1
0
mirror of https://github.com/golang/go synced 2024-11-13 18:30:26 -07:00

runtime/pprof: test multithreaded profile, remove OS X workarounds

This means that pprof will no longer report profiles on OS X.
That's unfortunate, but the profiles were often wrong and, worse,
it was difficult to tell whether the profile was wrong or not.

The workarounds were making the scheduler more complex,
possibly caused a deadlock (see issue 5519), and did not actually
deliver reliable results.

It may be possible for adventurous users to apply a patch to
their kernels to get working results, or perhaps having no results
will encourage someone to do the work of creating a profiling
thread like on Windows. Issue 6047 has details.

Fixes #5519.
Fixes #6047.

R=golang-dev, bradfitz, r
CC=golang-dev
https://golang.org/cl/12429045
This commit is contained in:
Russ Cox 2013-08-05 19:49:02 -04:00
parent d8e27db395
commit d3066e47b1
15 changed files with 68 additions and 136 deletions

View File

@ -331,7 +331,7 @@ getprofile(Profile *p)
if(p->wholding) {
// Release previous log to signal handling side.
// Loop because we are racing against setprofile(off).
// Loop because we are racing against SetCPUProfileRate(0).
for(;;) {
n = p->handoff;
if(n == 0) {

View File

@ -83,11 +83,7 @@ runtime·lock(Lock *l)
if(v == MUTEX_UNLOCKED)
return;
wait = MUTEX_SLEEPING;
if(m->profilehz > 0)
runtime·setprof(false);
runtime·futexsleep((uint32*)&l->key, MUTEX_SLEEPING, -1);
if(m->profilehz > 0)
runtime·setprof(true);
}
}
@ -133,12 +129,8 @@ runtime·notesleep(Note *n)
{
if(g != m->g0)
runtime·throw("notesleep not on g0");
if(m->profilehz > 0)
runtime·setprof(false);
while(runtime·atomicload((uint32*)&n->key) == 0)
runtime·futexsleep((uint32*)&n->key, 0, -1);
if(m->profilehz > 0)
runtime·setprof(true);
}
#pragma textflag 7
@ -179,16 +171,11 @@ runtime·notetsleep(Note *n, int64 ns)
if(g != m->g0 && !m->gcing)
runtime·throw("notetsleep not on g0");
if(m->profilehz > 0)
runtime·setprof(false);
res = notetsleep(n, ns, 0, 0);
if(m->profilehz > 0)
runtime·setprof(true);
return res;
}
// same as runtime·notetsleep, but called on user g (not g0)
// does not need to call runtime·setprof, because entersyscallblock does it
// calls only nosplit functions between entersyscallblock/exitsyscall
bool
runtime·notetsleepg(Note *n, int64 ns)

View File

@ -81,11 +81,7 @@ unlocked:
}
if(v&LOCKED) {
// Queued. Wait.
if(m->profilehz > 0)
runtime·setprof(false);
runtime·semasleep(-1);
if(m->profilehz > 0)
runtime·setprof(true);
i = 0;
}
}
@ -164,11 +160,7 @@ runtime·notesleep(Note *n)
return;
}
// Queued. Sleep.
if(m->profilehz > 0)
runtime·setprof(false);
runtime·semasleep(-1);
if(m->profilehz > 0)
runtime·setprof(true);
}
#pragma textflag 7
@ -240,16 +232,11 @@ runtime·notetsleep(Note *n, int64 ns)
if(m->waitsema == 0)
m->waitsema = runtime·semacreate();
if(m->profilehz > 0)
runtime·setprof(false);
res = notetsleep(n, ns, 0, nil);
if(m->profilehz > 0)
runtime·setprof(true);
return res;
}
// same as runtime·notetsleep, but called on user g (not g0)
// does not need to call runtime·setprof, because entersyscallblock does it
// calls only nosplit functions between entersyscallblock/exitsyscall
bool
runtime·notetsleepg(Note *n, int64 ns)

