runtime: cgo support for Linux/ARM

Part 2 of CL 5601044 (cgo: Linux/ARM support)

R=dave, rsc
CC=golang-dev
https://golang.org/cl/5989057

src/pkg/runtime/asm_arm.s

@@ -31,6 +31,12 @@ TEXT _rt0_arm(SB),7,$-4
 	MOVW	R13, g_stackbase(g)
 	BL	runtime·emptyfunc(SB)	// fault if stack check is wrong
 
+	// if there is an initcgo, call it.
+	MOVW	initcgo(SB), R2
+	CMP	$0, R2
+	MOVW.NE	g, R0 // first argument of initcgo is g
+	BL.NE	(R2) // will clobber R0-R3
+
 	BL	runtime·check(SB)
 
 	// saved argc, argv
@@ -86,9 +92,12 @@ TEXT runtime·gosave(SB), 7, $-4
 // restore state from Gobuf; longjmp
 TEXT runtime·gogo(SB), 7, $-4
 	MOVW	0(FP), R1		// gobuf
-	MOVW	4(FP), R0		// return 2nd arg
 	MOVW	gobuf_g(R1), g
 	MOVW	0(g), R2		// make sure g != nil
+	MOVW	cgo_save_gm(SB), R2
+	CMP 	$0, R2 // if in Cgo, we have to save g and m
+	BL.NE	(R2) // this call will clobber R0
+	MOVW	4(FP), R0		// return 2nd arg
 	MOVW	gobuf_sp(R1), SP	// restore SP
 	MOVW	gobuf_pc(R1), PC
 
@@ -97,13 +106,16 @@ TEXT runtime·gogo(SB), 7, $-4
 // (call fn, returning to state in Gobuf)
 // using frame size $-4 means do not save LR on stack.
 TEXT runtime·gogocall(SB), 7, $-4
-	MOVW	0(FP), R0		// gobuf
+	MOVW	0(FP), R3		// gobuf
 	MOVW	4(FP), R1		// fn
 	MOVW	8(FP), R2		// fp offset
-	MOVW	gobuf_g(R0), g
-	MOVW	0(g), R3		// make sure g != nil
-	MOVW	gobuf_sp(R0), SP	// restore SP
-	MOVW	gobuf_pc(R0), LR
+	MOVW	gobuf_g(R3), g
+	MOVW	0(g), R0		// make sure g != nil
+	MOVW	cgo_save_gm(SB), R0
+	CMP 	$0, R0 // if in Cgo, we have to save g and m
+	BL.NE	(R0) // this call will clobber R0
+	MOVW	gobuf_sp(R3), SP	// restore SP
+	MOVW	gobuf_pc(R3), LR
 	MOVW	R1, PC
 
 // void mcall(void (*fn)(G*))
@@ -224,11 +236,114 @@ TEXT runtime·jmpdefer(SB), 7, $0
 	MOVW	$-4(SP), SP	// SP is 4 below argp, due to saved LR
 	B		(R0)
 
-TEXT	runtime·asmcgocall(SB),7,$0
-	B	runtime·cgounimpl(SB)
+// Dummy function to use in saved gobuf.PC,
+// to match SP pointing at a return address.
+// The gobuf.PC is unused by the contortions here
+// but setting it to return will make the traceback code work.
+TEXT return<>(SB),7,$0
+	RET
 
