linux/386: use Xen-friendly ELF TLS instruction sequence

Fixes #465.

R=iant
CC=golang-dev
https://golang.org/cl/1665051

src/cmd/8l/list.c

@@ -102,7 +102,7 @@ Dconv(Fmt *fp)
 	i = a->type;
 	if(i >= D_INDIR && i < 2*D_INDIR) {
 		if(a->offset)
-			sprint(str, "%ld(%R)", a->offset, i-D_INDIR);
+			sprint(str, "%ld(%R)", (long)a->offset, i-D_INDIR);
 		else
 			sprint(str, "(%R)", i-D_INDIR);
 		goto brk;

src/cmd/8l/pass.c

@@ -378,7 +378,7 @@ patch(void)
 	s = lookup("exit", 0);
 	vexit = s->value;
 	for(p = firstp; p != P; p = p->link) {
-		if(HEADTYPE == 10) {
+		if(HEADTYPE == 10) {	// Windows
 			// Convert
 			//   op   n(GS), reg
 			// to
@@ -391,7 +391,7 @@ patch(void)
 			&& p->to.type >= D_AX && p->to.type <= D_DI) {
 				q = appendp(p);
 				q->from = p->from;
-				q->from.type += p->to.type-D_GS;
+				q->from.type = D_INDIR + p->to.type;
 				q->to = p->to;
 				q->as = p->as;
 				p->as = AMOVL;
@@ -399,6 +399,23 @@ patch(void)
 				p->from.offset = 0x2C;
 			}
 		}
+		if(HEADTYPE == 7) {	// Linux
+			// Running binaries under Xen requires using
+			//	MOVL 0(GS), reg
+			// and then off(reg) instead of saying off(GS) directly
+			// when the offset is negative.
+			if(p->from.type == D_INDIR+D_GS && p->from.offset < 0
+			&& p->to.type >= D_AX && p->to.type <= D_DI) {
+				q = appendp(p);
+				q->from = p->from;
+				q->from.type = D_INDIR + p->to.type;
+				q->to = p->to;
+				q->as = p->as;
+				p->as = AMOVL;
+				p->from.type = D_INDIR+D_GS;
+				p->from.offset = 0;
+			}
+		}
 		if(p->as == ATEXT)
 			curtext = p;
 		if(p->as == ACALL || (p->as == AJMP && p->to.type != D_BRANCH)) {
@@ -616,7 +633,8 @@ dostkoff(void)
 			if(pmorestack != P)
 			if(!(p->from.scale & NOSPLIT)) {
 				p = appendp(p);	// load g into CX
-				if(HEADTYPE == 10) {
+				switch(HEADTYPE) {
+				case 10:	// Windows
 					p->as = AMOVL;
 					p->from.type = D_INDIR+D_FS;
 					p->from.offset = 0x2c;
@@ -627,7 +645,22 @@ dostkoff(void)
 					p->from.type = D_INDIR+D_CX;
 					p->from.offset = 0;
 					p->to.type = D_CX;
-				} else {
+					break;
+				
+				case 7:	// Linux
+					p->as = AMOVL;
+					p->from.type = D_INDIR+D_GS;
+					p->from.offset = 0;
+					p->to.type = D_CX;
+
+					p = appendp(p);
+					p->as = AMOVL;
+					p->from.type = D_INDIR+D_CX;
+					p->from.offset = tlsoffset + 0;
+					p->to.type = D_CX;
+					break;
+
+				default:
 					p->as = AMOVL;
 					p->from.type = D_INDIR+D_GS;
 					p->from.offset = tlsoffset + 0;

src/libcgo/linux_386.c

@@ -45,11 +45,14 @@ threadentry(void *v)
 	 * Set specific keys.  On Linux/ELF, the thread local storage
 	 * is just before %gs:0.  Our dynamic 8.out's reserve 8 bytes
 	 * for the two words g and m at %gs:-8 and %gs:-4.
+	 * Xen requires us to access those words indirect from %gs:0
+	 * which points at itself.
 	 */
 	asm volatile (
-		"movl %0, %%gs:-8\n"	// MOVL g, -8(GS)
-		"movl %1, %%gs:-4\n"	// MOVL m, -4(GS)
-		:: "r"(ts.g), "r"(ts.m)
+		"movl %%gs:0, %%eax\n"		// MOVL 0(GS), tmp
+		"movl %0, -8(%%eax)\n"	// MOVL g, -8(GS)
+		"movl %1, -4(%%eax)\n"	// MOVL m, -4(GS)
+		:: "r"(ts.g), "r"(ts.m) : "%eax"
 	);
 