View File

@ -12,7 +12,6 @@ extern SigTab runtime·sigtab[];
static Sigset sigset_none;
static Sigset sigset_all = ~(Sigset)0;
static Sigset sigset_prof = 1<<(SIGPROF-1);
static void
unimplemented(int8 *name)
@ -129,7 +128,6 @@ runtime·minit(void)
runtime·signalstack((byte*)m->gsignal->stackguard - StackGuard, 32*1024);
runtime·sigprocmask(SIG_SETMASK, &sigset_none, nil);
runtime·setprof(m->profilehz > 0);
}
// Called from dropm to undo the effect of an minit.
@ -481,37 +479,6 @@ runtime·memlimit(void)
return 0;
}
// NOTE(rsc): On OS X, when the CPU profiling timer expires, the SIGPROF
// signal is not guaranteed to be sent to the thread that was executing to
// cause it to expire. It can and often does go to a sleeping thread, which is
// not interesting for our profile. This is filed Apple Bug Report #9177434,
// copied to http://code.google.com/p/go/source/detail?r=35b716c94225.
// To work around this bug, we disable receipt of the profiling signal on
// a thread while in blocking system calls. This forces the kernel to deliver
// the profiling signal to an executing thread.
//
// The workaround fails on OS X machines using a 64-bit Snow Leopard kernel.
// In that configuration, the kernel appears to want to deliver SIGPROF to the
// sleeping threads regardless of signal mask and, worse, does not deliver
// the signal until the thread wakes up on its own.
//
// If necessary, we can switch to using ITIMER_REAL for OS X and handle
// the kernel-generated SIGALRM by generating our own SIGALRMs to deliver
// to all the running threads. SIGALRM does not appear to be affected by
// the 64-bit Snow Leopard bug. However, as of this writing Mountain Lion
// is in preview, making Snow Leopard two versions old, so it is unclear how
// much effort we need to spend on one buggy kernel.
// Control whether profiling signal can be delivered to this thread.
void
runtime·setprof(bool on)
{
if(on)
runtime·sigprocmask(SIG_UNBLOCK, &sigset_prof, nil);
else
runtime·sigprocmask(SIG_BLOCK, &sigset_prof, nil);
}
void
runtime·setsig(int32 i, GoSighandler *fn, bool restart)
{

View File

@ -233,12 +233,6 @@ runtime·memlimit(void)
return rl.rlim_cur - used;
}
void
runtime·setprof(bool on)
{
USED(on);
}
extern void runtime·sigtramp(void);
typedef struct sigaction {

View File

@ -274,12 +274,6 @@ runtime·memlimit(void)
return rl.rlim_cur - used;
}
void
runtime·setprof(bool on)
{
USED(on);
}
#ifdef GOARCH_386
#define sa_handler k_sa_handler
#endif

View File

@ -271,12 +271,6 @@ runtime·memlimit(void)
return 0;
}
void
runtime·setprof(bool on)
{
USED(on);
}
extern void runtime·sigtramp(void);
typedef struct sigaction {

View File

@ -248,12 +248,6 @@ runtime·memlimit(void)
return 0;
}
void
runtime·setprof(bool on)
{
USED(on);
}
extern void runtime·sigtramp(void);
typedef struct sigaction {

View File

@ -322,12 +322,6 @@ runtime·memlimit(void)
return 0;
}
void
runtime·setprof(bool on)
{
USED(on);
}
#pragma dataflag 16 // no pointers
static int8 badsignal[] = "runtime: signal received on thread not created by Go.\n";

View File

@ -474,12 +474,6 @@ runtime·memlimit(void)
return 0;
}
void
runtime·setprof(bool on)
{
USED(on);
}
#pragma dataflag 16 // no pointers
int8 runtime·badsignalmsg[] = "runtime: signal received on thread not created by Go.\n";
int32 runtime·badsignallen = sizeof runtime·badsignalmsg - 1;

View File

@ -20,8 +20,7 @@ import (
"text/tabwriter"
)
// BUG(rsc): A bug in the OS X Snow Leopard 64-bit kernel prevents
// CPU profiling from giving accurate results on that system.
// BUG(rsc): Profiles are incomplete and inaccuate on OS X. See http://golang.org/issue/6047 for details.
// A Profile is a collection of stack traces showing the call sequences
// that led to instances of a particular event, such as allocation.

