runtime: use set_thread_area instead of modify_ldt on linux/386

linux/386 depends on modify_ldt system call, but recent Linux kernels
can disable this system call. Any Go programs built as linux/386
crash with the message 'Trace/breakpoint trap'.

The kernel config CONFIG_MODIFY_LDT_SYSCALL, which control
enable/disable modify_ldt, is disabled on Amazon Linux 2016.03.

This fixes this problem by using set_thread_area instead of modify_ldt
on linux/386.

Fixes #14795.

Change-Id: I0cc5139e40e9e5591945164156a77b6bdff2c7f1
Reviewed-on: https://go-review.googlesource.com/21190Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Minux Ma <minux@golang.org>

src/runtime/sys_linux_386.s

@@ -406,9 +406,18 @@ TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 #define SEG_NOT_PRESENT 0x20
 #define USEABLE 0x40
 
+// `-1` means the kernel will pick a TLS entry on the first setldt call,
+// which happens during runtime init, and that we'll store back the saved
+// entry and reuse that on subsequent calls when creating new threads.
+DATA  runtime·tls_entry_number+0(SB)/4, $-1
+GLOBL runtime·tls_entry_number(SB), NOPTR, $4
+
 // setldt(int entry, int address, int limit)
+// We use set_thread_area, which mucks with the GDT, instead of modify_ldt,
+// which would modify the LDT, but is disabled on some kernels.
+// The name, setldt, is a misnomer, although we leave this name as it is for
+// the compatibility with other platforms.
 TEXT runtime·setldt(SB),NOSPLIT,$32
-	MOVL	entry+0(FP), BX	// entry
 	MOVL	address+4(FP), DX	// base address
 
 #ifdef GOOS_android
@@ -437,18 +446,19 @@ TEXT runtime·setldt(SB),NOSPLIT,$32
 	MOVL	DX, 0(DX)
 #endif
 
+	// get entry number
+	MOVL	runtime·tls_entry_number(SB), CX
+
 	// set up user_desc
 	LEAL	16(SP), AX	// struct user_desc
-	MOVL	BX, 0(AX)
-	MOVL	DX, 4(AX)
-	MOVL	$0xfffff, 8(AX)
+	MOVL	CX, 0(AX)	// unsigned int entry_number
+	MOVL	DX, 4(AX)	// unsigned long base_addr
+	MOVL	$0xfffff, 8(AX)	// unsigned int limit
 	MOVL	$(SEG_32BIT|LIMIT_IN_PAGES|USEABLE|CONTENTS_DATA), 12(AX)	// flag bits
 
-	// call modify_ldt
-	MOVL	$1, BX	// func = 1 (write)
-	MOVL	AX, CX	// user_desc
-	MOVL	$16, DX	// sizeof(user_desc)
-	MOVL	$123, AX	// syscall - modify_ldt
+	// call set_thread_area
+	MOVL	AX, BX	// user_desc
+	MOVL	$243, AX	// syscall - set_thread_area
 	// We can't call this via 0x10(GS) because this is called from setldt0 to set that up.
 	INT     $0x80
 
@@ -457,10 +467,18 @@ TEXT runtime·setldt(SB),NOSPLIT,$32
 	JLS 2(PC)
 	INT $3
 
-	// compute segment selector - (entry*8+7)
-	MOVL	entry+0(FP), AX
+	// read allocated entry number back out of user_desc
+	LEAL	16(SP), AX	// get our user_desc back
+	MOVL	0(AX), AX
+
+	// store entry number if the kernel allocated it
+	CMPL	CX, $-1
+	JNE	2(PC)
+	MOVL	AX, runtime·tls_entry_number(SB)
+
+	// compute segment selector - (entry*8+3)
 	SHLL	$3, AX
-	ADDL	$7, AX
+	ADDL	$3, AX
 	MOVW	AX, GS
 
 	RET