zircon/kernel/arch/x86/start16.S

// Copyright 2016 The Fuchsia Authors
//
// Use of this source code is governed by a MIT-style
// license that can be found in the LICENSE file or at
// https://opensource.org/licenses/MIT

// This file provides real-mode entry points for
// 1) secondary CPU initialization
// 2) suspend-to-RAM wakeup

#include <asm.h>
#include <arch/x86/bootstrap16.h>
#include <arch/x86/descriptor.h>
#include <arch/x86/registers.h>
#include <arch/defines.h>

// This code's only non-PIC instructions are movabs, which can be fixed up
// safely (see gen-kaslr-fixups.sh).  This section name is specially known
// by kernel.ld and gen-kaslr-fixups.sh.
.section .text.bootstrap16,"ax",%progbits
.balign PAGE_SIZE

DATA(x86_bootstrap16_start)

.code16
FUNCTION_LABEL(x86_bootstrap16_entry)
    // Enter no-fill cache mode (allegedly this is the initial state
    // according to Intel 3A, but on at least one Broadwell the APs can
    // come up with caching enabled)
    mov %cr0, %ebx
    or $X86_CR0_CD, %ebx
    and $~X86_CR0_NW, %ebx
    mov %ebx, %cr0
0:

    // We cheat a little and don't switch off of our real mode segments in
    // protected mode.  In real mode and protected mode, all of our code
    // and data accesses are relative to %cs and %ss, using the real mode
    // segment calculations.

    // setup %ds/%ss to refer to the data region
    mov %cs, %si
    add $0x100, %si
    mov %si, %ds
    mov %si, %ss

    lgdtl BCD_PHYS_GDTR_OFFSET

    // enter protected mode (but without paging)
    mov %cr0, %ebx
    or $X86_CR0_PE, %ebx
    mov %ebx, %cr0

    // clear instruction prefetch queue
    jmp 0f
0:
    // enable PAE / PGE
    mov %cr4, %ecx
    or $(X86_CR4_PAE|X86_CR4_PGE), %ecx
    mov %ecx, %cr4

    // load CR3 with the bootstrap PML4
    mov BCD_PHYS_BOOTSTRAP_PML4_OFFSET, %ecx
    mov %ecx, %cr3

    // enable IA-32e mode and indicate support for NX pages.
    // need the latter for once we switch to the real kernel
    // address space.
    mov $X86_MSR_IA32_EFER, %ecx
    rdmsr
    or $X86_EFER_LME, %eax
    or $X86_EFER_NXE, %eax
    wrmsr

    // enable paging
    mov %cr0, %ebx
    or $X86_CR0_PG, %ebx
    mov %ebx, %cr0

    // Translate data page segment into full address
    mov %ds, %esi
    shl $4, %esi

    // Jump to 64-bit CS
    mov $BCD_PHYS_LM_ENTRY_OFFSET, %esp
    lretl

// Get the secondary cpu into 64-bit mode with interrupts disabled and no TSS
.code64
FUNCTION_LABEL(_x86_secondary_cpu_long_mode_entry)
    // When we get here, %rsi should contain the absolute address of our data
    // page.
    mov $1, %rdi
    LOCK xadd %edi, BCD_CPU_COUNTER_OFFSET(%esi)
    // %rdi is now the index this CPU should use to grab resources

    // Shift index by 2, since the per_cpu member contains two 64-bit values which
    // will be at offsets 8*(2n) and 8*(2n+1) relative to PER_CPU_BASE_OFFSET
    shl $1, %rdi
    // Retrieve the top of this CPUs initial kernel stack
    // Note: the stack is unusable until we switch cr3 below
    mov BCD_PER_CPU_BASE_OFFSET(%rsi, %rdi, 8), %rsp

    // Retrieve this CPUs initial thread
    // Note: the stack is unusable until we switch cr3 below
    add $1, %rdi
    mov BCD_PER_CPU_BASE_OFFSET(%rsi, %rdi, 8), %rdx

