cmd/trace: Record mark assists in execution traces

During the mark phase of garbage collection, goroutines that allocate
may be recruited to assist. This change creates trace events for mark
assists and displays them similarly to sweep assists in the trace
viewer.

Mark assists are different than sweeps in that they can be preempted, so
displaying them in the trace viewer is a little tricky -- we may need to
synthesize multiple slices for one mark assist. This could have been
done in the parser instead, but I thought it might be preferable to keep
the parser as true to the event stream as possible.

Change-Id: I381dcb1027a187a354b1858537851fa68a620ea7
Reviewed-on: https://go-review.googlesource.com/36015
Run-TryBot: Heschi Kreinick <heschi@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Hyang-Ah Hana Kim <hyangah@gmail.com>

src/cmd/trace/trace.go

@@ -288,6 +288,13 @@ const (
 	gStateCount
 )
 
+type gInfo struct {
+	state      gState       // current state
+	name       string       // name chosen for this goroutine at first EvGoStart
+	start      *trace.Event // most recent EvGoStart
+	markAssist *trace.Event // if non-nil, the mark assist currently running.
+}
+
 type ViewerData struct {
 	Events   []*ViewerEvent         `json:"traceEvents"`
 	Frames   map[string]ViewerFrame `json:"stackFrames"`
@@ -337,35 +344,47 @@ func generateTrace(params *traceParams) (ViewerData, error) {
 	ctx.data.Frames = make(map[string]ViewerFrame)
 	ctx.data.TimeUnit = "ns"
 	maxProc := 0
-	gnames := make(map[uint64]string)
-	gstates := make(map[uint64]gState)
+	ginfos := make(map[uint64]*gInfo)
+
+	getGInfo := func(g uint64) *gInfo {
+		info, ok := ginfos[g]
+		if !ok {
+			info = &gInfo{}
+			ginfos[g] = info
+		}
+		return info
+	}
+
 	// Since we make many calls to setGState, we record a sticky
 	// error in setGStateErr and check it after every event.
 	var setGStateErr error
 	setGState := func(ev *trace.Event, g uint64, oldState, newState gState) {
-		if oldState == gWaiting && gstates[g] == gWaitingGC {
+		info := getGInfo(g)
+		if oldState == gWaiting && info.state == gWaitingGC {
 			// For checking, gWaiting counts as any gWaiting*.
-			oldState = gstates[g]
+			oldState = info.state
 		}
-		if gstates[g] != oldState && setGStateErr == nil {
+		if info.state != oldState && setGStateErr == nil {
 			setGStateErr = fmt.Errorf("expected G %d to be in state %d, but got state %d", g, oldState, newState)
 		}
-		ctx.gstates[gstates[g]]--
+		ctx.gstates[info.state]--
 		ctx.gstates[newState]++
-		gstates[g] = newState
+		info.state = newState
 	}
 	for _, ev := range ctx.events {
 		// Handle state transitions before we filter out events.
 		switch ev.Type {
 		case trace.EvGoStart, trace.EvGoStartLabel:
 			setGState(ev, ev.G, gRunnable, gRunning)
-			if _, ok := gnames[ev.G]; !ok {
+			info := getGInfo(ev.G)
+			if info.name == "" {
 				if len(ev.Stk) > 0 {
-					gnames[ev.G] = fmt.Sprintf("G%v %s", ev.G, ev.Stk[0].Fn)
+					info.name = fmt.Sprintf("G%v %s", ev.G, ev.Stk[0].Fn)
 				} else {
-					gnames[ev.G] = fmt.Sprintf("G%v", ev.G)
+					info.name = fmt.Sprintf("G%v", ev.G)
 				}
 			}
+			info.start = ev
 		case trace.EvProcStart:
 			ctx.threadStats.prunning++
 		case trace.EvProcStop:
@@ -392,6 +411,10 @@ func generateTrace(params *traceParams) (ViewerData, error) {
 			setGState(ev, ev.G, gRunning, gWaiting)
 		case trace.EvGoBlockGC:
 			setGState(ev, ev.G, gRunning, gWaitingGC)
+		case trace.EvGCMarkAssistStart:
+			getGInfo(ev.G).markAssist = ev
+		case trace.EvGCMarkAssistDone:
+			getGInfo(ev.G).markAssist = nil
 		case trace.EvGoWaiting:
 			setGState(ev, ev.G, gRunnable, gWaiting)
 		case trace.EvGoInSyscall:
@@ -444,13 +467,41 @@ func generateTrace(params *traceParams) (ViewerData, error) {
 			}
 			ctx.emitSlice(ev, "MARK TERMINATION")
 		case trace.EvGCScanDone:
+		case trace.EvGCMarkAssistStart:
+			// Mark assists can continue past preemptions, so truncate to the
+			// whichever comes first. We'll synthesize another slice if
+			// necessary in EvGoStart.
+			markFinish := ev.Link
+			goFinish := getGInfo(ev.G).start.Link
+			fakeMarkStart := *ev
+			text := "MARK ASSIST"
+			if markFinish.Ts > goFinish.Ts {
+				fakeMarkStart.Link = goFinish
+				text = "MARK ASSIST (unfinished)"
+			}
+			ctx.emitSlice(&fakeMarkStart, text)
 		case trace.EvGCSweepStart:
 			ctx.emitSlice(ev, "SWEEP")
-		case trace.EvGCSweepDone:
-		case trace.EvGoStart:
-			ctx.emitSlice(ev, gnames[ev.G])
-		case trace.EvGoStartLabel:
-			ctx.emitSlice(ev, ev.SArgs[0])
+		case trace.EvGoStart, trace.EvGoStartLabel:
+			info := getGInfo(ev.G)
+			if ev.Type == trace.EvGoStartLabel {
+				ctx.emitSlice(ev, ev.SArgs[0])
+			} else {
+				ctx.emitSlice(ev, info.name)
+			}
+			if info.markAssist != nil {
+				// If we're in a mark assist, synthesize a new slice, ending
+				// either when the mark assist ends or when we're descheduled.
+				markFinish := info.markAssist.Link
+				goFinish := ev.Link
+				fakeMarkStart := *ev
+				text := "MARK ASSIST (resumed, unfinished)"
+				if markFinish.Ts < goFinish.Ts {
+					fakeMarkStart.Link = markFinish
+					text = "MARK ASSIST (resumed)"
+				}
+				ctx.emitSlice(&fakeMarkStart, text)
+			}
 		case trace.EvGoCreate:
 			ctx.emitArrow(ev, "go")
 		case trace.EvGoUnblock:
@@ -493,11 +544,11 @@ func generateTrace(params *traceParams) (ViewerData, error) {
 	}
 
