runtime: block sweep completion on all sweep paths

The runtime currently has two different notions of sweep completion:

1. All spans are either swept or have begun sweeping.

2. The sweeper has *finished* sweeping all spans.

Most things depend on condition 1. Notably, GC correctness depends on
condition 1, but since all sweep operations a non-preemptible, the STW
at the beginning of GC forces condition 1 to become condition 2.

runtime.GC(), however, depends on condition 2, since the intent is to
complete a complete GC cycle, and also update the heap profile (which
can only be done after sweeping is complete).

However, the way we compute condition 2 is racy right now and may in
fact only indicate condition 1. Specifically, sweepone blocks
condition 2 until all sweepone calls are done, but there are many
other ways to enter the sweeper that don't block this. Hence, sweepone
may see that there are no more spans in the sweep list and see that
it's the last sweepone and declare sweeping done, while there's some
other sweeper still working on a span.

Fix this by making sure every entry to the sweeper participates in the
protocol that blocks condition 2. To make sure we get this right, this
CL introduces a type to track sweep blocking and (lightly) enforces
span sweep ownership via the type system. This has the nice
side-effect of abstracting the pattern of acquiring sweep ownership
that's currently repeated in many different places.

Fixes #45315.

Change-Id: I7fab30170c5ae14c8b2f10998628735b8be6d901
Reviewed-on: https://go-review.googlesource.com/c/go/+/307915
Trust: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>

1 files changed, 10 insertions, 5 deletions

diff --git a/src/runtime/mgc.go b/src/runtime/mgc.go
index ef3436d1f4..8c1ff20936 100644
--- a/src/runtime/mgc.go
+++ b/src/runtime/mgc.go
@@ -1749,6 +1749,13 @@ func gcMarkTermination(nextTriggerRatio float64) {
 	// so events don't leak into the wrong cycle.
 	mProf_NextCycle()
 
+	// There may be stale spans in mcaches that need to be swept.
+	// Those aren't tracked in any sweep lists, so we need to
+	// count them against sweep completion until we ensure all
+	// those spans have been forced out.
+	sl := newSweepLocker()
+	sl.blockCompletion()
+
 	systemstack(func() { startTheWorldWithSema(true) })
 
 	// Flush the heap profile so we can start a new cycle next GC.
@@ -1772,6 +1779,9 @@ func gcMarkTermination(nextTriggerRatio float64) {
 			_p_.mcache.prepareForSweep()
 		})
 	})
+	// Now that we've swept stale spans in mcaches, they don't
+	// count against unswept spans.
+	sl.dispose()
 
 	// Print gctrace before dropping worldsema. As soon as we drop
 	// worldsema another cycle could start and smash the stats
@@ -2391,11 +2401,6 @@ func gcTestIsReachable(ptrs ...unsafe.Pointer) (mask uint64) {
 	// Force a full GC and sweep.
 	GC()
 
-	// TODO(austin): Work around issue #45315. One GC() can return
-	// without finishing the sweep. Do a second to force the sweep
-	// through.
-	GC()
-
 	// Process specials.
 	for i, s := range specials {
 		if !s.done {