func SaveState(vm *platform.Vm, model *machine.Model) (State, error) {
// Pause the vm.
// NOTE: Our model will also be stopped automatically
// with model.DeviceInfo() below, but we manually pause
// the Vcpus here in order to ensure they are completely
// stopped prior to saving all device state.
err := vm.Pause(false)
if err != nil {
return State{}, err
}
defer vm.Unpause(false)
// Grab our vcpu states.
vcpus, err := vm.VcpuInfo()
if err != nil {
return State{}, err
}
// Grab our devices.
// NOTE: This should block until devices have
// actually quiesed (finished processing outstanding
// requests generated by the VCPUs).
devices, err := model.DeviceInfo(vm)
if err != nil {
return State{}, err
}
// Done.
return State{Vcpus: vcpus, Devices: devices}, nil
}
func Loop(
vm *platform.Vm,
vcpu *platform.Vcpu,
model *machine.Model,
tracer *loader.Tracer) error {
// It's not really kosher to switch threads constantly when running a
// KVM VCPU. So we simply lock this goroutine to a single system
// thread. That way we know it won't be bouncing around.
runtime.LockOSThread()
defer runtime.UnlockOSThread()
log.Printf("Vcpu[%d] running.", vcpu.Id)
for {
// Enter the guest.
err := vcpu.Run()
// Trace if requested.
trace_err := tracer.Trace(vcpu, vcpu.IsStepping())
if trace_err != nil {
return trace_err
}
// No reason for exit?
if err == nil {
return ExitWithoutReason
}
// Handle the error.
switch err.(type) {
case *platform.ExitPio:
err = model.HandlePio(vm, err.(*platform.ExitPio))
case *platform.ExitMmio:
err = model.HandleMmio(vm, err.(*platform.ExitMmio))
case *platform.ExitDebug:
err = nil
case *platform.ExitShutdown:
// Vcpu shutdown.
return nil
}
// Error handling the exit.
if err != nil {
return err
}
}
// Unreachable.
return nil
}
func SetupLinux(
vcpu *platform.Vcpu,
model *machine.Model,
orig_boot_data []byte,
entry_point uint64,
is_64bit bool,
initrd_addr platform.Paddr,
initrd_len uint64,
cmdline_addr platform.Paddr) error {
// Copy in the GDT table.
// These match the segments below.
gdt_addr, gdt, err := model.Allocate(
machine.MemoryTypeUser,
0, // Start.
model.Max(), // End.
platform.PageSize, // Size.
false) // From bottom.
if err != nil {
return err
}
if is_64bit {
C.build_64bit_gdt(unsafe.Pointer(&gdt[0]))
} else {
C.build_32bit_gdt(unsafe.Pointer(&gdt[0]))
}
BootGdt := platform.DescriptorValue{
Base: uint64(gdt_addr),
Limit: uint16(platform.PageSize),
}
err = vcpu.SetDescriptor(platform.GDT, BootGdt, true)
if err != nil {
return err
}
// Set a null IDT.
BootIdt := platform.DescriptorValue{
Base: 0,
Limit: 0,
}
err = vcpu.SetDescriptor(platform.IDT, BootIdt, true)
if err != nil {
return err
}
// Enable protected-mode.
// This does not set any flags (e.g. paging) beyond the
// protected mode flag. This is according to Linux entry
// protocol for 32-bit protected mode.
cr0, err := vcpu.GetControlRegister(platform.CR0)
if err != nil {
return err
}
cr0 = cr0 | (1 << 0) // Protected mode.
err = vcpu.SetControlRegister(platform.CR0, cr0, true)
if err != nil {
return err
}
// Always have the PAE bit set.
cr4, err := vcpu.GetControlRegister(platform.CR4)
if err != nil {
return err
}
cr4 = cr4 | (1 << 5) // PAE enabled.
err = vcpu.SetControlRegister(platform.CR4, cr4, true)
if err != nil {
return err
}
// For 64-bit kernels, we need to enable long mode,
// and load an identity page table. This will require
// only a page of pages, as we use huge page sizes.
if is_64bit {
// Create our page tables.
pde_addr, pde, err := model.Allocate(
machine.MemoryTypeUser,
0, // Start.
model.Max(), // End.
platform.PageSize, // Size.
false) // From bottom.
