perf_counter: software counter event infrastructure

Provide generic software counter infrastructure that supports
software events.

This will be used to allow sample based profiling based on software
events such as pagefaults. The current infrastructure can only
provide a count of such events, no place information.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index dde564517b661c3ae39af23412a32a2e917444f2..3fefc3b8150dff9263157b5d888ae5e9a4b5f4b1 100644 (file)
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -126,6 +126,7 @@ struct hw_perf_counter {
        unsigned long                   counter_base;
        int                             nmi;
        unsigned int                    idx;
+       atomic64_t                      count; /* software */
        atomic64_t                      prev_count;
        u64                             irq_period;
        atomic64_t                      period_left;
@@ -283,6 +284,8 @@ static inline int is_software_counter(struct perf_counter *counter)
        return !counter->hw_event.raw && counter->hw_event.type < 0;
 }
 
+extern void perf_swcounter_event(enum hw_event_types, u64, int, struct pt_regs *);
+
 #else
 static inline void
 perf_counter_task_sched_in(struct task_struct *task, int cpu)          { }
@@ -295,10 +298,13 @@ static inline void perf_counter_exit_task(struct task_struct *child)      { }
 static inline void perf_counter_notify(struct pt_regs *regs)           { }
 static inline void perf_counter_print_debug(void)                      { }
 static inline void perf_counter_unthrottle(void)                       { }
-static inline void hw_perf_restore(u64 ctrl)                   { }
+static inline void hw_perf_restore(u64 ctrl)                           { }
 static inline u64 hw_perf_save_disable(void)                 { return 0; }
 static inline int perf_counter_task_disable(void)      { return -EINVAL; }
 static inline int perf_counter_task_enable(void)       { return -EINVAL; }
+
+static inline void perf_swcounter_event(enum hw_event_types event, u64 nr,
+                                       int nmi, struct pt_regs *regs)  { }
 #endif
 
 #endif /* __KERNEL__ */

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 0fe22c916e29401f9c1170f29c32bfefcd73e7f6..eeb1b46cf707016cf3fa466e895bc8677f591616 100644 (file)
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1328,6 +1328,185 @@ static const struct file_operations perf_fops = {
        .compat_ioctl           = perf_ioctl,
 };
 
