diff options
Diffstat (limited to 'pkg/ring0')
49 files changed, 7673 insertions, 0 deletions
diff --git a/pkg/ring0/defs_impl_amd64.go b/pkg/ring0/defs_impl_amd64.go new file mode 100644 index 000000000..933018c52 --- /dev/null +++ b/pkg/ring0/defs_impl_amd64.go @@ -0,0 +1,597 @@ +// +build amd64 +// +build amd64 +// +build 386 amd64 + +package ring0 + +import ( + "fmt" + "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/ring0/pagetables" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/usermem" + "io" + "reflect" +) + +// Kernel is a global kernel object. +// +// This contains global state, shared by multiple CPUs. +type Kernel struct { + // PageTables are the kernel pagetables; this must be provided. + PageTables *pagetables.PageTables + + KernelArchState +} + +// Hooks are hooks for kernel functions. +type Hooks interface { + // KernelSyscall is called for kernel system calls. + // + // Return from this call will restore registers and return to the kernel: the + // registers must be modified directly. + // + // If this function is not provided, a kernel exception results in halt. + // + // This must be go:nosplit, as this will be on the interrupt stack. + // Closures are permitted, as the pointer to the closure frame is not + // passed on the stack. + KernelSyscall() + + // KernelException handles an exception during kernel execution. + // + // Return from this call will restore registers and return to the kernel: the + // registers must be modified directly. + // + // If this function is not provided, a kernel exception results in halt. + // + // This must be go:nosplit, as this will be on the interrupt stack. + // Closures are permitted, as the pointer to the closure frame is not + // passed on the stack. + KernelException(Vector) +} + +// CPU is the per-CPU struct. +type CPU struct { + // self is a self reference. + // + // This is always guaranteed to be at offset zero. + self *CPU + + // kernel is reference to the kernel that this CPU was initialized + // with. This reference is kept for garbage collection purposes: CPU + // registers may refer to objects within the Kernel object that cannot + // be safely freed. + kernel *Kernel + + // CPUArchState is architecture-specific state. + CPUArchState + + // registers is a set of registers; these may be used on kernel system + // calls and exceptions via the Registers function. + registers arch.Registers + + // hooks are kernel hooks. + hooks Hooks +} + +// Registers returns a modifiable-copy of the kernel registers. +// +// This is explicitly safe to call during KernelException and KernelSyscall. +// +//go:nosplit +func (c *CPU) Registers() *arch.Registers { + return &c.registers +} + +// SwitchOpts are passed to the Switch function. +type SwitchOpts struct { + // Registers are the user register state. + Registers *arch.Registers + + // FloatingPointState is a byte pointer where floating point state is + // saved and restored. + FloatingPointState *byte + + // PageTables are the application page tables. + PageTables *pagetables.PageTables + + // Flush indicates that a TLB flush should be forced on switch. + Flush bool + + // FullRestore indicates that an iret-based restore should be used. + FullRestore bool + + // SwitchArchOpts are architecture-specific options. + SwitchArchOpts +} + +var ( + // UserspaceSize is the total size of userspace. + UserspaceSize = uintptr(1) << (VirtualAddressBits() - 1) + + // MaximumUserAddress is the largest possible user address. + MaximumUserAddress = (UserspaceSize - 1) & ^uintptr(usermem.PageSize-1) + + // KernelStartAddress is the starting kernel address. + KernelStartAddress = ^uintptr(0) - (UserspaceSize - 1) +) + +// Segment indices and Selectors. +const ( + // Index into GDT array. + _ = iota // Null descriptor first. + _ // Reserved (Linux is kernel 32). + segKcode // Kernel code (64-bit). + segKdata // Kernel data. + segUcode32 // User code (32-bit). + segUdata // User data. + segUcode64 // User code (64-bit). + segTss // Task segment descriptor. + segTssHi // Upper bits for TSS. + segLast // Last segment (terminal, not included). +) + +// Selectors. +const ( + Kcode Selector = segKcode << 3 + Kdata Selector = segKdata << 3 + Ucode32 Selector = (segUcode32 << 3) | 3 + Udata Selector = (segUdata << 3) | 3 + Ucode64 Selector = (segUcode64 << 3) | 3 + Tss Selector = segTss << 3 +) + +// Standard segments. +var ( + UserCodeSegment32 SegmentDescriptor + UserDataSegment SegmentDescriptor + UserCodeSegment64 SegmentDescriptor + KernelCodeSegment SegmentDescriptor + KernelDataSegment SegmentDescriptor +) + +// KernelArchState contains architecture-specific state. +type KernelArchState struct { + // cpuEntries is array of kernelEntry for all cpus. + cpuEntries []kernelEntry + + // globalIDT is our set of interrupt gates. + globalIDT *idt64 +} + +// kernelEntry contains minimal CPU-specific arch state +// that can be mapped at the upper of the address space. +// Malicious APP might steal info from it via CPU bugs. +type kernelEntry struct { + // stack is the stack used for interrupts on this CPU. + stack [256]byte + + // scratch space for temporary usage. + scratch0 uint64 + + // stackTop is the top of the stack. + stackTop uint64 + + // cpuSelf is back reference to CPU. + cpuSelf *CPU + + // kernelCR3 is the cr3 used for sentry kernel. + kernelCR3 uintptr + + // gdt is the CPU's descriptor table. + gdt descriptorTable + + // tss is the CPU's task state. + tss TaskState64 +} + +// CPUArchState contains CPU-specific arch state. +type CPUArchState struct { + // errorCode is the error code from the last exception. + errorCode uintptr + + // errorType indicates the type of error code here, it is always set + // along with the errorCode value above. + // + // It will either by 1, which indicates a user error, or 0 indicating a + // kernel error. If the error code below returns false (kernel error), + // then it cannot provide relevant information about the last + // exception. + errorType uintptr + + *kernelEntry +} + +// ErrorCode returns the last error code. +// +// The returned boolean indicates whether the error code corresponds to the +// last user error or not. If it does not, then fault information must be +// ignored. This is generally the result of a kernel fault while servicing a +// user fault. +// +//go:nosplit +func (c *CPU) ErrorCode() (value uintptr, user bool) { + return c.errorCode, c.errorType != 0 +} + +// ClearErrorCode resets the error code. +// +//go:nosplit +func (c *CPU) ClearErrorCode() { + c.errorCode = 0 + c.errorType = 1 +} + +// SwitchArchOpts are embedded in SwitchOpts. +type SwitchArchOpts struct { + // UserPCID indicates that the application PCID to be used on switch, + // assuming that PCIDs are supported. + // + // Per pagetables_x86.go, a zero PCID implies a flush. + UserPCID uint16 + + // KernelPCID indicates that the kernel PCID to be used on return, + // assuming that PCIDs are supported. + // + // Per pagetables_x86.go, a zero PCID implies a flush. + KernelPCID uint16 +} + +func init() { + KernelCodeSegment.setCode64(0, 0, 0) + KernelDataSegment.setData(0, 0xffffffff, 0) + UserCodeSegment32.setCode64(0, 0, 3) + UserDataSegment.setData(0, 0xffffffff, 3) + UserCodeSegment64.setCode64(0, 0, 3) +} + +// Emit prints architecture-specific offsets. +func Emit(w io.Writer) { + fmt.Fprintf(w, "// Automatically generated, do not edit.\n") + + c := &CPU{} + fmt.Fprintf(w, "\n// CPU offsets.\n") + fmt.Fprintf(w, "#define CPU_REGISTERS 0x%02x\n", reflect.ValueOf(&c.registers).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_ERROR_CODE 0x%02x\n", reflect.ValueOf(&c.errorCode).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_ERROR_TYPE 0x%02x\n", reflect.ValueOf(&c.errorType).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_ENTRY 0x%02x\n", reflect.ValueOf(&c.kernelEntry).Pointer()-reflect.ValueOf(c).Pointer()) + + e := &kernelEntry{} + fmt.Fprintf(w, "\n// CPU entry offsets.\n") + fmt.Fprintf(w, "#define ENTRY_SCRATCH0 0x%02x\n", reflect.ValueOf(&e.scratch0).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_STACK_TOP 0x%02x\n", reflect.ValueOf(&e.stackTop).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_CPU_SELF 0x%02x\n", reflect.ValueOf(&e.cpuSelf).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_KERNEL_CR3 0x%02x\n", reflect.ValueOf(&e.kernelCR3).Pointer()-reflect.ValueOf(e).Pointer()) + + fmt.Fprintf(w, "\n// Bits.\n") + fmt.Fprintf(w, "#define _RFLAGS_IF 0x%02x\n", _RFLAGS_IF) + fmt.Fprintf(w, "#define _RFLAGS_IOPL0 0x%02x\n", _RFLAGS_IOPL0) + fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) + + fmt.Fprintf(w, "\n// Vectors.\n") + fmt.Fprintf(w, "#define DivideByZero 0x%02x\n", DivideByZero) + fmt.Fprintf(w, "#define Debug 0x%02x\n", Debug) + fmt.Fprintf(w, "#define NMI 0x%02x\n", NMI) + fmt.Fprintf(w, "#define Breakpoint 0x%02x\n", Breakpoint) + fmt.Fprintf(w, "#define Overflow 0x%02x\n", Overflow) + fmt.Fprintf(w, "#define BoundRangeExceeded 0x%02x\n", BoundRangeExceeded) + fmt.Fprintf(w, "#define InvalidOpcode 0x%02x\n", InvalidOpcode) + fmt.Fprintf(w, "#define DeviceNotAvailable 0x%02x\n", DeviceNotAvailable) + fmt.Fprintf(w, "#define DoubleFault 0x%02x\n", DoubleFault) + fmt.Fprintf(w, "#define CoprocessorSegmentOverrun 0x%02x\n", CoprocessorSegmentOverrun) + fmt.Fprintf(w, "#define InvalidTSS 0x%02x\n", InvalidTSS) + fmt.Fprintf(w, "#define SegmentNotPresent 0x%02x\n", SegmentNotPresent) + fmt.Fprintf(w, "#define StackSegmentFault 0x%02x\n", StackSegmentFault) + fmt.Fprintf(w, "#define GeneralProtectionFault 0x%02x\n", GeneralProtectionFault) + fmt.Fprintf(w, "#define PageFault 0x%02x\n", PageFault) + fmt.Fprintf(w, "#define X87FloatingPointException 0x%02x\n", X87FloatingPointException) + fmt.Fprintf(w, "#define AlignmentCheck 0x%02x\n", AlignmentCheck) + fmt.Fprintf(w, "#define MachineCheck 0x%02x\n", MachineCheck) + fmt.Fprintf(w, "#define SIMDFloatingPointException 0x%02x\n", SIMDFloatingPointException) + fmt.Fprintf(w, "#define VirtualizationException 0x%02x\n", VirtualizationException) + fmt.Fprintf(w, "#define SecurityException 0x%02x\n", SecurityException) + fmt.Fprintf(w, "#define SyscallInt80 0x%02x\n", SyscallInt80) + fmt.Fprintf(w, "#define Syscall 0x%02x\n", Syscall) + + p := &arch.Registers{} + fmt.Fprintf(w, "\n// Ptrace registers.\n") + fmt.Fprintf(w, "#define PTRACE_R15 0x%02x\n", reflect.ValueOf(&p.R15).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R14 0x%02x\n", reflect.ValueOf(&p.R14).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R13 0x%02x\n", reflect.ValueOf(&p.R13).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R12 0x%02x\n", reflect.ValueOf(&p.R12).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RBP 0x%02x\n", reflect.ValueOf(&p.Rbp).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RBX 0x%02x\n", reflect.ValueOf(&p.Rbx).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R11 0x%02x\n", reflect.ValueOf(&p.R11).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R10 0x%02x\n", reflect.ValueOf(&p.R10).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R9 0x%02x\n", reflect.ValueOf(&p.R9).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R8 0x%02x\n", reflect.ValueOf(&p.R8).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RAX 0x%02x\n", reflect.ValueOf(&p.Rax).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RCX 0x%02x\n", reflect.ValueOf(&p.Rcx).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RDX 0x%02x\n", reflect.ValueOf(&p.Rdx).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RSI 0x%02x\n", reflect.ValueOf(&p.Rsi).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RDI 0x%02x\n", reflect.ValueOf(&p.Rdi).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_ORIGRAX 0x%02x\n", reflect.ValueOf(&p.Orig_rax).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RIP 0x%02x\n", reflect.ValueOf(&p.Rip).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_CS 0x%02x\n", reflect.ValueOf(&p.Cs).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_FLAGS 0x%02x\n", reflect.ValueOf(&p.Eflags).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_RSP 0x%02x\n", reflect.ValueOf(&p.Rsp).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_SS 0x%02x\n", reflect.ValueOf(&p.Ss).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_FS 0x%02x\n", reflect.ValueOf(&p.Fs_base).