runtime/cgo: store M for C-created thread in pthread key

This reapplies CL 481061, with the followup fixes in CL 482975, CL 485315, and CL 485316 incorporated. CL 481061, by doujiang24 <doujiang24@gmail.com>, speed up C to Go calls by binding the M to the C thread. See below for its description. CL 482975 is a followup fix to a C declaration in testprogcgo. CL 485315 is a followup fix for x_cgo_getstackbound on Illumos. CL 485316 is a followup cleanup for ppc64 assembly. [Original CL 481061 description] This reapplies CL 392854, with the followup fixes in CL 479255, CL 479915, and CL 481057 incorporated. CL 392854, by doujiang24 <doujiang24@gmail.com>, speed up C to Go calls by binding the M to the C thread. See below for its description. CL 479255 is a followup fix for a small bug in ARM assembly code. CL 479915 is another followup fix to address C to Go calls after the C code uses some stack, but that CL is also buggy. CL 481057, by Michael Knyszek, is a followup fix for a memory leak bug of CL 479915. [Original CL 392854 description] In a C thread, it's necessary to acquire an extra M by using needm while invoking a Go function from C. But, needm and dropm are heavy costs due to the signal-related syscalls. So, we change to not dropm while returning back to C, which means binding the extra M to the C thread until it exits, to avoid needm and dropm on each C to Go call. Instead, we only dropm while the C thread exits, so the extra M won't leak. When invoking a Go function from C: Allocate a pthread variable using pthread_key_create, only once per shared object, and register a thread-exit-time destructor. And store the g0 of the current m into the thread-specified value of the pthread key, only once per C thread, so that the destructor will put the extra M back onto the extra M list while the C thread exits. When returning back to C: Skip dropm in cgocallback, when the pthread variable has been created, so that the extra M will be reused the next time invoke a Go function from C. This is purely a performance optimization. The old version, in which needm & dropm happen on each cgo call, is still correct too, and we have to keep the old version on systems with cgo but without pthreads, like Windows. This optimization is significant, and the specific value depends on the OS system and CPU, but in general, it can be considered as 10x faster, for a simple Go function call from a C thread. For the newly added BenchmarkCGoInCThread, some benchmark results: 1. it's 28x faster, from 3395 ns/op to 121 ns/op, in darwin OS & Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz 2. it's 6.5x faster, from 1495 ns/op to 230 ns/op, in Linux OS & Intel(R) Xeon(R) CPU E5-2630 0 @ 2.30GHz [CL 479915 description] Currently, when C calls into Go the first time, we grab an M using needm, which sets m.g0's stack bounds using the SP. We don't know how big the stack is, so we simply assume 32K. Previously, when the Go function returns to C, we drop the M, and the next time C calls into Go, we put a new stack bound on the g0 based on the current SP. After CL 392854, we don't drop the M, and the next time C calls into Go, we reuse the same g0, without recomputing the stack bounds. If the C code uses quite a bit of stack space before calling into Go, the SP may be well below the 32K stack bound we assumed, so the runtime thinks the g0 stack overflows. This CL makes needm get a more accurate stack bound from pthread. (In some platforms this may still be a guess as we don't know exactly where we are in the C stack), but it is probably better than simply assuming 32K. [CL 485500 description] CL 479915 passed the G to _cgo_getstackbound for direct updates to gp.stack.lo. A G can be reused on a new thread after the previous thread exited. This could trigger the C TSAN race detector because it couldn't see the synchronization in Go (lockextra) preventing the same G from being used on multiple threads at the same time. We work around this by passing the address of a stack variable to _cgo_getstackbound rather than the G. The stack is generally unique per thread, so TSAN won't see the same address from multiple threads. Even if stacks are reused across threads by pthread, C TSAN should see the synchonization in the stack allocator. A regression test is added to misc/cgo/testsanitizer. Fixes golang#51676. Fixes golang#59294. Fixes golang#59678. Change-Id: Ic62be31a06ee83568215e875a891df37084e08ca Reviewed-on: https://go-review.googlesource.com/c/go/+/485500 TryBot-Result: Gopher Robot <gobot@golang.org> Reviewed-by: Cherry Mui <cherryyz@google.com> Run-TryBot: Michael Pratt <mpratt@google.com>
aytanet · Apr 26, 2023 · 7b87461 · 7b87461
1 parent d816f85
commit 7b87461
Showing 46 changed files with 1,012 additions and 67 deletions.
diff --git a/misc/cgo/test/cgo_test.go b/misc/cgo/test/cgo_test.go
@@ -104,6 +104,7 @@ func TestThreadLock(t *testing.T)            { testThreadLockFunc(t) }
 func TestUnsignedInt(t *testing.T)           { testUnsignedInt(t) }
 func TestZeroArgCallback(t *testing.T)       { testZeroArgCallback(t) }
 
