1d6a3aa133..7530588fa5c 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -154,6 +154,20 @@ static void unclone_ctx(struct perf_counter_context *ctx) } } +/* + * If we inherit counters we want to return the parent counter id + * to userspace. + */ +static u64 primary_counter_id(struct perf_counter *counter) +{ + u64 id = counter->id; + + if (counter->parent) + id = counter->parent->id; + + return id; +} + /* * Get the perf_counter_context for a task and lock it. * This has to cope with with the fact that until it is locked, @@ -1296,7 +1310,6 @@ static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu) #define MAX_INTERRUPTS (~0ULL) static void perf_log_throttle(struct perf_counter *counter, int enable); -static void perf_log_period(struct perf_counter *counter, u64 period); static void perf_adjust_period(struct perf_counter *counter, u64 events) { @@ -1315,8 +1328,6 @@ static void perf_adjust_period(struct perf_counter *counter, u64 events) if (!sample_period) sample_period = 1; - perf_log_period(counter, sample_period); - hwc->sample_period = sample_period; } @@ -1705,7 +1716,7 @@ perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count) values[n++] = counter->total_time_running + atomic64_read(&counter->child_total_time_running); if (counter->attr.read_format & PERF_FORMAT_ID) - values[n++] = counter->id; + values[n++] = primary_counter_id(counter); mutex_unlock(&counter->child_mutex); if (count < n * sizeof(u64)) @@ -1812,8 +1823,6 @@ static int perf_counter_period(struct perf_counter *counter, u64 __user *arg) counter->attr.sample_freq = value; } else { - perf_log_period(counter, value); - counter->attr.sample_period = value; counter->hw.sample_period = value; } @@ -2662,6 +2671,9 @@ static void perf_counter_output(struct perf_counter *counter, int nmi, if (sample_type & PERF_SAMPLE_ID) header.size += sizeof(u64); + if (sample_type & PERF_SAMPLE_STREAM_ID) + header.size += sizeof(u64); + if (sample_type & PERF_SAMPLE_CPU) { header.size += sizeof(cpu_entry); @@ -2705,7 +2717,13 @@ static void perf_counter_output(struct perf_counter *counter, int nmi, if (sample_type & PERF_SAMPLE_ADDR) perf_output_put(&handle, data->addr); - if (sample_type & PERF_SAMPLE_ID) + if (sample_type & PERF_SAMPLE_ID) { + u64 id = primary_counter_id(counter); + + perf_output_put(&handle, id); + } + + if (sample_type & PERF_SAMPLE_STREAM_ID) perf_output_put(&handle, counter->id); if (sample_type & PERF_SAMPLE_CPU) @@ -2728,7 +2746,7 @@ static void perf_counter_output(struct perf_counter *counter, int nmi, if (sub != counter) sub->pmu->read(sub); - group_entry.id = sub->id; + group_entry.id = primary_counter_id(sub); group_entry.counter = atomic64_read(&sub->count); perf_output_put(&handle, group_entry); @@ -2788,15 +2806,8 @@ perf_counter_read_event(struct perf_counter *counter, } if (counter->attr.read_format & PERF_FORMAT_ID) { - u64 id; - event.header.size += sizeof(u64); - if (counter->parent) - id = counter->parent->id; - else - id = counter->id; - - event.format[i++] = id; + event.format[i++] = primary_counter_id(counter); } ret = perf_output_begin(&handle, counter, event.header.size, 0, 0); @@ -3190,49 +3201,6 @@ void __perf_counter_mmap(struct vm_area_struct *vma) perf_counter_mmap_event(&mmap_event); } -/* - * Log sample_period changes so that analyzing tools can re-normalize the - * event flow. - */ - -struct freq_event { - struct perf_event_header header; - u64 time; - u64 id; - u64 period; -}; - -static void perf_log_period(struct perf_counter *counter, u64 period) -{ - struct perf_output_handle handle; - struct freq_event event; - int ret; - - if (counter->hw.sample_period == period) - return; - - if (counter->attr.sample_type & PERF_SAMPLE_PERIOD) - return; - - event = (struct freq_event) { - .header = { - .type = PERF_EVENT_PERIOD, - .misc = 0, - .size = sizeof(event), - }, - .time = sched_clock(), - .id = counter->id, - .period = period, - }; - - ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0); - if (ret) - return; - - perf_output_put(&handle, event); - perf_output_end(&handle); -} - /* * IRQ throttle logging */ @@ -3246,14 +3214,16 @@ static void perf_log_throttle(struct perf_counter *counter, int