From 8f7aeebfc21af051ecb914d286564a10492927a0 Mon Sep 17 00:00:00 2001 From: Remy Noel Date: Mon, 30 Apr 2018 20:55:28 +0200 Subject: perfmon: add polling overflow implementation in perfmon --- arch/x86/machine/pmu_amd.c | 38 ++++------------------------------ arch/x86/machine/pmu_intel.c | 49 +++++++++++--------------------------------- arch/x86/machine/trap.c | 1 + 3 files changed, 17 insertions(+), 71 deletions(-) (limited to 'arch/x86/machine') diff --git a/arch/x86/machine/pmu_amd.c b/arch/x86/machine/pmu_amd.c index 34df620..8e56bfa 100644 --- a/arch/x86/machine/pmu_amd.c +++ b/arch/x86/machine/pmu_amd.c @@ -21,6 +21,7 @@ #include #include +#include #include #include #include @@ -175,7 +176,6 @@ pmu_amd_start(unsigned int pmc_id, unsigned int raw_event_id) /* TODO Handle PERFMON_EF_KERN/PERFMON_EF_USER */ high = code->event_select >> 8; low = PMU_AMD_EVTSEL_EN - | PMU_AMD_EVTSEL_INT | PMU_AMD_EVTSEL_OS | PMU_AMD_EVTSEL_USR | (code->umask << 8) @@ -205,38 +205,6 @@ pmu_amd_write(unsigned int pmc_id, uint64_t value) cpu_set_msr64(PMU_AMD_MSR_PERCTR0 + pmc_id, value); } -/* - * TODO Make the perfmon module handle basic overflow handling by polling - * counters. - */ -static void -pmu_amd_handle_of_intr_v1(void) -{ - struct pmu_amd *pmu; - uint64_t value, prev; - unsigned int mask; - - pmu = pmu_amd_get(); - - for (unsigned int pmc_id = 0; pmc_id != PMU_AMD_NR_PMCS; pmc_id++) { - mask = (1U << pmc_id); - - if (pmu->pmc_bm & mask) { - continue; - } - - value = pmu_amd_read(pmc_id); - prev = perfmon_cpu_pmc_get_prev(pmc_id); - - if (prev > value) { - /* Overflow */ - perfmon_cpu_pmc_inc_of(pmc_id); - /* Prevents us from overflowing twice */ - perfmon_cpu_pmc_set_prev(pmc_id, value); - } - } -} - static int __init pmu_amd_setup(void) { @@ -259,7 +227,9 @@ pmu_amd_setup(void) pmu->pmc_bm = (1U << PMU_AMD_NR_PMCS) - 1; pmu_driver.pmc_width = PMU_AMD_PMC_WIDTH; - pmu_driver.of_max_ticks = 1UL << (pmu_driver.pmc_width - 1); + /* Set max_tick to half the number of instruction per seconds. */ + pmu_driver.of_max_ticks = + (1UL << (pmu_driver.pmc_width - 1)) / (cpu_get_freq() / CLOCK_FREQ); pmu_driver.ops.info = pmu_amd_info; pmu_driver.ops.translate = pmu_amd_translate; diff --git a/arch/x86/machine/pmu_intel.c b/arch/x86/machine/pmu_intel.c index ccc2294..e5d9ef7 100644 --- a/arch/x86/machine/pmu_intel.c +++ b/arch/x86/machine/pmu_intel.c @@ -21,6 +21,7 @@ #include #include +#include #include #include #include @@ -174,7 +175,7 @@ pmu_intel_info(void) pmu = pmu_intel_get(); nr_events = pmu_popcount(pmu->events); - log_info("pmu: driver: intel, architectural v%d\n" + log_info("pmu: driver: intel, architectural v%d " "pmu: nr_pmcs: %u, pmc_width: %u, events: %#x, nr_events: %u\n", pmu->version, pmu->nr_pmcs, pmu->pmc_width, pmu->events, nr_events); @@ -233,19 +234,23 @@ static void