runtime: don't acquirem on malloc fast path

Mallocgc must be atomic wrt GC, but for performance reasons
don't acquirem/releasem on fast path. The code does not have
split stack checks, so it can't be preempted by GC.
Functions like roundup/add are inlined. And onM/racemalloc are nosplit.
Also add debug code that checks these assumptions.

benchmark                     old ns/op     new ns/op     delta
BenchmarkMalloc8              20.5          17.2          -16.10%
BenchmarkMalloc16             29.5          27.0          -8.47%
BenchmarkMallocTypeInfo8      31.5          27.6          -12.38%
BenchmarkMallocTypeInfo16     34.7          30.9          -10.95%

LGTM=khr
R=golang-codereviews, khr
CC=golang-codereviews, rlh, rsc
https://golang.org/cl/123100043

src/cmd/api/goapi.go

@@ -378,7 +378,10 @@ func (w *Walker) parseFile(dir, file string) (*ast.File, error) {
 	}
 	if w.context != nil && file == fmt.Sprintf("zruntime_defs_%s_%s.go", w.context.GOOS, w.context.GOARCH) {
 		// Just enough to keep the api checker happy.
-		src := "package runtime; type maptype struct{}; type _type struct{}; type alg struct{}; type mspan struct{}; type m struct{}; type lock struct{}; type slicetype struct{}; type iface struct{}; type eface struct{}; type interfacetype struct{}; type itab struct{}"
+		src := "package runtime; type maptype struct{}; type _type struct{}; type alg struct{};" +
+			" type mspan struct{}; type m struct{}; type lock struct{}; type slicetype struct{};" +
+			" type iface struct{}; type eface struct{}; type interfacetype struct{}; type itab struct{};" +
+			" type mcache struct{}"
 		f, err = parser.ParseFile(fset, filename, src, 0)
 		if err != nil {
 			log.Fatalf("incorrect generated file: %s", err)

src/pkg/runtime/malloc.go

@@ -59,14 +59,25 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 	if size == 0 {
 		return unsafe.Pointer(&zeroObject)
 	}
-	mp := acquirem()
-	if mp.mallocing != 0 {
-		gothrow("malloc/free - deadlock")
-	}
-	mp.mallocing = 1
 	size0 := size
 