enable) struct perf_event_header header; u64 time; u64 id; + u64 stream_id; } throttle_event = { .header = { .type = PERF_EVENT_THROTTLE + 1, .misc = 0, .size = sizeof(throttle_event), }, - .time = sched_clock(), - .id = counter->id, + .time = sched_clock(), + .id = primary_counter_id(counter), + .stream_id = counter->id, }; ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0); diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c index 5f9eefecc57..1dba568e194 100644 --- a/tools/perf/builtin-annotate.c +++ b/tools/perf/builtin-annotate.c @@ -74,20 +74,12 @@ struct fork_event { u32 pid, ppid; }; -struct period_event { - struct perf_event_header header; - u64 time; - u64 id; - u64 sample_period; -}; - typedef union event_union { struct perf_event_header header; struct ip_event ip; struct mmap_event mmap; struct comm_event comm; struct fork_event fork; - struct period_event period; } event_t; @@ -997,19 +989,6 @@ process_fork_event(event_t *event, unsigned long offset, unsigned long head) return 0; } -static int -process_period_event(event_t *event, unsigned long offset, unsigned long head) -{ - dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n", - (void *)(offset + head), - (void *)(long)(event->header.size), - event->period.time, - event->period.id, - event->period.sample_period); - - return 0; -} - static int process_event(event_t *event, unsigned long offset, unsigned long head) { @@ -1025,9 +1004,6 @@ process_event(event_t *event, unsigned long offset, unsigned long head) case PERF_EVENT_FORK: return process_fork_event(event, offset, head); - - case PERF_EVENT_PERIOD: - return process_period_event(event, offset, head); /* * We dont process them right now but they are fine: */ diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index a118bc77286..b20a4b6e31b 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -101,13 +101,6 @@ struct fork_event { u32 pid, ppid; }; -struct period_event { - struct perf_event_header header; - u64 time; - u64 id; - u64 sample_period; -}; - struct lost_event { struct perf_event_header header; u64 id; @@ -127,7 +120,6 @@ typedef union event_union { struct mmap_event mmap; struct comm_event comm; struct fork_event fork; - struct period_event period; struct lost_event lost; struct read_event read; } event_t; @@ -1635,19 +1627,6 @@ process_fork_event(event_t *event, unsigned long offset, unsigned long head) return 0; } -static int -process_period_event(event_t *event, unsigned long offset, unsigned long head) -{ - dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n", - (void *)(offset + head), - (void *)(long)(event->header.size), - event->period.time, - event->period.id, - event->period.sample_period); - - return 0; -} - static int process_lost_event(event_t *event, unsigned long offset, unsigned long head) { @@ -1729,9 +1708,6 @@ process_event(event_t *event, unsigned long offset, unsigned long head) case PERF_EVENT_FORK: return process_fork_event(event, offset, head); - case PERF_EVENT_PERIOD: - return process_period_event(event, offset, head); - case PERF_EVENT_LOST: return process_lost_event(event, offset, head); From a0541234f89c93f313961ce7b28676e11488a5f0 Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Wed, 22 Jul 2009 23:04:12 +1000 Subject: [PATCH 24/31] perf_counter: Improve perf stat and perf record option parsing perf stat and perf record currently look for all options on the command line. This can lead to some confusion: # perf stat ls -l Error: unknown switch `l' While we can work around this by adding '--' before the command, the git option parsing code can stop at the first non option: # perf stat ls -l Performance counter stats for 'ls -l': .... Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090722130412.GD9029@kryten> --- tools/perf/builtin-record.c | 3 ++- tools/perf/builtin-stat.c | 3 ++- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index 68a9be0d151..6da09928130 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -664,7 +664,8 @@ int cmd_record(int argc, const char **argv, const char *prefix __used) { int counter; - argc = parse_options(argc, argv, options, record_usage, 0); + argc = parse_options(argc, argv, options, record_usage, + PARSE_OPT_STOP_AT_NON_OPTION); if (!argc && target_pid == -1 && !system_wide) usage_with_options(record_usage, options); diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 27921a8ce1a..f9510eeeb6c 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -511,7 +511,8 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used) { int status; - argc = parse_options(argc, argv, options, stat_usage, 0); + argc = parse_options(argc, argv, options, stat_usage, + PARSE_OPT_STOP_AT_NON_OPTION); if (!argc) usage_with_options(stat_usage, options); if (run_count <= 0 || run_count > MAX_RUN) From 966ee4d6b887c14159043ac80b8c3661d2bbe5e2 Mon Sep 17 00:00:00 2001 From: Anton Blanchard Date: Wed, 22 Jul 2009 23:05:46 +1000 Subject: [PATCH 25/31] perf_counter: Fix throttle/unthrottle event logging Right now we only print PERF_EVENT_THROTTLE + 1 (ie PERF_EVENT_UNTHROTTLE). Fix this to print both a throttle and unthrottle event. Signed-off-by: Anton Blanchard Signed-off-by: Peter Zijlstra LKML-Reference: <20090722130546.GE9029@kryten> --- kernel/perf_counter.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index 7530588fa5c..787d4daef18 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -3217,7 +3217,7 @@ static void perf_log_throttle(struct perf_counter *counter, int enable) u64 stream_id; } throttle_event = { .header = { - .type = PERF_EVENT_THROTTLE + 1, + .type = PERF_EVENT_THROTTLE, .misc = 0, .size = sizeof(throttle_event), }, @@ -3226,6 +3226,9 @@ static void perf_log_throttle(struct perf_counter *counter, int enable) .stream_id = counter->id, }; + if (enable) + throttle_event.header.type = PERF_EVENT_UNTHROTTLE; + ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0); if (ret) return; From dfe5a50461db90fab901cb697eff0d3d2e9fd229 Mon Sep 17 00:00:00 2001 From: Arjan van de Ven Date: Mon, 20 Jul 2009 22:54:26 -0700 Subject: [PATCH 26/31] perf: avoid structure size confusion by using a fixed size for some reason, this structure gets compiled as 36 bytes in some files (the ones that alloacte it) but 40 bytes in others (the ones that use it). The cause is an off_t type that gets a different size in different compilation units for some yet-to-be-explained reason. But the effect is disasterous; the size/offset members of the struct are at different offsets, and result in mostly complete garbage. The parser in perf is so robust that this all gets hidden, and after skipping an certain amount of samples, it recovers.... so this bug is not normally noticed. .... except when you want every sample to be exact. Fix this by just using an explicitly sized type. Signed-off-by: Arjan van de Ven Signed-off-by: Peter Zijlstra LKML-Reference: <4A655917.9080504@linux.intel.com> --- tools/perf/util/header.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/perf/util/header.h b/tools/perf/util/header.h index b5ef53ad4c7..bf280449fcf 100644 --- a/tools/perf/util/header.h +++ b/tools/perf/util/header.h @@ -16,7 +16,7 @@ struct perf_header { int frozen; int attrs, size; struct perf_header_attr **attr; - off_t attr_offset; + s64 attr_offset; u64 data_offset; u64 data_size; }; From 28ac909b49a155856c957d080f8a796b3c1d1f3e Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Mon, 20 Jul 2009 14:14:12 -0300 Subject: [PATCH 27/31] perf symbol: C++ demangling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit [acme@doppio ~]$ perf report -s comm,dso,symbol -C firefox -d /usr/lib64/xulrunner-1.9.1/libxul.so | grep :: | head 2.21% [.] nsDeque::Push(void*) 1.78% [.] GraphWalker::DoWalk(nsDeque&) 1.30% [.] GCGraphBuilder::AddNode(void*, nsCycleCollectionParticipant*) 1.27% [.] XPCWrappedNative::CallMethod(XPCCallContext&, XPCWrappedNative::CallMode) 1.18% [.] imgContainer::DrawFrameTo(gfxIImageFrame*, gfxIImageFrame*, nsRect&) 1.13% [.] nsDeque::PopFront() 1.11% [.] nsGlobalWindow::RunTimeout(nsTimeout*) 0.97% [.] nsXPConnect::Traverse(void*, nsCycleCollectionTraversalCallback&) 0.95% [.] nsJSEventListener::cycleCollection::Traverse(void*, nsCycleCollectionTraversalCallback&) 0.95% [.] nsCOMPtr_base::~nsCOMPtr_base() [acme@doppio ~]$ Cc: Pekka Enberg Cc: Vegard Nossum Cc: Paul Mackerras Cc: Frédéric Weisbecker Suggested-by: Clark Williams Signed-off-by: Arnaldo Carvalho de Melo Signed-off-by: Peter Zijlstra LKML-Reference: <20090720171412.GB10410@ghostprotocols.net> --- tools/perf/Makefile | 2 +- tools/perf/util/symbol.c | 21 +++++++++++++++++++-- 2 files changed, 20 insertions(+), 3 deletions(-) diff --git a/tools/perf/Makefile b/tools/perf/Makefile index 7822b3d6bac..a5e9b876ca0 100644 --- a/tools/perf/Makefile +++ b/tools/perf/Makefile @@ -345,7 +345,7 @@ BUILTIN_OBJS += builtin-stat.o BUILTIN_OBJS += builtin-top.o PERFLIBS = $(LIB_FILE) -EXTLIBS = +EXTLIBS = -lbfd # # Platform specific tweaks diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c index f40266b4845..98aee922acc 100644 --- a/tools/perf/util/symbol.c +++ b/tools/perf/util/symbol.c @@ -6,9 +6,15 @@ #include #include #include +#include const char *sym_hist_filter; +#ifndef DMGL_PARAMS +#define DMGL_PARAMS (1 << 0) /* Include function args */ +#define DMGL_ANSI (1 << 1) /* Include const, volatile, etc */ +#endif + static struct symbol *symbol__new(u64 start, u64 len, const char *name, unsigned int priv_size, u64 obj_start, int verbose) @@ -571,6 +577,8 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, NULL) != NULL); elf_symtab__for_each_symbol(syms, nr_syms, index, sym) { struct symbol *f; + const char *name; + char *demangled; u64 obj_start; struct section *section = NULL; int is_label = elf_sym__is_label(&sym); @@ -609,10 +617,19 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, goto out_elf_end; } } + /* + * We need to figure out if the object was created from C++ sources + * DWARF DW_compile_unit has this, but we don't always have access + * to it... + */ + name = elf_sym__name(&sym, symstrs); + demangled = bfd_demangle(NULL, name, DMGL_PARAMS | DMGL_ANSI); + if (demangled != NULL) + name = demangled; - f = symbol__new(sym.st_value, sym.st_size, - elf_sym__name(&sym, symstrs), + f = symbol__new(sym.st_value, sym.st_size, name, self->sym_priv_size, obj_start, verbose); + free(demangled); if (!f) goto out_elf_end; From f6bdafef2ab911f03321fa83d8da1df99878009e Mon Sep 17 00:00:00 2001 From: Jason Baron Date: Tue, 21 Jul 2009 12:20:22 -0400 Subject: [PATCH 28/31] perf_counter: Add tracepoint support to perf list, perf stat Add support to 'perf list' and 'perf stat' for kernel tracepoints. The implementation creates a 'for_each_subsystem' and 'for_each_event' for easy iteration over the tracepoints. Signed-off-by: Jason Baron Signed-off-by: Peter Zijlstra LKML-Reference: <426129bf9fcc8ee63bb094cf736e7316a7dcd77a.1248190728.git.jbaron@redhat.com> --- tools/perf/util/parse-events.c | 175 ++++++++++++++++++++++++++++++++- tools/perf/util/util.h | 2 + 2 files changed, 176 insertions(+), 1 deletion(-) diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index d18c98edd00..5a3cd3a34af 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -12,6 +12,8 @@ int nr_counters; struct perf_counter_attr attrs[MAX_COUNTERS]; +static char default_debugfs_path[] = "/sys/kernel/debug/tracing/events"; + struct event_symbol { u8 type; u64 config; @@ -110,6 +112,88 @@ static unsigned long hw_cache_stat[C(MAX)] = { [C(BPU)] = (CACHE_READ), }; +#define for_each_subsystem(sys_dir, sys_dirent, sys_next, file, st) \ + while (!readdir_r(sys_dir, &sys_dirent, &sys_next) && sys_next) \ + if (snprintf(file, MAXPATHLEN, "%s/%s", default_debugfs_path, \ + sys_dirent.d_name) && \ + (!stat(file, &st)) && (S_ISDIR(st.st_mode)) && \ + (strcmp(sys_dirent.d_name, ".")) && \ + (strcmp(sys_dirent.d_name, ".."))) + +#define for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next, file, st) \ + while (!readdir_r(evt_dir, &evt_dirent, &evt_next) && evt_next) \ + if (snprintf(file, MAXPATHLEN, "%s/%s/%s", default_debugfs_path, \ + sys_dirent.d_name, evt_dirent.d_name) && \ + (!stat(file, &st)) && (S_ISDIR(st.st_mode)) && \ + (strcmp(evt_dirent.d_name, ".")) && \ + (strcmp(evt_dirent.d_name, ".."))) + +#define MAX_EVENT_LENGTH 30 + +static int valid_debugfs_mount(void) +{ + struct statfs st_fs; + + if (statfs(default_debugfs_path, &st_fs) < 0) + return -ENOENT; + else if (st_fs.f_type != (long) DEBUGFS_MAGIC) + return -ENOENT; + return 0; +} + +static char *tracepoint_id_to_name(u64 config) +{ + static char tracepoint_name[2 * MAX_EVENT_LENGTH]; + DIR *sys_dir, *evt_dir; + struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent; + struct stat st; + char id_buf[4]; + int fd; + u64 id; + char evt_path[MAXPATHLEN]; + + if (valid_debugfs_mount()) + return "unkown"; + + sys_dir = opendir(default_debugfs_path); + if (!sys_dir) + goto cleanup; + + for_each_subsystem(sys_dir, sys_dirent, sys_next, evt_path, st) { + evt_dir = opendir(evt_path); + if (!evt_dir) + goto cleanup; + for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next, + evt_path, st) { + snprintf(evt_path, MAXPATHLEN, "%s/%s/%s/id", + default_debugfs_path, sys_dirent.d_name, + evt_dirent.d_name); + fd = open(evt_path, O_RDONLY); + if (fd < 0) + continue; + if (read(fd, id_buf, sizeof(id_buf)) < 0) { + close(fd); + continue; + } + close(fd); + id = atoll(id_buf); + if (id == config) { + closedir(evt_dir); + closedir(sys_dir); + snprintf(tracepoint_name, 2 * MAX_EVENT_LENGTH, + "%s:%s", sys_dirent.d_name, + evt_dirent.d_name); + return tracepoint_name; + } + } + closedir(evt_dir); + } + +cleanup: + closedir(sys_dir); + return "unkown"; +} + static int is_cache_op_valid(u8 cache_type, u8 cache_op) { if (hw_cache_stat[cache_type] & COP(cache_op)) @@ -177,6 +261,9 @@ char *event_name(int counter) return sw_event_names[config]; return "unknown-software"; + case PERF_TYPE_TRACEPOINT: + return tracepoint_id_to_name(config); + default: break; } @@ -265,6 +352,53 @@ parse_generic_hw_event(const char **str, struct perf_counter_attr *attr) return 1; } +static int parse_tracepoint_event(const char **strp, + struct perf_counter_attr *attr) +{ + const char *evt_name; + char sys_name[MAX_EVENT_LENGTH]; + char id_buf[4]; + int fd; + unsigned int sys_length, evt_length; + u64 id; + char evt_path[MAXPATHLEN]; + + if (valid_debugfs_mount()) + return 0; + + evt_name = strchr(*strp, ':'); + if (!evt_name) + return 0; + + sys_length = evt_name - *strp; + if (sys_length >= MAX_EVENT_LENGTH) + return 0; + + strncpy(sys_name, *strp, sys_length); + sys_name[sys_length] = '\0'; + evt_name = evt_name + 1; + evt_length = strlen(evt_name); + if (evt_length >= MAX_EVENT_LENGTH) + return 0; + + snprintf(evt_path, MAXPATHLEN, "%s/%s/%s/id", default_debugfs_path, + sys_name, evt_name); + fd = open(evt_path, O_RDONLY); + if (fd < 0) + return 0; + + if (read(fd, id_buf, sizeof(id_buf)) < 0) { + close(fd); + return 0; + } + close(fd); + id = atoll(id_buf); + attr->config = id; + attr->type = PERF_TYPE_TRACEPOINT; + *strp = evt_name + evt_length; + return 1; +} + static int check_events(const char *str, unsigned int i) { int n; @@ -374,7 +508,8 @@ parse_event_modifier(const char **strp, struct perf_counter_attr *attr) */ static int parse_event_symbols(const char **str, struct perf_counter_attr *attr) { - if (!(parse_raw_event(str, attr) || + if (!(parse_tracepoint_event(str, attr) || + parse_raw_event(str, attr) || parse_numeric_event(str, attr) || parse_symbolic_event(str, attr) || parse_generic_hw_event(str, attr))) @@ -422,6 +557,42 @@ static const char * const event_type_descriptors[] = { "Hardware cache event", }; +/* + * Print the events from /tracing/events + */ + +static void print_tracepoint_events(void) +{ + DIR *sys_dir, *evt_dir; + struct dirent *sys_next, *evt_next, sys_dirent, evt_dirent; + struct stat st; + char evt_path[MAXPATHLEN]; + + if (valid_debugfs_mount()) + return; + + sys_dir = opendir(default_debugfs_path); + if (!sys_dir) + goto cleanup; + + for_each_subsystem(sys_dir, sys_dirent, sys_next, evt_path, st) { + evt_dir = opendir(evt_path); + if (!evt_dir) + goto cleanup; + for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next, + evt_path, st) { + snprintf(evt_path, MAXPATHLEN, "%s:%s", + sys_dirent.d_name, evt_dirent.d_name); + fprintf(stderr, " %-40s [%s]\n", evt_path, + event_type_descriptors[PERF_TYPE_TRACEPOINT+1]); + } + closedir(evt_dir); + } + +cleanup: + closedir(sys_dir); +} + /* * Print the help text for the event symbols: */ @@ -472,5 +643,7 @@ void print_events(void) "rNNN"); fprintf(stderr, "\n"); + print_tracepoint_events(); + exit(129); } diff --git a/tools/perf/util/util.h b/tools/perf/util/util.h index b4be6071c10..68fe157d72f 100644 --- a/tools/perf/util/util.h +++ b/tools/perf/util/util.h @@ -50,6 +50,7 @@ #include #include #include +#include #include #include #include @@ -80,6 +81,7 @@ #include #include #include +#include "../../../include/linux/magic.h" #ifndef NO_ICONV #include From 5beeded123c5befa21f1c6e16219f2a3eb7dd197 Mon Sep 17 00:00:00 2001 From: Jason Baron Date: Tue, 21 Jul 2009 14:16:29 -0400 Subject: [PATCH 29/31] perf_counter: Detect debugfs location If "/sys/kernel/debug" is not a debugfs mount point, search for the debugfs filesystem in /proc/mounts, but also allows the user to specify '--debugfs-dir=blah' or set the environment variable: 'PERF_DEBUGFS_DIR' Signed-off-by: Jason Baron [ also made it probe "/debug" by default ] Signed-off-by: Peter Zijlstra LKML-Reference: <20090721181629.GA3094@redhat.com> --- tools/perf/perf.c | 77 +++++++++++++++++++++++++++++++++- tools/perf/util/cache.h | 1 + tools/perf/util/parse-events.c | 33 ++++++++------- tools/perf/util/parse-events.h | 5 +++ tools/perf/util/string.h | 3 ++ 5 files changed, 102 insertions(+), 17 deletions(-) diff --git a/tools/perf/perf.c b/tools/perf/perf.c index c5656784c61..31982ad064b 100644 --- a/tools/perf/perf.c +++ b/tools/perf/perf.c @@ -12,6 +12,8 @@ #include "util/cache.h" #include "util/quote.h" #include "util/run-command.h" +#include "util/parse-events.h" +#include "util/string.h" const char perf_usage_string[] = "perf [--version] [--help] COMMAND [ARGS]"; @@ -25,6 +27,8 @@ struct pager_config { int val; }; +static char debugfs_mntpt[MAXPATHLEN]; + static int pager_command_config(const char *var, const char *value, void *data) { struct pager_config *c = data; @@ -56,6 +60,15 @@ static void commit_pager_choice(void) { } } +static void set_debugfs_path(void) +{ + char *path; + + path = getenv(PERF_DEBUGFS_ENVIRONMENT); + snprintf(debugfs_path, MAXPATHLEN, "%s/%s", path ?: debugfs_mntpt, + "tracing/events"); +} + static int handle_options(const char*** argv, int* argc, int* envchanged) { int handled = 0; @@ -122,6 +135,22 @@ static int handle_options(const char*** argv, int* argc, int* envchanged) setenv(PERF_WORK_TREE_ENVIRONMENT, cmd + 12, 1); if (envchanged) *envchanged = 1; + } else if (!strcmp(cmd, "--debugfs-dir")) { + if (*argc < 2) { + fprintf(stderr, "No directory given for --debugfs-dir.\n"); + usage(perf_usage_string); + } + strncpy(debugfs_mntpt, (*argv)[1], MAXPATHLEN); + debugfs_mntpt[MAXPATHLEN - 1] = '\0'; + if (envchanged) + *envchanged = 1; + (*argv)++; + (*argc)--; + } else if (!prefixcmp(cmd, "--debugfs-dir=")) { + strncpy(debugfs_mntpt, cmd + 14, MAXPATHLEN); + debugfs_mntpt[MAXPATHLEN - 1] = '\0'; + if (envchanged) + *envchanged = 1; } else { fprintf(stderr, "Unknown option: %s\n", cmd); usage(perf_usage_string); @@ -228,6 +257,7 @@ static int run_builtin(struct cmd_struct *p, int argc, const char **argv) if (use_pager == -1 && p->option & USE_PAGER) use_pager = 1; commit_pager_choice(); + set_debugfs_path(); status = p->fn(argc, argv, prefix); if (status) @@ -346,6 +376,49 @@ static int run_argv(int *argcp, const char ***argv) return done_alias; } +/* mini /proc/mounts parser: searching for "^blah /mount/point debugfs" */ +static void get_debugfs_mntpt(void) +{ + FILE *file; + char fs_type[100]; + char debugfs[MAXPATHLEN]; + + /* + * try the standard location + */ + if (valid_debugfs_mount("/sys/kernel/debug/") == 0) { + strcpy(debugfs_mntpt, "/sys/kernel/debug/"); + return; + } + + /* + * try the sane location + */ + if (valid_debugfs_mount("/debug/") == 0) { + strcpy(debugfs_mntpt, "/debug/"); + return; + } + + /* + * give up and parse /proc/mounts + */ + file = fopen("/proc/mounts", "r"); + if (file == NULL) + return; + + while (fscanf(file, "%*s %" + STR(MAXPATHLEN) + "s %99s %*s %*d %*d\n", + debugfs, fs_type) == 2) { + if (strcmp(fs_type, "debugfs") == 0) + break; + } + fclose(file); + if (strcmp(fs_type, "debugfs") == 0) { + strncpy(debugfs_mntpt, debugfs, MAXPATHLEN); + debugfs_mntpt[MAXPATHLEN - 1] = '\0'; + } +} int main(int argc, const char **argv) { @@ -354,7 +427,8 @@ int main(int argc, const char **argv) cmd = perf_extract_argv0_path(argv[0]); if (!cmd) cmd = "perf-help"; - + /* get debugfs mount point from /proc/mounts */ + get_debugfs_mntpt(); /* * "perf-xxxx" is