Proper FPU init, fxsave/fxrstor for threads

arch/amd64/kernel.c

@@ -11,6 +11,7 @@
 #include "arch/amd64/hw/ps2.h"
 #include "arch/amd64/cpuid.h"
 #include "arch/amd64/mm/mm.h"
+#include "arch/amd64/fpu.h"
 #include "sys/block/ram.h"
 #include "sys/config.h"
 #include "sys/kernel.h"
@@ -77,6 +78,8 @@ void kernel_early_init(struct amd64_loader_data *data) {
 
     amd64_make_random_seed();
 
+    amd64_fpu_init();
+
 #if defined(AMD64_SMP)
     amd64_smp_init();
 #endif

arch/amd64/sched_s.S

@@ -62,6 +62,14 @@ context_exec_enter:
 context_switch_to:
     // %rdi - new thread
     // %rsi - from
+    pushq %rdi
+    pushq %rsi
+
+    call context_save_fpu
+
+    popq %rsi
+    popq %rdi
+
     pushq %r15
     pushq %r14
     pushq %r13
@@ -83,6 +91,14 @@ context_switch_first:
     popq %r14
     popq %r15
 
+    pushq %rdi
+    pushq %rsi
+
+    call context_restore_fpu
+
+    popq %rsi
+    popq %rdi
+
     // Load TSS.RSP0 for user -> kernel transition
     // %rax = top of task's stack
     movq THREAD_RSP0_TOP(%rdi), %rax

include/arch/amd64/asm/asm_thread.h

@@ -16,5 +16,7 @@ struct thread_data {
 
     uintptr_t rsp0_base, rsp0_size;
     uintptr_t rsp3_base, rsp3_size;
+
+    void *fxsave;
 };
 #endif

include/sys/thread.h

@@ -21,6 +21,7 @@ enum thread_state {
 
 #define THREAD_KERNEL           (1 << 0)
 #define THREAD_EMPTY            (1 << 1)
+#define THREAD_FPU_SAVED        (1 << 2)
 
 #define thread_signal_clear(thr, signum) \
     (thr)->sigq &= ~(1ULL << ((signum) - 1))

sys/thread.c

@@ -40,6 +40,26 @@ struct sys_fork_frame {
 LIST_HEAD(threads_all_head);
 static pid_t last_kernel_pid = 0;
 static pid_t last_user_pid = 0;
+static uint64_t fxsave_buf[FXSAVE_REGION / 8] __attribute__((aligned(16)));
+
+void context_save_fpu(struct thread *new, struct thread *old) {
+    _assert(old);
+    if (old->data.fxsave) {
+        asm volatile ("fxsave (%0)"::"r"(fxsave_buf):"memory");
+        memcpy(old->data.fxsave, fxsave_buf, FXSAVE_REGION);
+        old->flags |= THREAD_FPU_SAVED;
+    }
+
+}
+
+void context_restore_fpu(struct thread *new, struct thread *old) {
+    _assert(new);
+    if (new->flags & THREAD_FPU_SAVED) {
+        memcpy(fxsave_buf, new->data.fxsave, FXSAVE_REGION);
+        asm volatile ("fxrstor (%0)"::"r"(fxsave_buf):"memory");
+        new->flags &= ~THREAD_FPU_SAVED;
+    }
+}
 
 pid_t thread_alloc_pid(int is_user) {
     if (is_user) {
@@ -226,6 +246,13 @@ int thread_init(struct thread *thr, uintptr_t entry, void *arg, int user) {
     thr->name[0] = 0;
     thr->flags = user ? 0 : THREAD_KERNEL;
 
+    if (!(thr->flags & THREAD_KERNEL)) {
+        thr->data.fxsave = kmalloc(FXSAVE_REGION);
+        _assert(thr->data.fxsave);
+    } else {
+        thr->data.fxsave = NULL;
+    }
+
     list_link_init(&thr->g_link);
 
     uint64_t *stack = (uint64_t *) (thr->data.rsp0_base + thr->data.rsp0_size);
@@ -375,6 +402,16 @@ int sys_fork(struct sys_fork_frame *frame) {
 
     dst->signal_entry = src->signal_entry;
 
+    strcpy(dst->name, src->name);
+
+    dst->data.fxsave = kmalloc(FXSAVE_REGION);
+    _assert(dst->data.fxsave);
+    _assert(src->data.fxsave);
+
+    if (src->flags & THREAD_FPU_SAVED) {
+        memcpy(dst->data.fxsave, src->data.fxsave, FXSAVE_REGION);
+    }
+
     thread_ioctx_fork(dst, src);
 
     dst->state = THREAD_READY;
@@ -546,6 +583,9 @@ int sys_execve(const char *path, const char **argv, const char **envp) {
         thr->data.cr3 = MM_PHYS(thr->space);
         thr->flags = 0;
 
+        thr->data.fxsave = kmalloc(FXSAVE_REGION);
+        _assert(thr->data.fxsave);
+
         mm_space_clone(thr->space, mm_kernel, MM_CLONE_FLG_KERNEL);
     } else {
         mm_space_release(thr->space);