-	c := mp.mcache
+	// This function must be atomic wrt GC, but for performance reasons
+	// we don't acquirem/releasem on fast path. The code below does not have
+	// split stack checks, so it can't be preempted by GC.
+	// Functions like roundup/add are inlined. And onM/racemalloc are nosplit.
+	// If debugMalloc = true, these assumptions are checked below.
+	if debugMalloc {
+		mp := acquirem()
+		if mp.mallocing != 0 {
+			gothrow("malloc deadlock")
+		}
+		mp.mallocing = 1
+		if mp.curg != nil {
+			mp.curg.stackguard0 = ^uint(0xfff) | 0xbad
+		}
+	}
+
+	c := gomcache()
 	var s *mspan
 	var x unsafe.Pointer
 	if size <= maxSmallSize {
@@ -118,8 +129,18 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 					x = tiny
 					c.tiny = (*byte)(add(x, size))
 					c.tinysize -= uint(size1)
-					mp.mallocing = 0
-					releasem(mp)
+					if debugMalloc {
+						mp := acquirem()
+						if mp.mallocing == 0 {
+							gothrow("bad malloc")
+						}
+						mp.mallocing = 0
+						if mp.curg != nil {
+							mp.curg.stackguard0 = mp.curg.stackguard
+						}
+						releasem(mp)
+						releasem(mp)
+					}
 					return x
 				}
 			}
@@ -127,8 +148,10 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 			s = c.alloc[tinySizeClass]
 			v := s.freelist
 			if v == nil {
+				mp := acquirem()
 				mp.scalararg[0] = tinySizeClass
 				onM(&mcacheRefill_m)
+				releasem(mp)
 				s = c.alloc[tinySizeClass]
 				v = s.freelist
 			}
@@ -156,8 +179,10 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 			s = c.alloc[sizeclass]
 			v := s.freelist
 			if v == nil {
+				mp := acquirem()
 				mp.scalararg[0] = uint(sizeclass)
 				onM(&mcacheRefill_m)
+				releasem(mp)
 				s = c.alloc[sizeclass]
 				v = s.freelist
 			}
@@ -174,11 +199,13 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 		}
 		c.local_cachealloc += int(size)
 	} else {
+		mp := acquirem()
 		mp.scalararg[0] = uint(size)
 		mp.scalararg[1] = uint(flags)
 		onM(&largeAlloc_m)
 		s = (*mspan)(mp.ptrarg[0])
 		mp.ptrarg[0] = nil
+		releasem(mp)
 		x = unsafe.Pointer(uintptr(s.start << pageShift))
 		size = uintptr(s.elemsize)
 	}
@@ -221,18 +248,22 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 					// into the GC bitmap. It's 7 times slower than copying
 					// from the pre-unrolled mask, but saves 1/16 of type size
 					// memory for the mask.
+					mp := acquirem()
 					mp.ptrarg[0] = x
 					mp.ptrarg[1] = unsafe.Pointer(typ)
 					mp.scalararg[0] = uint(size)
 					mp.scalararg[1] = uint(size0)
 					onM(&unrollgcproginplace_m)
+					releasem(mp)
 					goto marked
 				}
 				ptrmask = (*uint8)(unsafe.Pointer(uintptr(typ.gc[0])))
 				// Check whether the program is already unrolled.
 				if uintptr(goatomicloadp(unsafe.Pointer(ptrmask)))&0xff == 0 {
+					mp := acquirem()
 					mp.ptrarg[0] = unsafe.Pointer(typ)
 					onM(&unrollgcprog_m)
+					releasem(mp)
 				}
 				ptrmask = (*uint8)(add(unsafe.Pointer(ptrmask), 1)) // skip the unroll flag byte
 			} else {
@@ -287,11 +318,23 @@ func gomallocgc(size uintptr, typ *_type, flags int) unsafe.Pointer {
 		}
 	}
 marked:
-	mp.mallocing = 0
-
 	if raceenabled {
 		racemalloc(x, size)
 	}
+
+	if debugMalloc {
+		mp := acquirem()
+		if mp.mallocing == 0 {
+			gothrow("bad malloc")
+		}
+		mp.mallocing = 0
+		if mp.curg != nil {
+			mp.curg.stackguard0 = mp.curg.stackguard
+		}
+		releasem(mp)
+		releasem(mp)
+	}
+
 	if debug.allocfreetrace != 0 {
 		tracealloc(x, size, typ)
 	}
@@ -300,12 +343,12 @@ marked:
 		if size < uintptr(rate) && int32(size) < c.next_sample {
 			c.next_sample -= int32(size)
 		} else {
+			mp := acquirem()
 			profilealloc(mp, x, size)
+			releasem(mp)
 		}
 	}
 
-	releasem(mp)
-
 	if memstats.heap_alloc >= memstats.next_gc {
 		gogc(0)
 	}

src/pkg/runtime/race.c

@@ -11,6 +11,7 @@
 #include "race.h"
 #include "type.h"
 #include "typekind.h"
+#include "../../cmd/ld/textflag.h"
 
 // Race runtime functions called via runtime·racecall.
 void __tsan_init(void);
@@ -106,6 +107,7 @@ runtime·racemapshadow(void *addr, uintptr size)
 	runtime·racecall(__tsan_map_shadow, addr, size);
 }
 
+#pragma textflag NOSPLIT
 void
 runtime·racemalloc(void *p, uintptr sz)
 {

src/pkg/runtime/stubs.go

@@ -43,6 +43,7 @@ func roundup(p unsafe.Pointer, n uintptr) unsafe.Pointer {
 // in stubs.goc
 func acquirem() *m
 func releasem(mp *m)
+func gomcache() *mcache
 
 // An mFunction represents a C function that runs on the M stack.  It
 // can be called from Go using mcall or onM.  Through the magic of

src/pkg/runtime/stubs.goc

@@ -83,6 +83,11 @@ func runtime·releasem(mp *M) {
 	}
 }
 
+#pragma textflag NOSPLIT
+func runtime·gomcache() (ret *MCache) {
+	ret = g->m->mcache;
+}
+
 // For testing.
 // TODO: find a better place for this.
 func GCMask(x Eface) (mask Slice) {