runtime: cpu profiling support

R=r
CC=golang-dev
https://golang.org/cl/4306043

src/pkg/runtime/Makefile

@@ -53,6 +53,7 @@ OFILES=\
 	cgocall.$O\
 	chan.$O\
 	closure.$O\
+	cpuprof.$O\
 	float.$O\
 	complex.$O\
 	hashmap.$O\

src/pkg/runtime/amd64/traceback.c

@@ -18,8 +18,8 @@ void runtime·morestack(void);
 // as well as the runtime.Callers function (pcbuf != nil).
 // A little clunky to merge the two but avoids duplicating
 // the code and all its subtlety.
-static int32
-gentraceback(byte *pc0, byte *sp, G *g, int32 skip, uintptr *pcbuf, int32 max)
+int32
+runtime·gentraceback(byte *pc0, byte *sp, byte *lr0, G *g, int32 skip, uintptr *pcbuf, int32 max)
 {
 	byte *p;
 	int32 i, n, iter, sawnewstack;
@@ -28,6 +28,7 @@ gentraceback(byte *pc0, byte *sp, G *g, int32 skip, uintptr *pcbuf, int32 max)
 	Stktop *stk;
 	Func *f;
 
+	USED(lr0);
 	pc = (uintptr)pc0;
 	lr = 0;
 	fp = nil;
@@ -199,7 +200,7 @@ gentraceback(byte *pc0, byte *sp, G *g, int32 skip, uintptr *pcbuf, int32 max)
 void
 runtime·traceback(byte *pc0, byte *sp, byte*, G *g)
 {
-	gentraceback(pc0, sp, g, 0, nil, 100);
+	runtime·gentraceback(pc0, sp, nil, g, 0, nil, 100);
 }
 
 int32
@@ -211,7 +212,7 @@ runtime·callers(int32 skip, uintptr *pcbuf, int32 m)
 	sp = (byte*)&skip;
 	pc = runtime·getcallerpc(&skip);
 
-	return gentraceback(pc, sp, g, skip, pcbuf, m);
+	return runtime·gentraceback(pc, sp, nil, g, skip, pcbuf, m);
 }
 
 static uintptr

src/pkg/runtime/arm/traceback.c

@@ -15,7 +15,7 @@ void _divu(void);
 void _modu(void);
 
 static int32
-gentraceback(byte *pc0, byte *sp, byte *lr0, G *g, int32 skip, uintptr *pcbuf, int32 max)
+runtime·gentraceback(byte *pc0, byte *sp, byte *lr0, G *g, int32 skip, uintptr *pcbuf, int32 max)
 {
 	int32 i, n, iter;
 	uintptr pc, lr, tracepc, x;
@@ -189,11 +189,10 @@ gentraceback(byte *pc0, byte *sp, byte *lr0, G *g, int32 skip, uintptr *pcbuf, i
 	return n;		
 }
 
-
 void
 runtime·traceback(byte *pc0, byte *sp, byte *lr, G *g)
 {
-	gentraceback(pc0, sp, lr, g, 0, nil, 100);
+	runtime·gentraceback(pc0, sp, lr, g, 0, nil, 100);
 }
 
 // func caller(n int) (pc uintptr, file string, line int, ok bool)
@@ -205,5 +204,5 @@ runtime·callers(int32 skip, uintptr *pcbuf, int32 m)
 	sp = runtime·getcallersp(&skip);
 	pc = runtime·getcallerpc(&skip);
 
-	return gentraceback(pc, sp, 0, g, skip, pcbuf, m);
+	return runtime·gentraceback(pc, sp, 0, g, skip, pcbuf, m);
 }