if err != nil {
return err
}
pgd_addr, pgd, err := model.Allocate(
machine.MemoryTypeUser,
0, // Start.
model.Max(), // End.
platform.PageSize, // Size.
false) // From bottom.
if err != nil {
return err
}
pml4_addr, pml4, err := model.Allocate(
machine.MemoryTypeUser,
0, // Start.
model.Max(), // End.
platform.PageSize, // Size.
false) // From bottom.
if err != nil {
return err
}
C.build_pde(unsafe.Pointer(&pde[0]), platform.PageSize)
C.build_pgd(unsafe.Pointer(&pgd[0]), C.__u64(pde_addr), platform.PageSize)
C.build_pml4(unsafe.Pointer(&pml4[0]), C.__u64(pgd_addr), platform.PageSize)
log.Printf("loader: Created PDE @ %08x.", pde_addr)
log.Printf("loader: Created PGD @ %08x.", pgd_addr)
log.Printf("loader: Created PML4 @ %08x.", pml4_addr)
// Set our newly build page table.
err = vcpu.SetControlRegister(
platform.CR3,
platform.ControlRegisterValue(pml4_addr),
true)
if err != nil {
return err
}
// Enable long mode.
efer, err := vcpu.GetControlRegister(platform.EFER)
if err != nil {
return err
}
efer = efer | (1 << 8) // Long-mode enable.
err = vcpu.SetControlRegister(platform.EFER, efer, true)
if err != nil {
return err
}
// Enable paging.
cr0, err = vcpu.GetControlRegister(platform.CR0)
if err != nil {
return err
}
cr0 = cr0 | (1 << 31) // Paging enable.
err = vcpu.SetControlRegister(platform.CR0, cr0, true)
if err != nil {
return err
}
}
// NOTE: For 64-bit kernels, we need to enable
// real 64-bit mode. This means that the L bit in
// the segments must be one, the Db bit must be
// zero, and we set the LME bit in EFER (above).
var lVal uint8
var dVal uint8
if is_64bit {
lVal = 1
dVal = 0
} else {
lVal = 0
dVal = 1
}
// Load the VMCS segments.
//
// NOTE: These values are loaded into the VMCS
// registers and are expected to match the descriptors
// we've used above. Unfortunately the API format doesn't
// match, so we need to duplicate some work here. Ah, well
// at least the below serves as an explanation for what
// the magic numbers in GDT_ENTRY() above mean.
BootCs := platform.SegmentValue{
Base: 0,
Limit: 0xffffffff,
Selector: uint16(C.BootCsSelector), // @ 0x10
Dpl: 0, // Privilege level (kernel).
Db: dVal, // 32-bit segment?
G: 1, // Granularity (page).
S: 1, // As per BOOT_CS (code/data).
L: lVal, // 64-bit extension.
Type: 0xb, // As per BOOT_CS (access must be set).
Present: 1,
}
BootDs := platform.SegmentValue{
Base: 0,
Limit: 0xffffffff,
Selector: uint16(C.BootDsSelector), // @ 0x18
Dpl: 0, // Privilege level (kernel).
Db: 1, // 32-bit segment?
G: 1, // Granularity (page).
S: 1, // As per BOOT_DS (code/data).
L: 0, // 64-bit extension.
Type: 0x3, // As per BOOT_DS (access must be set).
Present: 1,
}
BootTr := platform.SegmentValue{
Base: 0,
Limit: 0xffffffff,
Selector: uint16(C.BootTrSelector), // @ 0x20
Dpl: 0, // Privilege level (kernel).
Db: 1, // 32-bit segment?
G: 1, // Granularity (page).
S: 0, // As per BOOT_TR (system).
L: 0, // 64-bit extension.
Type: 0xb, // As per BOOT_TR.
Present: 1,
}
err = vcpu.SetSegment(platform.CS, BootCs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.DS, BootDs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.ES, BootDs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.FS, BootDs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.GS, BootDs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.SS, BootDs, true)
if err != nil {
return err
}
err = vcpu.SetSegment(platform.TR, BootTr, true)
if err != nil {
return err
}
// Create our boot parameters.
boot_addr, boot_data, err := model.Allocate(
machine.MemoryTypeUser,
0, // Start.
model.Max(), // End.
platform.PageSize, // Size.
false) // From bottom.
if err != nil {
return err
}
err = SetupLinuxBootParams(
model,
boot_data,
orig_boot_data,
cmdline_addr,
initrd_addr,
initrd_len)
if err != nil {
return err
}
// Set our registers.
