Add perf tracepoint collection flag (#1664)

* Add tracepoint collector option for perf collector

Signed-off-by: Daniel Hodges <hodges.daniel.scott@gmail.com>

README.md

@@ -96,6 +96,13 @@ configuration is zero indexed and can also take a stride value; e.g.
 `--collector.perf --collector.perf.cpus=1-10:5` would collect on CPUs
 1, 5, and 10.
 
+The perf collector is also able to collect
+[tracepoint](https://www.kernel.org/doc/html/latest/core-api/tracepoint.html)
+counts when using the `--collector.perf.tracepoint` flag. Tracepoints can be
+found using [`perf list`](http://man7.org/linux/man-pages/man1/perf.1.html) or
+from debugfs. And example usage of this would be
+`--collector.perf.tracepoint="sched:sched_process_exec"`.
+
 
 Name     | Description | OS
 ---------|-------------|----

collector/perf_linux.go

@@ -20,8 +20,10 @@ import (
 	"strings"
 
 	"github.com/go-kit/kit/log"
+	"github.com/go-kit/kit/log/level"
 	"github.com/hodgesds/perf-utils"
 	"github.com/prometheus/client_golang/prometheus"
+	"golang.org/x/sys/unix"
 	kingpin "gopkg.in/alecthomas/kingpin.v2"
 )
 
@@ -31,26 +33,28 @@ const (
 
 var (
 	perfCPUsFlag       = kingpin.Flag("collector.perf.cpus", "List of CPUs from which perf metrics should be collected").Default("").String()
+	perfTracepointFlag = kingpin.Flag("collector.perf.tracepoint", "perf tracepoint that should be collected").Strings()
 )
 
 func init() {
 	registerCollector(perfSubsystem, defaultDisabled, NewPerfCollector)
 }
 
-// perfCollector is a Collector that uses the perf subsystem to collect
+// perfTracepointFlagToTracepoints returns the set of configured tracepoints.
-// metrics. It uses perf_event_open an ioctls for profiling. Due to the fact
+func perfTracepointFlagToTracepoints(tracepointsFlag []string) ([]*perfTracepoint, error) {
-// that the perf subsystem is highly dependent on kernel configuration and
+	tracepoints := make([]*perfTracepoint, len(tracepointsFlag))
-// settings not all profiler values may be exposed on the target system at any
+
-// given time.
+	for i, tracepoint := range tracepointsFlag {
-type perfCollector struct {
+		split := strings.Split(tracepoint, ":")
-	hwProfilerCPUMap    map[*perf.HardwareProfiler]int
+		if len(split) != 2 {
-	swProfilerCPUMap    map[*perf.SoftwareProfiler]int
+			return nil, fmt.Errorf("Invalid tracepoint config %v", tracepoint)
-	cacheProfilerCPUMap map[*perf.CacheProfiler]int
+		}
-	perfHwProfilers     map[int]*perf.HardwareProfiler
+		tracepoints[i] = &perfTracepoint{
-	perfSwProfilers     map[int]*perf.SoftwareProfiler
+			subsystem: split[0],
-	perfCacheProfilers  map[int]*perf.CacheProfiler
+			event:     split[1],
-	desc                map[string]*prometheus.Desc
+		}
-	logger              log.Logger
+	}
+	return tracepoints, nil
 }
 
 // perfCPUFlagToCPUs returns a set of CPUs for the perf collectors to monitor.
@@ -98,6 +102,144 @@ func perfCPUFlagToCPUs(cpuFlag string) ([]int, error) {
 	return cpus, nil
 }
 