src/pkg/runtime/cpuprof.c

+// Copyright 2011 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// CPU profiling.
+// Based on algorithms and data structures used in
+// http://code.google.com/p/google-perftools/.
+//
+// The main difference between this code and the google-perftools
+// code is that this code is written to allow copying the profile data
+// to an arbitrary io.Writer, while the google-perftools code always
+// writes to an operating system file.
+//
+// The signal handler for the profiling clock tick adds a new stack trace
+// to a hash table tracking counts for recent traces.  Most clock ticks
+// hit in the cache.  In the event of a cache miss, an entry must be 
+// evicted from the hash table, copied to a log that will eventually be
+// written as profile data.  The google-perftools code flushed the
+// log itself during the signal handler.  This code cannot do that, because
+// the io.Writer might block or need system calls or locks that are not
+// safe to use from within the signal handler.  Instead, we split the log
+// into two halves and let the signal handler fill one half while a goroutine
+// is writing out the other half.  When the signal handler fills its half, it
+// offers to swap with the goroutine.  If the writer is not done with its half,
+// we lose the stack trace for this clock tick (and record that loss).
+// The goroutine interacts with the signal handler by calling getprofile() to
+// get the next log piece to write, implicitly handing back the last log
+// piece it obtained.
+//
+// The state of this dance between the signal handler and the goroutine
+// is encoded in the Profile.handoff field.  If handoff == 0, then the goroutine
+// is not using either log half and is waiting (or will soon be waiting) for
+// a new piece by calling notesleep(&p->wait).  If the signal handler
+// changes handoff from 0 to non-zero, it must call notewakeup(&p->wait)
+// to wake the goroutine.  The value indicates the number of entries in the
+// log half being handed off.  The goroutine leaves the non-zero value in
+// place until it has finished processing the log half and then flips the number
+// back to zero.  Setting the high bit in handoff means that the profiling is over, 
+// and the goroutine is now in charge of flushing the data left in the hash table
+// to the log and returning that data.  
+//
+// The handoff field is manipulated using atomic operations.
+// For the most part, the manipulation of handoff is orderly: if handoff == 0
+// then the signal handler owns it and can change it to non-zero.  
+// If handoff != 0 then the goroutine owns it and can change it to zero.
+// If that were the end of the story then we would not need to manipulate
+// handoff using atomic operations.  The operations are needed, however,
+// in order to let the log closer set the high bit to indicate "EOF" safely
+// in the situation when normally the goroutine "owns" handoff.
+
+#include "runtime.h"
+#include "malloc.h"
+
+enum
+{
+	HashSize = 1<<10,
+	LogSize = 1<<17,
+	Assoc = 4,
+	MaxStack = 64,
+};
+
+typedef struct Profile Profile;
+typedef struct Bucket Bucket;
+typedef struct Entry Entry;
+
+struct Entry {
+	uintptr count;
+	uintptr depth;
+	uintptr stack[MaxStack];
+};
+
+struct Bucket {
+	Entry entry[Assoc];
+};
+
+struct Profile {
+	bool on;		// profiling is on
+	Note wait;		// goroutine waits here
+	uintptr count;		// tick count
+	uintptr evicts;		// eviction count
+	uintptr lost;		// lost ticks that need to be logged
+	uintptr totallost;	// total lost ticks
+
+	// Active recent stack traces.
+	Bucket hash[HashSize];
+
+	// Log of traces evicted from hash.
+	// Signal handler has filled log[toggle][:nlog].
+	// Goroutine is writing log[1-toggle][:handoff].
+	uintptr log[2][LogSize/2];
+	uintptr nlog;
+	int32 toggle;
+	uint32 handoff;
+	
+	// Writer state.
+	// Writer maintains its own toggle to avoid races
