diff --git a/src/pkg/runtime/malloc.goc b/src/pkg/runtime/malloc.goc
index 0470211506..0e8a812641 100644
--- a/src/pkg/runtime/malloc.goc
+++ b/src/pkg/runtime/malloc.goc
@@ -329,6 +329,7 @@ runtime·free(void *v)
 		// it might coalesce v and other blocks into a bigger span
 		// and change the bitmap further.
 		c->local_nsmallfree[sizeclass]++;
+		c->local_cachealloc -= size;
 		if(c->alloc[sizeclass] == s) {
 			// We own the span, so we can just add v to the freelist
 			runtime·markfreed(v);
diff --git a/src/pkg/runtime/mcache.c b/src/pkg/runtime/mcache.c
index 0b4bbd90be..26e3db2dca 100644
--- a/src/pkg/runtime/mcache.c
+++ b/src/pkg/runtime/mcache.c
@@ -97,7 +97,6 @@ runtime·MCache_Free(MCache *c, MLink *p, int32 sizeclass, uintptr size)
 	p->next = l->list;
 	l->list = p;
 	l->nlist++;
-	c->local_cachealloc -= size;
 
 	// We transfer a span at a time from MCentral to MCache,
 	// so we'll do the same in the other direction.
diff --git a/src/pkg/runtime/mgc0.c b/src/pkg/runtime/mgc0.c
index aa3eddbccd..400149c26d 100644
--- a/src/pkg/runtime/mgc0.c
+++ b/src/pkg/runtime/mgc0.c
@@ -2346,8 +2346,12 @@ gc(struct gc_args *args)
 		runtime·printf("pause %D\n", t4-t0);
 
 	if(runtime·debug.gctrace) {
-		updatememstats(&stats);
 		heap1 = mstats.heap_alloc;
+		updatememstats(&stats);
+		if(heap1 != mstats.heap_alloc) {
+			runtime·printf("runtime: mstats skew: heap=%p/%p\n", heap1, mstats.heap_alloc);
+			runtime·throw("mstats skew");
+		}
 		obj = mstats.nmalloc - mstats.nfree;
 
 		stats.nprocyield += work.markfor->nprocyield;