-func BenchmarkCgoCall(b *testing.B)     { benchCgoCall(b) }
-func BenchmarkGoString(b *testing.B)    { benchGoString(b) }
-func BenchmarkCGoCallback(b *testing.B) { benchCallback(b) }
+func BenchmarkCgoCall(b *testing.B)      { benchCgoCall(b) }
+func BenchmarkGoString(b *testing.B)     { benchGoString(b) }
+func BenchmarkCGoCallback(b *testing.B)  { benchCallback(b) }
+func BenchmarkCGoInCThread(b *testing.B) { benchCGoInCthread(b) }
diff --git a/misc/cgo/test/cthread_unix.c b/misc/cgo/test/cthread_unix.c
@@ -32,3 +32,27 @@ doAdd(int max, int nthread)
 	for(i=0; i<nthread; i++)
 		pthread_join(thread_id[i], 0);		
 }
+
+static void*
+goDummyCallbackThread(void* p)
+{
+	int i, max;
+
+	max = *(int*)p;
+	for(i=0; i<max; i++)
+		goDummy();
+	return NULL;
+}
+
+int
+callGoInCThread(int max)
+{
+	pthread_t thread;
+
+	if (pthread_create(&thread, NULL, goDummyCallbackThread, (void*)(&max)) != 0)
+		return -1;
+	if (pthread_join(thread, NULL) != 0)
+		return -1;
+
+	return max;
+}
diff --git a/misc/cgo/test/cthread_windows.c b/misc/cgo/test/cthread_windows.c
@@ -35,3 +35,25 @@ doAdd(int max, int nthread)
 		CloseHandle((HANDLE)thread_id[i]);
 	}
 }
+
+__stdcall
+static unsigned int
+goDummyCallbackThread(void* p)
+{
+	int i, max;
+
+	max = *(int*)p;
+	for(i=0; i<max; i++)
+		goDummy();
+	return 0;
+}
+
+int
+callGoInCThread(int max)
+{
+	uintptr_t thread_id;
+	thread_id = _beginthreadex(0, 0, goDummyCallbackThread, &max, 0, 0);
+	WaitForSingleObject((HANDLE)thread_id, INFINITE);
+	CloseHandle((HANDLE)thread_id);
+	return max;
+}
diff --git a/misc/cgo/test/testx.go b/misc/cgo/test/testx.go
@@ -24,6 +24,7 @@ import (
 /*
 // threads
 extern void doAdd(int, int);
+extern int callGoInCThread(int);
 
 // issue 1328
 void IntoC(void);
@@ -146,6 +147,10 @@ func Add(x int) {
 	*p = 2
 }
 
+//export goDummy
+func goDummy() {
+}
+
 func testCthread(t *testing.T) {
 	if (runtime.GOOS == "darwin" || runtime.GOOS == "ios") && runtime.GOARCH == "arm64" {
 		t.Skip("the iOS exec wrapper is unable to properly handle the panic from Add")
@@ -159,6 +164,15 @@ func testCthread(t *testing.T) {
 	}
 }
 
+// Benchmark measuring overhead from C to Go in a C thread.
+// Create a new C thread and invoke Go function repeatedly in the new C thread.
+func benchCGoInCthread(b *testing.B) {
+	n := C.callGoInCThread(C.int(b.N))
+	if int(n) != b.N {
+		b.Fatal("unmatch loop times")
+	}
+}
+
 // issue 1328
 