+	// looking at signal handler's toggle.
+	uint32 wtoggle;
+	bool wholding;	// holding & need to release a log half
+	bool flushing;	// flushing hash table - profile is over
+};
+
+static Lock lk;
+static Profile *prof;
+
+static void tick(uintptr*, int32);
+static void add(Profile*, uintptr*, int32);
+static bool evict(Profile*, Entry*);
+static bool flushlog(Profile*);
+
+// LostProfileData is a no-op function used in profiles
+// to mark the number of profiling stack traces that were
+// discarded due to slow data writers.
+static void LostProfileData(void) {
+}
+
+// SetCPUProfileRate sets the CPU profiling rate.
+// The user documentation is in debug.go.
+void
+runtime·SetCPUProfileRate(int32 hz)
+{
+	uintptr *p;
+	uintptr n;
+
+	// Clamp hz to something reasonable.
+	if(hz < 0)
+		hz = 0;
+	if(hz > 1000000)
+		hz = 1000000;
+
+	runtime·lock(&lk);
+	if(hz > 0) {
+		if(prof == nil) {
+			prof = runtime·SysAlloc(sizeof *prof);
+			if(prof == nil) {
+				runtime·printf("runtime: cpu profiling cannot allocate memory\n");
+				runtime·unlock(&lk);
+				return;
+			}
+		}
+		if(prof->on || prof->handoff != 0) {
+			runtime·printf("runtime: cannot set cpu profile rate until previous profile has finished.\n");
+			runtime·unlock(&lk);
+			return;
+		}
+
+		prof->on = true;
+		p = prof->log[0];
+		// pprof binary header format.
+		// http://code.google.com/p/google-perftools/source/browse/trunk/src/profiledata.cc#117
+		*p++ = 0;  // count for header
+		*p++ = 3;  // depth for header
+		*p++ = 0;  // version number
+		*p++ = 1000000 / hz;  // period (microseconds)
+		*p++ = 0;
+		prof->nlog = p - prof->log[0];
+		prof->toggle = 0;
+		prof->wholding = false;
+		prof->wtoggle = 0;
+		prof->flushing = false;
+		runtime·noteclear(&prof->wait);
+
+		runtime·setcpuprofilerate(tick, hz);
+	} else if(prof->on) {
+		runtime·setcpuprofilerate(nil, 0);
+		prof->on = false;
+
+		// Now add is not running anymore, and getprofile owns the entire log.
+		// Set the high bit in prof->handoff to tell getprofile.
+		for(;;) {
+			n = prof->handoff;
+			if(n&0x80000000)
+				runtime·printf("runtime: setcpuprofile(off) twice");
+			if(runtime·cas(&prof->handoff, n, n|0x80000000))
+				break;
+		}
+		if(n == 0) {
+			// we did the transition from 0 -> nonzero so we wake getprofile
+			runtime·notewakeup(&prof->wait);
+		}
+	}
+	runtime·unlock(&lk);
+}
+
+static void
+tick(uintptr *pc, int32 n)
+{
+	add(prof, pc, n);
+}
+
+// add adds the stack trace to the profile.
+// It is called from signal handlers and other limited environments
+// and cannot allocate memory or acquire locks that might be
+// held at the time of the signal, nor can it use substantial amounts
+// of stack.  It is allowed to call evict.
+static void
+add(Profile *p, uintptr *pc, int32 n)
+{
+	int32 i, j;
+	uintptr h, x;
+	Bucket *b;
+	Entry *e;
+
+	if(n > MaxStack)
+		n = MaxStack;
+	
+	// Compute hash.
+	h = 0;
+	for(i=0; i<n; i++) {
+		h = h<<8 | (h>>(8*(sizeof(h)-1)));
+		x = pc[i];
+		h += x*31 + x*7 + x*3;
+	}
+	p->count++;
+
+	// Add to entry count if already present in table.
+	b = &p->hash[h%HashSize];
+	for(i=0; i<Assoc; i++) {
+		e = &b->entry[i];
+		if(e->depth != n)	
+			continue;
+		for(j=0; j<n; j++)
+			if(e->stack[j] != pc[j])
+				goto ContinueAssoc;
+		e->count++;
+		return;
+	ContinueAssoc:;
+	}
+
+	// Evict entry with smallest count.
+	e = &b->entry[0];
+	for(i=1; i<Assoc; i++)
+		if(b->entry[i].count < e->count)
+			e = &b->entry[i];
+	if(e->count > 0) {
+		if(!evict(p, e)) {
+			// Could not evict entry.  Record lost stack.
+			p->lost++;
+			p->totallost++;
+			return;
+		}
+		p->evicts++;
+	}
+	
+	// Reuse the newly evicted entry.
+	e->depth = n;
+	e->count = 1;
+	for(i=0; i<n; i++)
+		e->stack[i] = pc[i];
+}
+
+// evict copies the given entry's data into the log, so that
+// the entry can be reused.  evict is called from add, which
+// is called from the profiling signal handler, so it must not
