runtime/coverage: use atomic access for counter reads

Read counters using atomic ops so as to avoid problems with the race
detector if a goroutine happens to still be executing at the end of a
test run when we're writing out counter data. In theory we could guard
the atomic use on the counter mode, but it's better just to do it in
all cases, leaves us with a simpler implementation.

Fixes #56006.

Change-Id: I81c2234b5a1c3b00cff6c77daf2c2315451b7f6c
Reviewed-on: https://go-review.googlesource.com/c/go/+/438256
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Bryan Mills <bcmills@google.com>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>

src/runtime/coverage/apis.go

@@ -9,6 +9,7 @@ import (
 	"internal/coverage"
 	"io"
 	"reflect"
+	"sync/atomic"
 	"unsafe"
 )
 
@@ -151,7 +152,7 @@ func ClearCoverageCounters() error {
 	// inconsistency when reading the counter array from the thread
 	// running ClearCoverageCounters.
 
-	var sd []uint32
+	var sd []atomic.Uint32
 
 	bufHdr := (*reflect.SliceHeader)(unsafe.Pointer(&sd))
 	for _, c := range cl {
@@ -160,13 +161,14 @@ func ClearCoverageCounters() error {
 		bufHdr.Cap = int(c.Len)
 		for i := 0; i < len(sd); i++ {
 			// Skip ahead until the next non-zero value.
-			if sd[i] == 0 {
+			sdi := sd[i].Load()
+			if sdi == 0 {
 				continue
 			}
 			// We found a function that was executed; clear its counters.
-			nCtrs := sd[i]
+			nCtrs := sdi
 			for j := 0; j < int(nCtrs); j++ {
-				sd[i+coverage.FirstCtrOffset+j] = 0
+				sd[i+coverage.FirstCtrOffset+j].Store(0)
 			}
 			// Move to next function.
 			i += coverage.FirstCtrOffset + int(nCtrs) - 1