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_GS 0x%02x\n", reflect.ValueOf(&p.Gs_base).Pointer()-reflect.ValueOf(p).Pointer()) +} + +// Useful bits. +const ( + _CR0_PE = 1 << 0 + _CR0_ET = 1 << 4 + _CR0_AM = 1 << 18 + _CR0_PG = 1 << 31 + + _CR4_PSE = 1 << 4 + _CR4_PAE = 1 << 5 + _CR4_PGE = 1 << 7 + _CR4_OSFXSR = 1 << 9 + _CR4_OSXMMEXCPT = 1 << 10 + _CR4_FSGSBASE = 1 << 16 + _CR4_PCIDE = 1 << 17 + _CR4_OSXSAVE = 1 << 18 + _CR4_SMEP = 1 << 20 + + _RFLAGS_AC = 1 << 18 + _RFLAGS_NT = 1 << 14 + _RFLAGS_IOPL0 = 1 << 12 + _RFLAGS_IOPL1 = 1 << 13 + _RFLAGS_IOPL = _RFLAGS_IOPL0 | _RFLAGS_IOPL1 + _RFLAGS_DF = 1 << 10 + _RFLAGS_IF = 1 << 9 + _RFLAGS_STEP = 1 << 8 + _RFLAGS_RESERVED = 1 << 1 + + _EFER_SCE = 0x001 + _EFER_LME = 0x100 + _EFER_LMA = 0x400 + _EFER_NX = 0x800 + + _MSR_STAR = 0xc0000081 + _MSR_LSTAR = 0xc0000082 + _MSR_CSTAR = 0xc0000083 + _MSR_SYSCALL_MASK = 0xc0000084 + _MSR_PLATFORM_INFO = 0xce + _MSR_MISC_FEATURES = 0x140 + + _PLATFORM_INFO_CPUID_FAULT = 1 << 31 + + _MISC_FEATURE_CPUID_TRAP = 0x1 +) + +const ( + // KernelFlagsSet should always be set in the kernel. + KernelFlagsSet = _RFLAGS_RESERVED + + // UserFlagsSet are always set in userspace. + // + // _RFLAGS_IOPL is a set of two bits and it shows the I/O privilege + // level. The Current Privilege Level (CPL) of the task must be less + // than or equal to the IOPL in order for the task or program to access + // I/O ports. + // + // Here, _RFLAGS_IOPL0 is used only to determine whether the task is + // running in the kernel or userspace mode. In the user mode, the CPL is + // always 3 and it doesn't matter what IOPL is set if it is bellow CPL. + // + // We need to have one bit which will be always different in user and + // kernel modes. And we have to remember that even though we have + // KernelFlagsClear, we still can see some of these flags in the kernel + // mode. This can happen when the goruntime switches on a goroutine + // which has been saved in the host mode. On restore, the popf + // instruction is used to restore flags and this means that all flags + // what the goroutine has in the host mode will be restored in the + // kernel mode. + // + // _RFLAGS_IOPL0 is never set in host and kernel modes and we always set + // it in the user mode. So if this flag is set, the task is running in + // the user mode and if it isn't set, the task is running in the kernel + // mode. + UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF | _RFLAGS_IOPL0 + + // KernelFlagsClear should always be clear in the kernel. + KernelFlagsClear = _RFLAGS_STEP | _RFLAGS_IF | _RFLAGS_IOPL | _RFLAGS_AC | _RFLAGS_NT + + // UserFlagsClear are always cleared in userspace. + UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL1 +) + +// IsKernelFlags returns true if rflags coresponds to the kernel mode. +// +// go:nosplit +func IsKernelFlags(rflags uint64) bool { + return rflags&_RFLAGS_IOPL0 == 0 +} + +// Vector is an exception vector. +type Vector uintptr + +// Exception vectors. +const ( + DivideByZero Vector = iota + Debug + NMI + Breakpoint + Overflow + BoundRangeExceeded + InvalidOpcode + DeviceNotAvailable + DoubleFault + CoprocessorSegmentOverrun + InvalidTSS + SegmentNotPresent + StackSegmentFault + GeneralProtectionFault + PageFault + _ + X87FloatingPointException + AlignmentCheck + MachineCheck + SIMDFloatingPointException + VirtualizationException + SecurityException = 0x1e + SyscallInt80 = 0x80 + _NR_INTERRUPTS = 0x100 +) + +// System call vectors. +const ( + Syscall Vector = _NR_INTERRUPTS +) + +// VirtualAddressBits returns the number bits available for virtual addresses. +// +// Note that sign-extension semantics apply to the highest order bit. +// +// FIXME(b/69382326): This should use the cpuid passed to Init. +func VirtualAddressBits() uint32 { + ax, _, _, _ := cpuid.HostID(0x80000008, 0) + return (ax >> 8) & 0xff +} + +// PhysicalAddressBits returns the number of bits available for physical addresses. +// +// FIXME(b/69382326): This should use the cpuid passed to Init. +func PhysicalAddressBits() uint32 { + ax, _, _, _ := cpuid.HostID(0x80000008, 0) + return ax & 0xff +} + +// Selector is a segment Selector. +type Selector uint16 + +// SegmentDescriptor is a segment descriptor. +type SegmentDescriptor struct { + bits [2]uint32 +} + +// descriptorTable is a collection of descriptors. +type descriptorTable [32]SegmentDescriptor + +// SegmentDescriptorFlags are typed flags within a descriptor. +type SegmentDescriptorFlags uint32 + +// SegmentDescriptorFlag declarations. +const ( + SegmentDescriptorAccess SegmentDescriptorFlags = 1 << 8 // Access bit (always set). + SegmentDescriptorWrite = 1 << 9 // Write permission. + SegmentDescriptorExpandDown = 1 << 10 // Grows down, not used. + SegmentDescriptorExecute = 1 << 11 // Execute permission. + SegmentDescriptorSystem = 1 << 12 // Zero => system, 1 => user code/data. + SegmentDescriptorPresent = 1 << 15 // Present. + SegmentDescriptorAVL = 1 << 20 // Available. + SegmentDescriptorLong = 1 << 21 // Long mode. + SegmentDescriptorDB = 1 << 22 // 16 or 32-bit. + SegmentDescriptorG = 1 << 23 // Granularity: page or byte. +) + +// Base returns the descriptor's base linear address. +func (d *SegmentDescriptor) Base() uint32 { + return d.bits[1]&0xFF000000 | (d.bits[1]&0x000000FF)<<16 | d.bits[0]>>16 +} + +// Limit returns the descriptor size. +func (d *SegmentDescriptor) Limit() uint32 { + l := d.bits[0]&0xFFFF | d.bits[1]&0xF0000 + if d.bits[1]&uint32(SegmentDescriptorG) != 0 { + l <<= 12 + l |= 0xFFF + } + return l +} + +// Flags returns descriptor flags. +func (d *SegmentDescriptor) Flags() SegmentDescriptorFlags { + return SegmentDescriptorFlags(d.bits[1] & 0x00F09F00) +} + +// DPL returns the descriptor privilege level. +func (d *SegmentDescriptor) DPL() int { + return int((d.bits[1] >> 13) & 3) +} + +func (d *SegmentDescriptor) setNull() { + d.bits[0] = 0 + d.bits[1] = 0 +} + +func (d *SegmentDescriptor) set(base, limit uint32, dpl int, flags SegmentDescriptorFlags) { + flags |= SegmentDescriptorPresent + if limit>>12 != 0 { + limit >>= 12 + flags |= SegmentDescriptorG + } + d.bits[0] = base<<16 | limit&0xFFFF + d.bits[1] = base&0xFF000000 | (base>>16)&0xFF | limit&0x000F0000 | uint32(flags) | uint32(dpl)<<13 +} + +func (d *SegmentDescriptor) setCode32(base, limit uint32, dpl int) { + d.set(base, limit, dpl, + SegmentDescriptorDB| + SegmentDescriptorExecute| + SegmentDescriptorSystem) +} + +func (d *SegmentDescriptor) setCode64(base, limit uint32, dpl int) { + d.set(base, limit, dpl, + SegmentDescriptorG| + SegmentDescriptorLong| + SegmentDescriptorExecute| + SegmentDescriptorSystem) +} + +func (d *SegmentDescriptor) setData(base, limit uint32, dpl int) { + d.set(base, limit, dpl, + SegmentDescriptorWrite| + SegmentDescriptorSystem) +} + +// setHi is only used for the TSS segment, which is magically 64-bits. +func (d *SegmentDescriptor) setHi(base uint32) { + d.bits[0] = base + d.bits[1] = 0 +} + +// Gate64 is a 64-bit task, trap, or interrupt gate. +type Gate64 struct { + bits [4]uint32 +} + +// idt64 is a 64-bit interrupt descriptor table. +type idt64 [_NR_INTERRUPTS]Gate64 + +func (g *Gate64) setInterrupt(cs Selector, rip uint64, dpl int, ist int) { + g.bits[0] = uint32(cs)<<16 | uint32(rip)&0xFFFF + g.bits[1] = uint32(rip)&0xFFFF0000 | SegmentDescriptorPresent | uint32(dpl)<<13 | 14<<8 | uint32(ist)&0x7 + g.bits[2] = uint32(rip >> 32) +} + +func (g *Gate64) setTrap(cs Selector, rip uint64, dpl int, ist int) { + g.setInterrupt(cs, rip, dpl, ist) + g.bits[1] |= 1 << 8 +} + +// TaskState64 is a 64-bit task state structure. +type TaskState64 struct { + _ uint32 + rsp0Lo, rsp0Hi uint32 + rsp1Lo, rsp1Hi uint32 + rsp2Lo, rsp2Hi uint32 + _ [2]uint32 + ist1Lo, ist1Hi uint32 + ist2Lo, ist2Hi uint32 + ist3Lo, ist3Hi uint32 + ist4Lo, ist4Hi uint32 + ist5Lo, ist5Hi uint32 + ist6Lo, ist6Hi uint32 + ist7Lo, ist7Hi uint32 + _ [2]uint32 + _ uint16 + ioPerm uint16 +} diff --git a/pkg/ring0/defs_impl_arm64.go b/pkg/ring0/defs_impl_arm64.go new file mode 100644 index 000000000..bd18a11e7 --- /dev/null +++ b/pkg/ring0/defs_impl_arm64.go @@ -0,0 +1,427 @@ +// +build arm64 +// +build arm64 +// +build arm64 + +package ring0 + +import ( + "fmt" + "gvisor.dev/gvisor/pkg/ring0/pagetables" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/usermem" + "io" + "reflect" +) + +// Useful bits. +const ( + _PGD_PGT_BASE = 0x1000 + _PGD_PGT_SIZE = 0x1000 + _PUD_PGT_BASE = 0x2000 + _PUD_PGT_SIZE = 0x1000 + _PMD_PGT_BASE = 0x3000 + _PMD_PGT_SIZE = 0x4000 + _PTE_PGT_BASE = 0x7000 + _PTE_PGT_SIZE = 0x1000 +) + +const ( + // DAIF bits:debug, sError, IRQ, FIQ. + _PSR_D_BIT = 0x00000200 + _PSR_A_BIT = 0x00000100 + _PSR_I_BIT = 0x00000080 + _PSR_F_BIT = 0x00000040 + _PSR_DAIF_SHIFT = 6 + _PSR_DAIF_MASK = 0xf << _PSR_DAIF_SHIFT + + // PSR bits. + _PSR_MODE_EL0t = 0x00000000 + _PSR_MODE_EL1t = 0x00000004 + _PSR_MODE_EL1h = 0x00000005 + _PSR_MODE_MASK = 0x0000000f + + PsrFlagsClear = _PSR_MODE_MASK | _PSR_DAIF_MASK + PsrModeMask = _PSR_MODE_MASK + + // KernelFlagsSet should always be set in the kernel. + KernelFlagsSet = _PSR_MODE_EL1h | _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT + + // UserFlagsSet are always set in userspace. + UserFlagsSet = _PSR_MODE_EL0t +) + +// Vector is an exception vector. +type Vector uintptr + +// Exception vectors. +const ( + El1InvSync = iota + El1InvIrq + El1InvFiq + El1InvError + + El1Sync + El1Irq + El1Fiq + El1Err + + El0Sync + El0Irq + El0Fiq + El0Err + + El0InvSync + El0InvIrq + El0InvFiq + El0InvErr + + El1SyncDa + El1SyncIa + El1SyncSpPc + El1SyncUndef + El1SyncDbg + El1SyncInv + + El0SyncSVC + El0SyncDa + El0SyncIa + El0SyncFpsimdAcc + El0SyncSveAcc + El0SyncFpsimdExc + El0SyncSys + El0SyncSpPc + El0SyncUndef + El0SyncDbg + El0SyncWfx + El0SyncInv + + El0ErrNMI + El0ErrBounce + + _NR_INTERRUPTS +) + +// System call vectors. +const ( + Syscall Vector = El0SyncSVC + PageFault Vector = El0SyncDa + VirtualizationException Vector = El0ErrBounce +) + +// VirtualAddressBits returns the number bits available for virtual addresses. +func VirtualAddressBits() uint32 { + return 48 +} + +// PhysicalAddressBits returns the number of bits available for physical addresses. +func PhysicalAddressBits() uint32 { + return 40 +} + +// Kernel is a global kernel object. +// +// This contains global state, shared by multiple CPUs. +type Kernel struct { + // PageTables are the kernel pagetables; this must be provided. + PageTables *pagetables.PageTables + + KernelArchState +} + +// Hooks are hooks for kernel functions. +type Hooks interface { + // KernelSyscall is called for kernel system calls. + // + // Return from this call will restore registers and return to the kernel: the + // registers must be modified directly. + // + // If this function is not provided, a kernel exception results in halt. + // + // This must be go:nosplit, as this will be on the interrupt stack. + // Closures are permitted, as the pointer to the closure frame is not + // passed on the stack. + KernelSyscall() + + // KernelException handles an exception during kernel execution. + // + // Return from this call will restore registers and return to the kernel: the + // registers must be modified directly. + // + // If this function is not provided, a kernel exception results in halt. + // + // This must be go:nosplit, as this will be on the interrupt stack. + // Closures are permitted, as the pointer to the closure frame is not + // passed on the stack. + KernelException(Vector) +} + +// CPU is the per-CPU struct. +type CPU struct { + // self is a self reference. + // + // This is always guaranteed to be at offset zero. + self *CPU + + // kernel is reference to the kernel that this CPU was initialized + // with. This reference is kept for garbage collection purposes: CPU + // registers may refer to objects within the Kernel object that cannot + // be safely freed. + kernel *Kernel + + // CPUArchState is architecture-specific state. + CPUArchState + + // registers is a set of registers; these may be used on kernel system + // calls and exceptions via the Registers function. + registers arch.Registers + + // hooks are kernel hooks. + hooks Hooks +} + +// Registers returns a modifiable-copy of the kernel registers. +// +// This is explicitly safe to call during KernelException and KernelSyscall. +// +//go:nosplit +func (c *CPU) Registers() *arch.Registers { + return &c.registers +} + +// SwitchOpts are passed to the Switch function. +type SwitchOpts struct { + // Registers are the user register state. + Registers *arch.Registers + + // FloatingPointState is a byte pointer where floating point state is + // saved and restored. + FloatingPointState *byte + + // PageTables are the application page tables. + PageTables *pagetables.PageTables + + // Flush indicates that a TLB flush should be forced on switch. + Flush bool + + // FullRestore indicates that an iret-based restore should be used. + FullRestore bool + + // SwitchArchOpts are architecture-specific options. + SwitchArchOpts +} + +var ( + // UserspaceSize is the total size of userspace. + UserspaceSize = uintptr(1) << (VirtualAddressBits()) + + // MaximumUserAddress is the largest possible user address. + MaximumUserAddress = (UserspaceSize - 1) & ^uintptr(usermem.PageSize-1) + + // KernelStartAddress is the starting kernel address. + KernelStartAddress = ^uintptr(0) - (UserspaceSize - 1) +) + +// KernelArchState contains architecture-specific state. +type KernelArchState struct { +} + +// CPUArchState contains CPU-specific arch state. +type CPUArchState struct { + // stack is the stack used for interrupts on this CPU. + stack [512]byte + + // errorCode is the error code from the last exception. + errorCode uintptr + + // errorType indicates the type of error code here, it is always set + // along with the errorCode value above. + // + // It will either by 1, which indicates a user error, or 0 indicating a + // kernel error. If the error code below returns false (kernel error), + // then it cannot provide relevant information about the last + // exception. + errorType uintptr + + // faultAddr is the value of far_el1. + faultAddr uintptr + + // ttbr0Kvm is the value of ttbr0_el1 for sentry. + ttbr0Kvm uintptr + + // ttbr0App is the value of ttbr0_el1 for applicaton. + ttbr0App uintptr + + // exception vector. + vecCode Vector + + // application context pointer. + appAddr uintptr + + // lazyVFP is the value of cpacr_el1. + lazyVFP uintptr + + // appASID is the asid value of guest application. + appASID uintptr +} + +// ErrorCode returns the last error code. +// +// The returned boolean indicates whether the error code corresponds to the +// last user error or not. If it does not, then fault information must be +// ignored. This is generally the result of a kernel fault while servicing a +// user fault. +// +//go:nosplit +func (c *CPU) ErrorCode() (value uintptr, user bool) { + return c.errorCode, c.errorType != 0 +} + +// ClearErrorCode resets the error code. +// +//go:nosplit +func (c *CPU) ClearErrorCode() { + c.errorCode = 0 + c.errorType = 1 +} + +//go:nosplit +func (c *CPU) GetFaultAddr() (value uintptr) { + return c.faultAddr +} + +//go:nosplit +func (c *CPU) SetTtbr0Kvm(value uintptr) { + c.ttbr0Kvm = value +} + +//go:nosplit +func (c *CPU) SetTtbr0App(value uintptr) { + c.ttbr0App = value +} + +//go:nosplit +func (c *CPU) GetVector() (value Vector) { + return c.vecCode +} + +//go:nosplit +func (c *CPU) SetAppAddr(value uintptr) { + c.appAddr = value +} + +// GetLazyVFP returns the value of cpacr_el1. +//go:nosplit +func (c *CPU) GetLazyVFP() (value uintptr) { + return c.lazyVFP +} + +// SwitchArchOpts are embedded in SwitchOpts. +type SwitchArchOpts struct { + // UserASID indicates that the application ASID to be used on switch, + UserASID uint16 + + // KernelASID indicates that the kernel ASID to be used on return, + KernelASID uint16 +} + +func init() { +} + +// Emit prints architecture-specific offsets. +func Emit(w io.Writer) { + fmt.Fprintf(w, "// Automatically generated, do not edit.