Add SSE support

kernel/src/cpu/cpu.rs

@@ -23,9 +23,9 @@ extern "C" fn ap_main() -> ! {
     setup_gdt();
     setup_idt();
     setup_lapic(0);
+    set_cpu_flags();
     // TODO: Also calibrate other cores
     setup_lapic_timer(false);
-    set_cpu_flags();
     unsafe {
         sti();
     }
@@ -37,8 +37,10 @@ pub fn set_cpu_flags() {
     let features = cpuid.get_feature_info().unwrap();
     let extended_features = cpuid.get_extended_feature_info().unwrap();
     unsafe {
-        // Time Stamp Disable, SMEP and SMAP
-        asm!("mov rax, cr4; bts rax, 2; bts rax, 20; bts rax, 21; mov cr4, rax", out("rax") _);
+        // Numeric Error
+        asm!("mov rax, cr0; bts rax, 5; mov cr0, rax", out("rax") _);
+        // Time Stamp Disable, OSFXSR, OSXMMEXCPT, SMEP and SMAP
+        asm!("mov rax, cr4; bts rax, 2; bts rax, 9; bts rax, 10; bts rax, 20; bts rax, 21; mov cr4, rax", out("rax") _);
         if features.has_mce() {
             asm!("mov rax, cr4; bts rax, 6; mov cr4, rax", out("rax") _);
         }

kernel/src/cpu/usermode.s

@@ -30,4 +30,5 @@ jump_usermode:
   xor %r13, %r13
   xor %r14, %r14
   xor %r15, %r15
+  fninit
   iretq

kernel/src/main.rs

@@ -95,12 +95,12 @@ extern "C" fn early_main(temp_loader_struct: *const LoaderStruct) -> ! {
     }
     parse_madt(&early_acpi_tables);
     setup_hpet(&early_acpi_tables);
+    set_cpu_flags();
     setup_lapic_timer(true);
     setup_multitasking();
 }
 fn main() {
     info!("Starting main kernel task...");
-    set_cpu_flags();
     let mut status = unsafe { AcpiInitializeSubsystem() };
     assert_eq!(status, AE_OK);
     status = unsafe { AcpiInitializeTables(null_mut(), 0, 0) };

kernel/src/sys/task.rs

@@ -1,6 +1,9 @@
-use core::sync::atomic::{AtomicBool, AtomicI64, AtomicU64, AtomicUsize, Ordering};
+use core::{
+    arch::x86_64::{_fxrstor64, _fxsave64},
+    sync::atomic::{AtomicBool, AtomicI64, AtomicU64, AtomicUsize, Ordering},
+};
 
-use alloc::{sync::Arc, vec, vec::Vec};
+use alloc::{boxed::Box, sync::Arc, vec, vec::Vec};
 use kernel_common::instructions::{cli, get_rflags, hlt, pause, sti};
 
 use crate::{
@@ -30,6 +33,9 @@ pub enum TaskState {
     SemaphoreBlocked,
 }
 
+#[repr(align(16))]
+struct FXSaveRegion([u8; 512]);
+
 pub struct Task {
     pub id: usize,
     pub process: Arc<Spinlock<Process>>,
@@ -41,6 +47,7 @@ pub struct Task {
     pub block_on_semaphore: Option<Arc<RawSemaphore>>,
     pub semaphore_requested_count: usize,
     user_stack: u64,
+    fxsave_region: Box<FXSaveRegion>,
 }
 
 #[repr(C)]
@@ -69,6 +76,9 @@ pub fn switch_task(current_state: &mut ISRState, new_task: Task) {
     if let Some(mut current_task) = current_tasks[lapic_id].take() {
         current_task.state = *current_state;
         current_task.user_stack = cpu_data.user_stack;
+        unsafe {
+            _fxsave64(current_task.fxsave_region.0.as_mut_ptr());
+        }
         if current_task.task_state == TaskState::Terminated {
             unsafe {
                 get_kernel_process().lock().address_space.switch();
@@ -95,6 +105,9 @@ pub fn switch_task(current_state: &mut ISRState, new_task: Task) {
     let kernel_stack = new_task.kernel_stack.as_ptr() as u64 + STACK_SIZE as u64;
     cpu_data.kernel_stack = kernel_stack;
     cpu_data.user_stack = new_task.user_stack;
+    unsafe {
+        _fxrstor64(new_task.fxsave_region.0.as_ptr());
+    }
     current_tasks[get_current_lapic_id()] = Some(new_task);
     schedule_timer_interrupt();
 }
@@ -135,6 +148,7 @@ pub fn create_task(process: Arc<Spinlock<Process>>, func: fn()) -> Task {
         block_on_semaphore: None,
         semaphore_requested_count: 0,
         user_stack: 0,
+        fxsave_region: Box::new(FXSaveRegion([0; 512])),
     }
 }
 fn create_idle_task() {

loader/src/main.rs

@@ -58,6 +58,7 @@ fn main() -> Status {
     let extended_features = cpuid.get_extended_feature_info().unwrap();
     assert!(extended_features.has_smap());
     assert!(extended_features.has_smep());
+    assert!(features.has_sse42());
     assert!(cpuid.get_extended_processor_and_feature_identifiers().unwrap().has_syscall_sysret());
     let (kernel_start, kernel_entry) = load_kernel(KERNEL);
     let heap_start = allocate_pages(AllocateType::AnyPages, MemoryType::LOADER_DATA, KERNEL_HEAP_INITIAL_SIZE / 0x1000).unwrap().as_ptr() as u64;

qemu.sh

@@ -2,4 +2,4 @@
 # shellcheck disable=SC2068
 set -euo pipefail
 ./build.py
-qemu-system-x86_64 -M q35 -accel kvm -cpu qemu64,rdrand,smap,smep,umip -smp 4 -m 256M -L /usr/share/OVMF -bios OVMF_CODE.fd -drive file=img/os.img,format=raw $@
+qemu-system-x86_64 -M q35 -accel kvm -cpu host -smp 4 -m 256M -L /usr/share/OVMF -bios OVMF_CODE.fd -drive file=img/os.img,format=raw $@