 //export BackIntoGo

diff --git a/misc/cgo/testcarchive/carchive_test.go b/misc/cgo/testcarchive/carchive_test.go
@@ -1247,3 +1247,57 @@ func TestPreemption(t *testing.T) {
 		t.Error(err)
 	}
 }
+
+// Issue 59294. Test calling Go function from C after using some
+// stack space.
+func TestDeepStack(t *testing.T) {
+	t.Parallel()
+
+	if !testWork {
+		defer func() {
+			os.Remove("testp9" + exeSuffix)
+			os.Remove("libgo9.a")
+			os.Remove("libgo9.h")
+		}()
+	}
+
+	cmd := exec.Command("go", "build", "-buildmode=c-archive", "-o", "libgo9.a", "./libgo9")
+	out, err := cmd.CombinedOutput()
+	t.Logf("%v\n%s", cmd.Args, out)
+	if err != nil {
+		t.Fatal(err)
+	}
+	checkLineComments(t, "libgo9.h")
+	checkArchive(t, "libgo9.a")
+
+	// build with -O0 so the C compiler won't optimize out the large stack frame
+	ccArgs := append(cc, "-O0", "-o", "testp9"+exeSuffix, "main9.c", "libgo9.a")
+	out, err = exec.Command(ccArgs[0], ccArgs[1:]...).CombinedOutput()
+	t.Logf("%v\n%s", ccArgs, out)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	argv := cmdToRun("./testp9")
+	cmd = exec.Command(argv[0], argv[1:]...)
+	sb := new(strings.Builder)
+	cmd.Stdout = sb
+	cmd.Stderr = sb
+	if err := cmd.Start(); err != nil {
+		t.Fatal(err)
+	}
+
+	timer := time.AfterFunc(time.Minute,
+		func() {
+			t.Error("test program timed out")
+			cmd.Process.Kill()
+		},
+	)
+	defer timer.Stop()
+
+	err = cmd.Wait()
+	t.Logf("%v\n%s", cmd.Args, sb)
+	if err != nil {
+		t.Error(err)
+	}
+}
diff --git a/misc/cgo/testcarchive/testdata/libgo9/a.go b/misc/cgo/testcarchive/testdata/libgo9/a.go
@@ -0,0 +1,14 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+import "runtime"
+
+import "C"
+
+func main() {}
+
+//export GoF
+func GoF() { runtime.GC() }
diff --git a/misc/cgo/testcarchive/testdata/main9.c b/misc/cgo/testcarchive/testdata/main9.c
@@ -0,0 +1,24 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "libgo9.h"
+
+void use(int *x) { (*x)++; }
+
+void callGoFWithDeepStack() {
+	int x[10000];
+
+	use(&x[0]);
+	use(&x[9999]);
+
+	GoF();
+
+	use(&x[0]);
+	use(&x[9999]);
+}
+
+int main() {
+	GoF();                  // call GoF without using much stack
+	callGoFWithDeepStack(); // call GoF with a deep stack
+}
diff --git a/misc/cgo/testsanitizers/testdata/tsan14.go b/misc/cgo/testsanitizers/testdata/tsan14.go
@@ -0,0 +1,53 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package main
+
+// This program failed when run under the C/C++ ThreadSanitizer.
+//
+// cgocallback on a new thread calls into runtime.needm -> _cgo_getstackbound
+// to update gp.stack.lo with the stack bounds. If the G itself is passed to
+// _cgo_getstackbound, then writes to the same G can be seen on multiple
+// threads (when the G is reused after thread exit). This would trigger TSAN.
+
+/*
+#include <pthread.h>
+
+void go_callback();
+
+static void *thr(void *arg) {
+    go_callback();
+    return 0;
+}
+
+static void foo() {
+    pthread_t th;
+    pthread_attr_t attr;
+    pthread_attr_init(&attr);
+    pthread_attr_setstacksize(&attr, 256 << 10);
+    pthread_create(&th, &attr, thr, 0);
+    pthread_join(th, 0);
+}
+*/
+import "C"
+
+import (
+	"time"
+)
+
+//export go_callback
+func go_callback() {
+}
+
+func main() {
+	for i := 0; i < 2; i++ {
+		go func() {
+			for {
+				C.foo()
+			}
+		}()
+	}
+
+	time.Sleep(1000*time.Millisecond)
+}
diff --git a/misc/cgo/testsanitizers/tsan_test.go b/misc/cgo/testsanitizers/tsan_test.go
@@ -47,6 +47,7 @@ func TestTSAN(t *testing.T) {
 		{src: "tsan11.go", needsRuntime: true},
 		{src: "tsan12.go", needsRuntime: true},
 		{src: "tsan13.go", needsRuntime: true},
+		{src: "tsan14.go", needsRuntime: true},
 	}
 	for _, tc := range cases {
 		tc := tc

diff --git a/src/runtime/asm_386.s b/src/runtime/asm_386.s
@@ -689,7 +689,20 @@ nosave:
 TEXT ·cgocallback(SB),NOSPLIT,$12-12  // Frame size must match commented places below
 	NO_LOCAL_POINTERS
 
-	// If g is nil, Go did not create the current thread.
+	// Skip cgocallbackg, just dropm when fn is nil, and frame is the saved g.
+	// It is used to dropm while thread is exiting.
+	MOVL	fn+0(FP), AX
+	CMPL	AX, $0
+	JNE	loadg
+	// Restore the g from frame.
+	get_tls(CX)
+	MOVL	frame+4(FP), BX
+	MOVL	BX, g(CX)
+	JMP	dropm
+
+loadg:
+	// If g is nil, Go did not create the current thread,
+	// or if this thread never called into Go on pthread platforms.
 	// Call needm to obtain one for temporary use.
 	// In this case, we're running on the thread stack, so there's
 	// lots of space, but the linker doesn't know. Hide the call from
@@ -707,9 +720,9 @@ TEXT ·cgocallback(SB),NOSPLIT,$12-12  // Frame size must match commented places
 	MOVL	BP, savedm-4(SP) // saved copy of oldm
 	JMP	havem
 needm:
-	MOVL	$runtime·needm(SB), AX
+	MOVL	$runtime·needAndBindM(SB), AX
 	CALL	AX
-	MOVL	$0, savedm-4(SP) // dropm on return
+	MOVL	$0, savedm-4(SP)
 	get_tls(CX)
 	MOVL	g(CX), BP
 	MOVL	g_m(BP), BP
@@ -784,13 +797,29 @@ havem:
 	MOVL	0(SP), AX
 	MOVL	AX, (g_sched+gobuf_sp)(SI)
 
