x86_64: implement proper user/kernel %gs base switching

the %gs base address is now always set to the current cpu block while
in kernel-mode, and is switched back to the userspace %gs base
when returning to user-mode.

arch/x86_64/include/arch/msr.h

@@ -4,6 +4,7 @@
 #include <stdint.h>
 
 #define MSR_GS_BASE        0xC0000101
+#define MSR_KERNEL_GS_BASE 0xC0000102
 
 #ifdef __cplusplus
 extern "C" {

arch/x86_64/irqvec.S

@@ -333,11 +333,30 @@ IRQ 223, 255
 isr_common_stub:
 	PUSH_REGS
 
+	# When ISR occurs in Ring 3, CPU sets %ss (and other non-code selectors)
+	# to 0.
+	mov %ss, %ax
+	cmp $0, %ax
+	jne isr_skipgs1
+
+	mov $0x10, %ax
+	mov %ax, %ss
+	swapgs
+
+isr_skipgs1:
 	mov %rsp, %rdi
 	call isr_dispatch
 
 	POP_REGS
 	add $16, %rsp
+
+	cmpq $0x1b, 32(%rsp)
+	jne isr_skipgs2
+
+	swapgs
+
+isr_skipgs2:
+
 	iretq
 
 
@@ -347,11 +366,31 @@ isr_common_stub:
 irq_common_stub:
 	PUSH_REGS
 
+	# When IRQ occurs in Ring 3, CPU sets %ss (and other non-code selectors)
+	# to 0.
+	mov %ss, %ax
+	cmp $0, %ax
+	jne irq_skipgs1
+
+	mov $0x10, %ax
+	mov %ax, %ss
+	swapgs
+
+irq_skipgs1:
 	mov %rsp, %rdi
 	call irq_dispatch
 
+
 	POP_REGS
 	add $16, %rsp
+
+	cmpq $0x1b, 32(%rsp)
+	jne isr_skipgs2
+
+	swapgs
+
+irq_skipgs2:
+
 	iretq
 
 
@@ -363,12 +402,12 @@ irq_common_stub:
 
 syscall_gate:
 	swapgs
-	movq %rsp, %gs:20       # GS+20 = rsp2 in the current TSS block (user stack storage)
-	movq %gs:4, %rsp	# GS+4  = rsp0 in the current TSS block (per-thread kstack)
+	movq %rsp, %gs:94       # GS+20 = rsp2 in the current TSS block (user stack storage)
+	movq %gs:78, %rsp	# GS+4  = rsp0 in the current TSS block (per-thread kstack)
 
 	# start building a ml_cpu_context
 	pushq $0x1b
-	pushq %gs:20
+	pushq %gs:94
 	push %r11
 	push $0x23
 	push %rcx
@@ -380,10 +419,6 @@ syscall_gate:
 
 	mov %rsp, %rdi
 
-	# switch back to user gs while in syscall_dispatch. Interrupts are enabled in syscall_dispatch,
-	# and if the task gets pre-empted, the incoming task will expect %gs to have its usermode value.
-	swapgs
-
 	call syscall_dispatch
 
 	POP_REGS
@@ -394,8 +429,8 @@ syscall_gate:
 	pop %r11
 	add $16, %rsp
 
-	swapgs
-	movq %gs:20, %rsp       # GS+20 = rsp2 in the current TSS block
+	movq %gs:94, %rsp       # GS+20 = rsp2 in the current TSS block
+
 	swapgs
 
 	# back to usermode

arch/x86_64/thread_switch.S

@@ -73,4 +73,5 @@ ml_thread_switch_user:
 	pop %rax
 
 	add $16, %rsp
+	swapgs
 	iretq

arch/x86_64/tss.c

@@ -1,5 +1,3 @@
-#include "arch/msr.h"
-
 #include <arch/gdt.h>
 #include <arch/tss.h>
 #include <kernel/libc/string.h>
@@ -22,9 +20,6 @@ void tss_init(struct tss *tss, struct tss_ptr *ptr)
 void tss_load(struct tss *tss)
 {
 	tss_flush(TSS_GDT_INDEX);
-
-	uintptr_t kernel_gs_base_reg = 0xC0000102;
-	wrmsr(kernel_gs_base_reg, (uintptr_t)tss);
 }
 
 virt_addr_t tss_get_kstack(struct tss *tss)