runtime: speed up StartTrace with lots of blocked goroutines

In StartTrace we emit EvGoCreate for all existing goroutines.
This includes stack unwind to obtain current stack.
Real Go programs can contain hundreds of thousands of blocked goroutines.
For such programs StartTrace can take up to a second (few ms per goroutine).

Obtain current stack ID once and use it for all EvGoCreate events.

This speeds up StartTrace with 10K blocked goroutines from 20ms to 4 ms
(win for StartTrace called from net/http/pprof hander will be bigger
as stack is deeper).

Change-Id: I9e5ff9468331a840f8fdcdd56c5018c2cfde61fc
Reviewed-on: https://go-review.googlesource.com/25573
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Hyang-Ah Hana Kim <hyangah@gmail.com>

src/runtime/trace.go

@@ -184,10 +184,21 @@ func StartTrace() error {
 	// trace.enabled is set afterwards once we have emitted all preliminary events.
 	_g_ := getg()
 	_g_.m.startingtrace = true
+
+	// Obtain current stack ID to use in all traceEvGoCreate events below.
+	mp := acquirem()
+	stkBuf := make([]uintptr, traceStackSize)
+	stackID := traceStackID(mp, stkBuf, 2)
+	releasem(mp)
+
 	for _, gp := range allgs {
 		status := readgstatus(gp)
 		if status != _Gdead {
-			traceGoCreate(gp, gp.startpc) // also resets gp.traceseq/tracelastp
+			gp.traceseq = 0
+			gp.tracelastp = getg().m.p
+			// +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum.
+			id := trace.stackTab.put([]uintptr{gp.startpc + sys.PCQuantum})
+			traceEvent(traceEvGoCreate, -1, uint64(gp.goid), uint64(id), stackID)
 		}
 		if status == _Gwaiting {
 			// traceEvGoWaiting is implied to have seq=1.
@@ -513,28 +524,7 @@ func traceEvent(ev byte, skip int, args ...uint64) {
 	if skip == 0 {
 		buf.varint(0)
 	} else if skip > 0 {
-		_g_ := getg()
-		gp := mp.curg
-		var nstk int
-		if gp == _g_ {
-			nstk = callers(skip, buf.stk[:])
-		} else if gp != nil {
-			gp = mp.curg
-			// This may happen when tracing a system call,
-			// so we must lock the stack.
-			if gcTryLockStackBarriers(gp) {
-				nstk = gcallers(gp, skip, buf.stk[:])
-				gcUnlockStackBarriers(gp)
-			}
-		}
-		if nstk > 0 {
-			nstk-- // skip runtime.goexit
-		}
-		if nstk > 0 && gp.goid == 1 {
-			nstk-- // skip runtime.main
-		}
-		id := trace.stackTab.put(buf.stk[:nstk])
-		buf.varint(uint64(id))
+		buf.varint(traceStackID(mp, buf.stk[:], skip))
 	}
 	evSize := buf.pos - startPos
 	if evSize > maxSize {
@@ -547,6 +537,31 @@ func traceEvent(ev byte, skip int, args ...uint64) {
 	traceReleaseBuffer(pid)
 }
 
+func traceStackID(mp *m, buf []uintptr, skip int) uint64 {
+	_g_ := getg()
+	gp := mp.curg
+	var nstk int
+	if gp == _g_ {
+		nstk = callers(skip+1, buf[:])
+	} else if gp != nil {
+		gp = mp.curg
+		// This may happen when tracing a system call,
+		// so we must lock the stack.
+		if gcTryLockStackBarriers(gp) {
+			nstk = gcallers(gp, skip, buf[:])
+			gcUnlockStackBarriers(gp)
+		}
+	}
+	if nstk > 0 {
+		nstk-- // skip runtime.goexit
+	}
+	if nstk > 0 && gp.goid == 1 {
+		nstk-- // skip runtime.main
+	}
+	id := trace.stackTab.put(buf[:nstk])
+	return uint64(id)
+}
+
 // traceAcquireBuffer returns trace buffer to use and, if necessary, locks it.
 func traceAcquireBuffer() (mp *m, pid int32, bufp *traceBufPtr) {
 	mp = acquirem()