the same as "perf xxxx", but we obviously: * @@ -377,6 +451,7 @@ int main(int argc, const char **argv) argc--; handle_options(&argv, &argc, NULL); commit_pager_choice(); + set_debugfs_path(); if (argc > 0) { if (!prefixcmp(argv[0], "--")) argv[0] += 2; diff --git a/tools/perf/util/cache.h b/tools/perf/util/cache.h index 161d5f413e2..4b50c412b9c 100644 --- a/tools/perf/util/cache.h +++ b/tools/perf/util/cache.h @@ -18,6 +18,7 @@ #define PERFATTRIBUTES_FILE ".perfattributes" #define INFOATTRIBUTES_FILE "info/attributes" #define ATTRIBUTE_MACRO_PREFIX "[attr]" +#define PERF_DEBUGFS_ENVIRONMENT "PERF_DEBUGFS_DIR" typedef int (*config_fn_t)(const char *, const char *, void *); extern int perf_default_config(const char *, const char *, void *); diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c index 5a3cd3a34af..7bdad8df22a 100644 --- a/tools/perf/util/parse-events.c +++ b/tools/perf/util/parse-events.c @@ -5,6 +5,7 @@ #include "parse-events.h" #include "exec_cmd.h" #include "string.h" +#include "cache.h" extern char *strcasestr(const char *haystack, const char *needle); @@ -12,8 +13,6 @@ int nr_counters; struct perf_counter_attr attrs[MAX_COUNTERS]; -static char default_debugfs_path[] = "/sys/kernel/debug/tracing/events"; - struct event_symbol { u8 type; u64 config; @@ -21,6 +20,8 @@ struct event_symbol { char *alias; }; +char debugfs_path[MAXPATHLEN]; + #define CHW(x) .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_##x #define CSW(x) .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_##x @@ -114,27 +115,27 @@ static unsigned long hw_cache_stat[C(MAX)] = { #define for_each_subsystem(sys_dir, sys_dirent, sys_next, file, st) \ while (!readdir_r(sys_dir, &sys_dirent, &sys_next) && sys_next) \ - if (snprintf(file, MAXPATHLEN, "%s/%s", default_debugfs_path, \ - sys_dirent.d_name) && \ + if (snprintf(file, MAXPATHLEN, "%s/%s", debugfs_path, \ + sys_dirent.d_name) && \ (!stat(file, &st)) && (S_ISDIR(st.st_mode)) && \ (strcmp(sys_dirent.d_name, ".")) && \ (strcmp(sys_dirent.d_name, ".."))) #define for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next, file, st) \ while (!readdir_r(evt_dir, &evt_dirent, &evt_next) && evt_next) \ - if (snprintf(file, MAXPATHLEN, "%s/%s/%s", default_debugfs_path, \ - sys_dirent.d_name, evt_dirent.d_name) && \ + if (snprintf(file, MAXPATHLEN, "%s/%s/%s", debugfs_path, \ + sys_dirent.d_name, evt_dirent.d_name) && \ (!stat(file, &st)) && (S_ISDIR(st.st_mode)) && \ (strcmp(evt_dirent.d_name, ".")) && \ (strcmp(evt_dirent.d_name, ".."))) #define MAX_EVENT_LENGTH 30 -static int valid_debugfs_mount(void) +int valid_debugfs_mount(const char *debugfs) { struct statfs st_fs; - if (statfs(default_debugfs_path, &st_fs) < 0) + if (statfs(debugfs, &st_fs) < 0) return -ENOENT; else if (st_fs.f_type != (long) DEBUGFS_MAGIC) return -ENOENT; @@ -152,10 +153,10 @@ static char *tracepoint_id_to_name(u64 config) u64 id; char evt_path[MAXPATHLEN]; - if (valid_debugfs_mount()) + if (valid_debugfs_mount(debugfs_path)) return "unkown"; - sys_dir = opendir(default_debugfs_path); + sys_dir = opendir(debugfs_path); if (!sys_dir) goto cleanup; @@ -166,7 +167,7 @@ static char *tracepoint_id_to_name(u64 config) for_each_event(sys_dirent, evt_dir, evt_dirent, evt_next, evt_path, st) { snprintf(evt_path, MAXPATHLEN, "%s/%s/%s/id", - default_debugfs_path, sys_dirent.d_name, + debugfs_path, sys_dirent.d_name, evt_dirent.d_name); fd = open(evt_path, O_RDONLY); if (fd < 0) @@ -363,7 +364,7 @@ static int parse_tracepoint_event(const char **strp, u64 id; char evt_path[MAXPATHLEN]; - if (valid_debugfs_mount()) + if (valid_debugfs_mount(debugfs_path)) return 0; evt_name = strchr(*strp, ':'); @@ -381,8 +382,8 @@ static int parse_tracepoint_event(const char **strp, if (evt_length >= MAX_EVENT_LENGTH) return 0; - snprintf(evt_path, MAXPATHLEN, "%s/%s/%s/id", default_debugfs_path, - sys_name, evt_name); + snprintf(evt_path, MAXPATHLEN, "%s/%s/%s/id", debugfs_path, + sys_name, evt_name); fd = open(evt_path, O_RDONLY); if (fd < 0) return 0; @@ -568,10 +569,10 @@ static void print_tracepoint_events(void) struct stat st; char evt_path[MAXPATHLEN]; - if (valid_debugfs_mount()) + if (valid_debugfs_mount(debugfs_path)) return; - sys_dir = opendir(default_debugfs_path); + sys_dir = opendir(debugfs_path); if (!sys_dir) goto cleanup; diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h index e3d552908e6..1ea5d09b6eb 100644 --- a/tools/perf/util/parse-events.h +++ b/tools/perf/util/parse-events.h @@ -3,6 +3,8 @@ * Parse symbolic events/counts passed in as options: */ +struct option; + extern int nr_counters; extern struct perf_counter_attr attrs[MAX_COUNTERS]; @@ -15,3 +17,6 @@ extern int parse_events(const struct option *opt, const char *str, int unset); extern void print_events(void); +extern char debugfs_path[]; +extern int valid_debugfs_mount(const char *debugfs); + diff --git a/tools/perf/util/string.h b/tools/perf/util/string.h index 3dca2f654cd..bf39dfadfd2 100644 --- a/tools/perf/util/string.h +++ b/tools/perf/util/string.h @@ -5,4 +5,7 @@ int hex2u64(const char *ptr, u64 *val); +#define _STR(x) #x +#define STR(x) _STR(x) + #endif From d20ff6bd6bba2e7e6681fa17565347b410c46ab3 Mon Sep 17 00:00:00 2001 From: Mike Galbraith Date: Mon, 20 Jul 2009 14:01:38 +0200 Subject: [PATCH 30/31] perf_counter tools: Fix vmlinux symbol generation breakage vmlinux meets the criteria for symbol adjustment, which breaks vmlinux generated symbols. Fix this by exempting vmlinux. This is a bit fragile in that someone could change the kernel dso's name, but currently that name is also hardwired. Signed-off-by: Mike Galbraith Cc: Ingo Molnar Signed-off-by: Peter Zijlstra LKML-Reference: <1248091298.18702.18.camel@marge.simson.net> --- tools/perf/util/symbol.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c index 98aee922acc..28106059bf1 100644 --- a/tools/perf/util/symbol.c +++ b/tools/perf/util/symbol.c @@ -525,7 +525,7 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, GElf_Sym sym; Elf_Scn *sec, *sec_strndx; Elf *elf; - int nr = 0; + int nr = 0, kernel = !strcmp("[kernel]", self->name); elf = elf_begin(fd, ELF_C_READ_MMAP, NULL); if (elf == NULL) { @@ -571,10 +571,13 @@ static int dso__load_sym(struct dso *self, int fd, const char *name, nr_syms = shdr.sh_size / shdr.sh_entsize; memset(&sym, 0, sizeof(sym)); - self->adjust_symbols = (ehdr.e_type == ET_EXEC || + if (!kernel) { + self->adjust_symbols = (ehdr.e_type == ET_EXEC || elf_section_by_name(elf, &ehdr, &shdr, ".gnu.prelink_undo", NULL) != NULL); + } else self->adjust_symbols = 0; + elf_symtab__for_each_symbol(syms, nr_syms, index, sym) { struct symbol *f; const char *name; From 0fdc7e67dd312986e30b861adff48732bd33eb3f Mon Sep 17 00:00:00 2001 From: Mike Galbraith Date: Tue, 21 Jul 2009 10:30:36 +0200 Subject: [PATCH 31/31] perf_counter tools: Give perf top inherit option Currently, perf top -p only tracks the pid provided, which isn't very useful for watching forky loads, so give it an inherit option. Signed-off-by: Mike Galbraith Cc: Ingo Molnar Signed-off-by: Peter Zijlstra LKML-Reference: <1248165036.9795.10.camel@marge.simson.net> --- tools/perf/builtin-top.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c index 95d5c0ae375..c0a423004e1 100644 --- a/tools/perf/builtin-top.c +++ b/tools/perf/builtin-top.c @@ -58,6 +58,7 @@ static u64 count_filter = 5; static int print_entries = 15; static int target_pid = -1; +static int inherit = 0; static int profile_cpu = -1; static int nr_cpus = 0; static unsigned int realtime_prio = 0; @@ -549,7 +550,7 @@ int group_fd; static void start_counter(int i, int counter) { struct perf_counter_attr *attr; - unsigned int cpu; + int cpu; cpu = profile_cpu; if (target_pid == -1 && profile_cpu == -1) @@ -559,6 +560,7 @@ static void start_counter(int i, int counter) attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID; attr->freq = freq; + attr->inherit = (cpu < 0) && inherit; try_again: fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0); @@ -685,6 +687,8 @@ static const struct option options[] = { "only display functions with more events than this"), OPT_BOOLEAN('g', "group", &group, "put the counters into a counter group"), + OPT_BOOLEAN('i', "inherit", &inherit, + "child tasks inherit counters"), OPT_STRING('s', "sym-filter", &sym_filter, "pattern", "only display symbols matchig this pattern"), OPT_BOOLEAN('z', "zero", &zero,