// This is according to the Linux 32-bit boot protocol.
log.Printf("loader: boot_params @ %08x.", boot_addr)
err = vcpu.SetRegister(platform.RSI, platform.RegisterValue(boot_addr))
if err != nil {
return err
}
err = vcpu.SetRegister(platform.RBP, 0)
if err != nil {
return err
}
err = vcpu.SetRegister(platform.RDI, 0)
if err != nil {
return err
}
err = vcpu.SetRegister(platform.RBX, 0)
if err != nil {
return err
}
// Jump to our entry point.
err = vcpu.SetRegister(platform.RIP, platform.RegisterValue(entry_point))
if err != nil {
return err
}
// Make sure interrupts are disabled.
// This will actually clear out all other flags.
rflags, err := vcpu.GetRegister(platform.RFLAGS)
if err != nil {
return err
}
rflags = rflags &^ (1 << 9) // Interrupts off.
rflags = rflags | (1 << 1) // Reserved 1.
err = vcpu.SetRegister(
platform.RFLAGS,
platform.RegisterValue(rflags))
if err != nil {
return err
}
// We're done.
return nil
}
func LoadLinux(
vcpu *platform.Vcpu,
model *machine.Model,
boot_params string,
vmlinux string,
initrd string,
cmdline string,
system_map string) (SystemMap, *Convention, error) {
// Read the boot_params.
log.Print("loader: Reading kernel image...")
kernel_data, err := ioutil.ReadFile(boot_params)
log.Printf("loader: Kernel is %d bytes.", len(kernel_data))
if err != nil {
return nil, nil, err
}
// They may have passed the entire vmlinuz image as the
// parameter here. That's okay, we do an efficient mmap
// above. But we need to truncate the visible slice.
boot_params_data := kernel_data[0:platform.PageSize]
// Load the kernel.
log.Print("loader: Reading kernel binary...")
vmlinux_data, err := ioutil.ReadFile(vmlinux)
log.Printf("loader: Kernel binary is %d bytes.", len(vmlinux_data))
if err != nil {
return nil, nil, err
}
// Load the ramdisk.
log.Print("loader: Reading ramdisk...")
initrd_data, err := ioutil.ReadFile(initrd)
log.Printf("loader: Ramdisk is %d bytes.", len(initrd_data))
if err != nil {
return nil, nil, err
}
// Load the system map.
log.Print("loader: Loading system map...")
sysmap, err := LoadLinuxSystemMap(system_map)
if err != nil {
return nil, nil, err
}
// Load the kernel into memory.
log.Print("loader: Loading kernel...")
entry_point, is_64bit, err := ElfLoad(vmlinux_data, model)
if err != nil {
return nil, nil, err
}
if is_64bit {
log.Print("loader: 64-bit kernel found.")
} else {
log.Print("loader: 32-bit kernel found.")
}
log.Printf("loader: Entry point is %08x.", entry_point)
// Set our calling convention.
var convention *Convention
if is_64bit {
convention = &Linux64Convention
} else {
convention = &Linux32Convention
}
// Load the cmdline.
// NOTE: Here we create a full page with
// trailing zeros. This is the expected form
// for the command line.
full_cmdline := make(
[]byte,
platform.PageSize,
platform.PageSize)
copy(full_cmdline, []byte(cmdline))
cmdline_addr, err := model.MemoryMap.Load(
platform.Paddr(0),
model.Max(),
full_cmdline,
false)
if err != nil {
return nil, nil, err
}
log.Printf("loader: cmdline @ %08x: %s",
cmdline_addr,
cmdline)
// Load the initrd.
initrd_addr, err := model.MemoryMap.Load(
platform.Paddr(0),
model.Max(),
initrd_data,
true)
if err != nil {
return nil, nil, err
}
log.Printf("loader: initrd @ %08x.", initrd_addr)
// Create our setup page,
// and initialize the VCPU.
err = SetupLinux(
vcpu,
model,
boot_params_data,
entry_point,
is_64bit,
initrd_addr,
uint64(len(initrd_data)),
cmdline_addr)
// Everything is okay.
return sysmap, convention, err
}