 	if ctx.gtrace && ctx.gs != nil {
-		for k, v := range gnames {
+		for k, v := range ginfos {
 			if !ctx.gs[k] {
 				continue
 			}
-			ctx.emit(&ViewerEvent{Name: "thread_name", Phase: "M", Pid: 0, Tid: k, Arg: &NameArg{v}})
+			ctx.emit(&ViewerEvent{Name: "thread_name", Phase: "M", Pid: 0, Tid: k, Arg: &NameArg{v.name}})
 		}
 		ctx.emit(&ViewerEvent{Name: "thread_sort_index", Phase: "M", Pid: 0, Tid: ctx.maing, Arg: &SortIndexArg{-2}})
 		ctx.emit(&ViewerEvent{Name: "thread_sort_index", Phase: "M", Pid: 0, Tid: 0, Arg: &SortIndexArg{-1}})

src/cmd/trace/trace_test.go

@@ -2,6 +2,7 @@ package main
 
 import (
 	"internal/trace"
+	"strings"
 	"testing"
 )
 
@@ -99,3 +100,43 @@ func TestGoroutineFilter(t *testing.T) {
 		t.Fatalf("generateTrace failed: %v", err)
 	}
 }
+
+func TestPreemptedMarkAssist(t *testing.T) {
+	w := trace.NewWriter()
+	w.Emit(trace.EvBatch, 0, 0)  // start of per-P batch event [pid, timestamp]
+	w.Emit(trace.EvFrequency, 1) // [ticks per second]
+
+	// goroutine 9999: running -> mark assisting -> preempted -> assisting -> running -> block
+	w.Emit(trace.EvGoCreate, 1, 9999, 1, 1) // [timestamp, new goroutine id, new stack id, stack id]
+	w.Emit(trace.EvGoStartLocal, 1, 9999)   // [timestamp, goroutine id]
+	w.Emit(trace.EvGCMarkAssistStart, 1, 2) // [timestamp, stack]
+	w.Emit(trace.EvGoPreempt, 1, 3)         // [timestamp, stack]
+	w.Emit(trace.EvGoStartLocal, 1, 9999)   // [timestamp, goroutine id]
+	w.Emit(trace.EvGCMarkAssistDone, 1)     // [timestamp]
+	w.Emit(trace.EvGoBlock, 1, 4)           // [timestamp, stack]
+
+	events, err := trace.Parse(w, "")
+	if err != nil {
+		t.Fatalf("failed to parse test trace: %v", err)
+	}
+
+	params := &traceParams{
+		events:  events,
+		endTime: int64(1<<63 - 1),
+	}
+
+	viewerData, err := generateTrace(params)
+	if err != nil {
+		t.Fatalf("generateTrace failed: %v", err)
+	}
+
+	marks := 0
+	for _, ev := range viewerData.Events {
+		if strings.Contains(ev.Name, "MARK ASSIST") {
+			marks++
+		}
+	}
+	if marks != 2 {
+		t.Errorf("Got %v MARK ASSIST events, want %v", marks, 2)
+	}
+}