pmu_intel_start(unsigned int pmc_id, unsigned int raw_event_id) { const struct pmu_intel_event_code *code; + struct pmu_intel *pmu; uint32_t evtsel; assert(raw_event_id < ARRAY_SIZE(pmu_intel_event_codes)); code = &pmu_intel_event_codes[raw_event_id]; + pmu = pmu_intel_get(); /* TODO Handle PERFMON_EF_KERN/PERFMON_EF_USER */ evtsel = PMU_INTEL_EVTSEL_EN | PMU_INTEL_EVTSEL_OS | PMU_INTEL_EVTSEL_USR - | PMU_INTEL_EVTSEL_INT | (code->umask << 8) | code->event_select; + if (pmu->version >= 2) { + evtsel |= PMU_INTEL_EVTSEL_INT; + } cpu_set_msr(PMU_INTEL_MSR_EVTSEL0 + pmc_id, 0, evtsel); } @@ -267,38 +272,6 @@ pmu_intel_write(unsigned int pmc_id, uint64_t value) cpu_set_msr64(PMU_INTEL_MSR_PMC0 + pmc_id, value); } -/* - * TODO Make the perfmon module handle basic overflow handling by polling - * counters. - */ -static void -pmu_intel_handle_of_intr_v1(void) -{ - struct pmu_intel *pmu; - unsigned int mask; - uint64_t value; - uint64_t prev; - - pmu = pmu_intel_get(); - - for (unsigned int pmc_id = 0; pmc_id != pmu->nr_pmcs; pmc_id++) { - mask = (1U << pmc_id); - if (pmu->pmc_bm & mask) { - /* counter not enabled: can't overflow. */ - continue; - } - - value = pmu_intel_read(pmc_id); - prev = perfmon_cpu_pmc_get_prev(pmc_id); - if (prev > value) { - /* Overflow */ - perfmon_cpu_pmc_inc_of(pmc_id); - /* Prevents us from overflowing twice */ - perfmon_cpu_pmc_set_prev(pmc_id, value); - } - } -} - static int pmu_intel_consume_bits(uint64_t *bits) { @@ -330,7 +303,7 @@ pmu_intel_handle_of_intr_v2(void) pmu_intel_ack_status(status); pmu = pmu_intel_get(); - status &= ((1U << pmu->pmc_width) - 1); + status &= ((1ULL << pmu->pmc_width) - 1); for (;;) { pmc_id = pmu_intel_consume_bits(&status); @@ -366,7 +339,6 @@ pmu_intel_setup(void) cpu_cpuid(&eax, &ebx, &ecx, &edx); pmu->version = eax & PMU_INTEL_ID_VERSION_MASK; - /* TODO Check this */ if (pmu->version == 0) { return ENODEV; } @@ -396,7 +368,10 @@ pmu_intel_setup(void) pmu_driver.ops.handle_of_intr = pmu_intel_handle_of_intr_v2; pmu_driver.of_max_ticks = 0; } else { - pmu_driver.of_max_ticks = 1UL << (pmu_driver.pmc_width - 1); + /* Set max_tick to half the number of instruction per seconds. */ + pmu_driver.ops.handle_of_intr = NULL; + pmu_driver.of_max_ticks = + (1ULL << (pmu_driver.pmc_width - 1)) / (cpu_get_freq() / CLOCK_FREQ); } return perfmon_pmu_register(&pmu_driver); diff --git a/arch/x86/machine/trap.c b/arch/x86/machine/trap.c index 101adf8..a7a5cbd 100644 --- a/arch/x86/machine/trap.c +++ b/arch/x86/machine/trap.c @@ -214,6 +214,7 @@ trap_setup(void) trap_install(TRAP_LAPIC_TIMER, TRAP_HF_INTR, lapic_timer_intr); trap_install(TRAP_LAPIC_ERROR, TRAP_HF_INTR, lapic_error_intr); trap_install(TRAP_LAPIC_SPURIOUS, TRAP_HF_INTR, lapic_spurious_intr); + trap_install(TRAP_LAPIC_PMC_OF, TRAP_HF_INTR, lapic_pmc_of_intr); return 0; } -- cgit v1.2.3