blob: e1206f99f105ade38274b98cf04567696ecab70b [file] [log] [blame]
// 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
// https://github.com/google/pprof.
//
// 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.
package runtime
import (
"runtime/internal/atomic"
"unsafe"
)
const (
numBuckets = 1 << 10
logSize = 1 << 17
assoc = 4
maxCPUProfStack = 64
)
type cpuprofEntry struct {
count uintptr
depth int
stack [maxCPUProfStack]uintptr
}
//go:notinheap
type cpuProfile struct {
on bool // profiling is on
wait note // goroutine waits here
count uintptr // tick count
evicts uintptr // eviction count
lost uintptr // lost ticks that need to be logged
// Active recent stack traces.
hash [numBuckets]struct {
entry [assoc]cpuprofEntry
}
// Log of traces evicted from hash.
// Signal handler has filled log[toggle][:nlog].
// Goroutine is writing log[1-toggle][:handoff].
log [2][logSize / 2]uintptr
nlog int
toggle int32
handoff uint32
// Writer state.
// Writer maintains its own toggle to avoid races
// looking at signal handler's toggle.
wtoggle uint32
wholding bool // holding & need to release a log half
flushing bool // flushing hash table - profile is over
eodSent bool // special end-of-data record sent; => flushing
}
var (
cpuprofLock mutex
cpuprof *cpuProfile
eod = [3]uintptr{0, 1, 0}
)
func setcpuprofilerate(hz int32) {
systemstack(func() {
setcpuprofilerate_m(hz)
})
}
// 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.
func lostProfileData() {}
// 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) {
// Clamp hz to something reasonable.
if hz < 0 {
hz = 0
}
if hz > 1000000 {
hz = 1000000
}
lock(&cpuprofLock)
if hz > 0 {
if cpuprof == nil {
cpuprof = (*cpuProfile)(sysAlloc(unsafe.Sizeof(cpuProfile{}), &memstats.other_sys))
if cpuprof == nil {
print("runtime: cpu profiling cannot allocate memory\n")
unlock(&cpuprofLock)
return
}
}
if cpuprof.on || cpuprof.handoff != 0 {
print("runtime: cannot set cpu profile rate until previous profile has finished.\n")
unlock(&cpuprofLock)
return
}
cpuprof.on = true
// pprof binary header format.
// https://github.com/gperftools/gperftools/blob/master/src/profiledata.cc#L119
p := &cpuprof.log[0]
p[0] = 0 // count for header
p[1] = 3 // depth for header
p[2] = 0 // version number
p[3] = uintptr(1e6 / hz) // period (microseconds)
p[4] = 0
cpuprof.nlog = 5
cpuprof.toggle = 0
cpuprof.wholding = false
cpuprof.wtoggle = 0
cpuprof.flushing = false
cpuprof.eodSent = false
noteclear(&cpuprof.wait)
setcpuprofilerate(int32(hz))
} else if cpuprof != nil && cpuprof.on {
setcpuprofilerate(0)
cpuprof.on = false
// Now add is not running anymore, and getprofile owns the entire log.
// Set the high bit in cpuprof.handoff to tell getprofile.
for {
n := cpuprof.handoff
if n&0x80000000 != 0 {
print("runtime: setcpuprofile(off) twice\n")
}
if atomic.Cas(&cpuprof.handoff, n, n|0x80000000) {
if n == 0 {
// we did the transition from 0 -> nonzero so we wake getprofile
notewakeup(&cpuprof.wait)
}
break
}
}
}
unlock(&cpuprofLock)
}
// 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.
//go:nowritebarrierrec
func (p *cpuProfile) add(pc []uintptr) {
p.addWithFlushlog(pc, p.flushlog)
}
// addWithFlushlog implements add and addNonGo.
// 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 may be called by a signal handler with no g or m.
// It is allowed to call evict, passing the flushlog parameter.
//go:nosplit
//go:nowritebarrierrec
func (p *cpuProfile) addWithFlushlog(pc []uintptr, flushlog func() bool) {
if len(pc) > maxCPUProfStack {
pc = pc[:maxCPUProfStack]
}
// Compute hash.
h := uintptr(0)
for _, x := range pc {
h = h<<8 | (h >> (8 * (unsafe.Sizeof(h) - 1)))
h += x * 41
}
p.count++
// Add to entry count if already present in table.
b := &p.hash[h%numBuckets]
Assoc:
for i := range b.entry {
e := &b.entry[i]
if e.depth != len(pc) {
continue
}
for j := range pc {
if e.stack[j] != pc[j] {
continue Assoc
}
}
e.count++
return
}
// Evict entry with smallest count.
var e *cpuprofEntry
for i := range b.entry {
if e == nil || b.entry[i].count < e.count {
e = &b.entry[i]
}
}
if e.count > 0 {
if !p.evict(e, flushlog) {
// Could not evict entry. Record lost stack.
p.lost++
return
}
p.evicts++
}
// Reuse the newly evicted entry.
e.depth = len(pc)
e.count = 1
copy(e.stack[:], pc)
}
// 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, and it may be called with no g or m.