src/internal/trace/writer.go

@@ -9,7 +9,7 @@ type Writer struct {
 
 func NewWriter() *Writer {
 	w := new(Writer)
-	w.Write([]byte("go 1.7 trace\x00\x00\x00\x00"))
+	w.Write([]byte("go 1.9 trace\x00\x00\x00\x00"))
 	return w
 }
 

src/runtime/mgcmark.go

@@ -415,6 +415,10 @@ func gcAssistAlloc(gp *g) {
 		return
 	}
 
+	if trace.enabled {
+		traceGCMarkAssistStart()
+	}
+
 retry:
 	// Compute the amount of scan work we need to do to make the
 	// balance positive. When the required amount of work is low,
@@ -450,6 +454,9 @@ retry:
 		if scanWork == 0 {
 			// We were able to steal all of the credit we
 			// needed.
+			if trace.enabled {
+				traceGCMarkAssistDone()
+			}
 			return
 		}
 	}
@@ -496,6 +503,9 @@ retry:
 		// At this point either background GC has satisfied
 		// this G's assist debt, or the GC cycle is over.
 	}
+	if trace.enabled {
+		traceGCMarkAssistDone()
+	}
 }
 
 // gcAssistAlloc1 is the part of gcAssistAlloc that runs on the system

src/runtime/trace.go

@@ -19,50 +19,52 @@ import (
 
 // Event types in the trace, args are given in square brackets.
 const (
-	traceEvNone           = 0  // unused
-	traceEvBatch          = 1  // start of per-P batch of events [pid, timestamp]
-	traceEvFrequency      = 2  // contains tracer timer frequency [frequency (ticks per second)]
-	traceEvStack          = 3  // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}]
-	traceEvGomaxprocs     = 4  // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id]
-	traceEvProcStart      = 5  // start of P [timestamp, thread id]
-	traceEvProcStop       = 6  // stop of P [timestamp]
-	traceEvGCStart        = 7  // GC start [timestamp, seq, stack id]
-	traceEvGCDone         = 8  // GC done [timestamp]
-	traceEvGCScanStart    = 9  // GC mark termination start [timestamp]
-	traceEvGCScanDone     = 10 // GC mark termination done [timestamp]
-	traceEvGCSweepStart   = 11 // GC sweep start [timestamp, stack id]
-	traceEvGCSweepDone    = 12 // GC sweep done [timestamp]
-	traceEvGoCreate       = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id]
-	traceEvGoStart        = 14 // goroutine starts running [timestamp, goroutine id, seq]
-	traceEvGoEnd          = 15 // goroutine ends [timestamp]
-	traceEvGoStop         = 16 // goroutine stops (like in select{}) [timestamp, stack]
-	traceEvGoSched        = 17 // goroutine calls Gosched [timestamp, stack]
-	traceEvGoPreempt      = 18 // goroutine is preempted [timestamp, stack]
-	traceEvGoSleep        = 19 // goroutine calls Sleep [timestamp, stack]
-	traceEvGoBlock        = 20 // goroutine blocks [timestamp, stack]
-	traceEvGoUnblock      = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack]
-	traceEvGoBlockSend    = 22 // goroutine blocks on chan send [timestamp, stack]
-	traceEvGoBlockRecv    = 23 // goroutine blocks on chan recv [timestamp, stack]
-	traceEvGoBlockSelect  = 24 // goroutine blocks on select [timestamp, stack]
-	traceEvGoBlockSync    = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack]
-	traceEvGoBlockCond    = 26 // goroutine blocks on Cond [timestamp, stack]
-	traceEvGoBlockNet     = 27 // goroutine blocks on network [timestamp, stack]
-	traceEvGoSysCall      = 28 // syscall enter [timestamp, stack]
-	traceEvGoSysExit      = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp]
-	traceEvGoSysBlock     = 30 // syscall blocks [timestamp]
-	traceEvGoWaiting      = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id]
-	traceEvGoInSyscall    = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id]
-	traceEvHeapAlloc      = 33 // memstats.heap_live change [timestamp, heap_alloc]
-	traceEvNextGC         = 34 // memstats.next_gc change [timestamp, next_gc]
-	traceEvTimerGoroutine = 35 // denotes timer goroutine [timer goroutine id]
-	traceEvFutileWakeup   = 36 // denotes that the previous wakeup of this goroutine was futile [timestamp]
-	traceEvString         = 37 // string dictionary entry [ID, length, string]
-	traceEvGoStartLocal   = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id]
-	traceEvGoUnblockLocal = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack]
-	traceEvGoSysExitLocal = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp]
-	traceEvGoStartLabel   = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id]