+/*
+ * Generic software counter infrastructure
+ */
+
+static void perf_swcounter_update(struct perf_counter *counter)
+{
+       struct hw_perf_counter *hwc = &counter->hw;
+       u64 prev, now;
+       s64 delta;
+
+again:
+       prev = atomic64_read(&hwc->prev_count);
+       now = atomic64_read(&hwc->count);
+       if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
+               goto again;
+
+       delta = now - prev;
+
+       atomic64_add(delta, &counter->count);
+       atomic64_sub(delta, &hwc->period_left);
+}
+
+static void perf_swcounter_set_period(struct perf_counter *counter)
+{
+       struct hw_perf_counter *hwc = &counter->hw;
+       s64 left = atomic64_read(&hwc->period_left);
+       s64 period = hwc->irq_period;
+
+       if (unlikely(left <= -period)) {
+               left = period;
+               atomic64_set(&hwc->period_left, left);
+       }
+
+       if (unlikely(left <= 0)) {
+               left += period;
+               atomic64_add(period, &hwc->period_left);
+       }
+
+       atomic64_set(&hwc->prev_count, -left);
+       atomic64_set(&hwc->count, -left);
+}
+
+static void perf_swcounter_save_and_restart(struct perf_counter *counter)
+{
+       perf_swcounter_update(counter);
+       perf_swcounter_set_period(counter);
+}
+
+static void perf_swcounter_store_irq(struct perf_counter *counter, u64 data)
+{
+       struct perf_data *irqdata = counter->irqdata;
+
+       if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
+               irqdata->overrun++;
+       } else {
+               u64 *p = (u64 *) &irqdata->data[irqdata->len];
+
+               *p = data;
+               irqdata->len += sizeof(u64);
+       }
+}
+
+static void perf_swcounter_handle_group(struct perf_counter *sibling)
+{
+       struct perf_counter *counter, *group_leader = sibling->group_leader;
+
+       list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
+               perf_swcounter_update(counter);
+               perf_swcounter_store_irq(sibling, counter->hw_event.type);
+               perf_swcounter_store_irq(sibling, atomic64_read(&counter->count));
+       }
+}
+
+static void perf_swcounter_interrupt(struct perf_counter *counter,
+                                    int nmi, struct pt_regs *regs)
+{
+       perf_swcounter_save_and_restart(counter);
+
+       switch (counter->hw_event.record_type) {
+       case PERF_RECORD_SIMPLE:
+               break;
+
+       case PERF_RECORD_IRQ:
+               perf_swcounter_store_irq(counter, instruction_pointer(regs));
+               break;
+
+       case PERF_RECORD_GROUP:
+               perf_swcounter_handle_group(counter);
+               break;
+       }
+
+       if (nmi) {
+               counter->wakeup_pending = 1;
+               set_tsk_thread_flag(current, TIF_PERF_COUNTERS);
+       } else
+               wake_up(&counter->waitq);
+}
+
+static int perf_swcounter_match(struct perf_counter *counter,
+                               enum hw_event_types event,
+                               struct pt_regs *regs)
+{
+       if (counter->state != PERF_COUNTER_STATE_ACTIVE)
+               return 0;
+
+       if (counter->hw_event.raw)
+               return 0;
+
+       if (counter->hw_event.type != event)
+               return 0;
+
+       if (counter->hw_event.exclude_user && user_mode(regs))
+               return 0;
+
+       if (counter->hw_event.exclude_kernel && !user_mode(regs))
+               return 0;
+
+       return 1;
+}
+
+static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
+                                    enum hw_event_types event, u64 nr,
+                                    int nmi, struct pt_regs *regs)
+{
+       struct perf_counter *counter;
+       unsigned long flags;
+       int neg;
+
+       if (list_empty(&ctx->counter_list))
+               return;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+
+       /*
+        * XXX: make counter_list RCU safe
+        */
+       list_for_each_entry(counter, &ctx->counter_list, list_entry) {
+               if (perf_swcounter_match(counter, event, regs)) {
+                       neg = atomic64_add_negative(nr, &counter->hw.count);
+                       if (counter->hw.irq_period && !neg)
+                               perf_swcounter_interrupt(counter, nmi, regs);
+               }
+       }
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+}
+
+void perf_swcounter_event(enum hw_event_types event, u64 nr,
+                         int nmi, struct pt_regs *regs)
+{
+       struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
+
+       perf_swcounter_ctx_event(&cpuctx->ctx, event, nr, nmi, regs);
+       if (cpuctx->task_ctx)
+               perf_swcounter_ctx_event(cpuctx->task_ctx, event, nr, nmi, regs);
+
+       put_cpu_var(perf_cpu_context);
+}
+
+static void perf_swcounter_read(struct perf_counter *counter)
+{
+       perf_swcounter_update(counter);
+}
+
+static int perf_swcounter_enable(struct perf_counter *counter)
+{
+       perf_swcounter_set_period(counter);
+       return 0;
+}
+
+static void perf_swcounter_disable(struct perf_counter *counter)
+{
+       perf_swcounter_update(counter);
+}
+
+/*
+ * Software counter: cpu wall time clock
+ */
+
 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
 {
        int cpu = raw_smp_processor_id();
@@ -1364,6 +1543,10 @@ static const struct hw_perf_counter_ops perf_ops_cpu_clock = {
        .read           = cpu_clock_perf_counter_read,
 };
 
+/*
+ * Software counter: task time clock
+ */
+
 /*
  * Called from within the scheduler:
  */
@@ -1420,6 +1603,10 @@ static const struct hw_perf_counter_ops perf_ops_task_clock = {
        .read           = task_clock_perf_counter_read,
 };
 
+/*
+ * Software counter: page faults
+ */
+
 #ifdef CONFIG_VM_EVENT_COUNTERS
 #define cpu_page_faults()      __get_cpu_var(vm_event_states).event[PGFAULT]
 #else
@@ -1473,6 +1660,10 @@ static const struct hw_perf_counter_ops perf_ops_page_faults = {
        .read           = page_faults_perf_counter_read,
 };
 
+/*
+ * Software counter: context switches
+ */
+
 static u64 get_context_switches(struct perf_counter *counter)
 {
        struct task_struct *curr = counter->ctx->task;
@@ -1521,6 +1712,10 @@ static const struct hw_perf_counter_ops perf_ops_context_switches = {
        .read           = context_switches_perf_counter_read,
 };
 
+/*
+ * Software counter: cpu migrations
+ */
+
 static inline u64 get_cpu_migrations(struct perf_counter *counter)
 {
        struct task_struct *curr = counter->ctx->task;
@@ -1572,7 +1767,9 @@ static const struct hw_perf_counter_ops perf_ops_cpu_migrations = {
 static const struct hw_perf_counter_ops *
 sw_perf_counter_init(struct perf_counter *counter)
 {
+       struct perf_counter_hw_event *hw_event = &counter->hw_event;
        const struct hw_perf_counter_ops *hw_ops = NULL;
+       struct hw_perf_counter *hwc = &counter->hw;
 
        /*
         * Software counters (currently) can't in general distinguish
@@ -1618,6 +1815,10 @@ sw_perf_counter_init(struct perf_counter *counter)
        default:
                break;
        }
+
+       if (hw_ops)
+               hwc->irq_period = hw_event->irq_period;
+
        return hw_ops;
 }