A draft of PMU support authored by Pengfei Su

api/except.txt

@@ -459,3 +459,4 @@ pkg syscall (freebsd-arm-cgo), type Statfs_t struct, Mntfromname [88]int8
 pkg syscall (freebsd-arm-cgo), type Statfs_t struct, Mntonname [88]int8
 pkg text/scanner, const GoTokens = 1012
 pkg unicode, const Version = "10.0.0"
+pkg runtime/pprof, func StartCPUProfile(io.Writer) error

src/cmd/go/internal/test/test.go

@@ -313,6 +313,30 @@ profile the tests during execution:
 	    Write a CPU profile to the specified file before exiting.
 	    Writes test binary as -c would.
 
+	-pmuprofile pmu.out
+	    Write a PMU profile to the specified file before exiting.
+	    Writes test binary as -c would.
+
+	-pmuevent event
+	    Specify which pmu event to sample.
+	    To sample PMU cycles, use -test.pmuevent=cycles.
+	    To sample PMU instructions, use -test.pmuevent=instructions.
+	    To sample PMU cache misses, use -test.pmuevent=cacheMisses.
+	    To sample PMU last-level cache read accesses, use -test.pmuevent=cacheLLReadAccesses.
+	    To sample PMU last-level cache read misses, use -test.pmuevent=cacheLLReadMisses.
+
+	-pmuperiod n
+	    Set the sampling period for the pmu event specified by -pmuevent.
+
+	-pmupreciseip n
+	    Set the precise ip level for the pmu event specified by -pmuevent.
+
+	-pmukernelincl
+	    Count the kernel or not.
+
+	-pmuhvincl
+	    Count the hypervisor or not.
+
 	-memprofile mem.out
 	    Write an allocation profile to the file after all tests have passed.
 	    Writes test binary as -c would.

src/cmd/go/internal/test/testflag.go

@@ -46,6 +46,12 @@ var testFlagDefn = []*cmdflag.Defn{
 	{Name: "coverprofile", PassToTest: true},
 	{Name: "cpu", PassToTest: true},
 	{Name: "cpuprofile", PassToTest: true},
+	{Name: "pmuprofile", PassToTest: true},
+	{Name: "pmuevent", PassToTest: true},
+	{Name: "pmuperiod", PassToTest: true},
+	{Name: "pmupreciseip", PassToTest: true},
+	{Name: "pmukernelincl", BoolVar: new(bool), PassToTest: true},
+	{Name: "pmuhvincl", BoolVar: new(bool), PassToTest: true},
 	{Name: "failfast", BoolVar: new(bool), PassToTest: true},
 	{Name: "list", PassToTest: true},
 	{Name: "memprofile", PassToTest: true},
@@ -157,7 +163,7 @@ func testFlags(usage func(), args []string) (packageNames, passToTest []string)
 				testList = true
 			case "timeout":
 				testTimeout = value
-			case "blockprofile", "cpuprofile", "memprofile", "mutexprofile":
+			case "blockprofile", "cpuprofile", "pmuprofile", "memprofile", "mutexprofile":
 				testProfile = "-" + f.Name
 				testNeedBinary = true
 			case "trace":

src/net/http/pprof/pprof.go

@@ -110,6 +110,52 @@ func serveError(w http.ResponseWriter, status int, txt string) {
 	fmt.Fprintln(w, txt)
 }
 