+// allocate memory or block.  It is safe to call flushLog.
+// evict returns true if the entry was copied to the log,
+// false if there was no room available.
+static bool
+evict(Profile *p, Entry *e)
+{
+	int32 i, d, nslot;
+	uintptr *log, *q;
+	
+	d = e->depth;
+	nslot = d+2;
+	log = p->log[p->toggle];
+	if(p->nlog+nslot > nelem(p->log[0])) {
+		if(!flushlog(p))
+			return false;
+		log = p->log[p->toggle];
+	}
+	
+	q = log+p->nlog;
+	*q++ = e->count;
+	*q++ = d;
+	for(i=0; i<d; i++)
+		*q++ = e->stack[i];
+	p->nlog = q - log;
+	e->count = 0;
+	return true;
+}
+
+// flushlog tries to flush the current log and switch to the other one.
+// flushlog is called from evict, called from add, called from the signal handler,
+// so it cannot allocate memory or block.  It can try to swap logs with
+// the writing goroutine, as explained in the comment at the top of this file.
+static bool
+flushlog(Profile *p)
+{
+	uintptr *log, *q;
+
+	if(!runtime·cas(&p->handoff, 0, p->nlog))
+		return false;
+	runtime·notewakeup(&p->wait);
+
+	p->toggle = 1 - p->toggle;
+	log = p->log[p->toggle];
+	q = log;
+	if(p->lost > 0) {
+		*q++ = p->lost;
+		*q++ = 1;
+		*q++ = (uintptr)LostProfileData;
+	}
+	p->nlog = q - log;
+	return true;
+}
+
+// getprofile blocks until the next block of profiling data is available
+// and returns it as a []byte.  It is called from the writing goroutine.
+Slice
+getprofile(Profile *p)
+{
+	uint32 i, j, n;
+	Slice ret;
+	Bucket *b;
+	Entry *e;
+
+	ret.array = nil;
+	ret.len = 0;
+	ret.cap = 0;
+	
+	if(p == nil)	
+		return ret;
+
+	if(p->wholding) {
+		// Release previous log to signal handling side.
+		// Loop because we are racing against setprofile(off).
+		for(;;) {
+			n = p->handoff;
+			if(n == 0) {
+				runtime·printf("runtime: phase error during cpu profile handoff\n");
+				return ret;
+			}
+			if(n & 0x80000000) {
+				p->wtoggle = 1 - p->wtoggle;
+				p->wholding = false;
+				p->flushing = true;
+				goto flush;
+			}
+			if(runtime·cas(&p->handoff, n, 0))
+				break;
+		}
+		p->wtoggle = 1 - p->wtoggle;
+		p->wholding = false;
+	}
+	
+	if(p->flushing)
+		goto flush;
+	
+	if(!p->on && p->handoff == 0)
+		return ret;
+
+	// Wait for new log.
+	runtime·entersyscall();
+	runtime·notesleep(&p->wait);
+	runtime·exitsyscall();
+	runtime·noteclear(&p->wait);
+
+	n = p->handoff;
+	if(n == 0) {
+		runtime·printf("runtime: phase error during cpu profile wait\n");
+		return ret;
+	}
+	if(n == 0x80000000) {
+		p->flushing = true;
+		goto flush;
+	}
+	n &= ~0x80000000;
+
+	// Return new log to caller.
+	p->wholding = true;
+
+	ret.array = (byte*)p->log[p->wtoggle];
+	ret.len = n*sizeof(uintptr);
+	ret.cap = ret.len;
+	return ret;
+
+flush:
+	// In flush mode.
+	// Add is no longer being called.  We own the log.
+	// Also, p->handoff is non-zero, so flushlog will return false.
+	// Evict the hash table into the log and return it.
+	for(i=0; i<HashSize; i++) {
+		b = &p->hash[i];
+		for(j=0; j<Assoc; j++) {
+			e = &b->entry[j];
+			if(e->count > 0 && !evict(p, e)) {
+				// Filled the log.  Stop the loop and return what we've got.
+				goto breakflush;
+			}
+		}
+	}
+breakflush:
+
+	// Return pending log data.
+	if(p->nlog > 0) {
+		// Note that we're using toggle now, not wtoggle,
+		// because we're working on the log directly.
+		ret.array = (byte*)p->log[p->toggle];
+		ret.len = p->nlog*sizeof(uintptr);
+		ret.cap = ret.len;
+		p->nlog = 0;
+		return ret;
+	}
+
+	// Made it through the table without finding anything to log.
+	// Finally done.  Clean up and return nil.
+	p->flushing = false;
+	if(!runtime·cas(&p->handoff, p->handoff, 0))
+		runtime·printf("runtime: profile flush racing with something\n");
+	return ret;  // set to nil at top of function
+}
+
+// CPUProfile returns the next cpu profile block as a []byte.
+// The user documentation is in debug.go.
+void
+runtime·CPUProfile(Slice ret)
+{
+	ret = getprofile(prof);
+	FLUSH(&ret);
+}

