runtime: don't call traceReadCPU on the system stack

traceReadCPU calls profBuf.read, which does a raceacquire. g0 does not
have a race context, so this crashes when running on the system stack.

We could borrow a race context, but it is simpler to just move
traceReadCPU off of the system stack.

Fixes golang#65607.

Change-Id: I335155b96d683aebb92b2f4e1eea063dd139f2d5
Reviewed-on: https://go-review.googlesource.com/c/go/+/562996
Auto-Submit: Michael Pratt <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>

Author: prattmic

src/runtime/trace2.go

@@ -516,6 +516,9 @@ func traceAdvance(stopTrace bool) {
 	}
 	statusWriter.flush().end()
 
+	// Read everything out of the last gen's CPU profile buffer.
+	traceReadCPU(gen)
+
 	systemstack(func() {
 		// Flush CPU samples, stacks, and strings for the last generation. This is safe,
 		// because we're now certain no M is writing to the last generation.

src/runtime/trace2cpu.go

@@ -112,6 +112,9 @@ func traceStopReadCPU() {
 //
 // No more than one goroutine may be in traceReadCPU for the same
 // profBuf at a time.
+//
+// Must not run on the system stack because profBuf.read performs race
+// operations.
 func traceReadCPU(gen uintptr) bool {
 	var pcBuf [traceStackSize]uintptr
 
@@ -198,9 +201,6 @@ func traceReadCPU(gen uintptr) bool {
 //
 //go:systemstack
 func traceCPUFlush(gen uintptr) {
-	// Read everything out of the last gen's CPU profile buffer.
-	traceReadCPU(gen)
-
 	// Flush any remaining trace buffers containing CPU samples.
 	if buf := trace.cpuBuf[gen%2]; buf != nil {
 		lock(&trace.lock)