-	// If the m on entry was nil, we called needm above to borrow an m
-	// for the duration of the call. Since the call is over, return it with dropm.
+	// If the m on entry was nil, we called needm above to borrow an m,
+	// 1. for the duration of the call on non-pthread platforms,
+	// 2. or the duration of the C thread alive on pthread platforms.
+	// If the m on entry wasn't nil,
+	// 1. the thread might be a Go thread,
+	// 2. or it's wasn't the first call from a C thread on pthread platforms,
+	//    since the we skip dropm to resue the m in the first call.
 	MOVL	savedm-4(SP), DX
 	CMPL	DX, $0
-	JNE 3(PC)
+	JNE	droppedm
+
+	// Skip dropm to reuse it in the next call, when a pthread key has been created.
+	MOVL	_cgo_pthread_key_created(SB), DX
+	// It means cgo is disabled when _cgo_pthread_key_created is a nil pointer, need dropm.
+	CMPL	DX, $0
+	JEQ	dropm
+	CMPL	(DX), $0
+	JNE	droppedm
+
+dropm:
 	MOVL	$runtime·dropm(SB), AX
 	CALL	AX
+droppedm:
 
 	// Done!
 	RET

diff --git a/src/runtime/asm_amd64.s b/src/runtime/asm_amd64.s
@@ -918,7 +918,20 @@ GLOBL zeroTLS<>(SB),RODATA,$const_tlsSize
 TEXT ·cgocallback(SB),NOSPLIT,$24-24
 	NO_LOCAL_POINTERS
 
-	// If g is nil, Go did not create the current thread.
+	// Skip cgocallbackg, just dropm when fn is nil, and frame is the saved g.
+	// It is used to dropm while thread is exiting.
+	MOVQ	fn+0(FP), AX
+	CMPQ	AX, $0
+	JNE	loadg
+	// Restore the g from frame.
+	get_tls(CX)
+	MOVQ	frame+8(FP), BX
+	MOVQ	BX, g(CX)
+	JMP	dropm
+
+loadg:
+	// If g is nil, Go did not create the current thread,
+	// or if this thread never called into Go on pthread platforms.
 	// Call needm to obtain one m for temporary use.
 	// In this case, we're running on the thread stack, so there's
 	// lots of space, but the linker doesn't know. Hide the call from
@@ -956,9 +969,9 @@ needm:
 	// a bad value in there, in case needm tries to use it.
 	XORPS	X15, X15
 	XORQ    R14, R14
-	MOVQ	$runtime·needm<ABIInternal>(SB), AX
+	MOVQ	$runtime·needAndBindM<ABIInternal>(SB), AX
 	CALL	AX
-	MOVQ	$0, savedm-8(SP) // dropm on return
+	MOVQ	$0, savedm-8(SP)
 	get_tls(CX)
 	MOVQ	g(CX), BX
 	MOVQ	g_m(BX), BX
@@ -1047,11 +1060,26 @@ havem:
 	MOVQ	0(SP), AX
 	MOVQ	AX, (g_sched+gobuf_sp)(SI)
 
-	// If the m on entry was nil, we called needm above to borrow an m
-	// for the duration of the call. Since the call is over, return it with dropm.
+	// If the m on entry was nil, we called needm above to borrow an m,
+	// 1. for the duration of the call on non-pthread platforms,
+	// 2. or the duration of the C thread alive on pthread platforms.
+	// If the m on entry wasn't nil,
+	// 1. the thread might be a Go thread,
+	// 2. or it's wasn't the first call from a C thread on pthread platforms,
+	//    since the we skip dropm to resue the m in the first call.
 	MOVQ	savedm-8(SP), BX
 	CMPQ	BX, $0
 	JNE	done
+
+	// Skip dropm to reuse it in the next call, when a pthread key has been created.
+	MOVQ	_cgo_pthread_key_created(SB), AX
+	// It means cgo is disabled when _cgo_pthread_key_created is a nil pointer, need dropm.
+	CMPQ	AX, $0
+	JEQ	dropm
+	CMPQ	(AX), $0
+	JNE	done
+
+dropm:
 	MOVQ	$runtime·dropm(SB), AX
 	CALL	AX
 #ifdef GOOS_windows