+func pmuProfile(w http.ResponseWriter, r *http.Request) error {
+	var eventConfig pprof.PMUEventConfig
+	eventConfig.Period = 1e7 // default value
+
+	if period, err := strconv.ParseInt(r.FormValue("pmuperiod"), 10, 64); err == nil {
+		eventConfig.Period = period
+	}
+	if preciseIP, err := strconv.ParseInt(r.FormValue("pmupreciseip"), 10, 8); err == nil {
+		eventConfig.PreciseIP = int8(preciseIP)
+	}
+	if isKernelIncluded, err := strconv.ParseBool(r.FormValue("pmukernelincl")); err == nil {
+		eventConfig.IsKernelIncluded = isKernelIncluded
+	}
+	if isHvIncluded, err := strconv.ParseBool(r.FormValue("pmuhvincl")); err == nil {
+		eventConfig.IsHvIncluded = isHvIncluded
+	}
+
+	switch eventName := r.FormValue("pmuevent"); eventName {
+	case "":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCycles(w, &eventConfig))
+	case "cycles":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCycles(w, &eventConfig))
+	case "instructions":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUInstructions(w, &eventConfig))
+	case "cacheReferences":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCacheReferences(w, &eventConfig))
+	case "cacheMisses":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCacheMisses(w, &eventConfig))
+	case "cacheLLReadAccesses":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCacheLLReadAccesses(w, &eventConfig))
+	case "cacheLLReadMisses":
+		return pprof.StartPMUProfile(pprof.WithProfilingPMUCacheLLReadMisses(w, &eventConfig))
+	default:
+		// Is this a raw event?
+		if strings.HasPrefix(eventName, "r") {
+			if rawHexEvent, err := strconv.ParseInt(eventName[1:], 16, 64); err == nil {
+				eventConfig.RawEvent = rawHexEvent
+				return pprof.StartPMUProfile(pprof.WithProfilingPMURaw(w, &eventConfig))
+			}
+			return fmt.Errorf("Incorrect hex format for raw event")
+		} else {
+			return fmt.Errorf("Unknown or not yet implemented event")
+		}
+	}
+}
+
 // Profile responds with the pprof-formatted cpu profile.
 // Profiling lasts for duration specified in seconds GET parameter, or for 30 seconds if not specified.
 // The package initialization registers it as /debug/pprof/profile.
@@ -129,14 +175,31 @@ func Profile(w http.ResponseWriter, r *http.Request) {
 	// because if it does it starts writing.
 	w.Header().Set("Content-Type", "application/octet-stream")
 	w.Header().Set("Content-Disposition", `attachment; filename="profile"`)
-	if err := pprof.StartCPUProfile(w); err != nil {
-		// StartCPUProfile failed, so no writes yet.
-		serveError(w, http.StatusInternalServerError,
-			fmt.Sprintf("Could not enable CPU profiling: %s", err))
-		return
+
+	isPMUEnabled, err := strconv.ParseBool(r.FormValue("pmu"))
+
+	if (err != nil) || (!isPMUEnabled) {
+		if err = pprof.StartCPUProfile(w); err != nil {
+			serveError(w, http.StatusInternalServerError,
+				fmt.Sprintf("Could not enable CPU profiling: %s", err))
+			return
+		}
+	} else {
+		// err == nil and isPMUEnabled == true
+		if err = pmuProfile(w, r); err != nil {
+			serveError(w, http.StatusInternalServerError,
+				fmt.Sprintf("Could not enable PMU profiling: %s", err))
+			return
+		}
 	}
+
 	sleep(w, time.Duration(sec)*time.Second)
-	pprof.StopCPUProfile()
+
+	if isPMUEnabled {
+		pprof.StopPMUProfile()
+	} else {
+		pprof.StopCPUProfile()
+	}
 }
 
 // Trace responds with the execution trace in binary form.

src/runtime/asm_386.s

@@ -1566,6 +1566,18 @@ TEXT runtime·panicExtendSlice3CU(SB),NOSPLIT,$0-12
 	MOVL	CX, y+8(FP)
 	JMP	runtime·goPanicExtendSlice3CU(SB)
 
+TEXT runtime·mb(SB),NOSPLIT,$0-0
+	MFENCE
+	RET
+
+TEXT runtime·rmb(SB),NOSPLIT,$0-0
+	LFENCE
+	RET
+
+TEXT runtime·wmb(SB),NOSPLIT,$0-0
+	SFENCE
+	RET
+
 #ifdef GOOS_android
 // Use the free TLS_SLOT_APP slot #2 on Android Q.
 // Earlier androids are set up in gcc_android.c.

src/runtime/asm_amd64.s

@@ -1721,6 +1721,18 @@ TEXT runtime·panicSlice3CU(SB),NOSPLIT,$0-16
 	MOVQ	CX, y+8(FP)
 	JMP	runtime·goPanicSlice3CU(SB)
 
+TEXT runtime·mb(SB),NOSPLIT,$0-0
+	MFENCE
+	RET
+
+TEXT runtime·rmb(SB),NOSPLIT,$0-0
+	LFENCE
+	RET
+
+TEXT runtime·wmb(SB),NOSPLIT,$0-0
+	SFENCE
+	RET
+
 #ifdef GOOS_android
 // Use the free TLS_SLOT_APP slot #2 on Android Q.
 // Earlier androids are set up in gcc_android.c.