    // Retrieve the new PML4 address before our data page becomes unreachable
    mov BCD_PHYS_KERNEL_PML4_OFFSET(%esi), %ecx
    // Similarly for the CPU waiting mask
    mov BCD_CPU_WAITING_OFFSET(%esi), %rdi

    // Switch out of the copied code page and into the kernel's
    // version of it
    movabs $.Lhighaddr, %rbx
    jmp  *%rbx
.Lhighaddr:
    // Switch to the kernel's PML4
    mov %rcx, %cr3
    // As of this point, %esi is invalid

    // Reload the GDT with one based off of non-identity mapping
    lgdt _temp_gdtr(%rip)

    // Zero our data segments
    xor %eax, %eax
    mov %eax, %ds
    mov %eax, %es
    mov %eax, %fs
    mov %eax, %gs
    mov %eax, %ss

    // Load the IDT
    call load_startup_idt

    mov %rdx, %rsi
    // Do an indirect call to keep this position independent
    // x86_secondary_entry(CPU ready counter, thread)
    movabs $x86_secondary_entry, %rbx
    call *%rbx

// If x86_secondary_entry returns, hang.
0:
    hlt
    jmp 0b

// Get the cpu into 64-bit mode with interrupts disabled and no TSS.  This must
// only be called on the bootstrap processor.
FUNCTION_LABEL(_x86_suspend_wakeup)
    // Retrieve the new PML4 address before our data page becomes unreachable
    mov BCD_PHYS_KERNEL_PML4_OFFSET(%esi), %ecx

    // Stash register pointer so that we can read it after we change
    // address spaces
    mov RED_REGISTERS_OFFSET(%esi), %rdi

    // Switch out of the copied code page and into the kernel's
    // version of it
    movabs $.Lwakeup_highaddr, %rbx
    jmp  *%rbx
.Lwakeup_highaddr:
    // Switch to the kernel's PML4
    mov %rcx, %cr3
    // As of this point, %esi is invalid

    // Reload the GDT with one based off of non-identity mapping
    lgdt _temp_gdtr(%rip)

    // Zero our data segments
    xor %eax, %eax
    mov %eax, %ds
    mov %eax, %es
    mov %eax, %fs
    mov %eax, %gs
    mov %eax, %ss

    // Restore %gs.base to &bp_percpu.  We need to do this before
    // returning to C code, since the C code might use safe-stack
    // and/or stack-protector.
    // TODO(teisenbe):  There is a small performance gain that could be made here
    // by switching from wrmsr to wrgsbase, if wrgsbase is supported.  Currently
    // this is omitted for simplicity.
    lea bp_percpu(%rip), %rax
    mov %rax, %rdx
    shr $32, %rdx
    mov $X86_MSR_IA32_GS_BASE, %ecx
    wrmsr

    // Restore the stack pointer first, so we can use the stack right now.
    mov 120(%rdi), %rsp

    // Load the IDT.  Note this uses the stack and clobbers %rax, but not %rdi.
    call load_startup_idt

    mov 8(%rdi), %rsi
    mov 16(%rdi), %rbp
    mov 24(%rdi), %rbx
    mov 32(%rdi), %rdx
    mov 40(%rdi), %rcx
    mov 48(%rdi), %rax
    mov 56(%rdi), %r8
    mov 64(%rdi), %r9
    mov 72(%rdi), %r10
    mov 80(%rdi), %r11
    mov 88(%rdi), %r12
    mov 96(%rdi), %r13
    mov 104(%rdi), %r14
    mov 112(%rdi), %r15

    // Note: %rdi is not restored, but it is a caller-save register anyway.
    // If we want to restore %rdi, we could potentially use the stack here
    // to do something like "push 128(%rdi); mov (%rdi), %rdi; ret".

    // Restore RIP
    jmp *128(%rdi)

DATA(x86_bootstrap16_end)
    nop