+// perfTracepoint is a struct for holding tracepoint information.
+type perfTracepoint struct {
+	subsystem string
+	event     string
+}
+
+// label returns the tracepoint name in the format of subsystem_tracepoint.
+func (t *perfTracepoint) label() string {
+	return t.subsystem + "_" + t.event
+}
+
+// tracepoint returns the tracepoint name in the format of subsystem:tracepoint.
+func (t *perfTracepoint) tracepoint() string {
+	return t.subsystem + ":" + t.event
+}
+
+// perfCollector is a Collector that uses the perf subsystem to collect
+// metrics. It uses perf_event_open an ioctls for profiling. Due to the fact
+// that the perf subsystem is highly dependent on kernel configuration and
+// settings not all profiler values may be exposed on the target system at any
+// given time.
+type perfCollector struct {
+	hwProfilerCPUMap    map[*perf.HardwareProfiler]int
+	swProfilerCPUMap    map[*perf.SoftwareProfiler]int
+	cacheProfilerCPUMap map[*perf.CacheProfiler]int
+	perfHwProfilers     map[int]*perf.HardwareProfiler
+	perfSwProfilers     map[int]*perf.SoftwareProfiler
+	perfCacheProfilers  map[int]*perf.CacheProfiler
+	desc                map[string]*prometheus.Desc
+	logger              log.Logger
+	tracepointCollector *perfTracepointCollector
+}
+
+type perfTracepointCollector struct {
+	// desc is the mapping of subsystem to tracepoint *prometheus.Desc.
+	descs map[string]map[string]*prometheus.Desc
+	// collection order is the sorted configured collection order of the profiler.
+	collectionOrder []string
+
+	logger    log.Logger
+	profilers map[int]perf.GroupProfiler
+}
+
+// update is used collect all tracepoints across all tracepoint profilers.
+func (c *perfTracepointCollector) update(ch chan<- prometheus.Metric) error {
+	for cpu := range c.profilers {
+		if err := c.updateCPU(cpu, ch); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// updateCPU is used to update metrics per CPU profiler.
+func (c *perfTracepointCollector) updateCPU(cpu int, ch chan<- prometheus.Metric) error {
+	cpuStr := fmt.Sprintf("%d", cpu)
+	profiler := c.profilers[cpu]
+	p, err := profiler.Profile()
+	if err != nil {
+		level.Error(c.logger).Log("msg", "Failed to collect tracepoint profile", "err", err)
+		return err
+	}
+
+	for i, value := range p.Values {
+		// Get the Desc from the ordered group value.
+		descKey := c.collectionOrder[i]
+		descKeySlice := strings.Split(descKey, ":")
+		ch <- prometheus.MustNewConstMetric(
+			c.descs[descKeySlice[0]][descKeySlice[1]],
+			prometheus.CounterValue,
+			float64(value),
+			cpuStr,
+		)
+	}
+	return nil
+}
+
+// newPerfTracepointCollector returns a configured perfTracepointCollector.
+func newPerfTracepointCollector(
+	logger log.Logger,
+	tracepointsFlag []string,
+	cpus []int,
+) (*perfTracepointCollector, error) {
+	tracepoints, err := perfTracepointFlagToTracepoints(tracepointsFlag)
+	if err != nil {
+		return nil, err
+	}
+
+	collectionOrder := make([]string, len(tracepoints))
+	descs := map[string]map[string]*prometheus.Desc{}
+	eventAttrs := make([]unix.PerfEventAttr, len(tracepoints))
+
+	for i, tracepoint := range tracepoints {
+		eventAttr, err := perf.TracepointEventAttr(tracepoint.subsystem, tracepoint.event)
+		if err != nil {
+			return nil, err
+		}
+		eventAttrs[i] = *eventAttr
+		collectionOrder[i] = tracepoint.tracepoint()
+		if _, ok := descs[tracepoint.subsystem]; !ok {
+			descs[tracepoint.subsystem] = map[string]*prometheus.Desc{}
+		}
+		descs[tracepoint.subsystem][tracepoint.event] = prometheus.NewDesc(
+			prometheus.BuildFQName(
+				namespace,
+				perfSubsystem,
+				tracepoint.label(),
+			),
+			"Perf tracepoint "+tracepoint.tracepoint(),
+			[]string{"cpu"},
+			nil,
+		)
+	}
+
+	profilers := make(map[int]perf.GroupProfiler, len(cpus))
+	for _, cpu := range cpus {
+		profiler, err := perf.NewGroupProfiler(-1, cpu, 0, eventAttrs...)
+		if err != nil {
+			return nil, err
+		}
+		profilers[cpu] = profiler
+	}
+
+	c := &perfTracepointCollector{
+		descs:           descs,
+		collectionOrder: collectionOrder,
+		profilers:       profilers,
+		logger:          logger,
+	}
+
+	for _, profiler := range c.profilers {
+		if err := profiler.Start(); err != nil {
+			return nil, err
+		}
+	}
+	return c, nil
+}
+
 // NewPerfCollector returns a new perf based collector, it creates a profiler
 // per CPU.
 func NewPerfCollector(logger log.Logger) (Collector, error) {
@@ -127,6 +269,16 @@ func NewPerfCollector(logger log.Logger) (Collector, error) {
 		}
 	}
 