src/pkg/runtime/darwin/386/signal.c

@@ -46,6 +46,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	mc = uc->uc_mcontext;
 	r = &mc->ss;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->eip, (uint8*)r->esp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Work around Leopard bug that doesn't set FPE_INTDIV.
 		// Look at instruction to see if it is a divide.
@@ -126,31 +131,58 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_SIGINFO|SA_ONSTACK;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0U;
+	sa.sa_tramp = (uintptr)runtime·sigtramp;	// runtime·sigtramp's job is to call into real handler
+	sa.__sigaction_u.__sa_sigaction = (uintptr)fn;
+	runtime·sigaction(i, &sa, nil);
+}
+
 void
 runtime·initsig(int32 queue)
 {
 	int32 i;
-	static Sigaction sa;
+	void *fn;
 
 	runtime·siginit();
 
-	sa.sa_flags |= SA_SIGINFO|SA_ONSTACK;
-	sa.sa_mask = 0xFFFFFFFFU;
-	sa.sa_tramp = runtime·sigtramp;	// runtime·sigtramp's job is to call into real handler
 	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
-			if(runtime·sigtab[i].flags & (SigCatch | SigQueue)) {
-				sa.__sigaction_u.__sa_sigaction = runtime·sighandler;
-			} else {
-				sa.__sigaction_u.__sa_sigaction = runtime·sigignore;
-			}
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
+			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
+				fn = runtime·sighandler;
 			else
-				sa.sa_flags &= ~SA_RESTART;
-			runtime·sigaction(i, &sa, nil);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Sigaction sa;
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/darwin/amd64/signal.c

@@ -54,6 +54,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	mc = uc->uc_mcontext;
 	r = &mc->ss;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->rip, (uint8*)r->rsp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Work around Leopard bug that doesn't set FPE_INTDIV.
 		// Look at instruction to see if it is a divide.
@@ -136,31 +141,58 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_SIGINFO|SA_ONSTACK;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.sa_tramp = (uintptr)runtime·sigtramp;	// runtime·sigtramp's job is to call into real handler
+	sa.__sigaction_u.__sa_sigaction = (uintptr)fn;
+	runtime·sigaction(i, &sa, nil);
+}
+
 void
 runtime·initsig(int32 queue)
 {
 	int32 i;
-	static Sigaction sa;
+	void *fn;
 
 	runtime·siginit();
 
-	sa.sa_flags |= SA_SIGINFO|SA_ONSTACK;
-	sa.sa_mask = 0xFFFFFFFFU;
-	sa.sa_tramp = runtime·sigtramp;	// runtime·sigtramp's job is to call into real handler
 	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
-			if(runtime·sigtab[i].flags & (SigCatch | SigQueue)) {
-				sa.__sigaction_u.__sa_sigaction = runtime·sighandler;
-			} else {
-				sa.__sigaction_u.__sa_sigaction = runtime·sigignore;
-			}
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
+			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
+				fn = runtime·sighandler;
 			else
-				sa.sa_flags &= ~SA_RESTART;
-			runtime·sigaction(i, &sa, nil);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Sigaction sa;
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/debug.go

@@ -92,4 +92,24 @@ func (r *MemProfileRecord) Stack() []uintptr {
 // where r.AllocBytes > 0 but r.AllocBytes == r.FreeBytes.
 // These are sites where memory was allocated, but it has all
 // been released back to the runtime.
+// Most clients should use the runtime/pprof package or
+// the testing package's -test.memprofile flag instead
+// of calling MemProfile directly.
 func MemProfile(p []MemProfileRecord, inuseZero bool) (n int, ok bool)
+
+// CPUProfile returns the next chunk of binary CPU profiling stack trace data,
+// blocking until data is available.  If profiling is turned off and all the profile
+// data accumulated while it was on has been returned, CPUProfile returns nil.
+// The caller must save the returned data before calling CPUProfile again.
+// Most clients should use the runtime/pprof package or
+// the testing package's -test.cpuprofile flag instead of calling
+// CPUProfile directly.
+func CPUProfile() []byte
+
+// SetCPUProfileRate sets the CPU profiling rate to hz samples per second.
+// If hz <= 0, SetCPUProfileRate turns off profiling.
+// If the profiler is on, the rate cannot be changed without first turning it off.
+// Most clients should use the runtime/pprof package or
+// the testing package's -test.cpuprofile flag instead of calling
+// SetCPUProfileRate directly.
+func SetCPUProfileRate(hz int)

src/pkg/runtime/freebsd/386/signal.c

@@ -54,6 +54,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	uc = context;
 	r = &uc->uc_mcontext;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->mc_eip, (uint8*)r->mc_esp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Make it look like a call to the signal func.
 		// Have to pass arguments out of band since
@@ -122,32 +127,57 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_SIGINFO|SA_ONSTACK;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.__sigaction_u.__sa_sigaction = (uintptr)fn;
+	runtime·sigaction(i, &sa, nil);
+}
+
 void
 runtime·initsig(int32 queue)
 {
-	static Sigaction sa;
+	int32 i;
+	void *fn;
 
 	runtime·siginit();
 
