x86: encode cpu# in IA32_TSC_AUX and size of GDTe#15

arch/x86/kernel/context.S

@@ -10,7 +10,7 @@
  * HISTORY
  */
 
-#define X86_CPU_LOCAL_OFFSET_TSS    128
+#define X86_CPU_LOCAL_OFFSET_TSS    176
 #define X86_TSS_OFFSET_SP0          4
 #define X86_CPU_LOCAL_OFFSET_SP0 \
 		(X86_CPU_LOCAL_OFFSET_TSS + X86_TSS_OFFSET_SP0)

arch/x86/kernel/cpu.c

@@ -103,6 +103,12 @@ static uint64_t gdt[] __attribute__((aligned(16))) = {
 	0x00aff3000000ffff, /* 56 : USER_DS */
 	0x0000890000000067, /* 64 : TSS */
 	0,                  /* (72: TSS) */
+	0,                  /* 80 */
+	0,                  /* 88 */
+	0,                  /* 96 */
+	0,                  /* 104 */
+	0,                  /* 112 */
+	0x0000f10000000000, /* 120 : GETCPU */
 };
 
 struct tss64 tss __attribute__((aligned(16)));
@@ -570,6 +576,7 @@ static void init_smp_processor(void)
 {
 	struct x86_cpu_local_variables *v;
 	unsigned long tss_addr;
+	unsigned node_cpu;
 
 	v = get_x86_this_cpu_local();
 	tss_addr = (unsigned long)&v->tss;
@@ -590,6 +597,9 @@ static void init_smp_processor(void)
 		| (0x89UL << 40) | ((tss_addr & 0xff000000) << 32);
 	v->gdt[GLOBAL_TSS_ENTRY + 1] = (tss_addr >> 32);
 
+	node_cpu = v->processor_id;	/* assumes NUMA node 0 */
+	v->gdt[GETCPU_ENTRY] |= node_cpu;
+
 	v->gdt_ptr.size = sizeof(v->gdt) - 1;
 	v->gdt_ptr.address = (unsigned long)v->gdt;
         
@@ -597,6 +607,8 @@ static void init_smp_processor(void)
 	reload_gdt(&v->gdt_ptr);
 
 	set_kstack((unsigned long)get_x86_this_cpu_kstack());
+#define MSR_IA32_TSC_AUX 0xc0000103
+	wrmsr(MSR_IA32_TSC_AUX, node_cpu);
 }
 
 static char *trampoline_va, *first_page_va;

arch/x86/kernel/include/arch-memory.h

@@ -22,6 +22,7 @@
 #define USER_CS_ENTRY      6
 #define USER_DS_ENTRY      7
 #define GLOBAL_TSS_ENTRY   8
+#define GETCPU_ENTRY       15
 
 #define KERNEL_CS          (KERNEL_CS_ENTRY * 8)
 #define KERNEL_DS          (KERNEL_DS_ENTRY * 8)

arch/x86/kernel/include/cpulocal.h

@@ -22,7 +22,7 @@
  * - 4096 : kernel stack
  */
 
-#define X86_CPU_LOCAL_OFFSET_TSS    128
+#define X86_CPU_LOCAL_OFFSET_TSS    176
 #define X86_CPU_LOCAL_OFFSET_KSTACK 16
 #define X86_CPU_LOCAL_OFFSET_USTACK 24
 
@@ -39,13 +39,13 @@ struct x86_cpu_local_variables {
 	struct x86_desc_ptr gdt_ptr;
 	unsigned short pad[3];
 /* 48 */
-	uint64_t gdt[10];
+	uint64_t gdt[16];
-/* 128 */
+/* 176 */
 	struct tss64 tss;
-/* 232 */
+/* 280 */
 	unsigned long paniced;
 	uint64_t panic_regs[21];
-/* 408 */
+/* 456 */
 } __attribute__((packed));
 
 struct x86_cpu_local_variables *get_x86_cpu_local_variable(int id);

arch/x86/kernel/interrupt.S

@@ -13,7 +13,7 @@
  * 2013/?? - bgerofi + shimosawa: handle rsp correctly for nested interrupts
  */
 
-#define X86_CPU_LOCAL_OFFSET_TSS    128
+#define X86_CPU_LOCAL_OFFSET_TSS    176
 #define X86_TSS_OFFSET_SP0          4
 #define X86_CPU_LOCAL_OFFSET_SP0 \
 		(X86_CPU_LOCAL_OFFSET_TSS + X86_TSS_OFFSET_SP0)

arch/x86/kernel/syscall.c

@@ -216,7 +216,9 @@ SYSCALL_DECLARE(rt_sigreturn)
 	struct sigsp ksigsp;
 	struct sigsp *sigsp;
 
-	asm("movq %%gs:132, %0" : "=r" (regs));
+	asm ("movq %%gs:(%1),%0"
+			: "=r"(regs)
+			: "r"(offsetof(struct x86_cpu_local_variables, tss.rsp0)));
 	--regs;
 
 	sigsp = (struct sigsp *)regs->gpr.rsp;
@@ -680,7 +682,9 @@ do_signal(unsigned long rc, void *regs0, struct thread *thread, struct sig_pendi
 	}
 
 	if(regs == NULL){ /* call from syscall */
-		asm("movq %%gs:132, %0" : "=r" (regs));
+		asm ("movq %%gs:(%1),%0"
+				: "=r"(regs)
+				: "r"(offsetof(struct x86_cpu_local_variables, tss.rsp0)));
 		--regs;
 	}
 	else{