-TEXT	runtime·cgocallback(SB),7,$0
-	B	runtime·cgounimpl(SB)
+// asmcgocall(void(*fn)(void*), void *arg)
+// Call fn(arg) on the scheduler stack,
+// aligned appropriately for the gcc ABI.
+// See cgocall.c for more details.
+TEXT	runtime·asmcgocall(SB),7,$0
+	MOVW	fn+0(FP), R1
+	MOVW	arg+4(FP), R0
+	MOVW	R13, R2
+	MOVW	g, R5
+
+	// Figure out if we need to switch to m->g0 stack.
+	// We get called to create new OS threads too, and those
+	// come in on the m->g0 stack already.
+	MOVW	m_g0(m), R3
+	CMP	R3, g
+	BEQ	7(PC)
+	MOVW	R13, (g_sched + gobuf_sp)(g)
+	MOVW	$return<>(SB), R4
+	MOVW	R4, (g_sched+gobuf_pc)(g)
+	MOVW	g, (g_sched+gobuf_g)(g)
+	MOVW	R3, g
+	MOVW	(g_sched+gobuf_sp)(g), R13
+
+	// Now on a scheduling stack (a pthread-created stack).
+	SUB	$24, R13
+	BIC	$0x7, R13	// alignment for gcc ABI
+	MOVW	R5, 20(R13) // save old g
+	MOVW	R2, 16(R13)	// save old SP
+	// R0 already contains the first argument
+	BL	(R1)
+
+	// Restore registers, g, stack pointer.
+	MOVW	20(R13), g
+	MOVW	16(R13), R13
+	RET
+
+// cgocallback(void (*fn)(void*), void *frame, uintptr framesize)
+// See cgocall.c for more details.
+TEXT	runtime·cgocallback(SB),7,$16
+	MOVW	fn+0(FP), R0
+	MOVW	frame+4(FP), R1
+	MOVW	framesize+8(FP), R2
+
+	// Save current m->g0->sched.sp on stack and then set it to SP.
+	MOVW	m_g0(m), R3
+	MOVW	(g_sched+gobuf_sp)(R3), R4
+	MOVW.W	R4, -4(SP)
+	MOVW	R13, (g_sched+gobuf_sp)(R3)
+
+	// Switch to m->curg stack and call runtime.cgocallbackg
+	// with the three arguments.  Because we are taking over
+	// the execution of m->curg but *not* resuming what had
+	// been running, we need to save that information (m->curg->gobuf)
+	// so that we can restore it when we're done. 
+	// We can restore m->curg->gobuf.sp easily, because calling
+	// runtime.cgocallbackg leaves SP unchanged upon return.
+	// To save m->curg->gobuf.pc, we push it onto the stack.
+	// This has the added benefit that it looks to the traceback
+	// routine like cgocallbackg is going to return to that
+	// PC (because we defined cgocallbackg to have
+	// a frame size of 16, the same amount that we use below),
+	// so that the traceback will seamlessly trace back into
+	// the earlier calls.
+	MOVW	m_curg(m), g
+	MOVW	(g_sched+gobuf_sp)(g), R4 // prepare stack as R4
+
+	// Push gobuf.pc
+	MOVW	(g_sched+gobuf_pc)(g), R5
+	SUB	$4, R4
+	MOVW	R5, 0(R4)
+
+	// Push arguments to cgocallbackg.
+	// Frame size here must match the frame size above
+	// to trick traceback routines into doing the right thing.
+	SUB	$16, R4
+	MOVW	R0, 4(R4)
+	MOVW	R1, 8(R4)
+	MOVW	R2, 12(R4)
+	
+	// Switch stack and make the call.
+	MOVW	R4, R13
+	BL	runtime·cgocallbackg(SB)
+
+	// Restore g->gobuf (== m->curg->gobuf) from saved values.
+	MOVW	16(R13), R5
+	MOVW	R5, (g_sched+gobuf_pc)(g)
+	ADD	$(16+4), R13 // SP clobbered! It is ok!
+	MOVW	R13, (g_sched+gobuf_sp)(g)
+
+	// Switch back to m->g0's stack and restore m->g0->sched.sp.
+	// (Unlike m->curg, the g0 goroutine never uses sched.pc,
+	// so we do not have to restore it.)
+	MOVW	m_g0(m), g
+	MOVW	(g_sched+gobuf_sp)(g), R13
+	// POP R6
+	MOVW	0(R13), R6
+	ADD	$4, R13
+	MOVW	R6, (g_sched+gobuf_sp)(g)
+
+	// Done!
+	RET
 