-	int32 i;
-	sa.sa_flags |= SA_ONSTACK | SA_SIGINFO;
-	sa.sa_mask = ~0x0ull;
-	
-	for(i = 0; i < NSIG; i++) {
+	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
 			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
-				sa.__sigaction_u.__sa_sigaction = (void*) runtime·sigtramp;
+				fn = runtime·sighandler;
 			else
-				sa.__sigaction_u.__sa_sigaction = (void*) runtime·sigignore;
-
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
-			else
-				sa.sa_flags &= ~SA_RESTART;
-
-			runtime·sigaction(i, &sa, nil);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Sigaction sa;
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/freebsd/amd64/signal.c

@@ -62,6 +62,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	uc = context;
 	r = &uc->uc_mcontext;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->mc_rip, (uint8*)r->mc_rsp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Make it look like a call to the signal func.
 		// Have to pass arguments out of band since
@@ -130,32 +135,57 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_SIGINFO|SA_ONSTACK;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.__sigaction_u.__sa_sigaction = (uintptr)fn;
+	runtime·sigaction(i, &sa, nil);
+}
+
 void
 runtime·initsig(int32 queue)
 {
-	static Sigaction sa;
+	int32 i;
+	void *fn;
 
 	runtime·siginit();
 
-	int32 i;
-	sa.sa_flags |= SA_ONSTACK | SA_SIGINFO;
-	sa.sa_mask = ~0x0ull;
-	
-	for(i = 0; i < NSIG; i++) {
+	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
 			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
-				sa.__sigaction_u.__sa_sigaction = (void*) runtime·sigtramp;
+				fn = runtime·sighandler;
 			else
-				sa.__sigaction_u.__sa_sigaction = (void*) runtime·sigignore;
-
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
-			else
-				sa.sa_flags &= ~SA_RESTART;
-
-			runtime·sigaction(i, &sa, nil);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Sigaction sa;
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/linux/386/signal.c

@@ -51,6 +51,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	uc = context;
 	r = &uc->uc_mcontext;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->eip, (uint8*)r->esp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Make it look like a call to the signal func.
 		// Have to pass arguments out of band since
@@ -114,30 +119,59 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.sa_restorer = (void*)runtime·sigreturn;
+	if(fn == runtime·sighandler)
+		fn = (void*)runtime·sigtramp;
+	sa.k_sa_handler = fn;
+	runtime·rt_sigaction(i, &sa, nil, 8);
+}
+
 void
 runtime·initsig(int32 queue)
 {
-	static Sigaction sa;
+	int32 i;
+	void *fn;
 
 	runtime·siginit();
 
-	int32 i;
-	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
-	sa.sa_mask = 0xFFFFFFFFFFFFFFFFULL;
-	sa.sa_restorer = (void*)runtime·sigreturn;
 	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
 			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
-				sa.k_sa_handler = (void*)runtime·sigtramp;
+				fn = runtime·sighandler;
 			else
-				sa.k_sa_handler = (void*)runtime·sigignore;
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
-			else
-				sa.sa_flags &= ~SA_RESTART;
-			runtime·rt_sigaction(i, &sa, nil, 8);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/linux/amd64/signal.c

@@ -61,6 +61,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	mc = &uc->uc_mcontext;
 	r = (Sigcontext*)mc;	// same layout, more conveient names
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->rip, (uint8*)r->rsp, nil, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Make it look like a call to the signal func.
 		// Have to pass arguments out of band since
@@ -124,30 +129,59 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.sa_restorer = (void*)runtime·sigreturn;
+	if(fn == runtime·sighandler)
+		fn = (void*)runtime·sigtramp;
+	sa.sa_handler = fn;
+	runtime·rt_sigaction(i, &sa, nil, 8);
+}
+
 void
 runtime·initsig(int32 queue)
 {
-	static Sigaction sa;
+	int32 i;
+	void *fn;
 
 	runtime·siginit();
 