-	traceEvGoBlockGC      = 42 // goroutine blocks on GC assist [timestamp, stack]
-	traceEvCount          = 43
+	traceEvNone              = 0  // unused
+	traceEvBatch             = 1  // start of per-P batch of events [pid, timestamp]
+	traceEvFrequency         = 2  // contains tracer timer frequency [frequency (ticks per second)]
+	traceEvStack             = 3  // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}]
+	traceEvGomaxprocs        = 4  // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id]
+	traceEvProcStart         = 5  // start of P [timestamp, thread id]
+	traceEvProcStop          = 6  // stop of P [timestamp]
+	traceEvGCStart           = 7  // GC start [timestamp, seq, stack id]
+	traceEvGCDone            = 8  // GC done [timestamp]
+	traceEvGCScanStart       = 9  // GC mark termination start [timestamp]
+	traceEvGCScanDone        = 10 // GC mark termination done [timestamp]
+	traceEvGCSweepStart      = 11 // GC sweep start [timestamp, stack id]
+	traceEvGCSweepDone       = 12 // GC sweep done [timestamp]
+	traceEvGoCreate          = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id]
+	traceEvGoStart           = 14 // goroutine starts running [timestamp, goroutine id, seq]
+	traceEvGoEnd             = 15 // goroutine ends [timestamp]
+	traceEvGoStop            = 16 // goroutine stops (like in select{}) [timestamp, stack]
+	traceEvGoSched           = 17 // goroutine calls Gosched [timestamp, stack]
+	traceEvGoPreempt         = 18 // goroutine is preempted [timestamp, stack]
+	traceEvGoSleep           = 19 // goroutine calls Sleep [timestamp, stack]
+	traceEvGoBlock           = 20 // goroutine blocks [timestamp, stack]
+	traceEvGoUnblock         = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack]
+	traceEvGoBlockSend       = 22 // goroutine blocks on chan send [timestamp, stack]
+	traceEvGoBlockRecv       = 23 // goroutine blocks on chan recv [timestamp, stack]
+	traceEvGoBlockSelect     = 24 // goroutine blocks on select [timestamp, stack]
+	traceEvGoBlockSync       = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack]
+	traceEvGoBlockCond       = 26 // goroutine blocks on Cond [timestamp, stack]
+	traceEvGoBlockNet        = 27 // goroutine blocks on network [timestamp, stack]
+	traceEvGoSysCall         = 28 // syscall enter [timestamp, stack]
+	traceEvGoSysExit         = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp]
+	traceEvGoSysBlock        = 30 // syscall blocks [timestamp]
+	traceEvGoWaiting         = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id]
+	traceEvGoInSyscall       = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id]
+	traceEvHeapAlloc         = 33 // memstats.heap_live change [timestamp, heap_alloc]
+	traceEvNextGC            = 34 // memstats.next_gc change [timestamp, next_gc]
+	traceEvTimerGoroutine    = 35 // denotes timer goroutine [timer goroutine id]
+	traceEvFutileWakeup      = 36 // denotes that the previous wakeup of this goroutine was futile [timestamp]
+	traceEvString            = 37 // string dictionary entry [ID, length, string]
+	traceEvGoStartLocal      = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id]
+	traceEvGoUnblockLocal    = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack]
+	traceEvGoSysExitLocal    = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp]
+	traceEvGoStartLabel      = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id]
+	traceEvGoBlockGC         = 42 // goroutine blocks on GC assist [timestamp, stack]
+	traceEvGCMarkAssistStart = 43 // GC mark assist start [timestamp, stack]
+	traceEvGCMarkAssistDone  = 44 // GC mark assist done [timestamp]
+	traceEvCount             = 45
 )
 
 const (
@@ -380,7 +382,7 @@ func ReadTrace() []byte {
 		trace.headerWritten = true
 		trace.lockOwner = nil
 		unlock(&trace.lock)
-		return []byte("go 1.8 trace\x00\x00\x00\x00")
+		return []byte("go 1.9 trace\x00\x00\x00\x00")
 	}
 	// Wait for new data.
 	if trace.fullHead == 0 && !trace.shutdown {
@@ -939,6 +941,14 @@ func traceGCSweepDone() {
 	traceEvent(traceEvGCSweepDone, -1)
 }
 
+func traceGCMarkAssistStart() {
+	traceEvent(traceEvGCMarkAssistStart, 1)
+}
+
+func traceGCMarkAssistDone() {
+	traceEvent(traceEvGCMarkAssistDone, -1)
+}
+
 func traceGoCreate(newg *g, pc uintptr) {
 	newg.traceseq = 0
 	newg.tracelastp = getg().m.p