 TEXT runtime·memclr(SB),7,$20
 	MOVW	0(FP), R0

src/pkg/runtime/cgo/gcc_arm.S

@@ -1 +1,57 @@
-/* unimplemented */
+// Copyright 2012 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.
+
+/*
+ * Apple still insists on underscore prefixes for C function names.
+ */
+#if defined(__APPLE__)
+#define EXT(s) _##s
+#else
+#define EXT(s) s
+#endif
+
+/*
+ * void crosscall_arm2(void (*fn)(void), void *g, void *m)
+ *
+ * Calling into the 5c tool chain, where all registers are caller save.
+ * Called from standard ARM EABI, where r4-r11 are callee-save, so they
+ * must be saved explicitly.
+ */
+.globl EXT(crosscall_arm2)
+EXT(crosscall_arm2):
+	push {r4, r5, r6, r7, r8, r9, r10, r11, ip, lr}
+	mov r10, r1 // g
+	mov r9, r2 // m
+	mov r3, r0 // save r0, cgo_tls_set_gm will clobber it
+	bl EXT(cgo_tls_set_gm) // save current g and m into TLS variable
+	mov lr, pc
+	mov pc, r3
+	pop {r4, r5, r6, r7, r8, r9, r10, r11, ip, pc}
+
+/*
+ * void crosscall2(void (*fn)(void*, int32), void*, int32)
+ *
+ * Save registers and call fn with two arguments.
+ */
+.globl EXT(crosscall2)
+EXT(crosscall2):
+	/* 
+	 * We still need to save all callee save register as before, and then
+	 *  push 2 args for fn (R1 and R2).
+	 * Also note that at procedure entry in 5c/5g world, 4(R13) will be the
+	 *  first arg, so we must push another dummy reg (R0) for 0(R13).
+	 *  Additionally, cgo_tls_set_gm will clobber R0, so we need to save R0
+	 *  nevertheless.
+	 */
+	push {r0, r1, r2, r4, r5, r6, r7, r8, r9, r10, r11, ip, lr}
+	bl EXT(cgo_tls_get_gm) // set up g and m from TLS
+	mov lr, pc
+	ldr pc, [sp, #0]
+	pop {r0, r1, r2, r4, r5, r6, r7, r8, r9, r10, r11, ip, pc}
+
+.globl EXT(__stack_chk_fail_local)
+EXT(__stack_chk_fail_local):
+1:
+	b 1b
+

src/pkg/runtime/cgo/gcc_linux_arm.c

@@ -2,12 +2,64 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+#include <pthread.h>
+#include <string.h>
 #include "libcgo.h"
 
+static void *threadentry(void*);
+
+// We have to resort to TLS variable to save g(R10) and
+// m(R9). One reason is that external code might trigger
+// SIGSEGV, and our runtime.sigtramp don't even know we
+// are in external code, and will continue to use R10/R9,
+// this might as well result in another SIGSEGV.
+// Note: all three functions will clobber R0, and the last
+// two can be called from 5c ABI code.
+void __aeabi_read_tp(void) __attribute__((naked));
+void cgo_tls_set_gm(void) __attribute__((naked));
+void cgo_tls_get_gm(void) __attribute__((naked));
+void __aeabi_read_tp(void) {
+	// b __kuser_get_tls @ 0xffff0fe0
+	__asm__ __volatile__ (
+		"mvn r0, #0xf000\n\t"
+		"sub pc, r0, #31\n\t"
+		"nop\n\tnop\n\t"
+	);
+}
+// g (R10) at 8(TP), m (R9) at 12(TP)
+void cgo_tls_get_gm(void) {
+	__asm__ __volatile__ (
+		"push {lr}\n\t"
+		"bl __aeabi_read_tp\n\t"
+		"ldr r10, [r0, #8]\n\t"
+		"ldr r9, [r0, #12]\n\t"
+		"pop {pc}\n\t"
+	);
+}
+void cgo_tls_set_gm(void) {
+	__asm__ __volatile__ (
+		"push {lr}\n\t"
+		"bl __aeabi_read_tp\n\t"
+		"str r10, [r0, #8]\n\t"
+		"str r9, [r0, #12]\n\t"
+		"pop {pc}\n\t"
+	);
+}
+// both cgo_tls_{get,set}_gm can be called from runtime
+void (*cgo_load_gm)(void) = cgo_tls_get_gm;
+void (*cgo_save_gm)(void) = cgo_tls_set_gm;
+
 static void
 xinitcgo(G *g)
 {
-	// unimplemented
+	pthread_attr_t attr;
+	size_t size;
+	cgo_tls_set_gm(); // save g and m for the initial thread
+
+	pthread_attr_init(&attr);
+	pthread_attr_getstacksize(&attr, &size);
+	g->stackguard = (uintptr)&attr - size + 4096;
+	pthread_attr_destroy(&attr);
 }
 