// It is safe to call flushlog. evict returns true if the entry was
// copied to the log, false if there was no room available.
//go:nosplit
//go:nowritebarrierrec
func (p *cpuProfile) evict(e *cpuprofEntry, flushlog func() bool) bool {
d := e.depth
nslot := d + 2
log := &p.log[p.toggle]
if p.nlog+nslot > len(log) {
if !flushlog() {
return false
}
log = &p.log[p.toggle]
}
q := p.nlog
log[q] = e.count
q++
log[q] = uintptr(d)
q++
copy(log[q:], e.stack[:d])
q += d
p.nlog = q
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.
//go:nowritebarrierrec
func (p *cpuProfile) flushlog() bool {
if !atomic.Cas(&p.handoff, 0, uint32(p.nlog)) {
return false
}
notewakeup(&p.wait)
p.toggle = 1 - p.toggle
log := &p.log[p.toggle]
q := 0
if p.lost > 0 {
lostPC := funcPC(lostProfileData)
log[0] = p.lost
log[1] = 1
log[2] = lostPC
q = 3
p.lost = 0
}
p.nlog = q
return true
}
// addNonGo is like add, but runs on a non-Go thread.
// It can't do anything that might need a g or an m.
// With this entry point, we don't try to flush the log when evicting an
// old entry. Instead, we just drop the stack trace if we're out of space.
//go:nosplit
//go:nowritebarrierrec
func (p *cpuProfile) addNonGo(pc []uintptr) {
p.addWithFlushlog(pc, func() bool { return false })
}
// getprofile blocks until the next block of profiling data is available
// and returns it as a []byte. It is called from the writing goroutine.
func (p *cpuProfile) getprofile() []byte {
if p == nil {
return nil
}
if p.wholding {
// Release previous log to signal handling side.
// Loop because we are racing against SetCPUProfileRate(0).
for {
n := p.handoff
if n == 0 {
print("runtime: phase error during cpu profile handoff\n")
return nil
}
if n&0x80000000 != 0 {
p.wtoggle = 1 - p.wtoggle
p.wholding = false
p.flushing = true
goto Flush
}
if atomic.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 nil
}
// Wait for new log.
notetsleepg(&p.wait, -1)
noteclear(&p.wait)
switch n := p.handoff; {
case n == 0:
print("runtime: phase error during cpu profile wait\n")
return nil
case n == 0x80000000:
p.flushing = true
goto Flush
default:
n &^= 0x80000000
// Return new log to caller.
p.wholding = true
return uintptrBytes(p.log[p.wtoggle][:n])
}
// 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.
Flush:
for i := range p.hash {
b := &p.hash[i]
for j := range b.entry {
e := &b.entry[j]
if e.count > 0 && !p.evict(e, p.flushlog) {
// Filled the log. Stop the loop and return what we've got.
break Flush
}
}
}
// 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.
n := p.nlog
p.nlog = 0
return uintptrBytes(p.log[p.toggle][:n])
}
// Made it through the table without finding anything to log.
if !p.eodSent {
// We may not have space to append this to the partial log buf,
// so we always return a new slice for the end-of-data marker.
p.eodSent = true
return uintptrBytes(eod[:])
}
// Finally done. Clean up and return nil.
p.flushing = false
if !atomic.Cas(&p.handoff, p.handoff, 0) {
print("runtime: profile flush racing with something\n")
}
return nil
}
func uintptrBytes(p []uintptr) (ret []byte) {
pp := (*slice)(unsafe.Pointer(&p))
rp := (*slice)(unsafe.Pointer(&ret))
rp.array = pp.array
rp.len = pp.len * int(unsafe.Sizeof(p[0]))
rp.cap = rp.len
return
}
// 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 {
return cpuprof.getprofile()
}
//go:linkname runtime_pprof_runtime_cyclesPerSecond runtime_pprof.runtime_cyclesPerSecond
func runtime_pprof_runtime_cyclesPerSecond() int64 {
return tickspersecond()
}