runtime: add bitmap-based markrootSpans implementation

Currently markrootSpans, the scanning routine which scans span specials
(particularly finalizers) as roots, uses sweepSpans to shard work and
find spans to mark.

However, as part of a future CL to change span ownership and how
mcentral works, we want to avoid having markrootSpans use the sweep bufs
to find specials, so in this change we introduce a new mechanism.

Much like for the page reclaimer, we set up a per-page bitmap where the
first page for a span is marked if the span contains any specials, and
unmarked if it has no specials. This bitmap is updated by addspecial,
removespecial, and during sweeping.

markrootSpans then shards this bitmap into mark work and markers iterate
over the bitmap looking for spans with specials to mark. Unlike the page
reclaimer, we don't need to use the pageInUse bits because having a
special implies that a span is in-use.

While in terms of computational complexity this design is technically
worse, because it needs to iterate over the mapped heap, in practice
this iteration is very fast (we can skip over large swathes of the heap
very quickly) and we only look at spans that have any specials at all,
rather than having to touch each span.

This new implementation of markrootSpans is behind a feature flag called
go115NewMarkrootSpans.

Updates #37487.

Change-Id: I8ea07b6c11059f6d412fe419e0ab512d989377b8
Reviewed-on: https://go-review.googlesource.com/c/go/+/221178
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>

1 files changed, 119 insertions, 14 deletions

diff --git a/src/runtime/mgcmark.go b/src/runtime/mgcmark.go
index 301d8020f1..ea73ccc1b1 100644
--- a/src/runtime/mgcmark.go
+++ b/src/runtime/mgcmark.go
@@ -21,10 +21,6 @@ const (
 	// BSS root.
 	rootBlockBytes = 256 << 10
 
-	// rootBlockSpans is the number of spans to scan per span
-	// root.
-	rootBlockSpans = 8 * 1024 // 64MB worth of spans
-
 	// maxObletBytes is the maximum bytes of an object to scan at
 	// once. Larger objects will be split up into "oblets" of at
 	// most this size. Since we can scan 1–2 MB/ms, 128 KB bounds
@@ -41,14 +37,26 @@ const (
 	// a syscall, so its overhead is nontrivial). Higher values
 	// make the system less responsive to incoming work.
 	drainCheckThreshold = 100000
+
+	// pagesPerSpanRoot indicates how many pages to scan from a span root
+	// at a time. Used by special root marking.
+	//
+	// Higher values improve throughput by increasing locality, but
+	// increase the minimum latency of a marking operation.
+	//
+	// Must be a multiple of the pageInUse bitmap element size and
+	// must also evenly divide pagesPerArena.
+	pagesPerSpanRoot = 512
+
+	// go115NewMarkrootSpans is a feature flag that indicates whether
+	// to use the new bitmap-based markrootSpans implementation.
+	go115NewMarkrootSpans = true
 )
 
 // gcMarkRootPrepare queues root scanning jobs (stacks, globals, and
 // some miscellany) and initializes scanning-related state.
 //
 // The world must be stopped.
-//
-//go:nowritebarrier
 func gcMarkRootPrepare() {
 	work.nFlushCacheRoots = 0
 
@@ -79,13 +87,24 @@ func gcMarkRootPrepare() {
 	//
 	// We depend on addfinalizer to mark objects that get
 	// finalizers after root marking.
-	//
-	// We're only interested in scanning the in-use spans,
-	// which will all be swept at this point. More spans
-	// may be added to this list during concurrent GC, but
-	// we only care about spans that were allocated before
-	// this mark phase.
-	work.nSpanRoots = mheap_.sweepSpans[mheap_.sweepgen/2%2].numBlocks()
+	if go115NewMarkrootSpans {
+		// We're going to scan the whole heap (that was available at the time the
+		// mark phase started, i.e. markArenas) for in-use spans which have specials.
+		//
+		// Break up the work into arenas, and further into chunks.
+		//
+		// Snapshot allArenas as markArenas. This snapshot is safe because allArenas
+		// is append-only.
+		mheap_.markArenas = mheap_.allArenas[:len(mheap_.allArenas):len(mheap_.allArenas)]
+		work.nSpanRoots = len(mheap_.markArenas) * (pagesPerArena / pagesPerSpanRoot)
+	} else {
+		// We're only interested in scanning the in-use spans,
+		// which will all be swept at this point. More spans
+		// may be added to this list during concurrent GC, but
+		// we only care about spans that were allocated before
+		// this mark phase.
+		work.nSpanRoots = mheap_.sweepSpans[mheap_.sweepgen/2%2].numBlocks()
+	}
 