View File

@ -19,6 +19,38 @@ import (
)
func TestCPUProfile(t *testing.T) {
buf := make([]byte, 100000)
testCPUProfile(t, []string{"crc32.ChecksumIEEE"}, func() {
// This loop takes about a quarter second on a 2 GHz laptop.
// We only need to get one 100 Hz clock tick, so we've got
// a 25x safety buffer.
for i := 0; i < 1000; i++ {
crc32.ChecksumIEEE(buf)
}
})
}
func TestCPUProfileMultithreaded(t *testing.T) {
buf := make([]byte, 100000)
defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(2))
testCPUProfile(t, []string{"crc32.ChecksumIEEE", "crc32.Update"}, func() {
c := make(chan int)
go func() {
for i := 0; i < 2000; i++ {
crc32.Update(0, crc32.IEEETable, buf)
}
c <- 1
}()
// This loop takes about a quarter second on a 2 GHz laptop.
// We only need to get one 100 Hz clock tick, so we've got
// a 25x safety buffer.
for i := 0; i < 2000; i++ {
crc32.ChecksumIEEE(buf)
}
})
}
func testCPUProfile(t *testing.T, need []string, f func()) {
switch runtime.GOOS {
case "darwin":
out, err := exec.Command("uname", "-a").CombinedOutput()
@ -27,26 +59,16 @@ func TestCPUProfile(t *testing.T) {
}
vers := string(out)
t.Logf("uname -a: %v", vers)
// Lion uses "Darwin Kernel Version 11".
if strings.Contains(vers, "Darwin Kernel Version 10") && strings.Contains(vers, "RELEASE_X86_64") {
t.Skip("skipping test on known-broken kernel (64-bit Leopard / Snow Leopard)")
}
case "plan9":
// unimplemented
return
}
buf := make([]byte, 100000)
var prof bytes.Buffer
if err := StartCPUProfile(&prof); err != nil {
t.Fatal(err)
}
// This loop takes about a quarter second on a 2 GHz laptop.
// We only need to get one 100 Hz clock tick, so we've got
// a 25x safety buffer.
for i := 0; i < 1000; i++ {
crc32.ChecksumIEEE(buf)
}
f()
StopCPUProfile()
// Convert []byte to []uintptr.
@ -56,6 +78,10 @@ func TestCPUProfile(t *testing.T) {
val = val[:l]
if l < 13 {
if runtime.GOOS == "darwin" {
t.Logf("ignoring failure on OS X; see golang.org/issue/6047")
return
}
t.Fatalf("profile too short: %#x", val)
}
@ -69,7 +95,7 @@ func TestCPUProfile(t *testing.T) {
}
// Check that profile is well formed and contains ChecksumIEEE.
found := false
have := make([]uintptr, len(need))
for len(val) > 0 {
if len(val) < 2 || val[0] < 1 || val[1] < 1 || uintptr(len(val)) < 2+val[1] {
t.Fatalf("malformed profile. leftover: %#x", val)
@ -79,14 +105,38 @@ func TestCPUProfile(t *testing.T) {
if f == nil {
continue
}
if strings.Contains(f.Name(), "ChecksumIEEE") {
found = true
for i, name := range need {
if strings.Contains(f.Name(), name) {
have[i] += val[0]
}
}
}
val = val[2+val[1]:]
}
if !found {
t.Fatal("did not find ChecksumIEEE in the profile")
var total uintptr
for i, name := range need {
total += have[i]
t.Logf("%s: %d\n", name, have[i])
}
ok := true
if total == 0 {
t.Logf("no CPU profile samples collected")
ok = false
}
min := total / uintptr(len(have)) / 2
for i, name := range need {
if have[i] < min {
t.Logf("%s has %d samples out of %d, want at least %d, ideally %d", name, have[i], total, min, total/uintptr(len(have)))
ok = false
}
}
if !ok {
if runtime.GOOS == "darwin" {
t.Logf("ignoring failure on OS X; see golang.org/issue/6047")
return
}
t.FailNow()
}
}

View File

@ -1384,9 +1384,6 @@ void
// but can have inconsistent g->sched, do not let GC observe it.
m->locks++;
if(m->profilehz > 0)
runtime·setprof(false);
// Leave SP around for gc and traceback.
save(runtime·getcallerpc(&dummy), runtime·getcallersp(&dummy));
@ -1439,9 +1436,6 @@ void
m->locks++; // see comment in entersyscall
if(m->profilehz > 0)
runtime·setprof(false);
// Leave SP around for gc and traceback.
save(runtime·getcallerpc(&dummy), runtime·getcallersp(&dummy));
g->gcsp = g->sched.sp;
@ -1477,10 +1471,6 @@ runtime·exitsyscall(void)
{
m->locks++; // see comment in entersyscall
// Check whether the profiler needs to be turned on.
if(m->profilehz > 0)
runtime·setprof(true);
if(g->isbackground) // do not consider blocked scavenger for deadlock detection
inclocked(-1);

View File

@ -1043,16 +1043,6 @@ void runtime·ifaceE2I(InterfaceType*, Eface, Iface*);
uintptr runtime·memlimit(void);
// If appropriate, ask the operating system to control whether this
// thread should receive profiling signals. This is only necessary on OS X.
// An operating system should not deliver a profiling signal to a
// thread that is not actually executing (what good is that?), but that's
// what OS X prefers to do. When profiling is turned on, we mask
// away the profiling signal when threads go to sleep, so that OS X
// is forced to deliver the signal to a thread that's actually running.
// This is a no-op on other systems.
void runtime·setprof(bool);
// float.c
extern float64 runtime·nan;
extern float64 runtime·posinf;

View File

@ -83,13 +83,11 @@ runtime·resetcpuprofiler(int32 hz)
runtime·memclr((byte*)&it, sizeof it);
if(hz == 0) {
runtime·setitimer(ITIMER_PROF, &it, nil);
runtime·setprof(false);
} else {
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);
runtime·setprof(true);
}
m->profilehz = hz;
}