runtime: break up large calls to memclrNoHeapPointers to allow preemp…

…tion

If something "huge" is allocated, and the zeroing is trivial (no pointers
involved) then zero it by chunks in a loop so that preemption can occur,
not all in a single non-preemptible call.

Benchmarking suggests that 256K is the best chunk size.

Updates #42642.

Change-Id: I94015e467eaa098c59870e479d6d83bc88efbfb4
Reviewed-on: https://go-review.googlesource.com/c/go/+/270943
Trust: David Chase <[email protected]>
Run-TryBot: David Chase <[email protected]>
TryBot-Result: Go Bot <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>

src/runtime/malloc.go

@@ -979,6 +979,9 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 	var span *mspan
 	var x unsafe.Pointer
 	noscan := typ == nil || typ.ptrdata == 0
+	// In some cases block zeroing can profitably (for latency reduction purposes)
+	// be delayed till preemption is possible; isZeroed tracks that state.
+	isZeroed := true
 	if size <= maxSmallSize {
 		if noscan && size < maxTinySize {
 			// Tiny allocator.
@@ -1074,7 +1077,9 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 		}
 	} else {
 		shouldhelpgc = true
-		span = c.allocLarge(size, needzero, noscan)
+		// For large allocations, keep track of zeroed state so that
+		// bulk zeroing can be happen later in a preemptible context.
+		span, isZeroed = c.allocLarge(size, needzero && !noscan, noscan)
 		span.freeindex = 1
 		span.allocCount = 1
 		x = unsafe.Pointer(span.base())
@@ -1133,6 +1138,12 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 	mp.mallocing = 0
 	releasem(mp)
 
+	// Pointerfree data can be zeroed late in a context where preemption can occur.
+	// x will keep the memory alive.
+	if !isZeroed && needzero {
+		memclrNoHeapPointersChunked(size, x)
+	}
+
 	if debug.malloc {
 		if debug.allocfreetrace != 0 {
 			tracealloc(x, size, typ)
@@ -1185,6 +1196,33 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 	return x
 }
 
+// memclrNoHeapPointersChunked repeatedly calls memclrNoHeapPointers
+// on chunks of the buffer to be zeroed, with opportunities for preemption
+// along the way.  memclrNoHeapPointers contains no safepoints and also
+// cannot be preemptively scheduled, so this provides a still-efficient
+// block copy that can also be preempted on a reasonable granularity.
+//
+// Use this with care; if the data being cleared is tagged to contain
+// pointers, this allows the GC to run before it is all cleared.
+func memclrNoHeapPointersChunked(size uintptr, x unsafe.Pointer) {
+	v := uintptr(x)
+	// got this from benchmarking. 128k is too small, 512k is too large.
+	const chunkBytes = 256 * 1024
+	vsize := v + size
+	for voff := v; voff < vsize; voff = voff + chunkBytes {
+		if getg().preempt {
+			// may hold locks, e.g., profiling
+			goschedguarded()
+		}
+		// clear min(avail, lump) bytes
+		n := vsize - voff
+		if n > chunkBytes {
+			n = chunkBytes
+		}
+		memclrNoHeapPointers(unsafe.Pointer(voff), n)
+	}
+}
+
 // implementation of new builtin
 // compiler (both frontend and SSA backend) knows the signature
 // of this function

src/runtime/mcache.go

@@ -206,7 +206,10 @@ func (c *mcache) refill(spc spanClass) {
 }
 
 // allocLarge allocates a span for a large object.
-func (c *mcache) allocLarge(size uintptr, needzero bool, noscan bool) *mspan {
+// The boolean result indicates whether the span is known-zeroed.
+// If it did not need to be zeroed, it may not have been zeroed;
+// but if it came directly from the OS, it is already zeroed.
+func (c *mcache) allocLarge(size uintptr, needzero bool, noscan bool) (*mspan, bool) {
 	if size+_PageSize < size {
 		throw("out of memory")
 	}
@@ -221,7 +224,7 @@ func (c *mcache) allocLarge(size uintptr, needzero bool, noscan bool) *mspan {
 	deductSweepCredit(npages*_PageSize, npages)
 
 	spc := makeSpanClass(0, noscan)
-	s := mheap_.alloc(npages, spc, needzero)
+	s, isZeroed := mheap_.alloc(npages, spc, needzero)
 	if s == nil {
 		throw("out of memory")
 	}
@@ -245,7 +248,7 @@ func (c *mcache) allocLarge(size uintptr, needzero bool, noscan bool) *mspan {
 	mheap_.central[spc].mcentral.fullSwept(mheap_.sweepgen).push(s)
 	s.limit = s.base() + size
 	heapBitsForAddr(s.base()).initSpan(s)
-	return s
+	return s, isZeroed
 }
 
 func (c *mcache) releaseAll() {

src/runtime/mcentral.go

@@ -238,7 +238,7 @@ func (c *mcentral) grow() *mspan {
 	npages := uintptr(class_to_allocnpages[c.spanclass.sizeclass()])
 	size := uintptr(class_to_size[c.spanclass.sizeclass()])
 
-	s := mheap_.alloc(npages, c.spanclass, true)
+	s, _ := mheap_.alloc(npages, c.spanclass, true)
 	if s == nil {
 		return nil
 	}

src/runtime/mheap.go

@@ -897,7 +897,8 @@ func (s spanAllocType) manual() bool {
 // spanclass indicates the span's size class and scannability.
 //
 // If needzero is true, the memory for the returned span will be zeroed.
-func (h *mheap) alloc(npages uintptr, spanclass spanClass, needzero bool) *mspan {
+// The boolean returned indicates whether the returned span is zeroed.
+func (h *mheap) alloc(npages uintptr, spanclass spanClass, needzero bool) (*mspan, bool) {
 	// Don't do any operations that lock the heap on the G stack.
 	// It might trigger stack growth, and the stack growth code needs
 	// to be able to allocate heap.
@@ -911,13 +912,15 @@ func (h *mheap) alloc(npages uintptr, spanclass spanClass, needzero bool) *mspan
 		s = h.allocSpan(npages, spanAllocHeap, spanclass)
 	})
 
+	isZeroed := s.needzero == 0
 	if s != nil {
 		if needzero && s.needzero != 0 {
 			memclrNoHeapPointers(unsafe.Pointer(s.base()), s.npages<<_PageShift)
+			isZeroed = true
 		}
 		s.needzero = 0
 	}
-	return s
+	return s, isZeroed
 }
 
 // allocManual allocates a manually-managed span of npage pages.