 	crosscall_386(ts.fn);

src/pkg/runtime/linux/386/sys.s

@@ -161,12 +161,12 @@ TEXT clone(SB),7,$0
 	// In child on new stack.  Reload registers (paranoia).
 	MOVL	0(SP), BX	// m
 	MOVL	4(SP), DX	// g
-	MOVL	8(SP), CX	// fn
+	MOVL	8(SP), SI	// fn
 
 	MOVL	AX, m_procid(BX)	// save tid as m->procid
 
 	// set up ldt 7+id to point at m->tls.
-	// m->tls is at m+40.  newosproc left the id in tls[0].
+	// newosproc left the id in tls[0].
 	LEAL	m_tls(BX), BP
 	MOVL	0(BP), DI
 	ADDL	$7, DI	// m0 is LDT#7. count up.
@@ -186,7 +186,7 @@ TEXT clone(SB),7,$0
 	MOVL	DX, g(AX)
 	MOVL	BX, m(AX)
 
-	CALL	stackcheck(SB)	// smashes AX
+	CALL	stackcheck(SB)	// smashes AX, CX
 	MOVL	0(DX), DX	// paranoia; check they are not nil
 	MOVL	0(BX), BX
 
@@ -195,7 +195,7 @@ TEXT clone(SB),7,$0
 	CALL	emptyfunc(SB)
 	POPAL
 
-	CALL	CX	// fn()
+	CALL	SI	// fn()
 	CALL	exit1(SB)
 	MOVL	$0x1234, 0x1005
 	RET
@@ -247,8 +247,11 @@ TEXT setldt(SB),7,$32
 	 * To accommodate that rewrite, we translate
 	 * the address here and bump the limit to 0xffffffff (no limit)
 	 * so that -8(GS) maps to 0(address).
+	 * Also, the final 0(GS) (current 8(CX)) has to point
+	 * to itself, to mimic ELF.
 	 */
 	ADDL	$0x8, CX	// address
+	MOVL	CX, 0(CX)
 
 	// set up user_desc
 	LEAL	16(SP), AX	// struct user_desc

src/pkg/runtime/mkasmh.sh

@@ -24,6 +24,31 @@ case "$GOARCH" in
 		echo '#define	g(r)	0(r)'
 		echo '#define	m(r)	4(r)'
 		;;
+	linux)
+		# On Linux systems, what we call 0(GS) and 4(GS) for g and m
+		# turn into %gs:-8 and %gs:-4 (using gcc syntax to denote
+		# what the machine sees as opposed to 8l input).
+		# 8l rewrites 0(GS) and 4(GS) into these.
+		#
+		# On Linux Xen, it is not allowed to use %gs:-8 and %gs:-4
+		# directly.  Instead, we have to store %gs:0 into a temporary
+		# register and then use -8(%reg) and -4(%reg).  This kind
+		# of addressing is correct even when not running Xen.
+		#
+		# 8l can rewrite MOVL 0(GS), CX into the appropriate pair
+		# of mov instructions, using CX as the intermediate register
+		# (safe because CX is about to be written to anyway).
+		# But 8l cannot handle other instructions, like storing into 0(GS),
+		# which is where these macros come into play.
+		# get_tls sets up the temporary and then g and r use it.
+		#
+		# The final wrinkle is that get_tls needs to read from %gs:0,
+		# but in 8l input it's called 8(GS), because 8l is going to
+		# subtract 8 from all the offsets, as described above.
+		echo '#define	get_tls(r)	MOVL 8(GS), r'
+		echo '#define	g(r)	-8(r)'
+		echo '#define	m(r)	-4(r)'
+		;;
 	*)
 		echo '#define	get_tls(r)'
 		echo '#define	g(r)	0(GS)'