+	// First configure any tracepoints.
+	if *perfTracepointFlag != nil && len(*perfTracepointFlag) > 0 {
+		tracepointCollector, err := newPerfTracepointCollector(logger, *perfTracepointFlag, cpus)
+		if err != nil {
+			return nil, err
+		}
+		collector.tracepointCollector = tracepointCollector
+	}
+
+	// Configure all profilers for the specified CPUs.
 	for _, cpu := range cpus {
 		// Use -1 to profile all processes on the CPU, see:
 		// man perf_event_open
@@ -411,6 +563,9 @@ func (c *perfCollector) Update(ch chan<- prometheus.Metric) error {
 	if err := c.updateCacheStats(ch); err != nil {
 		return err
 	}
+	if c.tracepointCollector != nil {
+		return c.tracepointCollector.update(ch)
+	}
 
 	return nil
 }

collector/perf_linux_test.go

@@ -180,8 +180,74 @@ func TestPerfCPUFlagToCPUs(t *testing.T) {
 				if test.exCpus[i] != cpus[i] {
 					t.Fatalf(
 						"expected cpus %v, got %v",
-						test.exCpus,
+						test.exCpus[i],
-						cpus,
+						cpus[i],
+					)
+				}
+			}
+		})
+	}
+}
+
+func TestPerfTracepointFlagToTracepoints(t *testing.T) {
+	tests := []struct {
+		name          string
+		flag          []string
+		exTracepoints []*perfTracepoint
+		errStr        string
+	}{
+		{
+			name: "valid single tracepoint",
+			flag: []string{"sched:sched_kthread_stop"},
+			exTracepoints: []*perfTracepoint{
+				{
+					subsystem: "sched",
+					event:     "sched_kthread_stop",
+				},
+			},
+		},
+		{
+			name: "valid multiple tracepoints",
+			flag: []string{"sched:sched_kthread_stop", "sched:sched_process_fork"},
+			exTracepoints: []*perfTracepoint{
+				{
+					subsystem: "sched",
+					event:     "sched_kthread_stop",
+				},
+				{
+					subsystem: "sched",
+					event:     "sched_process_fork",
+				},
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			tracepoints, err := perfTracepointFlagToTracepoints(test.flag)
+			if test.errStr != "" {
+				if err != nil {
+					t.Fatal("expected error to not be nil")
+				}
+				if test.errStr != err.Error() {
+					t.Fatalf(
+						"expected error %q, got %q",
+						test.errStr,
+						err.Error(),
+					)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatal(err)
+			}
+			for i := range tracepoints {
+				if test.exTracepoints[i].event != tracepoints[i].event &&
+					test.exTracepoints[i].subsystem != tracepoints[i].subsystem {
+					t.Fatalf(
+						"expected tracepoint %v, got %v",
+						test.exTracepoints[i],
+						tracepoints[i],
 					)
 				}
 			}