runtime: add per-p mspan cache

This change adds a per-p mspan object cache similar to the sudog cache.
Unfortunately this cache can't quite operate like the sudog cache, since
it is used in contexts where write barriers are disallowed (i.e.
allocation codepaths), so rather than managing an array and a slice,
it's just an array and a length. A little bit more unsafe, but avoids
any write barriers.

The purpose of this change is to reduce the number of operations which
require the heap lock in allocation, paving the way for a lockless fast
path.

Updates #35112.

Change-Id: I32cfdcd8528fb7be985640e4f3a13cb98ffb7865
Reviewed-on: https://go-review.googlesource.com/c/go/+/196642
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>

src/runtime/mheap.go

@@ -973,6 +973,84 @@ func (h *mheap) allocNeedsZero(base, npage uintptr) (needZero bool) {
 	return
 }
 
+// tryAllocMSpan attempts to allocate an mspan object from
+// the P-local cache, but may fail.
+//
+// h need not be locked.
+//
+// This caller must ensure that its P won't change underneath
+// it during this function. Currently to ensure that we enforce
+// that the function is run on the system stack, because that's
+// the only place it is used now. In the future, this requirement
+// may be relaxed if its use is necessary elsewhere.
+//
+//go:systemstack
+func (h *mheap) tryAllocMSpan() *mspan {
+	pp := getg().m.p.ptr()
+	// If we don't have a p or the cache is empty, we can't do
+	// anything here.
+	if pp == nil || pp.mspancache.len == 0 {
+		return nil
+	}
+	// Pull off the last entry in the cache.
+	s := pp.mspancache.buf[pp.mspancache.len-1]
+	pp.mspancache.len--
+	return s
+}
+
+// allocMSpanLocked allocates an mspan object.
+//
+// h must be locked.
+//
+// allocMSpanLocked must be called on the system stack because
+// its caller holds the heap lock. See mheap for details.
+// Running on the system stack also ensures that we won't
+// switch Ps during this function. See tryAllocMSpan for details.
+//
+//go:systemstack
+func (h *mheap) allocMSpanLocked() *mspan {
+	pp := getg().m.p.ptr()
+	if pp == nil {
+		// We don't have a p so just do the normal thing.
+		return (*mspan)(h.spanalloc.alloc())
+	}
+	// Refill the cache if necessary.
+	if pp.mspancache.len == 0 {
+		const refillCount = len(pp.mspancache.buf) / 2
+		for i := 0; i < refillCount; i++ {
+			pp.mspancache.buf[i] = (*mspan)(h.spanalloc.alloc())
+		}
+		pp.mspancache.len = refillCount
+	}
+	// Pull off the last entry in the cache.
+	s := pp.mspancache.buf[pp.mspancache.len-1]
+	pp.mspancache.len--
+	return s
+}
+
+// freeMSpanLocked free an mspan object.
+//
+// h must be locked.
+//
+// freeMSpanLocked must be called on the system stack because
+// its caller holds the heap lock. See mheap for details.
+// Running on the system stack also ensures that we won't
+// switch Ps during this function. See tryAllocMSpan for details.
+//
+//go:systemstack
+func (h *mheap) freeMSpanLocked(s *mspan) {
+	pp := getg().m.p.ptr()
+	// First try to free the mspan directly to the cache.
+	if pp != nil && pp.mspancache.len < len(pp.mspancache.buf) {
+		pp.mspancache.buf[pp.mspancache.len] = s
+		pp.mspancache.len++
+		return
+	}
+	// Failing that (or if we don't have a p), just free it to
+	// the heap.
+	h.spanalloc.free(unsafe.Pointer(s))
+}
+
 // allocSpan allocates an mspan which owns npages worth of memory.
 //
 // If manual == false, allocSpan allocates a heap span of class spanclass
@@ -995,6 +1073,9 @@ func (h *mheap) allocSpan(npages uintptr, manual bool, spanclass spanClass, sysS
 	gp := getg()
 	base, scav := uintptr(0), uintptr(0)
 
+	// Try to allocate a cached span.
+	s = h.tryAllocMSpan()
+
 	// We failed to do what we need to do without the lock.
 	lock(&h.lock)
 
@@ -1014,6 +1095,11 @@ func (h *mheap) allocSpan(npages uintptr, manual bool, spanclass spanClass, sysS
 	throw("grew heap, but no adequate free space found")
 
 HaveBase:
+	if s == nil {
+		// We failed to get an mspan earlier, so grab
+		// one now that we have the heap lock.
+		s = h.allocMSpanLocked()
+	}
 	if !manual {
 		// This is a heap span, so we should do some additional accounting
 		// which may only be done with the heap locked.
@@ -1036,9 +1122,6 @@ HaveBase:
 			gcController.revise()
 		}
 	}
-
-	// Allocate an mspan object before releasing the lock.
-	s = (*mspan)(h.spanalloc.alloc())
 	unlock(&h.lock)
 
 	// Initialize the span.
@@ -1294,7 +1377,7 @@ func (h *mheap) freeSpanLocked(s *mspan, acctinuse, acctidle bool) {
 
 	// Free the span structure. We no longer have a use for it.
 	s.state.set(mSpanDead)
-	h.spanalloc.free(unsafe.Pointer(s))
+	h.freeMSpanLocked(s)
 }
 
 // scavengeAll visits each node in the free treap and scavenges the

src/runtime/proc.go

@@ -4082,6 +4082,13 @@ func (pp *p) destroy() {
 		}
 		pp.deferpool[i] = pp.deferpoolbuf[i][:0]
 	}
+	systemstack(func() {
+		for i := 0; i < pp.mspancache.len; i++ {
+			// Safe to call since the world is stopped.
+			mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
+		}
+		pp.mspancache.len = 0
+	})
 	freemcache(pp.mcache)
 	pp.mcache = nil
 	gfpurge(pp)

src/runtime/runtime2.go

@@ -588,6 +588,17 @@ type p struct {
 	sudogcache []*sudog
 	sudogbuf   [128]*sudog
 
+	// Cache of mspan objects from the heap.
+	mspancache struct {
+		// We need an explicit length here because this field is used
+		// in allocation codepaths where write barriers are not allowed,
+		// and eliminating the write barrier/keeping it eliminated from
+		// slice updates is tricky, moreso than just managing the length
+		// ourselves.
+		len int
+		buf [128]*mspan
+	}
+
 	tracebuf traceBufPtr
 
 	// traceSweep indicates the sweep events should be traced.
@@ -600,7 +611,7 @@ type p struct {
 
 	palloc persistentAlloc // per-P to avoid mutex
 
-	_ uint32 // Alignment for atomic fields below
+	// _ uint32 // Alignment for atomic fields below
 
 	// Per-P GC state
 	gcAssistTime         int64    // Nanoseconds in assistAlloc