-	int32 i;
-	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
-	sa.sa_mask = 0xFFFFFFFFFFFFFFFFULL;
-	sa.sa_restorer = (void*)runtime·sigreturn;
 	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
 			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
-				sa.sa_handler = (void*)runtime·sigtramp;
+				fn = runtime·sighandler;
 			else
-				sa.sa_handler = (void*)runtime·sigignore;
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
-			else
-				sa.sa_flags &= ~SA_RESTART;
-			runtime·rt_sigaction(i, &sa, nil, 8);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/linux/arm/signal.c

@@ -58,6 +58,11 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 	uc = context;
 	r = &uc->uc_mcontext;
 
+	if(sig == SIGPROF) {
+		runtime·sigprof((uint8*)r->arm_pc, (uint8*)r->arm_sp, (uint8*)r->arm_lr, gp);
+		return;
+	}
+
 	if(gp != nil && (runtime·sigtab[sig].flags & SigPanic)) {
 		// Make it look like a call to the signal func.
 		// Have to pass arguments out of band since
@@ -119,31 +124,58 @@ runtime·signalstack(byte *p, int32 n)
 	runtime·sigaltstack(&st, nil);
 }
 
+static void
+sigaction(int32 i, void (*fn)(int32, Siginfo*, void*, G*), bool restart)
+{
+	Sigaction sa;
+
+	runtime·memclr((byte*)&sa, sizeof sa);
+	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
+	if(restart)
+		sa.sa_flags |= SA_RESTART;
+	sa.sa_mask = ~0ULL;
+	sa.sa_restorer = (void*)runtime·sigreturn;
+	sa.k_sa_handler = fn;
+	runtime·rt_sigaction(i, &sa, nil, 8);
+}
+
 void
 runtime·initsig(int32 queue)
 {
-	static Sigaction sa;
+	int32 i;
+	void *fn;
 
 	runtime·siginit();
 
-	int32 i;
-	sa.sa_flags = SA_ONSTACK | SA_SIGINFO | SA_RESTORER;
-	sa.sa_mask.sig[0] = 0xFFFFFFFF;
-	sa.sa_mask.sig[1] = 0xFFFFFFFF;
-	sa.sa_restorer = (void*)runtime·sigreturn;
 	for(i = 0; i<NSIG; i++) {
 		if(runtime·sigtab[i].flags) {
 			if((runtime·sigtab[i].flags & SigQueue) != queue)
 				continue;
 			if(runtime·sigtab[i].flags & (SigCatch | SigQueue))
-				sa.sa_handler = (void*)runtime·sigtramp;
+				fn = runtime·sighandler;
 			else
-				sa.sa_handler = (void*)runtime·sigignore;
-			if(runtime·sigtab[i].flags & SigRestart)
-				sa.sa_flags |= SA_RESTART;
-			else
-				sa.sa_flags &= ~SA_RESTART;
-			runtime·rt_sigaction(i, &sa, nil, 8);
+				fn = runtime·sigignore;
+			sigaction(i, fn, (runtime·sigtab[i].flags & SigRestart) != 0);
 		}
 	}
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	Sigaction sa;
+	Itimerval it;
+	
+	runtime·memclr((byte*)&it, sizeof it);
+	if(hz == 0) {
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+		sigaction(SIGPROF, SIG_IGN, true);
+	} else {
+		sigaction(SIGPROF, runtime·sighandler, true);
+		it.it_interval.tv_sec = 0;
+		it.it_interval.tv_usec = 1000000 / hz;
+		it.it_value = it.it_interval;
+		runtime·setitimer(ITIMER_PROF, &it, nil);
+	}
+	m->profilehz = hz;
+}

src/pkg/runtime/plan9/386/signal.c

@@ -14,3 +14,11 @@ runtime·signame(int32)
 {
 	return runtime·emptystring;
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	// TODO: Enable profiling interrupts.
+	
+	m->profilehz = hz;
+}

src/pkg/runtime/proc.c

@@ -70,6 +70,7 @@ struct Sched {
 	int32 msyscall;	// number of ms in system calls
 
 	int32 predawn;	// running initialization, don't run new gs.
+	int32 profilehz;	// cpu profiling rate
 
 	Note	stopped;	// one g can wait here for ms to stop
 	int32 waitstop;	// after setting this flag
@@ -96,9 +97,6 @@ static void matchmg(void);	// match ms to gs
 static void readylocked(G*);	// ready, but sched is locked
 static void mnextg(M*, G*);
 
