perf: add struct hw_perf_event

Change-Id: I0938e2b18064ad805a9edb6e15d26cf438bf0a59

arch/arm64/kernel/perfctr.c

@@ -215,3 +215,20 @@ int ihk_mc_perfctr_set_extra(struct mc_perf_event *event)
 	/* Nothing to do. */
 	return 0;
 }
+
+static inline uint64_t arm_pmu_event_max_period(struct mc_perf_event *event)
+{
+	return 0xFFFFFFFF;
+}
+
+int hw_perf_event_init(struct mc_perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+
+	if (!is_sampling_event(event)) {
+		hwc->sample_period  = arm_pmu_event_max_period(event) >> 1;
+		hwc->last_period    = hwc->sample_period;
+		ihk_atomic64_set(&hwc->period_left, hwc->sample_period);
+	}
+	return 0;
+}

arch/x86_64/kernel/perfctr.c

@@ -503,3 +503,8 @@ int ihk_mc_perf_get_num_counters(void)
 {
 	return NUM_PERF_COUNTERS;
 }
+
+int hw_perf_event_init(struct mc_perf_event *event)
+{
+	return 0;
+}

kernel/syscall.c

@@ -4519,6 +4519,7 @@ static int mc_perf_event_alloc(struct mc_perf_event **out,
 	unsigned long val = 0, extra_config = 0;
 	struct mc_perf_event *event = NULL;
 	int ereg_id;
+	struct hw_perf_event *hwc;
 
 	if (!attr) {
 		ret = -EINVAL;
@@ -4542,6 +4543,21 @@ static int mc_perf_event_alloc(struct mc_perf_event **out,
 	event->child_count_total = 0;
 	event->parent = NULL;
 
+	hwc = &event->hw;
+	hwc->sample_period = attr->sample_period;
+	if (attr->freq && attr->sample_freq) {
+		/*
+		 * Mark struct perf_event_attr::sample_freq is set by user.
+		 * Note that it's okay to use
+		 * struct hw_perf_event::sample_period for this purpose
+		 * because it's not union and not used when
+		 * struct perf_event_attr::freq is one.
+		 */
+		hwc->sample_period = 1;
+	}
+	hwc->last_period = hwc->sample_period;
+	ihk_atomic64_set(&hwc->period_left, hwc->sample_period);
+
 	if (attr->type == PERF_TYPE_HARDWARE &&
 		attr->config == PERF_COUNT_HW_REF_CPU_CYCLES) {
 		event->use_invariant_tsc = 1;
@@ -4588,6 +4604,8 @@ static int mc_perf_event_alloc(struct mc_perf_event **out,
 		event->extra_reg.idx = ihk_mc_get_extra_reg_idx(ereg_id);
 	}
 
+	ret = hw_perf_event_init(event);
+
 	*out = event;
 
 out:
@@ -4646,6 +4664,14 @@ SYSCALL_DECLARE(perf_event_open)
 		not_supported_flag = 1;
 	}
 
+	if (attr->freq) {
+		not_supported_flag = 1;
+	} else {
+		if (attr->sample_period & (1ULL << 63)) {
+			return -EINVAL;
+		}
+	}
+
 	if (not_supported_flag) {
 		return -ENOENT;
 	}

lib/include/ihk/cpu.h

@@ -17,6 +17,7 @@
 #include <list.h>
 #include <ihk/context.h>
 #include <arch/cpu.h>
+#include <mc_perf_event.h>
 
 extern int num_processors;
 
@@ -89,6 +90,12 @@ unsigned long ihk_mc_hw_cache_event_map(unsigned long hw_cache_event);
 unsigned long ihk_mc_hw_cache_extra_reg_map(unsigned long hw_cache_event);
 unsigned long ihk_mc_raw_event_map(unsigned long raw_event);
 int ihk_mc_validate_event(unsigned long hw_config);
+int hw_perf_event_init(struct mc_perf_event *event);
+
+static inline int is_sampling_event(struct mc_perf_event *event)
+{
+	return event->attr.sample_period != 0;
+}
 
 /* returns the 'prev' argument of the call that caused the switch to the context returned. */
 void *ihk_mc_switch_context(ihk_mc_kernel_context_t *old_ctx,

lib/include/mc_perf_event.h

@@ -257,6 +257,31 @@ struct hw_perf_event_extra {
 	int idx;
 };
 
+struct hw_perf_event {
+	/*
+	 * The last observed hardware counter value, updated with a
+	 * local64_cmpxchg() such that pmu::read() can be called nested.
+	 */
+	ihk_atomic64_t			prev_count;
+
+	/*
+	 * The period to start the next sample with.
+	 */
+	uint64_t			sample_period;
+
+	/*
+	 * The period we started this sample with.
+	 */
+	uint64_t			last_period;
+
+	/*
+	 * However much is left of the current period; note that this is
+	 * a full 64bit value and allows for generation of periods longer
+	 * than hardware might allow.
+	 */
+	ihk_atomic64_t			period_left;
+};
+
 /**
  * enum perf_event_state - the states of an event:
  */
@@ -291,6 +316,7 @@ struct mc_perf_event {
 	long long			base_system_tsc;
 	long long			stopped_system_tsc;
 	long long			system_accum_count;
+	struct hw_perf_event		hw;
 };
 
 struct perf_event_mmap_page {