 void (*initcgo)(G*) = xinitcgo;
@@ -15,6 +67,43 @@ void (*initcgo)(G*) = xinitcgo;
 void
 libcgo_sys_thread_start(ThreadStart *ts)
 {
-	// unimplemented
-	*(int*)0 = 0;
+	pthread_attr_t attr;
+	pthread_t p;
+	size_t size;
+	int err;
+
+	// Not sure why the memset is necessary here,
+	// but without it, we get a bogus stack size
+	// out of pthread_attr_getstacksize.  C'est la Linux.
+	memset(&attr, 0, sizeof attr);
+	pthread_attr_init(&attr);
+	size = 0;
+	pthread_attr_getstacksize(&attr, &size);
+	ts->g->stackguard = size;
+	err = pthread_create(&p, &attr, threadentry, ts);
+	if (err != 0) {
+		fprintf(stderr, "runtime/cgo: pthread_create failed: %s\n", strerror(err));
+		abort();
+	}
+}
+
+extern void crosscall_arm2(void (*fn)(void), void *g, void *m);
+static void*
+threadentry(void *v)
+{
+	ThreadStart ts;
+
+	ts = *(ThreadStart*)v;
+	free(v);
+
+	ts.g->stackbase = (uintptr)&ts;
+
+	/*
+	 * libcgo_sys_thread_start set stackguard to stack size;
+	 * change to actual guard pointer.
+	 */
+	ts.g->stackguard = (uintptr)&ts - ts.g->stackguard + 4096 * 2;
+
+	crosscall_arm2(ts.fn, (void *)ts.g, (void *)ts.m);
+	return nil;
 }

src/pkg/runtime/cgocall.c

@@ -84,6 +84,11 @@
 
 void *initcgo;	/* filled in by dynamic linker when Cgo is available */
 
+// These two are only used by the architecture where TLS based storage isn't
+// the default for g and m (e.g., ARM)
+void *cgo_load_gm; /* filled in by dynamic linker when Cgo is available */
+void *cgo_save_gm; /* filled in by dynamic linker when Cgo is available */
+
 static void unlockm(void);
 static void unwindm(void);
 
@@ -229,6 +234,7 @@ unwindm(void)
 		runtime·throw("runtime: unwindm not implemented");
 	case '8':
 	case '6':
+	case '5':
 		m->g0->sched.sp = *(void**)m->g0->sched.sp;
 		break;
 	}

src/pkg/runtime/signal_linux_arm.c

@@ -75,6 +75,9 @@ runtime·sighandler(int32 sig, Siginfo *info, void *context, G *gp)
 		// old link register is more useful in the stack trace.
 		if(r->arm_pc != 0)
 			r->arm_lr = r->arm_pc;
+		// In case we are panicking from external C code
+		r->arm_r10 = (uintptr)gp;
+		r->arm_r9 = (uintptr)m;
 		r->arm_pc = (uintptr)runtime·sigpanic;
 		return;
 	}

src/pkg/runtime/sys_linux_arm.s

@@ -293,6 +293,14 @@ TEXT runtime·sigaltstack(SB),7,$0
 	RET
 
 TEXT runtime·sigtramp(SB),7,$24
+	// this might be called in external code context,
+	// where g and m are not set.
+	// first save R0, becuase cgo_load_gm will clobber it
+	MOVW	R0, 4(R13)
+	MOVW	cgo_load_gm(SB), R0
+	CMP 	$0, R0
+	BL.NE	(R0)
+
 	// save g
 	MOVW	g, R3
 	MOVW	g, 20(R13)
@@ -301,7 +309,7 @@ TEXT runtime·sigtramp(SB),7,$24
 	MOVW	m_gsignal(m), g
 
 	// copy arguments for call to sighandler
-	MOVW	R0, 4(R13)
+	// R0 is already saved above
 	MOVW	R1, 8(R13)
 	MOVW	R2, 12(R13)
 	MOVW	R3, 16(R13)