-// Scheduler loop.
-static void scheduler(void);
-
 // The bootstrap sequence is:
 //
 //	call osinit
@@ -529,6 +527,8 @@ matchmg(void)
 static void
 schedule(G *gp)
 {
+	int32 hz;
+
 	schedlock();
 	if(gp != nil) {
 		if(runtime·sched.predawn)
@@ -574,6 +574,12 @@ schedule(G *gp)
 	gp->status = Grunning;
 	m->curg = gp;
 	gp->m = m;
+	
+	// Check whether the profiler needs to be turned on or off.
+	hz = runtime·sched.profilehz;
+	if(m->profilehz != hz)
+		runtime·resetcpuprofiler(hz);
+
 	if(gp->sched.pc == (byte*)runtime·goexit) {	// kickoff
 		runtime·gogocall(&gp->sched, (void(*)(void))gp->entry);
 	}
@@ -640,7 +646,7 @@ runtime·exitsyscall(void)
 	runtime·sched.msyscall--;
 	runtime·sched.mcpu++;
 	// Fast path - if there's room for this m, we're done.
-	if(runtime·sched.mcpu <= runtime·sched.mcpumax) {
+	if(m->profilehz == runtime·sched.profilehz && runtime·sched.mcpu <= runtime·sched.mcpumax) {
 		g->status = Grunning;
 		schedunlock();
 		return;
@@ -1251,3 +1257,57 @@ runtime·badmcall2(void)  // called from assembly
 {
 	runtime·throw("runtime: mcall function returned");
 }
+
+static struct {
+	Lock;
+	void (*fn)(uintptr*, int32);
+	int32 hz;
+	uintptr pcbuf[100];
+} prof;
+
+void
+runtime·sigprof(uint8 *pc, uint8 *sp, uint8 *lr, G *gp)
+{
+	int32 n;
+	
+	if(prof.fn == nil || prof.hz == 0)
+		return;
+	
+	runtime·lock(&prof);
+	if(prof.fn == nil) {
+		runtime·unlock(&prof);
+		return;
+	}
+	n = runtime·gentraceback(pc, sp, lr, gp, 0, prof.pcbuf, nelem(prof.pcbuf));
+	if(n > 0)
+		prof.fn(prof.pcbuf, n);
+	runtime·unlock(&prof);
+}
+
+void
+runtime·setcpuprofilerate(void (*fn)(uintptr*, int32), int32 hz)
+{
+	// Force sane arguments.
+	if(hz < 0)
+		hz = 0;
+	if(hz == 0)
+		fn = nil;
+	if(fn == nil)
+		hz = 0;
+
+	// Stop profiler on this cpu so that it is safe to lock prof.
+	// if a profiling signal came in while we had prof locked,
+	// it would deadlock.
+	runtime·resetcpuprofiler(0);
+
+	runtime·lock(&prof);
+	prof.fn = fn;
+	prof.hz = hz;
+	runtime·unlock(&prof);
+	runtime·lock(&runtime·sched);
+	runtime·sched.profilehz = hz;
+	runtime·unlock(&runtime·sched);
+	
+	if(hz != 0)
+		runtime·resetcpuprofiler(hz);
+}

src/pkg/runtime/runtime.h

@@ -225,6 +225,7 @@ struct	M
 	int32	nomemprof;
 	int32	waitnextg;
 	int32	dying;
+	int32	profilehz;
 	Note	havenextg;
 	G*	nextg;
 	M*	alllink;	// on allm
@@ -453,9 +454,13 @@ void	runtime·siginit(void);
 bool	runtime·sigsend(int32 sig);
 void	runtime·gettime(int64*, int32*);
 int32	runtime·callers(int32, uintptr*, int32);
+int32	runtime·gentraceback(byte*, byte*, byte*, G*, int32, uintptr*, int32);
 int64	runtime·nanotime(void);
 void	runtime·dopanic(int32);
 void	runtime·startpanic(void);
+void	runtime·sigprof(uint8 *pc, uint8 *sp, uint8 *lr, G *gp);
+void	runtime·resetcpuprofiler(int32);
+void	runtime·setcpuprofilerate(void(*)(uintptr*, int32), int32);
 
 #pragma	varargck	argpos	runtime·printf	1
 #pragma	varargck	type	"d"	int32

src/pkg/runtime/windows/386/signal.c

@@ -88,3 +88,11 @@ runtime·sighandler(ExceptionRecord *info, void *frame, Context *r)
 	runtime·exit(2);
 	return 0;
 }
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	// TODO: Enable profiling interrupts.
+	
+	m->profilehz = hz;
+}