runtime/pprof: remove hard-coded sleep in CPU profile reader

The CPU profiler reader goroutine has a hard-coded 100ms sleep between
reads of the CPU profile sample buffer. This is done because waking up
the CPU profile reader is not signal-safe on some platforms. As a
consequence, stopping the profiler takes 200ms (one iteration to read
the last samples and one to see the "eof"), and on many-core systems the
reader does not wake up frequently enought to keep up with incoming
data.

This CL removes the sleep where it is safe to do so, following a
suggestion by Austin Clements in the comments on CL 445375. We let the
reader fully block, and wake up the reader when the buffer is over
half-full.

Fixes #63043
Updates #56029

Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest,gotip-linux-arm64-longtest,gotip-linux-386-longtest
Change-Id: I9f7e7e9918a4a6f16e80f6aaf33103126568a81f
Reviewed-on: https://go-review.googlesource.com/c/go/+/610815
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Mark Freeman <markfreeman@google.com>

3 files changed, 103 insertions, 3 deletions

diff --git a/src/runtime/pprof/pprof.go b/src/runtime/pprof/pprof.go
index 78d00af6ca..c617a8b26a 100644
--- a/src/runtime/pprof/pprof.go
+++ b/src/runtime/pprof/pprof.go
@@ -924,7 +924,10 @@ func profileWriter(w io.Writer) {
 	b := newProfileBuilder(w)
 	var err error
 	for {
-		time.Sleep(100 * time.Millisecond)
+		if runtime.GOOS == "darwin" || runtime.GOOS == "ios" {
+			// see runtime_pprof_readProfile
+			time.Sleep(100 * time.Millisecond)
+		}
 		data, tags, eof := readProfile()
 		if e := b.addCPUData(data, tags); e != nil && err == nil {
 			err = e
diff --git a/src/runtime/profbuf.go b/src/runtime/profbuf.go
index 40aff9e1d2..147212206f 100644
--- a/src/runtime/profbuf.go
+++ b/src/runtime/profbuf.go
@@ -38,6 +38,10 @@ import (
 // be returned by read. By definition, r ≤ rNext ≤ w (before wraparound),
 // and rNext is only used by the reader, so it can be accessed without atomics.
 //
+// If the reader is blocked waiting for more data, the writer will wake it up if
+// either the buffer is more than half full, or when the writer sets the eof
+// marker or writes overflow entries (described below.)
+//
 // If the writer gets ahead of the reader, so that the buffer fills,
 // future writes are discarded and replaced in the output stream by an
 // overflow entry, which has size 2+hdrsize+1, time set to the time of
@@ -378,11 +382,28 @@ func (b *profBuf) write(tagPtr *unsafe.Pointer, now int64, hdr []uint64, stk []u
 		// Racing with reader setting flag bits in b.w, to avoid lost wakeups.
 		old := b.w.load()
 		new := old.addCountsAndClearFlags(skip+2+len(stk)+int(b.hdrsize), 1)
+		// We re-load b.r here to reduce the likelihood of early wakeups
+		// if the reader already consumed some data between the last
+		// time we read b.r and now. This isn't strictly necessary.
+		unread := countSub(new.dataCount(), b.r.load().dataCount())
+		if unread < 0 {
+			// The new count overflowed and wrapped around.
+			unread += len(b.data)
+		}
+		wakeupThreshold := len(b.data) / 2
+		if unread < wakeupThreshold {
+			// Carry over the sleeping flag since we're not planning
+			// to wake the reader yet
+			new |= old & profReaderSleeping
+		}
 		if !b.w.cas(old, new) {
 			continue
 		}
-		// If there was a reader, wake it up.
-		if old&profReaderSleeping != 0 {
+		// If we've hit our high watermark for data in the buffer,
+		// and there is a reader, wake it up.
+		if unread >= wakeupThreshold && old&profReaderSleeping != 0 {
+			// NB: if we reach this point, then the sleeping bit is
+			// cleared in the new b.w value
 			notewakeup(&b.wait)
 		}
 		break
diff --git a/src/runtime/profbuf_test.go b/src/runtime/profbuf_test.go
index 7a435c9e5f..2f068ac386 100644
--- a/src/runtime/profbuf_test.go
+++ b/src/runtime/profbuf_test.go
@@ -5,8 +5,12 @@
 package runtime_test
 
 import (
+	"fmt"
+	"regexp"
+	"runtime"
 	. "runtime"
 	"slices"
+	"sync"
 	"testing"
 	"time"
 	"unsafe"
@@ -190,3 +194,75 @@ func TestProfBufDoubleWakeup(t *testing.T) {
 		}
 	}
 }
+
+func TestProfBufWakeup(t *testing.T) {
+	b := NewProfBuf(2, 16, 2)
+	var wg sync.WaitGroup
+	wg.Go(func() {
+		read := 0
+		for {
+			rdata, _, eof := b.Read(ProfBufBlocking)
+			if read == 0 && len(rdata) < 8 {
+				t.Errorf("first wake up at less than half full, got %x", rdata)
+			}
+			read += len(rdata)
+			if eof {
+				return
+			}
+		}
+	})
+
+	// Under the hood profBuf uses notetsleepg when the reader blocks.
+	// Different platforms have different implementations, leading to
+	// different statuses we need to look for to determine whether the
+	// reader is blocked.
+	var waitStatus string
+	switch runtime.GOOS {
+	case "js":
+		waitStatus = "waiting"
+	case "wasip1":
+		waitStatus = "runnable"
+	default:
+		waitStatus = "syscall"
+	}
+
+	// Ensure that the reader is blocked
+	awaitBlockedGoroutine(waitStatus, "TestProfBufWakeup.func1")
+	// NB: this writes 6 words not 4. Fine for the test.
+	// The reader shouldn't wake up for this
+	b.Write(nil, 1, []uint64{1, 2}, []uintptr{3, 4})
+
+	// The reader should still be blocked
+	//
+	// TODO(nick): this is racy. We could Gosched and still have the reader
+	// blocked in a buggy implementation because it just didn't get a chance
+	// to run
+	awaitBlockedGoroutine(waitStatus, "TestProfBufWakeup.func1")
+	b.Write(nil, 1, []uint64{5, 6}, []uintptr{7, 8})
+	b.Close()
+
+	// Wait here so we can be sure the reader got the data
+	wg.Wait()
+}
+
+// see also runtime/pprof tests
+func awaitBlockedGoroutine(state, fName string) {
+	re := fmt.Sprintf(`(?m)^goroutine \d+ \[%s\]:\n(?:.+\n\t.+\n)*runtime_test\.%s`, regexp.QuoteMeta(state), fName)
+	r := regexp.MustCompile(re)
+
+	buf := make([]byte, 64<<10)
+	for {
+		Gosched()
+		n := Stack(buf, true)
+		if n == len(buf) {
+			// Buffer wasn't large enough for a full goroutine dump.
+			// Resize it and try again.
+			buf = make([]byte, 2*len(buf))
+			continue
+		}
+		const count = 1
+		if len(r.FindAll(buf[:n], -1)) >= count {
+			return
+		}
+	}
+}