 	// Scan stacks.
 	//
@@ -293,10 +312,96 @@ func markrootFreeGStacks() {
 	unlock(&sched.gFree.lock)
 }
 
-// markrootSpans marks roots for one shard of work.spans.
+// markrootSpans marks roots for one shard of markArenas.
 //
 //go:nowritebarrier
 func markrootSpans(gcw *gcWork, shard int) {
+	if !go115NewMarkrootSpans {
+		oldMarkrootSpans(gcw, shard)
+		return
+	}
+	// Objects with finalizers have two GC-related invariants:
+	//
+	// 1) Everything reachable from the object must be marked.
+	// This ensures that when we pass the object to its finalizer,
+	// everything the finalizer can reach will be retained.
+	//
+	// 2) Finalizer specials (which are not in the garbage
+	// collected heap) are roots. In practice, this means the fn
+	// field must be scanned.
+	sg := mheap_.sweepgen
+
+	// Find the arena and page index into that arena for this shard.
+	ai := mheap_.markArenas[shard/(pagesPerArena/pagesPerSpanRoot)]
+	ha := mheap_.arenas[ai.l1()][ai.l2()]
+	arenaPage := uint(uintptr(shard) * pagesPerSpanRoot % pagesPerArena)
+
+	// Construct slice of bitmap which we'll iterate over.
+	specialsbits := ha.pageSpecials[arenaPage/8:]
+	specialsbits = specialsbits[:pagesPerSpanRoot/8]
+	for i := range specialsbits {
+		// Find set bits, which correspond to spans with specials.
+		specials := atomic.Load8(&specialsbits[i])
+		if specials == 0 {
+			continue
+		}
+		for j := uint(0); j < 8; j++ {
+			if specials&(1<<j) == 0 {
+				continue
+			}
+			// Find the span for this bit.
+			//
+			// This value is guaranteed to be non-nil because having
+			// specials implies that the span is in-use, and since we're
+			// currently marking we can be sure that we don't have to worry
+			// about the span being freed and re-used.
+			s := ha.spans[arenaPage+uint(i)*8+j]
+
+			// The state must be mSpanInUse if the specials bit is set, so
+			// sanity check that.
+			if state := s.state.get(); state != mSpanInUse {
+				print("s.state = ", state, "\n")
+				throw("non in-use span found with specials bit set")
+			}
+			// Check that this span was swept (it may be cached or uncached).
+			if !useCheckmark && !(s.sweepgen == sg || s.sweepgen == sg+3) {
+				// sweepgen was updated (+2) during non-checkmark GC pass
+				print("sweep ", s.sweepgen, " ", sg, "\n")
+				throw("gc: unswept span")
+			}
+
+			// Lock the specials to prevent a special from being
+			// removed from the list while we're traversing it.
+			lock(&s.speciallock)
+			for sp := s.specials; sp != nil; sp = sp.next {
+				if sp.kind != _KindSpecialFinalizer {
+					continue
+				}
+				// don't mark finalized object, but scan it so we
+				// retain everything it points to.
+				spf := (*specialfinalizer)(unsafe.Pointer(sp))
+				// A finalizer can be set for an inner byte of an object, find object beginning.
+				p := s.base() + uintptr(spf.special.offset)/s.elemsize*s.elemsize
+
+				// Mark everything that can be reached from
+				// the object (but *not* the object itself or
+				// we'll never collect it).
+				scanobject(p, gcw)
+
+				// The special itself is a root.
+				scanblock(uintptr(unsafe.Pointer(&spf.fn)), sys.PtrSize, &oneptrmask[0], gcw, nil)
+			}
+			unlock(&s.speciallock)
+		}
+	}
+}
+
+// oldMarkrootSpans marks roots for one shard of work.spans.
+//
+// For go115NewMarkrootSpans = false.
+//
+//go:nowritebarrier
+func oldMarkrootSpans(gcw *gcWork, shard int) {
 	// Objects with finalizers have two GC-related invariants:
 	//
 	// 1) Everything reachable from the object must be marked.