\n") + + c := &CPU{} + fmt.Fprintf(w, "\n// CPU offsets.\n") + fmt.Fprintf(w, "#define CPU_SELF 0x%02x\n", reflect.ValueOf(&c.self).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_REGISTERS 0x%02x\n", reflect.ValueOf(&c.registers).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_STACK_TOP 0x%02x\n", reflect.ValueOf(&c.stack[0]).Pointer()-reflect.ValueOf(c).Pointer()+uintptr(len(c.stack))) + fmt.Fprintf(w, "#define CPU_ERROR_CODE 0x%02x\n", reflect.ValueOf(&c.errorCode).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_ERROR_TYPE 0x%02x\n", reflect.ValueOf(&c.errorType).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_FAULT_ADDR 0x%02x\n", reflect.ValueOf(&c.faultAddr).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_TTBR0_KVM 0x%02x\n", reflect.ValueOf(&c.ttbr0Kvm).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_TTBR0_APP 0x%02x\n", reflect.ValueOf(&c.ttbr0App).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_VECTOR_CODE 0x%02x\n", reflect.ValueOf(&c.vecCode).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_APP_ADDR 0x%02x\n", reflect.ValueOf(&c.appAddr).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_LAZY_VFP 0x%02x\n", reflect.ValueOf(&c.lazyVFP).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_APP_ASID 0x%02x\n", reflect.ValueOf(&c.appASID).Pointer()-reflect.ValueOf(c).Pointer()) + + fmt.Fprintf(w, "\n// Bits.\n") + fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) + + fmt.Fprintf(w, "\n// Vectors.\n") + + fmt.Fprintf(w, "#define El1Sync 0x%02x\n", El1Sync) + fmt.Fprintf(w, "#define El1Irq 0x%02x\n", El1Irq) + fmt.Fprintf(w, "#define El1Fiq 0x%02x\n", El1Fiq) + fmt.Fprintf(w, "#define El1Err 0x%02x\n", El1Err) + + fmt.Fprintf(w, "#define El0Sync 0x%02x\n", El0Sync) + fmt.Fprintf(w, "#define El0Irq 0x%02x\n", El0Irq) + fmt.Fprintf(w, "#define El0Fiq 0x%02x\n", El0Fiq) + fmt.Fprintf(w, "#define El0Err 0x%02x\n", El0Err) + + fmt.Fprintf(w, "#define El1SyncDa 0x%02x\n", El1SyncDa) + fmt.Fprintf(w, "#define El1SyncIa 0x%02x\n", El1SyncIa) + fmt.Fprintf(w, "#define El1SyncSpPc 0x%02x\n", El1SyncSpPc) + fmt.Fprintf(w, "#define El1SyncUndef 0x%02x\n", El1SyncUndef) + fmt.Fprintf(w, "#define El1SyncDbg 0x%02x\n", El1SyncDbg) + fmt.Fprintf(w, "#define El1SyncInv 0x%02x\n", El1SyncInv) + + fmt.Fprintf(w, "#define El0SyncSVC 0x%02x\n", El0SyncSVC) + fmt.Fprintf(w, "#define El0SyncDa 0x%02x\n", El0SyncDa) + fmt.Fprintf(w, "#define El0SyncIa 0x%02x\n", El0SyncIa) + fmt.Fprintf(w, "#define El0SyncFpsimdAcc 0x%02x\n", El0SyncFpsimdAcc) + fmt.Fprintf(w, "#define El0SyncSveAcc 0x%02x\n", El0SyncSveAcc) + fmt.Fprintf(w, "#define El0SyncFpsimdExc 0x%02x\n", El0SyncFpsimdExc) + fmt.Fprintf(w, "#define El0SyncSys 0x%02x\n", El0SyncSys) + fmt.Fprintf(w, "#define El0SyncSpPc 0x%02x\n", El0SyncSpPc) + fmt.Fprintf(w, "#define El0SyncUndef 0x%02x\n", El0SyncUndef) + fmt.Fprintf(w, "#define El0SyncDbg 0x%02x\n", El0SyncDbg) + fmt.Fprintf(w, "#define El0SyncWfx 0x%02x\n", El0SyncWfx) + fmt.Fprintf(w, "#define El0SyncInv 0x%02x\n", El0SyncInv) + + fmt.Fprintf(w, "#define El0ErrNMI 0x%02x\n", El0ErrNMI) + + fmt.Fprintf(w, "#define PageFault 0x%02x\n", PageFault) + fmt.Fprintf(w, "#define Syscall 0x%02x\n", Syscall) + fmt.Fprintf(w, "#define VirtualizationException 0x%02x\n", VirtualizationException) + + p := &arch.Registers{} + fmt.Fprintf(w, "\n// Ptrace registers.\n") + fmt.Fprintf(w, "#define PTRACE_R0 0x%02x\n", reflect.ValueOf(&p.Regs[0]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R1 0x%02x\n", reflect.ValueOf(&p.Regs[1]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R2 0x%02x\n", reflect.ValueOf(&p.Regs[2]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R3 0x%02x\n", reflect.ValueOf(&p.Regs[3]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R4 0x%02x\n", reflect.ValueOf(&p.Regs[4]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R5 0x%02x\n", reflect.ValueOf(&p.Regs[5]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R6 0x%02x\n", reflect.ValueOf(&p.Regs[6]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R7 0x%02x\n", reflect.ValueOf(&p.Regs[7]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R8 0x%02x\n", reflect.ValueOf(&p.Regs[8]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R9 0x%02x\n", reflect.ValueOf(&p.Regs[9]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R10 0x%02x\n", reflect.ValueOf(&p.Regs[10]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R11 0x%02x\n", reflect.ValueOf(&p.Regs[11]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R12 0x%02x\n", reflect.ValueOf(&p.Regs[12]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R13 0x%02x\n", reflect.ValueOf(&p.Regs[13]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R14 0x%02x\n", reflect.ValueOf(&p.Regs[14]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R15 0x%02x\n", reflect.ValueOf(&p.Regs[15]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R16 0x%02x\n", reflect.ValueOf(&p.Regs[16]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R17 0x%02x\n", reflect.ValueOf(&p.Regs[17]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R18 0x%02x\n", reflect.ValueOf(&p.Regs[18]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R19 0x%02x\n", reflect.ValueOf(&p.Regs[19]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R20 0x%02x\n", reflect.ValueOf(&p.Regs[20]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R21 0x%02x\n", reflect.ValueOf(&p.Regs[21]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R22 0x%02x\n", reflect.ValueOf(&p.Regs[22]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R23 0x%02x\n", reflect.ValueOf(&p.Regs[23]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R24 0x%02x\n", reflect.ValueOf(&p.Regs[24]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R25 0x%02x\n", reflect.ValueOf(&p.Regs[25]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R26 0x%02x\n", reflect.ValueOf(&p.Regs[26]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R27 0x%02x\n", reflect.ValueOf(&p.Regs[27]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R28 0x%02x\n", reflect.ValueOf(&p.Regs[28]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R29 0x%02x\n", reflect.ValueOf(&p.Regs[29]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_R30 0x%02x\n", reflect.ValueOf(&p.Regs[30]).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_SP 0x%02x\n", reflect.ValueOf(&p.Sp).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_PC 0x%02x\n", reflect.ValueOf(&p.Pc).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_PSTATE 0x%02x\n", reflect.ValueOf(&p.Pstate).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_TLS 0x%02x\n", reflect.ValueOf(&p.TPIDR_EL0).Pointer()-reflect.ValueOf(p).Pointer()) +} diff --git a/pkg/ring0/entry_amd64.go b/pkg/ring0/entry_amd64.go new file mode 100644 index 000000000..d87b1fd00 --- /dev/null +++ b/pkg/ring0/entry_amd64.go @@ -0,0 +1,131 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package ring0 + +import ( + "gvisor.dev/gvisor/pkg/sentry/arch" +) + +// This is an assembly function. +// +// The sysenter function is invoked in two situations: +// +// (1) The guest kernel has executed a system call. +// (2) The guest application has executed a system call. +// +// The interrupt flag is examined to determine whether the system call was +// executed from kernel mode or not and the appropriate stub is called. +func sysenter() + +// swapgs swaps the current GS value. +// +// This must be called prior to sysret/iret. +func swapgs() + +// jumpToKernel jumps to the kernel version of the current RIP. +func jumpToKernel() + +// sysret returns to userspace from a system call. +// +// The return code is the vector that interrupted execution. +// +// See stubs.go for a note regarding the frame size of this function. +func sysret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector + +// "iret is the cadillac of CPL switching." +// +// -- Neel Natu +// +// iret is nearly identical to sysret, except an iret is used to fully restore +// all user state. This must be called in cases where all registers need to be +// restored. +func iret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector + +// exception is the generic exception entry. +// +// This is called by the individual stub definitions. +func exception() + +// resume is a stub that restores the CPU kernel registers. +// +// This is used when processing kernel exceptions and syscalls. +func resume() + +// Start is the CPU entrypoint. +// +// The following start conditions must be satisfied: +// +// * AX should contain the CPU pointer. +// * c.GDT() should be loaded as the GDT. +// * c.IDT() should be loaded as the IDT. +// * c.CR0() should be the current CR0 value. +// * c.CR3() should be set to the kernel PageTables. +// * c.CR4() should be the current CR4 value. +// * c.EFER() should be the current EFER value. +// +// The CPU state will be set to c.Registers(). +func Start() + +// Exception stubs. +func divideByZero() +func debug() +func nmi() +func breakpoint() +func overflow() +func boundRangeExceeded() +func invalidOpcode() +func deviceNotAvailable() +func doubleFault() +func coprocessorSegmentOverrun() +func invalidTSS() +func segmentNotPresent() +func stackSegmentFault() +func generalProtectionFault() +func pageFault() +func x87FloatingPointException() +func alignmentCheck() +func machineCheck() +func simdFloatingPointException() +func virtualizationException() +func securityException() +func syscallInt80() + +// Exception handler index. +var handlers = map[Vector]func(){ + DivideByZero: divideByZero, + Debug: debug, + NMI: nmi, + Breakpoint: breakpoint, + Overflow: overflow, + BoundRangeExceeded: boundRangeExceeded, + InvalidOpcode: invalidOpcode, + DeviceNotAvailable: deviceNotAvailable, + DoubleFault: doubleFault, + CoprocessorSegmentOverrun: coprocessorSegmentOverrun, + InvalidTSS: invalidTSS, + SegmentNotPresent: segmentNotPresent, + StackSegmentFault: stackSegmentFault, + GeneralProtectionFault: generalProtectionFault, + PageFault: pageFault, + X87FloatingPointException: x87FloatingPointException, + AlignmentCheck: alignmentCheck, + MachineCheck: machineCheck, + SIMDFloatingPointException: simdFloatingPointException, + VirtualizationException: virtualizationException, + SecurityException: securityException, + SyscallInt80: syscallInt80, +} diff --git a/pkg/ring0/entry_arm64.go b/pkg/ring0/entry_arm64.go new file mode 100644 index 000000000..62a93f3d6 --- /dev/null +++ b/pkg/ring0/entry_arm64.go @@ -0,0 +1,60 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ring0 + +// This is an assembly function. +// +// The sysenter function is invoked in two situations: +// +// (1) The guest kernel has executed a system call. +// (2) The guest application has executed a system call. +// +// The interrupt flag is examined to determine whether the system call was +// executed from kernel mode or not and the appropriate stub is called. + +func El1_sync_invalid() +func El1_irq_invalid() +func El1_fiq_invalid() +func El1_error_invalid() + +func El1_sync() +func El1_irq() +func El1_fiq() +func El1_error() + +func El0_sync() +func El0_irq() +func El0_fiq() +func El0_error() + +func El0_sync_invalid() +func El0_irq_invalid() +func El0_fiq_invalid() +func El0_error_invalid() + +func Vectors() + +// Start is the CPU entrypoint. +// +// The CPU state will be set to c.Registers(). +func Start() +func kernelExitToEl1() + +func kernelExitToEl0() + +// Shutdown execution +func Shutdown() diff --git a/pkg/ring0/entry_impl_amd64.s b/pkg/ring0/entry_impl_amd64.s new file mode 100644 index 000000000..9a5d6c064 --- /dev/null +++ b/pkg/ring0/entry_impl_amd64.s @@ -0,0 +1,441 @@ +// build +amd64 + +// Automatically generated, do not edit. + +// CPU offsets. +#define CPU_REGISTERS 0x28 +#define CPU_ERROR_CODE 0x10 +#define CPU_ERROR_TYPE 0x18 +#define CPU_ENTRY 0x20 + +// CPU entry offsets. +#define ENTRY_SCRATCH0 0x100 +#define ENTRY_STACK_TOP 0x108 +#define ENTRY_CPU_SELF 0x110 +#define ENTRY_KERNEL_CR3 0x118 + +// Bits. +#define _RFLAGS_IF 0x200 +#define _RFLAGS_IOPL0 0x1000 +#define _KERNEL_FLAGS 0x02 + +// Vectors. +#define DivideByZero 0x00 +#define Debug 0x01 +#define NMI 0x02 +#define Breakpoint 0x03 +#define Overflow 0x04 +#define BoundRangeExceeded 0x05 +#define InvalidOpcode 0x06 +#define DeviceNotAvailable 0x07 +#define DoubleFault 0x08 +#define CoprocessorSegmentOverrun 0x09 +#define InvalidTSS 0x0a +#define SegmentNotPresent 0x0b +#define StackSegmentFault 0x0c +#define GeneralProtectionFault 0x0d +#define PageFault 0x0e +#define X87FloatingPointException 0x10 +#define AlignmentCheck 0x11 +#define MachineCheck 0x12 +#define SIMDFloatingPointException 0x13 +#define VirtualizationException 0x14 +#define SecurityException 0x1e +#define SyscallInt80 0x80 +#define Syscall 0x100 + +// Ptrace registers. +#define PTRACE_R15 0x00 +#define PTRACE_R14 0x08 +#define PTRACE_R13 0x10 +#define PTRACE_R12 0x18 +#define PTRACE_RBP 0x20 +#define PTRACE_RBX 0x28 +#define PTRACE_R11 0x30 +#define PTRACE_R10 0x38 +#define PTRACE_R9 0x40 +#define PTRACE_R8 0x48 +#define PTRACE_RAX 0x50 +#define PTRACE_RCX 0x58 +#define PTRACE_RDX 0x60 +#define PTRACE_RSI 0x68 +#define PTRACE_RDI 0x70 +#define PTRACE_ORIGRAX 0x78 +#define PTRACE_RIP 0x80 +#define PTRACE_CS 0x88 +#define PTRACE_FLAGS 0x90 +#define PTRACE_RSP 0x98 +#define PTRACE_SS 0xa0 +#define PTRACE_FS 0xa8 +#define PTRACE_GS 0xb0 +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "funcdata.h" +#include "textflag.h" + +// NB: Offsets are programmatically generated (see BUILD). +// +// This file is concatenated with the definitions. + +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not saved: AX, SP, IP, FLAGS, all segments. +#define REGISTERS_SAVE(reg, offset) \ + MOVQ R15, offset+PTRACE_R15(reg); \ + MOVQ R14, offset+PTRACE_R14(reg); \ + MOVQ R13, offset+PTRACE_R13(reg); \ + MOVQ R12, offset+PTRACE_R12(reg); \ + MOVQ BP, offset+PTRACE_RBP(reg); \ + MOVQ BX, offset+PTRACE_RBX(reg); \ + MOVQ CX, offset+PTRACE_RCX(reg); \ + MOVQ DX, offset+PTRACE_RDX(reg); \ + MOVQ R11, offset+PTRACE_R11(reg); \ + MOVQ R10, offset+PTRACE_R10(reg); \ + MOVQ R9, offset+PTRACE_R9(reg); \ + MOVQ R8, offset+PTRACE_R8(reg); \ + MOVQ SI, offset+PTRACE_RSI(reg); \ + MOVQ DI, offset+PTRACE_RDI(reg); + +// Loads a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not loaded: AX, SP, IP, FLAGS, all segments. +#define REGISTERS_LOAD(reg, offset) \ + MOVQ offset+PTRACE_R15(reg), R15; \ + MOVQ offset+PTRACE_R14(reg), R14; \ + MOVQ offset+PTRACE_R13(reg), R13; \ + MOVQ offset+PTRACE_R12(reg), R12; \ + MOVQ offset+PTRACE_RBP(reg), BP; \ + MOVQ offset+PTRACE_RBX(reg), BX; \ + MOVQ offset+PTRACE_RCX(reg), CX; \ + MOVQ offset+PTRACE_RDX(reg), DX; \ + MOVQ offset+PTRACE_R11(reg), R11; \ + MOVQ offset+PTRACE_R10(reg), R10; \ + MOVQ offset+PTRACE_R9(reg), R9; \ + MOVQ offset+PTRACE_R8(reg), R8; \ + MOVQ offset+PTRACE_RSI(reg), SI; \ + MOVQ offset+PTRACE_RDI(reg), DI; + +// WRITE_CR3() writes the given CR3 value. +// +// The code corresponds to: +// +// mov %rax, %cr3 +// +#define WRITE_CR3() \ + BYTE $0x0f; BYTE $0x22; BYTE $0xd8; + +// SWAP_GS swaps the kernel GS (CPU). +#define SWAP_GS() \ + BYTE $0x0F; BYTE $0x01; BYTE $0xf8; + +// IRET returns from an interrupt frame. +#define IRET() \ + BYTE $0x48; BYTE $0xcf; + +// SYSRET64 executes the sysret instruction. +#define SYSRET64() \ + BYTE $0x48; BYTE $0x0f; BYTE $0x07; + +// LOAD_KERNEL_STACK loads the kernel stack. +#define LOAD_KERNEL_STACK(entry) \ + MOVQ ENTRY_STACK_TOP(entry), SP; + +// See kernel.go. +TEXT ·Halt(SB),NOSPLIT,$0 + HLT + RET + +// See entry_amd64.go. +TEXT ·swapgs(SB),NOSPLIT,$0 + SWAP_GS() + RET + +// jumpToKernel changes execution to the kernel address space. +// +// This works by changing the return value to the kernel version. +TEXT ·jumpToKernel(SB),NOSPLIT,$0 + MOVQ 0(SP), AX + ORQ ·KernelStartAddress(SB), AX // Future return value. + MOVQ AX, 0(SP) + RET + +// See entry_amd64.go. +TEXT ·sysret(SB),NOSPLIT,$0-24 + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX + MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX) + MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX) + MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX) + + // save SP AX userCR3 on the kernel stack. + MOVQ CPU_ENTRY(BX), BX + LOAD_KERNEL_STACK(BX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_RAX(AX) + PUSHQ CX + + // Restore user register state. + REGISTERS_LOAD(AX, 0) + MOVQ PTRACE_RIP(AX), CX // Needed for SYSRET. + MOVQ PTRACE_FLAGS(AX), R11 // Needed for SYSRET. + + // restore userCR3, AX, SP. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. + POPQ SP // Restore SP. + SYSRET64() + +// See entry_amd64.go. +TEXT ·iret(SB),NOSPLIT,$0-24 + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX + MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX) + MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX) + MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX) + + // Build an IRET frame & restore state. + MOVQ CPU_ENTRY(BX), BX + LOAD_KERNEL_STACK(BX) + PUSHQ PTRACE_SS(AX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_FLAGS(AX) + PUSHQ PTRACE_CS(AX) + PUSHQ PTRACE_RIP(AX) + PUSHQ PTRACE_RAX(AX) // Save AX on kernel stack. + PUSHQ CX // Save userCR3 on kernel stack. + REGISTERS_LOAD(AX, 0) // Restore most registers. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. + IRET() + +// See entry_amd64.go. +TEXT ·resume(SB),NOSPLIT,$0 + // See iret, above. + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + PUSHQ CPU_REGISTERS+PTRACE_SS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RSP(AX) + PUSHQ CPU_REGISTERS+PTRACE_FLAGS(AX) + PUSHQ CPU_REGISTERS+PTRACE_CS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RIP(AX) + REGISTERS_LOAD(AX, CPU_REGISTERS) + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX + IRET() + +// See entry_amd64.go. +TEXT ·Start(SB),NOSPLIT,$0 + PUSHQ $0x0 // Previous frame pointer. + MOVQ SP, BP // Set frame pointer. + PUSHQ AX // First argument (CPU). + CALL ·start(SB) // Call Go hook. + JMP ·resume(SB) // Restore to registers. + +// See entry_amd64.go. +TEXT ·sysenter(SB),NOSPLIT,$0 + // _RFLAGS_IOPL0 is always set in the user mode and it is never set in + // the kernel mode. See the comment of UserFlagsSet for more details. + TESTL $_RFLAGS_IOPL0, R11 + JZ kernel +user: + SWAP_GS() + MOVQ AX, ENTRY_SCRATCH0(GS) // Save user AX on scratch. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user regs. + REGISTERS_SAVE(AX, 0) // Save all except IP, FLAGS, SP, AX. + MOVQ CX, PTRACE_RIP(AX) + MOVQ R11, PTRACE_FLAGS(AX) + MOVQ SP, PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), CX // Load saved user AX value. + MOVQ CX, PTRACE_RAX(AX) // Save everything else. + MOVQ CX, PTRACE_ORIGRAX(AX) + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Get stacks. + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. + + // Return to the kernel, where the frame is: + // + // vector (sp+32) + // userCR3 (sp+24) + // regs (sp+16) + // cpu (sp+8) + // vcpu.Switch (sp+0) + // + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ $Syscall, 32(SP) // Output vector. + RET + +kernel: + // We can't restore the original stack, but we can access the registers + // in the CPU state directly. No need for temporary juggling. + MOVQ AX, ENTRY_SCRATCH0(GS) + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + MOVQ CX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ R11, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ SP, CPU_REGISTERS+PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), BX + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. + + // Call the syscall trampoline. + LOAD_KERNEL_STACK(GS) + PUSHQ AX // First argument (vCPU). + CALL ·kernelSyscall(SB) // Call the trampoline. + POPQ AX // Pop vCPU. + JMP ·resume(SB) + +// exception is a generic exception handler. +// +// There are two cases handled: +// +// 1) An exception in kernel mode: this results in saving the state at the time +// of the exception and calling the defined hook. +// +// 2) An exception in guest mode: the original kernel frame is restored, and +// the vector & error codes are pushed as return values. +// +// See below for the stubs that call exception. +TEXT ·exception(SB),NOSPLIT,$0 + // Determine whether the exception occurred in kernel mode or user + // mode, based on the flags. We expect the following stack: + // + // SS (sp+48) + // SP (sp+40) + // FLAGS (sp+32) + // CS (sp+24) + // IP (sp+16) + // ERROR_CODE (sp+8) + // VECTOR (sp+0) + // + TESTL $_RFLAGS_IOPL0, 32(SP) + JZ kernel + +user: + SWAP_GS() + ADDQ $-8, SP // Adjust for flags. + MOVQ $_KERNEL_FLAGS, 0(SP); BYTE $0x9d; // Reset flags (POPFQ). + PUSHQ AX // Save user AX on stack. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user regs. + REGISTERS_SAVE(AX, 0) // Save all except IP, FLAGS, SP, AX. + POPQ BX // Restore original AX. + MOVQ BX, PTRACE_RAX(AX) // Save it. + MOVQ BX, PTRACE_ORIGRAX(AX) + MOVQ 16(SP), BX; MOVQ BX, PTRACE_RIP(AX) + MOVQ 24(SP), CX; MOVQ CX, PTRACE_CS(AX) + MOVQ 32(SP), DX; MOVQ DX, PTRACE_FLAGS(AX) + MOVQ 40(SP), DI; MOVQ DI, PTRACE_RSP(AX) + MOVQ 48(SP), SI; MOVQ SI, PTRACE_SS(AX) + + // Copy out and return. + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ 0(SP), BX // Load vector. + MOVQ 8(SP), CX // Load error code. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Original stack (kernel version). + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ CX, CPU_ERROR_CODE(AX) // Set error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. + MOVQ BX, 32(SP) // Output vector. + RET + +kernel: + // As per above, we can save directly. + PUSHQ AX + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + POPQ BX + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ 16(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ 32(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ 40(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RSP(AX) + + // Set the error code and adjust the stack. + MOVQ 8(SP), BX // Load the error code. + MOVQ BX, CPU_ERROR_CODE(AX) // Copy out to the CPU. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. + MOVQ 0(SP), BX // BX contains the vector. + + // Call the exception trampoline. + LOAD_KERNEL_STACK(GS) + PUSHQ BX // Second argument (vector). + PUSHQ AX // First argument (vCPU). + CALL ·kernelException(SB) // Call the trampoline. + POPQ BX // Pop vector. + POPQ AX // Pop vCPU. + JMP ·resume(SB) + +#define EXCEPTION_WITH_ERROR(value, symbol) \ +TEXT symbol,NOSPLIT,$0; \ + PUSHQ $value; \ + JMP ·exception(SB); + +#define EXCEPTION_WITHOUT_ERROR(value, symbol) \ +TEXT symbol,NOSPLIT,$0; \ + PUSHQ $0x0; \ + PUSHQ $value; \ + JMP ·exception(SB); + +EXCEPTION_WITHOUT_ERROR(DivideByZero, ·divideByZero(SB)) +EXCEPTION_WITHOUT_ERROR(Debug, ·debug(SB)) +EXCEPTION_WITHOUT_ERROR(NMI, ·nmi(SB)) +EXCEPTION_WITHOUT_ERROR(Breakpoint, ·breakpoint(SB)) +EXCEPTION_WITHOUT_ERROR(Overflow, ·overflow(SB)) +EXCEPTION_WITHOUT_ERROR(BoundRangeExceeded, ·boundRangeExceeded(SB)) +EXCEPTION_WITHOUT_ERROR(InvalidOpcode, ·invalidOpcode(SB)) +EXCEPTION_WITHOUT_ERROR(DeviceNotAvailable, ·deviceNotAvailable(SB)) +EXCEPTION_WITH_ERROR(DoubleFault, ·doubleFault(SB)) +EXCEPTION_WITHOUT_ERROR(CoprocessorSegmentOverrun, ·coprocessorSegmentOverrun(SB)) +EXCEPTION_WITH_ERROR(InvalidTSS, ·invalidTSS(SB)) +EXCEPTION_WITH_ERROR(SegmentNotPresent, ·segmentNotPresent(SB)) +EXCEPTION_WITH_ERROR(StackSegmentFault, ·stackSegmentFault(SB)) +EXCEPTION_WITH_ERROR(GeneralProtectionFault, ·generalProtectionFault(SB)) +EXCEPTION_WITH_ERROR(PageFault, ·pageFault(SB)) +EXCEPTION_WITHOUT_ERROR(X87FloatingPointException, ·x87FloatingPointException(SB)) +EXCEPTION_WITH_ERROR(AlignmentCheck, ·alignmentCheck(SB)) +EXCEPTION_WITHOUT_ERROR(MachineCheck, ·machineCheck(SB)) +EXCEPTION_WITHOUT_ERROR(SIMDFloatingPointException, ·simdFloatingPointException(SB)) +EXCEPTION_WITHOUT_ERROR(VirtualizationException, ·virtualizationException(SB)) +EXCEPTION_WITH_ERROR(SecurityException, ·securityException(SB)) +EXCEPTION_WITHOUT_ERROR(SyscallInt80, ·syscallInt80(SB)) diff --git a/pkg/ring0/entry_impl_arm64.s b/pkg/ring0/entry_impl_arm64.s new file mode 100644 index 000000000..b3d300fb7 --- /dev/null +++ b/pkg/ring0/entry_impl_arm64.s @@ -0,0 +1,858 @@ +// build +arm64 + +// Automatically generated, do not edit. + +// CPU offsets. +#define CPU_SELF 0x00 +#define CPU_REGISTERS 0x258 +#define CPU_STACK_TOP 0x210 +#define CPU_ERROR_CODE 0x210 +#define CPU_ERROR_TYPE 0x218 +#define CPU_FAULT_ADDR 0x220 +#define CPU_TTBR0_KVM 0x228 +#define CPU_TTBR0_APP 0x230 +#define CPU_VECTOR_CODE 0x238 +#define CPU_APP_ADDR 0x240 +#define CPU_LAZY_VFP 0x248 +#define CPU_APP_ASID 0x250 + +// Bits. +#define _KERNEL_FLAGS 0x3c5 + +// Vectors. +#define El1Sync 0x04 +#define El1Irq 0x05 +#define El1Fiq 0x06 +#define El1Err 0x07 +#define El0Sync 0x08 +#define El0Irq 0x09 +#define El0Fiq 0x0a +#define El0Err 0x0b +#define El1SyncDa 0x10 +#define El1SyncIa 0x11 +#define El1SyncSpPc 0x12 +#define El1SyncUndef 0x13 +#define El1SyncDbg 0x14 +#define El1SyncInv 0x15 +#define El0SyncSVC 0x16 +#define El0SyncDa 0x17 +#define El0SyncIa 0x18 +#define El0SyncFpsimdAcc 0x19 +#define El0SyncSveAcc 0x1a +#define El0SyncFpsimdExc 0x1b +#define El0SyncSys 0x1c +#define El0SyncSpPc 0x1d +#define El0SyncUndef 0x1e +#define El0SyncDbg 0x1f +#define El0SyncWfx 0x20 +#define El0SyncInv 0x21 +#define El0ErrNMI 0x22 +#define PageFault 0x17 +#define Syscall 0x16 +#define VirtualizationException 0x23 + +// Ptrace registers. +#define PTRACE_R0 0x00 +#define PTRACE_R1 0x08 +#define PTRACE_R2 0x10 +#define PTRACE_R3 0x18 +#define PTRACE_R4 0x20 +#define PTRACE_R5 0x28 +#define PTRACE_R6 0x30 +#define PTRACE_R7 0x38 +#define PTRACE_R8 0x40 +#define PTRACE_R9 0x48 +#define PTRACE_R10 0x50 +#define PTRACE_R11 0x58 +#define PTRACE_R12 0x60 +#define PTRACE_R13 0x68 +#define PTRACE_R14 0x70 +#define PTRACE_R15 0x78 +#define PTRACE_R16 0x80 +#define PTRACE_R17 0x88 +#define PTRACE_R18 0x90 +#define PTRACE_R19 0x98 +#define PTRACE_R20 0xa0 +#define PTRACE_R21 0xa8 +#define PTRACE_R22 0xb0 +#define PTRACE_R23 0xb8 +#define PTRACE_R24 0xc0 +#define PTRACE_R25 0xc8 +#define PTRACE_R26 0xd0 +#define PTRACE_R27 0xd8 +#define PTRACE_R28 0xe0 +#define PTRACE_R29 0xe8 +#define PTRACE_R30 0xf0 +#define PTRACE_SP 0xf8 +#define PTRACE_PC 0x100 +#define PTRACE_PSTATE 0x108 +#define PTRACE_TLS 0x110 +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "funcdata.h" +#include "textflag.h" + +// NB: Offsets are programatically generated (see BUILD). +// +// This file is concatenated with the definitions. + +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// + +// ERET returns using the ELR and SPSR for the current exception level. +#define ERET() \ + WORD $0xd69f03e0; \ + DSB $7; \ + ISB $15; + +// RSV_REG is a register that holds el1 information temporarily. +#define RSV_REG R18_PLATFORM + +// RSV_REG_APP is a register that holds el0 information temporarily. +#define RSV_REG_APP R9 + +#define FPEN_NOTRAP 0x3 +#define FPEN_SHIFT 20 + +#define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT) + +// sctlr_el1: system control register el1. +#define SCTLR_M 1 << 0 +#define SCTLR_C 1 << 2 +#define SCTLR_I 1 << 12 +#define SCTLR_DZE 1 << 14 +#define SCTLR_UCT 1 << 15 +#define SCTLR_UCI 1 << 26 + +#define SCTLR_EL1_DEFAULT (SCTLR_M | SCTLR_C | SCTLR_I | SCTLR_UCT | SCTLR_UCI | SCTLR_DZE) + +// cntkctl_el1: counter-timer kernel control register el1. +#define CNTKCTL_EL0PCTEN 1 << 0 +#define CNTKCTL_EL0VCTEN 1 << 1 + +#define CNTKCTL_EL1_DEFAULT (CNTKCTL_EL0PCTEN | CNTKCTL_EL0VCTEN) + +// Saves a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not saved: R9, R18. +#define REGISTERS_SAVE(reg, offset) \ + MOVD R0, offset+PTRACE_R0(reg); \ + MOVD R1, offset+PTRACE_R1(reg); \ + MOVD R2, offset+PTRACE_R2(reg); \ + MOVD R3, offset+PTRACE_R3(reg); \ + MOVD R4, offset+PTRACE_R4(reg); \ + MOVD R5, offset+PTRACE_R5(reg); \ + MOVD R6, offset+PTRACE_R6(reg); \ + MOVD R7, offset+PTRACE_R7(reg); \ + MOVD R8, offset+PTRACE_R8(reg); \ + MOVD R10, offset+PTRACE_R10(reg); \ + MOVD R11, offset+PTRACE_R11(reg); \ + MOVD R12, offset+PTRACE_R12(reg); \ + MOVD R13, offset+PTRACE_R13(reg); \ + MOVD R14, offset+PTRACE_R14(reg); \ + MOVD R15, offset+PTRACE_R15(reg); \ + MOVD R16, offset+PTRACE_R16(reg); \ + MOVD R17, offset+PTRACE_R17(reg); \ + MOVD R19, offset+PTRACE_R19(reg); \ + MOVD R20, offset+PTRACE_R20(reg); \ + MOVD R21, offset+PTRACE_R21(reg); \ + MOVD R22, offset+PTRACE_R22(reg); \ + MOVD R23, offset+PTRACE_R23(reg); \ + MOVD R24, offset+PTRACE_R24(reg); \ + MOVD R25, offset+PTRACE_R25(reg); \ + MOVD R26, offset+PTRACE_R26(reg); \ + MOVD R27, offset+PTRACE_R27(reg); \ + MOVD g, offset+PTRACE_R28(reg); \ + MOVD R29, offset+PTRACE_R29(reg); \ + MOVD R30, offset+PTRACE_R30(reg); + +// Loads a register set. +// +// This is a macro because it may need to executed in contents where a stack is +// not available for calls. +// +// The following registers are not loaded: R9, R18. +#define REGISTERS_LOAD(reg, offset) \ + MOVD offset+PTRACE_R0(reg), R0; \ + MOVD offset+PTRACE_R1(reg), R1; \ + MOVD offset+PTRACE_R2(reg), R2; \ + MOVD offset+PTRACE_R3(reg), R3; \ + MOVD offset+PTRACE_R4(reg), R4; \ + MOVD offset+PTRACE_R5(reg), R5; \ + MOVD offset+PTRACE_R6(reg), R6; \ + MOVD offset+PTRACE_R7(reg), R7; \ + MOVD offset+PTRACE_R8(reg), R8; \ + MOVD offset+PTRACE_R10(reg), R10; \ + MOVD offset+PTRACE_R11(reg), R11; \ + MOVD offset+PTRACE_R12(reg), R12; \ + MOVD offset+PTRACE_R13(reg), R13; \ + MOVD offset+PTRACE_R14(reg), R14; \ + MOVD offset+PTRACE_R15(reg), R15; \ + MOVD offset+PTRACE_R16(reg), R16; \ + MOVD offset+PTRACE_R17(reg), R17; \ + MOVD offset+PTRACE_R19(reg), R19; \ + MOVD offset+PTRACE_R20(reg), R20; \ + MOVD offset+PTRACE_R21(reg), R21; \ + MOVD offset+PTRACE_R22(reg), R22; \ + MOVD offset+PTRACE_R23(reg), R23; \ + MOVD offset+PTRACE_R24(reg), R24; \ + MOVD offset+PTRACE_R25(reg), R25; \ + MOVD offset+PTRACE_R26(reg), R26; \ + MOVD offset+PTRACE_R27(reg), R27; \ + MOVD offset+PTRACE_R28(reg), g; \ + MOVD offset+PTRACE_R29(reg), R29; \ + MOVD offset+PTRACE_R30(reg), R30; + +#define ESR_ELx_EC_UNKNOWN (0x00) +#define ESR_ELx_EC_WFx (0x01) +/* Unallocated EC: 0x02 */ +#define ESR_ELx_EC_CP15_32 (0x03) +#define ESR_ELx_EC_CP15_64 (0x04) +#define ESR_ELx_EC_CP14_MR (0x05) +#define ESR_ELx_EC_CP14_LS (0x06) +#define ESR_ELx_EC_FP_ASIMD (0x07) +#define ESR_ELx_EC_CP10_ID (0x08) /* EL2 only */ +#define ESR_ELx_EC_PAC (0x09) /* EL2 and above */ +/* Unallocated EC: 0x0A - 0x0B */ +#define ESR_ELx_EC_CP14_64 (0x0C) +/* Unallocated EC: 0x0d */ +#define ESR_ELx_EC_ILL (0x0E) +/* Unallocated EC: 0x0F - 0x10 */ +#define ESR_ELx_EC_SVC32 (0x11) +#define ESR_ELx_EC_HVC32 (0x12) /* EL2 only */ +#define ESR_ELx_EC_SMC32 (0x13) /* EL2 and above */ +/* Unallocated EC: 0x14 */ +#define ESR_ELx_EC_SVC64 (0x15) +#define ESR_ELx_EC_HVC64 (0x16) /* EL2 and above */ +#define ESR_ELx_EC_SMC64 (0x17) /* EL2 and above */ +#define ESR_ELx_EC_SYS64 (0x18) +#define ESR_ELx_EC_SVE (0x19) +/* Unallocated EC: 0x1A - 0x1E */ +#define ESR_ELx_EC_IMP_DEF (0x1f) /* EL3 only */ +#define ESR_ELx_EC_IABT_LOW (0x20) +#define ESR_ELx_EC_IABT_CUR (0x21) +#define ESR_ELx_EC_PC_ALIGN (0x22) +/* Unallocated EC: 0x23 */ +#define ESR_ELx_EC_DABT_LOW (0x24) +#define ESR_ELx_EC_DABT_CUR (0x25) +#define ESR_ELx_EC_SP_ALIGN (0x26) +/* Unallocated EC: 0x27 */ +#define ESR_ELx_EC_FP_EXC32 (0x28) +/* Unallocated EC: 0x29 - 0x2B */ +#define ESR_ELx_EC_FP_EXC64 (0x2C) +/* Unallocated EC: 0x2D - 0x2E */ +#define ESR_ELx_EC_SERROR (0x2F) +#define ESR_ELx_EC_BREAKPT_LOW (0x30) +#define ESR_ELx_EC_BREAKPT_CUR (0x31) +#define ESR_ELx_EC_SOFTSTP_LOW (0x32) +#define ESR_ELx_EC_SOFTSTP_CUR (0x33) +#define ESR_ELx_EC_WATCHPT_LOW (0x34) +#define ESR_ELx_EC_WATCHPT_CUR (0x35) +/* Unallocated EC: 0x36 - 0x37 */ +#define ESR_ELx_EC_BKPT32 (0x38) +/* Unallocated EC: 0x39 */ +#define ESR_ELx_EC_VECTOR32 (0x3A) /* EL2 only */ +/* Unallocted EC: 0x3B */ +#define ESR_ELx_EC_BRK64 (0x3C) +/* Unallocated EC: 0x3D - 0x3F */ +#define ESR_ELx_EC_MAX (0x3F) + +#define ESR_ELx_EC_SHIFT (26) +#define ESR_ELx_EC_MASK (UL(0x3F) << ESR_ELx_EC_SHIFT) +#define ESR_ELx_EC(esr) (((esr) & ESR_ELx_EC_MASK) >> ESR_ELx_EC_SHIFT) + +#define ESR_ELx_IL_SHIFT (25) +#define ESR_ELx_IL (UL(1) << ESR_ELx_IL_SHIFT) +#define ESR_ELx_ISS_MASK (ESR_ELx_IL - 1) + +/* ISS field definitions shared by different classes */ +#define ESR_ELx_WNR_SHIFT (6) +#define ESR_ELx_WNR (UL(1) << ESR_ELx_WNR_SHIFT) + +/* Asynchronous Error Type */ +#define ESR_ELx_IDS_SHIFT (24) +#define ESR_ELx_IDS (UL(1) << ESR_ELx_IDS_SHIFT) +#define ESR_ELx_AET_SHIFT (10) +#define ESR_ELx_AET (UL(0x7) << ESR_ELx_AET_SHIFT) + +#define ESR_ELx_AET_UC (UL(0) << ESR_ELx_AET_SHIFT) +#define ESR_ELx_AET_UEU (UL(1) << ESR_ELx_AET_SHIFT) +#define ESR_ELx_AET_UEO (UL(2) << ESR_ELx_AET_SHIFT) +#define ESR_ELx_AET_UER (UL(3) << ESR_ELx_AET_SHIFT) +#define ESR_ELx_AET_CE (UL(6) << ESR_ELx_AET_SHIFT) + +/* Shared ISS field definitions for Data/Instruction aborts */ +#define ESR_ELx_SET_SHIFT (11) +#define ESR_ELx_SET_MASK (UL(3) << ESR_ELx_SET_SHIFT) +#define ESR_ELx_FnV_SHIFT (10) +#define ESR_ELx_FnV (UL(1) << ESR_ELx_FnV_SHIFT) +#define ESR_ELx_EA_SHIFT (9) +#define ESR_ELx_EA (UL(1) << ESR_ELx_EA_SHIFT) +#define ESR_ELx_S1PTW_SHIFT (7) +#define ESR_ELx_S1PTW (UL(1) << ESR_ELx_S1PTW_SHIFT) + +/* Shared ISS fault status code(IFSC/DFSC) for Data/Instruction aborts */ +#define ESR_ELx_FSC (0x3F) +#define ESR_ELx_FSC_TYPE (0x3C) +#define ESR_ELx_FSC_EXTABT (0x10) +#define ESR_ELx_FSC_SERROR (0x11) +#define ESR_ELx_FSC_ACCESS (0x08) +#define ESR_ELx_FSC_FAULT (0x04) +#define ESR_ELx_FSC_PERM (0x0C) + +/* ISS field definitions for Data Aborts */ +#define ESR_ELx_ISV_SHIFT (24) +#define ESR_ELx_ISV (UL(1) << ESR_ELx_ISV_SHIFT) +#define ESR_ELx_SAS_SHIFT (22) +#define ESR_ELx_SAS (UL(3) << ESR_ELx_SAS_SHIFT) +#define ESR_ELx_SSE_SHIFT (21) +#define ESR_ELx_SSE (UL(1) << ESR_ELx_SSE_SHIFT) +#define ESR_ELx_SRT_SHIFT (16) +#define ESR_ELx_SRT_MASK (UL(0x1F) << ESR_ELx_SRT_SHIFT) +#define ESR_ELx_SF_SHIFT (15) +#define ESR_ELx_SF (UL(1) << ESR_ELx_SF_SHIFT) +#define ESR_ELx_AR_SHIFT (14) +#define ESR_ELx_AR (UL(1) << ESR_ELx_AR_SHIFT) +#define ESR_ELx_CM_SHIFT (8) +#define ESR_ELx_CM (UL(1) << ESR_ELx_CM_SHIFT) + +/* ISS field definitions for exceptions taken in to Hyp */ +#define ESR_ELx_CV (UL(1) << 24) +#define ESR_ELx_COND_SHIFT (20) +#define ESR_ELx_COND_MASK (UL(0xF) << ESR_ELx_COND_SHIFT) +#define ESR_ELx_WFx_ISS_TI (UL(1) << 0) +#define ESR_ELx_WFx_ISS_WFI (UL(0) << 0) +#define ESR_ELx_WFx_ISS_WFE (UL(1) << 0) +#define ESR_ELx_xVC_IMM_MASK ((1UL << 16) - 1) + +/* ISS field definitions for system error */ +#define ESR_ELx_SERR_MASK (0x1) +#define ESR_ELx_SERR_NMI (0x1) + +// LOAD_KERNEL_ADDRESS loads a kernel address. +#define LOAD_KERNEL_ADDRESS(from, to) \ + MOVD from, to; \ + ORR $0xffff000000000000, to, to; + +// LOAD_KERNEL_STACK loads the kernel temporary stack. +#define LOAD_KERNEL_STACK(from) \ + LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \ + MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \ + MOVD RSV_REG, RSP; \ + WORD $0xd538d092; //MRS TPIDR_EL1, R18 + +// SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. +#define SWITCH_TO_APP_PAGETABLE() \ + MOVD CPU_APP_ASID(RSV_REG), RSV_REG_APP; \ + MOVD CPU_TTBR0_APP(RSV_REG), RSV_REG; \ + BFI $48, RSV_REG_APP, $16, RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; \ + ISB $15; + +// SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. +#define SWITCH_TO_KVM_PAGETABLE() \ + MOVD CPU_TTBR0_KVM(RSV_REG), RSV_REG; \ + MOVD $1, RSV_REG_APP; \ + BFI $48, RSV_REG_APP, $16, RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; \ + ISB $15; + +TEXT ·EnableVFP(SB),NOSPLIT,$0 + MOVD $FPEN_ENABLE, R0 + WORD $0xd5181040 //MSR R0, CPACR_EL1 + ISB $15 + RET + +TEXT ·DisableVFP(SB),NOSPLIT,$0 + MOVD $0, R0 + WORD $0xd5181040 //MSR R0, CPACR_EL1 + ISB $15 + RET + +#define VFP_ENABLE \ + MOVD $FPEN_ENABLE, R0; \ + WORD $0xd5181040; \ //MSR R0, CPACR_EL1 + ISB $15; + +#define VFP_DISABLE \ + MOVD $0x0, R0; \ + WORD $0xd5181040; \ //MSR R0, CPACR_EL1 + ISB $15; + +// KERNEL_ENTRY_FROM_EL0 is the entry code of the vcpu from el0 to el1. +#define KERNEL_ENTRY_FROM_EL0 \ + SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. + STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ + WORD $0xd538d092; \ // MRS TPIDR_EL1, R18 + MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP; \ // step2, load app context pointer. + REGISTERS_SAVE(RSV_REG_APP, 0); \ // step3, save app context. + MOVD RSV_REG_APP, R20; \ + LDP 16*0(RSP), (RSV_REG, RSV_REG_APP); \ + ADD $16, RSP, RSP; \ + MOVD RSV_REG, PTRACE_R18(R20); \ + MOVD RSV_REG_APP, PTRACE_R9(R20); \ + MRS TPIDR_EL0, R3; \ + MOVD R3, PTRACE_TLS(R20); \ + WORD $0xd5384003; \ // MRS SPSR_EL1, R3 + MOVD R3, PTRACE_PSTATE(R20); \ + MRS ELR_EL1, R3; \ + MOVD R3, PTRACE_PC(R20); \ + WORD $0xd5384103; \ // MRS SP_EL0, R3 + MOVD R3, PTRACE_SP(R20); + +// KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. +#define KERNEL_ENTRY_FROM_EL1 \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 + REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // Save sentry context. + MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \ + MRS TPIDR_EL0, R4; \ + MOVD R4, CPU_REGISTERS+PTRACE_TLS(RSV_REG); \ + WORD $0xd5384004; \ // MRS SPSR_EL1, R4 + MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \ + MRS ELR_EL1, R4; \ + MOVD R4, CPU_REGISTERS+PTRACE_PC(RSV_REG); \ + MOVD RSP, R4; \ + MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ + LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. + +// EXCEPTION_EL0 is a common el0 exception handler function. +#define EXCEPTION_EL0(vector) \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 + WORD $0xd538601a; \ //MRS FAR_EL1, R26 + MOVD R26, CPU_FAULT_ADDR(RSV_REG); \ + MOVD $1, R3; \ + MOVD R3, CPU_ERROR_TYPE(RSV_REG); \ // Set error type to user. + MOVD $vector, R3; \ + MOVD R3, CPU_VECTOR_CODE(RSV_REG); \ + MRS ESR_EL1, R3; \ + MOVD R3, CPU_ERROR_CODE(RSV_REG); \ + B ·kernelExitToEl1(SB); + +// EXCEPTION_EL1 is a common el1 exception handler function. +#define EXCEPTION_EL1(vector) \ + MOVD $vector, R3; \ + MOVD R3, 8(RSP); \ + B ·HaltEl1ExceptionAndResume(SB); + +// storeAppASID writes the application's asid value. +TEXT ·storeAppASID(SB),NOSPLIT,$0-8 + MOVD asid+0(FP), R1 + MRS TPIDR_EL1, RSV_REG + MOVD R1, CPU_APP_ASID(RSV_REG) + RET + +// Halt halts execution. +TEXT ·Halt(SB),NOSPLIT,$0 + // Clear bluepill. + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + CMP RSV_REG, R9 + BNE mmio_exit + MOVD $0, CPU_REGISTERS+PTRACE_R9(RSV_REG) + +mmio_exit: + // Disable fpsimd. + WORD $0xd5381041 // MRS CPACR_EL1, R1 + MOVD R1, CPU_LAZY_VFP(RSV_REG) + VFP_DISABLE + + // Trigger MMIO_EXIT/_KVM_HYPERCALL_VMEXIT. + // + // To keep it simple, I used the address of exception table as the + // MMIO base address, so that I can trigger a MMIO-EXIT by forcibly writing + // a read-only space. + // Also, the length is engough to match a sufficient number of hypercall ID. + // Then, in host user space, I can calculate this address to find out + // which hypercall. + MRS VBAR_EL1, R9 + MOVD R0, 0x0(R9) + + RET + +// HaltAndResume halts execution and point the pointer to the resume function. +TEXT ·HaltAndResume(SB),NOSPLIT,$0 + BL ·Halt(SB) + B ·kernelExitToEl1(SB) // Resume. + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +TEXT ·HaltEl1SvcAndResume(SB),NOSPLIT,$0 + WORD $0xd538d092 // MRS TPIDR_EL1, R18 + MOVD CPU_SELF(RSV_REG), R3 // Load vCPU. + MOVD R3, 8(RSP) // First argument (vCPU). + CALL ·kernelSyscall(SB) // Call the trampoline. + B ·kernelExitToEl1(SB) // Resume. + +// HaltEl1ExceptionAndResume calls Hooks.KernelException and resume. +TEXT ·HaltEl1ExceptionAndResume(SB),NOSPLIT,$0-8 + WORD $0xd538d092 // MRS TPIDR_EL1, R18 + MOVD CPU_SELF(RSV_REG), R3 // Load vCPU. + MOVD R3, 8(RSP) // First argument (vCPU). + MOVD vector+0(FP), R3 + MOVD R3, 16(RSP) // Second argument (vector). + CALL ·kernelException(SB) // Call the trampoline. + B ·kernelExitToEl1(SB) // Resume. + +// Shutdown stops the guest. +TEXT ·Shutdown(SB),NOSPLIT,$0 + // PSCI EVENT. + MOVD $0x84000009, R0 + HVC $0 + +// See kernel.go. +TEXT ·Current(SB),NOSPLIT,$0-8 + MOVD CPU_SELF(RSV_REG), R8 + MOVD R8, ret+0(FP) + RET + +#define STACK_FRAME_SIZE 32 + +// kernelExitToEl0 is the entrypoint for application in guest_el0. +// Prepare the vcpu environment for container application. +TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 + // Step1, save sentry context into memory. + MRS TPIDR_EL1, RSV_REG + REGISTERS_SAVE(RSV_REG, CPU_REGISTERS) + MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG) + MRS TPIDR_EL0, R3 + MOVD R3, CPU_REGISTERS+PTRACE_TLS(RSV_REG) + + WORD $0xd5384003 // MRS SPSR_EL1, R3 + MOVD R3, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG) + MOVD R30, CPU_REGISTERS+PTRACE_PC(RSV_REG) + MOVD RSP, R3 + MOVD R3, CPU_REGISTERS+PTRACE_SP(RSV_REG) + + MOVD CPU_REGISTERS+PTRACE_R3(RSV_REG), R3 + + // Step2, switch to temporary stack. + LOAD_KERNEL_STACK(RSV_REG) + + // Step3, load app context pointer. + MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP + + // Step4, prepare the environment for container application. + // set sp_el0. + MOVD PTRACE_SP(RSV_REG_APP), R1 + WORD $0xd5184101 //MSR R1, SP_EL0 + // set pc. + MOVD PTRACE_PC(RSV_REG_APP), R1 + MSR R1, ELR_EL1 + // set pstate. + MOVD PTRACE_PSTATE(RSV_REG_APP), R1 + WORD $0xd5184001 //MSR R1, SPSR_EL1 + + // need use kernel space address to excute below code, since + // after SWITCH_TO_APP_PAGETABLE the ASID is changed to app's + // ASID. + WORD $0x10000061 // ADR R1, do_exit_to_el0 + ORR $0xffff000000000000, R1, R1 + JMP (R1) + +do_exit_to_el0: + // RSV_REG & RSV_REG_APP will be loaded at the end. + REGISTERS_LOAD(RSV_REG_APP, 0) + MOVD PTRACE_TLS(RSV_REG_APP), RSV_REG + MSR RSV_REG, TPIDR_EL0 + + // switch to user pagetable. + MOVD PTRACE_R18(RSV_REG_APP), RSV_REG + MOVD PTRACE_R9(RSV_REG_APP), RSV_REG_APP + + SUB $STACK_FRAME_SIZE, RSP, RSP + STP (RSV_REG, RSV_REG_APP), 16*0(RSP) + STP (R0, R1), 16*1(RSP) + + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + + SWITCH_TO_APP_PAGETABLE() + + LDP 16*1(RSP), (R0, R1) + LDP 16*0(RSP), (RSV_REG, RSV_REG_APP) + ADD $STACK_FRAME_SIZE, RSP, RSP + + ERET() + +// kernelExitToEl1 is the entrypoint for sentry in guest_el1. +// Prepare the vcpu environment for sentry. +TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + MOVD CPU_REGISTERS+PTRACE_PSTATE(RSV_REG), R1 + WORD $0xd5184001 //MSR R1, SPSR_EL1 + + MOVD CPU_REGISTERS+PTRACE_PC(RSV_REG), R1 + MSR R1, ELR_EL1 + + // restore sentry's tls. + MOVD CPU_REGISTERS+PTRACE_TLS(RSV_REG), R1 + MSR R1, TPIDR_EL0 + + MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R1 + MOVD R1, RSP + + REGISTERS_LOAD(RSV_REG, CPU_REGISTERS) + SWITCH_TO_KVM_PAGETABLE() + MRS TPIDR_EL1, RSV_REG + + MOVD CPU_REGISTERS+PTRACE_R9(RSV_REG), RSV_REG_APP + + ERET() + +// Start is the CPU entrypoint. +TEXT ·Start(SB),NOSPLIT,$0 + // Init. + WORD $0xd508871f // __tlbi(vmalle1) + DSB $7 // dsb(nsh) + + MOVD $1<<12, R1 // Reset mdscr_el1 and disable + MSR R1, MDSCR_EL1 // access to the DCC from EL0 + ISB $15 + + MRS TTBR1_EL1, R1 + MSR R1, TTBR0_EL1 + ISB $15 + + MOVD $CNTKCTL_EL1_DEFAULT, R1 + MSR R1, CNTKCTL_EL1 + + MOVD R8, RSV_REG + ORR $0xffff000000000000, RSV_REG, RSV_REG + WORD $0xd518d092 //MSR R18, TPIDR_EL1 + + // Init. + MOVD $SCTLR_EL1_DEFAULT, R1 // re-enable the mmu. + MSR R1, SCTLR_EL1 + ISB $15 + WORD $0xd508751f // ic iallu + + DSB $7 // dsb(nsh) + ISB $15 + + B ·kernelExitToEl1(SB) + +// El1_sync_invalid is the handler for an invalid EL1_sync. +TEXT ·El1_sync_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_irq_invalid is the handler for an invalid El1_irq. +TEXT ·El1_irq_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_fiq_invalid is the handler for an invalid El1_fiq. +TEXT ·El1_fiq_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_error_invalid is the handler for an invalid El1_error. +TEXT ·El1_error_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_sync is the handler for El1_sync. +TEXT ·El1_sync(SB),NOSPLIT,$0 + KERNEL_ENTRY_FROM_EL1 + MRS ESR_EL1, R25 // read the syndrome register + LSR $ESR_ELx_EC_SHIFT, R25, R24 // exception class + CMP $ESR_ELx_EC_DABT_CUR, R24 + BEQ el1_da // data abort in EL1 + CMP $ESR_ELx_EC_IABT_CUR, R24 + BEQ el1_ia // instruction abort in EL1 + CMP $ESR_ELx_EC_SP_ALIGN, R24 + BEQ el1_sp_pc // stack alignment exception + CMP $ESR_ELx_EC_PC_ALIGN, R24 + BEQ el1_sp_pc // pc alignment exception + CMP $ESR_ELx_EC_UNKNOWN, R24 + BEQ el1_undef // unknown exception in EL1 + CMP $ESR_ELx_EC_SVC64, R24 + BEQ el1_svc // SVC in 64-bit state + CMP $ESR_ELx_EC_BREAKPT_CUR, R24 + BEQ el1_dbg // debug exception in EL1 + CMP $ESR_ELx_EC_FP_ASIMD, R24 + BEQ el1_fpsimd_acc // FP/ASIMD access + CMP $ESR_ELx_EC_SVE, R24 + BEQ el1_sve_acc // SVE access + B el1_invalid + +el1_da: + EXCEPTION_EL1(El1SyncDa) +el1_ia: + EXCEPTION_EL1(El1SyncIa) +el1_sp_pc: + EXCEPTION_EL1(El1SyncSpPc) +el1_undef: + EXCEPTION_EL1(El1SyncUndef) +el1_svc: + B ·HaltEl1SvcAndResume(SB) +el1_dbg: + EXCEPTION_EL1(El1SyncDbg) +el1_fpsimd_acc: +el1_sve_acc: + VFP_ENABLE + B ·kernelExitToEl1(SB) // Resume. +el1_invalid: + EXCEPTION_EL1(El1SyncInv) + +// El1_irq is the handler for El1_irq. +TEXT ·El1_irq(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_fiq is the handler for El1_fiq. +TEXT ·El1_fiq(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El1_error is the handler for El1_error. +TEXT ·El1_error(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// El0_sync is the handler for El0_sync. +TEXT ·El0_sync(SB),NOSPLIT,$0 + KERNEL_ENTRY_FROM_EL0 + MRS ESR_EL1, R25 // read the syndrome register + LSR $ESR_ELx_EC_SHIFT, R25, R24 // exception class + CMP $ESR_ELx_EC_SVC64, R24 + BEQ el0_svc // SVC in 64-bit state + CMP $ESR_ELx_EC_DABT_LOW, R24 + BEQ el0_da // data abort in EL0 + CMP $ESR_ELx_EC_IABT_LOW, R24 + BEQ el0_ia // instruction abort in EL0 + CMP $ESR_ELx_EC_FP_ASIMD, R24 + BEQ el0_fpsimd_acc // FP/ASIMD access + CMP $ESR_ELx_EC_SVE, R24 + BEQ el0_sve_acc // SVE access + CMP $ESR_ELx_EC_FP_EXC64, R24 + BEQ el0_fpsimd_exc // FP/ASIMD exception + CMP $ESR_ELx_EC_SP_ALIGN, R24 + BEQ el0_sp_pc // stack alignment exception + CMP $ESR_ELx_EC_PC_ALIGN, R24 + BEQ el0_sp_pc // pc alignment exception + CMP $ESR_ELx_EC_UNKNOWN, R24 + BEQ el0_undef // unknown exception in EL0 + CMP $ESR_ELx_EC_BREAKPT_LOW, R24 + BEQ el0_dbg // debug exception in EL0 + CMP $ESR_ELx_EC_SYS64, R24 + BEQ el0_sys // configurable trap + CMP $ESR_ELx_EC_WFx, R24 + BEQ el0_wfx // WFX trap + B el0_invalid + +el0_svc: + WORD $0xd538d092 //MRS TPIDR_EL1, R18 + + MOVD $0, CPU_ERROR_CODE(RSV_REG) // Clear error code. + + MOVD $1, R3 + MOVD R3, CPU_ERROR_TYPE(RSV_REG) // Set error type to user. + + MOVD $Syscall, R3 + MOVD R3, CPU_VECTOR_CODE(RSV_REG) + + B ·kernelExitToEl1(SB) + +el0_da: +el0_ia: + EXCEPTION_EL0(PageFault) +el0_fpsimd_acc: + EXCEPTION_EL0(El0SyncFpsimdAcc) +el0_sve_acc: + EXCEPTION_EL0(El0SyncSveAcc) +el0_fpsimd_exc: + EXCEPTION_EL0(El0SyncFpsimdExc) +el0_sp_pc: + EXCEPTION_EL0(El0SyncSpPc) +el0_undef: + EXCEPTION_EL0(El0SyncUndef) +el0_dbg: + EXCEPTION_EL0(El0SyncDbg) +el0_sys: + EXCEPTION_EL0(El0SyncSys) +el0_wfx: + EXCEPTION_EL0(El0SyncWfx) +el0_invalid: + EXCEPTION_EL0(El0SyncInv) + +TEXT ·El0_irq(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +TEXT ·El0_fiq(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +TEXT ·El0_error(SB),NOSPLIT,$0 + KERNEL_ENTRY_FROM_EL0 + WORD $0xd5385219 // MRS ESR_EL1, R25 + AND $ESR_ELx_SERR_MASK, R25, R24 + CMP $ESR_ELx_SERR_NMI, R24 + BEQ el0_nmi + B el0_bounce + +el0_nmi: + EXCEPTION_EL0(El0ErrNMI) +el0_bounce: + EXCEPTION_EL0(VirtualizationException) + +TEXT ·El0_sync_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +TEXT ·El0_irq_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +TEXT ·El0_error_invalid(SB),NOSPLIT,$0 + B ·Shutdown(SB) + +// Vectors implements exception vector table. +// The start address of exception vector table should be 11-bits aligned. +// For detail, please refer to arm developer document: +// https://developer.arm.com/documentation/100933/0100/AArch64-exception-vector-table +// Also can refer to the code in linux kernel: arch/arm64/kernel/entry.S +TEXT ·Vectors(SB),NOSPLIT,$0 + PCALIGN $2048 + B ·El1_sync_invalid(SB) + PCALIGN $128 + B ·El1_irq_invalid(SB) + PCALIGN $128 + B ·El1_fiq_invalid(SB) + PCALIGN $128 + B ·El1_error_invalid(SB) + + PCALIGN $128 + B ·El1_sync(SB) + PCALIGN $128 + B ·El1_irq(SB) + PCALIGN $128 + B ·El1_fiq(SB) + PCALIGN $128 + B ·El1_error(SB) + + PCALIGN $128 + B ·El0_sync(SB) + PCALIGN $128 + B ·El0_irq(SB) + PCALIGN $128 + B ·El0_fiq(SB) + PCALIGN $128 + B ·El0_error(SB) + + PCALIGN $128 + B ·El0_sync_invalid(SB) + PCALIGN $128 + B ·El0_irq_invalid(SB) + PCALIGN $128 + B ·El0_fiq_invalid(SB) + PCALIGN $128 + B ·El0_error_invalid(SB) diff --git a/pkg/ring0/kernel.go b/pkg/ring0/kernel.go new file mode 100644 index 000000000..292f9d0cc --- /dev/null +++ b/pkg/ring0/kernel.go @@ -0,0 +1,90 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package ring0 + +// Init initializes a new kernel. +// +//go:nosplit +func (k *Kernel) Init(maxCPUs int) { + k.init(maxCPUs) +} + +// Halt halts execution. +func Halt() + +// defaultHooks implements hooks. +type defaultHooks struct{} + +// KernelSyscall implements Hooks.KernelSyscall. +// +// +checkescape:all +// +//go:nosplit +func (defaultHooks) KernelSyscall() { + Halt() +} + +// KernelException implements Hooks.KernelException. +// +// +checkescape:all +// +//go:nosplit +func (defaultHooks) KernelException(Vector) { + Halt() +} + +// kernelSyscall is a trampoline. +// +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// +// +checkescape:hard,stack +// +//go:nosplit +func kernelSyscall(c *CPU) { + c.hooks.KernelSyscall() +} + +// kernelException is a trampoline. +// +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// +// +checkescape:hard,stack +// +//go:nosplit +func kernelException(c *CPU, vector Vector) { + c.hooks.KernelException(vector) +} + +// Init initializes a new CPU. +// +// Init allows embedding in other objects. +func (c *CPU) Init(k *Kernel, cpuID int, hooks Hooks) { + c.self = c // Set self reference. + c.kernel = k // Set kernel reference. + c.init(cpuID) // Perform architectural init. + + // Require hooks. + if hooks != nil { + c.hooks = hooks + } else { + c.hooks = defaultHooks{} + } +} diff --git a/pkg/ring0/kernel_amd64.go b/pkg/ring0/kernel_amd64.go new file mode 100644 index 000000000..36a60700e --- /dev/null +++ b/pkg/ring0/kernel_amd64.go @@ -0,0 +1,323 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package ring0 + +import ( + "encoding/binary" + "reflect" + + "gvisor.dev/gvisor/pkg/usermem" +) + +// init initializes architecture-specific state. +func (k *Kernel) init(maxCPUs int) { + entrySize := reflect.TypeOf(kernelEntry{}).Size() + var ( + entries []kernelEntry + padding = 1 + ) + for { + entries = make([]kernelEntry, maxCPUs+padding-1) + totalSize := entrySize * uintptr(maxCPUs+padding-1) + addr := reflect.ValueOf(&entries[0]).Pointer() + if addr&(usermem.PageSize-1) == 0 && totalSize >= usermem.PageSize { + // The runtime forces power-of-2 alignment for allocations, and we are therefore + // safe once the first address is aligned and the chunk is at least a full page. + break + } + padding = padding << 1 + } + k.cpuEntries = entries + + k.globalIDT = &idt64{} + if reflect.TypeOf(idt64{}).Size() != usermem.PageSize { + panic("Size of globalIDT should be PageSize") + } + if reflect.ValueOf(k.globalIDT).Pointer()&(usermem.PageSize-1) != 0 { + panic("Allocated globalIDT should be page aligned") + } + + // Setup the IDT, which is uniform. + for v, handler := range handlers { + // Allow Breakpoint and Overflow to be called from all + // privilege levels. + dpl := 0 + if v == Breakpoint || v == Overflow { + dpl = 3 + } + // Note that we set all traps to use the interrupt stack, this + // is defined below when setting up the TSS. + k.globalIDT[v].setInterrupt(Kcode, uint64(kernelFunc(handler)), dpl, 1 /* ist */) + } +} + +// EntryRegions returns the set of kernel entry regions (must be mapped). +func (k *Kernel) EntryRegions() map[uintptr]uintptr { + regions := make(map[uintptr]uintptr) + + addr := reflect.ValueOf(&k.cpuEntries[0]).Pointer() + size := reflect.TypeOf(kernelEntry{}).Size() * uintptr(len(k.cpuEntries)) + end, _ := usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + addr = reflect.ValueOf(k.globalIDT).Pointer() + size = reflect.TypeOf(idt64{}).Size() + end, _ = usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + return regions +} + +// init initializes architecture-specific state. +func (c *CPU) init(cpuID int) { + c.kernelEntry = &c.kernel.cpuEntries[cpuID] + c.cpuSelf = c + // Null segment. + c.gdt[0].setNull() + + // Kernel & user segments. + c.gdt[segKcode] = KernelCodeSegment + c.gdt[segKdata] = KernelDataSegment + c.gdt[segUcode32] = UserCodeSegment32 + c.gdt[segUdata] = UserDataSegment + c.gdt[segUcode64] = UserCodeSegment64 + + // The task segment, this spans two entries. + tssBase, tssLimit, _ := c.TSS() + c.gdt[segTss].set( + uint32(tssBase), + uint32(tssLimit), + 0, // Privilege level zero. + SegmentDescriptorPresent| + SegmentDescriptorAccess| + SegmentDescriptorWrite| + SegmentDescriptorExecute) + c.gdt[segTssHi].setHi(uint32((tssBase) >> 32)) + + // Set the kernel stack pointer in the TSS (virtual address). + stackAddr := c.StackTop() + c.stackTop = stackAddr + c.tss.rsp0Lo = uint32(stackAddr) + c.tss.rsp0Hi = uint32(stackAddr >> 32) + c.tss.ist1Lo = uint32(stackAddr) + c.tss.ist1Hi = uint32(stackAddr >> 32) + + // Set the I/O bitmap base address beyond the last byte in the TSS + // to block access to the entire I/O address range. + // + // From section 18.5.2 "I/O Permission Bit Map" from Intel SDM vol1: + // I/O addresses not spanned by the map are treated as if they had set + // bits in the map. + c.tss.ioPerm = tssLimit + 1 + + // Permanently set the kernel segments. + c.registers.Cs = uint64(Kcode) + c.registers.Ds = uint64(Kdata) + c.registers.Es = uint64(Kdata) + c.registers.Ss = uint64(Kdata) + c.registers.Fs = uint64(Kdata) + c.registers.Gs = uint64(Kdata) + + // Set mandatory flags. + c.registers.Eflags = KernelFlagsSet +} + +// StackTop returns the kernel's stack address. +// +//go:nosplit +func (c *CPU) StackTop() uint64 { + return uint64(kernelAddr(&c.stack[0])) + uint64(len(c.stack)) +} + +// IDT returns the CPU's IDT base and limit. +// +//go:nosplit +func (c *CPU) IDT() (uint64, uint16) { + return uint64(kernelAddr(&c.kernel.globalIDT[0])), uint16(binary.Size(&c.kernel.globalIDT) - 1) +} + +// GDT returns the CPU's GDT base and limit. +// +//go:nosplit +func (c *CPU) GDT() (uint64, uint16) { + return uint64(kernelAddr(&c.gdt[0])), uint16(8*segLast - 1) +} + +// TSS returns the CPU's TSS base, limit and value. +// +//go:nosplit +func (c *CPU) TSS() (uint64, uint16, *SegmentDescriptor) { + return uint64(kernelAddr(&c.tss)), uint16(binary.Size(&c.tss) - 1), &c.gdt[segTss] +} + +// CR0 returns the CPU's CR0 value. +// +//go:nosplit +func (c *CPU) CR0() uint64 { + return _CR0_PE | _CR0_PG | _CR0_AM | _CR0_ET +} + +// CR4 returns the CPU's CR4 value. +// +//go:nosplit +func (c *CPU) CR4() uint64 { + cr4 := uint64(_CR4_PAE | _CR4_PSE | _CR4_OSFXSR | _CR4_OSXMMEXCPT) + if hasPCID { + cr4 |= _CR4_PCIDE + } + if hasXSAVE { + cr4 |= _CR4_OSXSAVE + } + if hasSMEP { + cr4 |= _CR4_SMEP + } + if hasFSGSBASE { + cr4 |= _CR4_FSGSBASE + } + return cr4 +} + +// EFER returns the CPU's EFER value. +// +//go:nosplit +func (c *CPU) EFER() uint64 { + return _EFER_LME | _EFER_LMA | _EFER_SCE | _EFER_NX +} + +// IsCanonical indicates whether addr is canonical per the amd64 spec. +// +//go:nosplit +func IsCanonical(addr uint64) bool { + return addr <= 0x00007fffffffffff || addr > 0xffff800000000000 +} + +// SwitchToUser performs either a sysret or an iret. +// +// The return value is the vector that interrupted execution. +// +// This function will not split the stack. Callers will probably want to call +// runtime.entersyscall (and pair with a call to runtime.exitsyscall) prior to +// calling this function. +// +// When this is done, this region is quite sensitive to things like system +// calls. After calling entersyscall, any memory used must have been allocated +// and no function calls without go:nosplit are permitted. Any calls made here +// are protected appropriately (e.g. IsCanonical and CR3). +// +// Also note that this function transitively depends on the compiler generating +// code that uses IP-relative addressing inside of absolute addresses. That's +// the case for amd64, but may not be the case for other architectures. +// +// Precondition: the Rip, Rsp, Fs and Gs registers must be canonical. +// +// +checkescape:all +// +//go:nosplit +func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { + userCR3 := switchOpts.PageTables.CR3(!switchOpts.Flush, switchOpts.UserPCID) + c.kernelCR3 = uintptr(c.kernel.PageTables.CR3(true, switchOpts.KernelPCID)) + + // Sanitize registers. + regs := switchOpts.Registers + regs.Eflags &= ^uint64(UserFlagsClear) + regs.Eflags |= UserFlagsSet + regs.Cs = uint64(Ucode64) // Required for iret. + regs.Ss = uint64(Udata) // Ditto. + + // Perform the switch. + swapgs() // GS will be swapped on return. + WriteFS(uintptr(regs.Fs_base)) // escapes: no. Set application FS. + WriteGS(uintptr(regs.Gs_base)) // escapes: no. Set application GS. + LoadFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy in floating point. + if switchOpts.FullRestore { + vector = iret(c, regs, uintptr(userCR3)) + } else { + vector = sysret(c, regs, uintptr(userCR3)) + } + SaveFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy out floating point. + WriteFS(uintptr(c.registers.Fs_base)) // escapes: no. Restore kernel FS. + return +} + +// start is the CPU entrypoint. +// +// This is called from the Start asm stub (see entry_amd64.go); on return the +// registers in c.registers will be restored (not segments). +// +//go:nosplit +func start(c *CPU) { + // Save per-cpu & FS segment. + WriteGS(kernelAddr(c.kernelEntry)) + WriteFS(uintptr(c.registers.Fs_base)) + + // Initialize floating point. + // + // Note that on skylake, the valid XCR0 mask reported seems to be 0xff. + // This breaks down as: + // + // bit0 - x87 + // bit1 - SSE + // bit2 - AVX + // bit3-4 - MPX + // bit5-7 - AVX512 + // + // For some reason, enabled MPX & AVX512 on platforms that report them + // seems to be cause a general protection fault. (Maybe there are some + // virtualization issues and these aren't exported to the guest cpuid.) + // This needs further investigation, but we can limit the floating + // point operations to x87, SSE & AVX for now. + fninit() + xsetbv(0, validXCR0Mask&0x7) + + // Set the syscall target. + wrmsr(_MSR_LSTAR, kernelFunc(sysenter)) + wrmsr(_MSR_SYSCALL_MASK, KernelFlagsClear|_RFLAGS_DF) + + // NOTE: This depends on having the 64-bit segments immediately + // following the 32-bit user segments. This is simply the way the + // sysret instruction is designed to work (it assumes they follow). + wrmsr(_MSR_STAR, uintptr(uint64(Kcode)<<32|uint64(Ucode32)<<48)) + wrmsr(_MSR_CSTAR, kernelFunc(sysenter)) +} + +// SetCPUIDFaulting sets CPUID faulting per the boolean value. +// +// True is returned if faulting could be set. +// +//go:nosplit +func SetCPUIDFaulting(on bool) bool { + // Per the SDM (Vol 3, Table 2-43), PLATFORM_INFO bit 31 denotes support + // for CPUID faulting, and we enable and disable via the MISC_FEATURES MSR. + if rdmsr(_MSR_PLATFORM_INFO)&_PLATFORM_INFO_CPUID_FAULT != 0 { + features := rdmsr(_MSR_MISC_FEATURES) + if on { + features |= _MISC_FEATURE_CPUID_TRAP + } else { + features &^= _MISC_FEATURE_CPUID_TRAP + } + wrmsr(_MSR_MISC_FEATURES, features) + return true // Setting successful. + } + return false +} + +// ReadCR2 reads the current CR2 value. +// +//go:nosplit +func ReadCR2() uintptr { + return readCR2() +} diff --git a/pkg/ring0/kernel_arm64.go b/pkg/ring0/kernel_arm64.go new file mode 100644 index 000000000..c05284641 --- /dev/null +++ b/pkg/ring0/kernel_arm64.go @@ -0,0 +1,85 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ring0 + +// HaltAndResume halts execution and point the pointer to the resume function. +//go:nosplit +func HaltAndResume() + +// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume. +//go:nosplit +func HaltEl1SvcAndResume() + +// HaltEl1ExceptionAndResume calls Hooks.KernelException and resume. +//go:nosplit +func HaltEl1ExceptionAndResume() + +// init initializes architecture-specific state. +func (k *Kernel) init(maxCPUs int) { +} + +// init initializes architecture-specific state. +func (c *CPU) init(cpuID int) { + // Set the kernel stack pointer(virtual address). + c.registers.Sp = uint64(c.StackTop()) + +} + +// StackTop returns the kernel's stack address. +// +//go:nosplit +func (c *CPU) StackTop() uint64 { + return uint64(kernelAddr(&c.stack[0])) + uint64(len(c.stack)) +} + +// IsCanonical indicates whether addr is canonical per the arm64 spec. +// +//go:nosplit +func IsCanonical(addr uint64) bool { + return addr <= 0x0000ffffffffffff || addr > 0xffff000000000000 +} + +// SwitchToUser performs an eret. +// +// The return value is the exception vector. +// +// +checkescape:all +// +//go:nosplit +func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { + storeAppASID(uintptr(switchOpts.UserASID)) + if switchOpts.Flush { + FlushTlbByASID(uintptr(switchOpts.UserASID)) + } + + regs := switchOpts.Registers + + regs.Pstate &= ^uint64(PsrFlagsClear) + regs.Pstate |= UserFlagsSet + + EnableVFP() + LoadFloatingPoint(switchOpts.FloatingPointState) + + kernelExitToEl0() + + SaveFloatingPoint(switchOpts.FloatingPointState) + DisableVFP() + + vector = c.vecCode + + return +} diff --git a/pkg/ring0/kernel_unsafe.go b/pkg/ring0/kernel_unsafe.go new file mode 100644 index 000000000..16955ad91 --- /dev/null +++ b/pkg/ring0/kernel_unsafe.go @@ -0,0 +1,41 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package ring0 + +import ( + "unsafe" +) + +// eface mirrors runtime.eface. +type eface struct { + typ uintptr + data unsafe.Pointer +} + +// kernelAddr returns the kernel virtual address for the given object. +// +//go:nosplit +func kernelAddr(obj interface{}) uintptr { + e := (*eface)(unsafe.Pointer(&obj)) + return KernelStartAddress | uintptr(e.data) +} + +// kernelFunc returns the address of the given function. +// +//go:nosplit +func kernelFunc(fn func()) uintptr { + fnptr := (**uintptr)(unsafe.Pointer(&fn)) + return KernelStartAddress | **fnptr +} diff --git a/pkg/ring0/lib_amd64.go b/pkg/ring0/lib_amd64.go new file mode 100644 index 000000000..0ec5c3bc5 --- /dev/null +++ b/pkg/ring0/lib_amd64.go @@ -0,0 +1,119 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package ring0 + +import ( + "gvisor.dev/gvisor/pkg/cpuid" +) + +// LoadFloatingPoint loads floating point state by the most efficient mechanism +// available (set by Init). +var LoadFloatingPoint func(*byte) + +// SaveFloatingPoint saves floating point state by the most efficient mechanism +// available (set by Init). +var SaveFloatingPoint func(*byte) + +// fxrstor uses fxrstor64 to load floating point state. +func fxrstor(*byte) + +// xrstor uses xrstor to load floating point state. +func xrstor(*byte) + +// fxsave uses fxsave64 to save floating point state. +func fxsave(*byte) + +// xsave uses xsave to save floating point state. +func xsave(*byte) + +// xsaveopt uses xsaveopt to save floating point state. +func xsaveopt(*byte) + +// WriteFS sets the GS address (set by init). +var WriteFS func(addr uintptr) + +// wrfsbase writes to the GS base address. +func wrfsbase(addr uintptr) + +// wrfsmsr writes to the GS_BASE MSR. +func wrfsmsr(addr uintptr) + +// WriteGS sets the GS address (set by init). +var WriteGS func(addr uintptr) + +// wrgsbase writes to the GS base address. +func wrgsbase(addr uintptr) + +// wrgsmsr writes to the GS_BASE MSR. +func wrgsmsr(addr uintptr) + +// readCR2 reads the current CR2 value. +func readCR2() uintptr + +// fninit initializes the floating point unit. +func fninit() + +// xsetbv writes to an extended control register. +func xsetbv(reg, value uintptr) + +// xgetbv reads an extended control register. +func xgetbv(reg uintptr) uintptr + +// wrmsr reads to the given MSR. +func wrmsr(reg, value uintptr) + +// rdmsr reads the given MSR. +func rdmsr(reg uintptr) uintptr + +// Mostly-constants set by Init. +var ( + hasSMEP bool + hasPCID bool + hasXSAVEOPT bool + hasXSAVE bool + hasFSGSBASE bool + validXCR0Mask uintptr +) + +// Init sets function pointers based on architectural features. +// +// This must be called prior to using ring0. +func Init(featureSet *cpuid.FeatureSet) { + hasSMEP = featureSet.HasFeature(cpuid.X86FeatureSMEP) + hasPCID = featureSet.HasFeature(cpuid.X86FeaturePCID) + hasXSAVEOPT = featureSet.UseXsaveopt() + hasXSAVE = featureSet.UseXsave() + hasFSGSBASE = featureSet.HasFeature(cpuid.X86FeatureFSGSBase) + validXCR0Mask = uintptr(featureSet.ValidXCR0Mask()) + if hasXSAVEOPT { + SaveFloatingPoint = xsaveopt + LoadFloatingPoint = xrstor + } else if hasXSAVE { + SaveFloatingPoint = xsave + LoadFloatingPoint = xrstor + } else { + SaveFloatingPoint = fxsave + LoadFloatingPoint = fxrstor + } + if hasFSGSBASE { + WriteFS = wrfsbase + WriteGS = wrgsbase + } else { + WriteFS = wrfsmsr + WriteGS = wrgsmsr + } +} diff --git a/pkg/ring0/lib_amd64.s b/pkg/ring0/lib_amd64.s new file mode 100644 index 000000000..2fe83568a --- /dev/null +++ b/pkg/ring0/lib_amd64.s @@ -0,0 +1,200 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "funcdata.h" +#include "textflag.h" + +// fxrstor loads floating point state. +// +// The code corresponds to: +// +// fxrstor64 (%rbx) +// +TEXT ·fxrstor(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), BX + MOVL $0xffffffff, AX + MOVL $0xffffffff, DX + BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x0b; + RET + +// xrstor loads floating point state. +// +// The code corresponds to: +// +// xrstor (%rdi) +// +TEXT ·xrstor(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), DI + MOVL $0xffffffff, AX + MOVL $0xffffffff, DX + BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x2f; + RET + +// fxsave saves floating point state. +// +// The code corresponds to: +// +// fxsave64 (%rbx) +// +TEXT ·fxsave(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), BX + MOVL $0xffffffff, AX + MOVL $0xffffffff, DX + BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x03; + RET + +// xsave saves floating point state. +// +// The code corresponds to: +// +// xsave (%rdi) +// +TEXT ·xsave(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), DI + MOVL $0xffffffff, AX + MOVL $0xffffffff, DX + BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x27; + RET + +// xsaveopt saves floating point state. +// +// The code corresponds to: +// +// xsaveopt (%rdi) +// +TEXT ·xsaveopt(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), DI + MOVL $0xffffffff, AX + MOVL $0xffffffff, DX + BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x37; + RET + +// wrfsbase writes to the FS base. +// +// The code corresponds to: +// +// wrfsbase %rax +// +TEXT ·wrfsbase(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), AX + BYTE $0xf3; BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0xd0; + RET + +// wrfsmsr writes to the FSBASE MSR. +// +// The code corresponds to: +// +// wrmsr (writes EDX:EAX to the MSR in ECX) +// +TEXT ·wrfsmsr(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), AX + MOVQ AX, DX + SHRQ $32, DX + MOVQ $0xc0000100, CX // MSR_FS_BASE + BYTE $0x0f; BYTE $0x30; + RET + +// wrgsbase writes to the GS base. +// +// The code corresponds to: +// +// wrgsbase %rax +// +TEXT ·wrgsbase(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), AX + BYTE $0xf3; BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0xd8; + RET + +// wrgsmsr writes to the GSBASE MSR. +// +// See wrfsmsr. +TEXT ·wrgsmsr(SB),NOSPLIT,$0-8 + MOVQ addr+0(FP), AX + MOVQ AX, DX + SHRQ $32, DX + MOVQ $0xc0000101, CX // MSR_GS_BASE + BYTE $0x0f; BYTE $0x30; // WRMSR + RET + +// readCR2 reads the current CR2 value. +// +// The code corresponds to: +// +// mov %cr2, %rax +// +TEXT ·readCR2(SB),NOSPLIT,$0-8 + BYTE $0x0f; BYTE $0x20; BYTE $0xd0; + MOVQ AX, ret+0(FP) + RET + +// fninit initializes the floating point unit. +// +// The code corresponds to: +// +// fninit +TEXT ·fninit(SB),NOSPLIT,$0 + BYTE $0xdb; BYTE $0xe3; + RET + +// xsetbv writes to an extended control register. +// +// The code corresponds to: +// +// xsetbv +// +TEXT ·xsetbv(SB),NOSPLIT,$0-16 + MOVL reg+0(FP), CX + MOVL value+8(FP), AX + MOVL value+12(FP), DX + BYTE $0x0f; BYTE $0x01; BYTE $0xd1; + RET + +// xgetbv reads an extended control register. +// +// The code corresponds to: +// +// xgetbv +// +TEXT ·xgetbv(SB),NOSPLIT,$0-16 + MOVL reg+0(FP), CX + BYTE $0x0f; BYTE $0x01; BYTE $0xd0; + MOVL AX, ret+8(FP) + MOVL DX, ret+12(FP) + RET + +// wrmsr writes to a control register. +// +// The code corresponds to: +// +// wrmsr +// +TEXT ·wrmsr(SB),NOSPLIT,$0-16 + MOVL reg+0(FP), CX + MOVL value+8(FP), AX + MOVL value+12(FP), DX + BYTE $0x0f; BYTE $0x30; + RET + +// rdmsr reads a control register. +// +// The code corresponds to: +// +// rdmsr +// +TEXT ·rdmsr(SB),NOSPLIT,$0-16 + MOVL reg+0(FP), CX + BYTE $0x0f; BYTE $0x32; + MOVL AX, ret+8(FP) + MOVL DX, ret+12(FP) + RET diff --git a/pkg/ring0/lib_arm64.go b/pkg/ring0/lib_arm64.go new file mode 100644 index 000000000..a490bf3af --- /dev/null +++ b/pkg/ring0/lib_arm64.go @@ -0,0 +1,71 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package ring0 + +// storeAppASID writes the application's asid value. +func storeAppASID(asid uintptr) + +// LocalFlushTlbAll same as FlushTlbAll, but only applies to the calling CPU. +func LocalFlushTlbAll() + +// FlushTlbByVA invalidates tlb by VA/Last-level/Inner-Shareable. +func FlushTlbByVA(addr uintptr) + +// FlushTlbByASID invalidates tlb by ASID/Inner-Shareable. +func FlushTlbByASID(asid uintptr) + +// FlushTlbAll invalidates all tlb. +func FlushTlbAll() + +// CPACREL1 returns the value of the CPACR_EL1 register. +func CPACREL1() (value uintptr) + +// GetFPCR returns the value of FPCR register. +func GetFPCR() (value uintptr) + +// SetFPCR writes the FPCR value. +func SetFPCR(value uintptr) + +// GetFPSR returns the value of FPSR register. +func GetFPSR() (value uintptr) + +// SetFPSR writes the FPSR value. +func SetFPSR(value uintptr) + +// SaveVRegs saves V0-V31 registers. +// V0-V31: 32 128-bit registers for floating point and simd. +func SaveVRegs(*byte) + +// LoadVRegs loads V0-V31 registers. +func LoadVRegs(*byte) + +// LoadFloatingPoint loads floating point state. +func LoadFloatingPoint(*byte) + +// SaveFloatingPoint saves floating point state. +func SaveFloatingPoint(*byte) + +// EnableVFP enables fpsimd. +func EnableVFP() + +// DisableVFP disables fpsimd. +func DisableVFP() + +// Init sets function pointers based on architectural features. +// +// This must be called prior to using ring0. +func Init() {} diff --git a/pkg/ring0/lib_arm64.s b/pkg/ring0/lib_arm64.s new file mode 100644 index 000000000..e39b32841 --- /dev/null +++ b/pkg/ring0/lib_arm64.s @@ -0,0 +1,180 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "funcdata.h" +#include "textflag.h" + +#define TLBI_ASID_SHIFT 48 + +TEXT ·FlushTlbByVA(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + DSB $10 // dsb(ishst) + WORD $0xd50883a1 // tlbi vale1is, x1 + DSB $11 // dsb(ish) + RET + +TEXT ·FlushTlbByASID(SB),NOSPLIT,$0-8 + MOVD asid+0(FP), R1 + LSL $TLBI_ASID_SHIFT, R1, R1 + DSB $10 // dsb(ishst) + WORD $0xd5088341 // tlbi aside1is, x1 + DSB $11 // dsb(ish) + RET + +TEXT ·LocalFlushTlbAll(SB),NOSPLIT,$0 + DSB $6 // dsb(nshst) + WORD $0xd508871f // __tlbi(vmalle1) + DSB $7 // dsb(nsh) + ISB $15 + RET + +TEXT ·FlushTlbAll(SB),NOSPLIT,$0 + DSB $10 // dsb(ishst) + WORD $0xd508831f // __tlbi(vmalle1is) + DSB $11 // dsb(ish) + ISB $15 + RET + +TEXT ·CPACREL1(SB),NOSPLIT,$0-8 + WORD $0xd5381041 // MRS CPACR_EL1, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·GetFPCR(SB),NOSPLIT,$0-8 + MOVD FPCR, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·GetFPSR(SB),NOSPLIT,$0-8 + MOVD FPSR, R1 + MOVD R1, ret+0(FP) + RET + +TEXT ·SetFPCR(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + MOVD R1, FPCR + RET + +TEXT ·SetFPSR(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + MOVD R1, FPSR + RET + +TEXT ·SaveVRegs(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + // Skip aarch64_ctx, fpsr, fpcr. + ADD $16, R0, R0 + + WORD $0xad000400 // stp q0, q1, [x0] + WORD $0xad010c02 // stp q2, q3, [x0, #32] + WORD $0xad021404 // stp q4, q5, [x0, #64] + WORD $0xad031c06 // stp q6, q7, [x0, #96] + WORD $0xad042408 // stp q8, q9, [x0, #128] + WORD $0xad052c0a // stp q10, q11, [x0, #160] + WORD $0xad06340c // stp q12, q13, [x0, #192] + WORD $0xad073c0e // stp q14, q15, [x0, #224] + WORD $0xad084410 // stp q16, q17, [x0, #256] + WORD $0xad094c12 // stp q18, q19, [x0, #288] + WORD $0xad0a5414 // stp q20, q21, [x0, #320] + WORD $0xad0b5c16 // stp q22, q23, [x0, #352] + WORD $0xad0c6418 // stp q24, q25, [x0, #384] + WORD $0xad0d6c1a // stp q26, q27, [x0, #416] + WORD $0xad0e741c // stp q28, q29, [x0, #448] + WORD $0xad0f7c1e // stp q30, q31, [x0, #480] + + RET + +TEXT ·LoadVRegs(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + // Skip aarch64_ctx, fpsr, fpcr. + ADD $16, R0, R0 + + WORD $0xad400400 // ldp q0, q1, [x0] + WORD $0xad410c02 // ldp q2, q3, [x0, #32] + WORD $0xad421404 // ldp q4, q5, [x0, #64] + WORD $0xad431c06 // ldp q6, q7, [x0, #96] + WORD $0xad442408 // ldp q8, q9, [x0, #128] + WORD $0xad452c0a // ldp q10, q11, [x0, #160] + WORD $0xad46340c // ldp q12, q13, [x0, #192] + WORD $0xad473c0e // ldp q14, q15, [x0, #224] + WORD $0xad484410 // ldp q16, q17, [x0, #256] + WORD $0xad494c12 // ldp q18, q19, [x0, #288] + WORD $0xad4a5414 // ldp q20, q21, [x0, #320] + WORD $0xad4b5c16 // ldp q22, q23, [x0, #352] + WORD $0xad4c6418 // ldp q24, q25, [x0, #384] + WORD $0xad4d6c1a // ldp q26, q27, [x0, #416] + WORD $0xad4e741c // ldp q28, q29, [x0, #448] + WORD $0xad4f7c1e // ldp q30, q31, [x0, #480] + + RET + +TEXT ·LoadFloatingPoint(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + MOVD 0(R0), R1 + MOVD R1, FPSR + MOVD 8(R0), R1 + MOVD R1, FPCR + + ADD $16, R0, R0 + + WORD $0xad400400 // ldp q0, q1, [x0] + WORD $0xad410c02 // ldp q2, q3, [x0, #32] + WORD $0xad421404 // ldp q4, q5, [x0, #64] + WORD $0xad431c06 // ldp q6, q7, [x0, #96] + WORD $0xad442408 // ldp q8, q9, [x0, #128] + WORD $0xad452c0a // ldp q10, q11, [x0, #160] + WORD $0xad46340c // ldp q12, q13, [x0, #192] + WORD $0xad473c0e // ldp q14, q15, [x0, #224] + WORD $0xad484410 // ldp q16, q17, [x0, #256] + WORD $0xad494c12 // ldp q18, q19, [x0, #288] + WORD $0xad4a5414 // ldp q20, q21, [x0, #320] + WORD $0xad4b5c16 // ldp q22, q23, [x0, #352] + WORD $0xad4c6418 // ldp q24, q25, [x0, #384] + WORD $0xad4d6c1a // ldp q26, q27, [x0, #416] + WORD $0xad4e741c // ldp q28, q29, [x0, #448] + WORD $0xad4f7c1e // ldp q30, q31, [x0, #480] + + RET + +TEXT ·SaveFloatingPoint(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R0 + + MOVD FPSR, R1 + MOVD R1, 0(R0) + MOVD FPCR, R1 + MOVD R1, 8(R0) + + ADD $16, R0, R0 + + WORD $0xad000400 // stp q0, q1, [x0] + WORD $0xad010c02 // stp q2, q3, [x0, #32] + WORD $0xad021404 // stp q4, q5, [x0, #64] + WORD $0xad031c06 // stp q6, q7, [x0, #96] + WORD $0xad042408 // stp q8, q9, [x0, #128] + WORD $0xad052c0a // stp q10, q11, [x0, #160] + WORD $0xad06340c // stp q12, q13, [x0, #192] + WORD $0xad073c0e // stp q14, q15, [x0, #224] + WORD $0xad084410 // stp q16, q17, [x0, #256] + WORD $0xad094c12 // stp q18, q19, [x0, #288] + WORD $0xad0a5414 // stp q20, q21, [x0, #320] + WORD $0xad0b5c16 // stp q22, q23, [x0, #352] + WORD $0xad0c6418 // stp q24, q25, [x0, #384] + WORD $0xad0d6c1a // stp q26, q27, [x0, #416] + WORD $0xad0e741c // stp q28, q29, [x0, #448] + WORD $0xad0f7c1e // stp q30, q31, [x0, #480] + + RET diff --git a/pkg/ring0/pagetables/allocator.go b/pkg/ring0/pagetables/allocator.go new file mode 100644 index 000000000..8d75b7599 --- /dev/null +++ b/pkg/ring0/pagetables/allocator.go @@ -0,0 +1,127 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +// Allocator is used to allocate and map PTEs. +// +// Note that allocators may be called concurrently. +type Allocator interface { + // NewPTEs returns a new set of PTEs and their physical address. + NewPTEs() *PTEs + + // PhysicalFor gives the physical address for a set of PTEs. + PhysicalFor(ptes *PTEs) uintptr + + // LookupPTEs looks up PTEs by physical address. + LookupPTEs(physical uintptr) *PTEs + + // FreePTEs marks a set of PTEs a freed, although they may not be available + // for use again until Recycle is called, below. + FreePTEs(ptes *PTEs) + + // Recycle makes freed PTEs available for use again. + Recycle() +} + +// RuntimeAllocator is a trivial allocator. +type RuntimeAllocator struct { + // used is the set of PTEs that have been allocated. This includes any + // PTEs that may be in the pool below. PTEs are only freed from this + // map by the Drain call. + // + // This exists to prevent accidental garbage collection. + used map[*PTEs]struct{} + + // pool is the set of free-to-use PTEs. + pool []*PTEs + + // freed is the set of recently-freed PTEs. + freed []*PTEs +} + +// NewRuntimeAllocator returns an allocator that uses runtime allocation. +func NewRuntimeAllocator() *RuntimeAllocator { + r := new(RuntimeAllocator) + r.Init() + return r +} + +// Init initializes a RuntimeAllocator. +func (r *RuntimeAllocator) Init() { + r.used = make(map[*PTEs]struct{}) +} + +// Recycle returns freed pages to the pool. +func (r *RuntimeAllocator) Recycle() { + r.pool = append(r.pool, r.freed...) + r.freed = r.freed[:0] +} + +// Drain empties the pool. +func (r *RuntimeAllocator) Drain() { + r.Recycle() + for i, ptes := range r.pool { + // Zap the entry in the underlying array to ensure that it can + // be properly garbage collected. + r.pool[i] = nil + // Similarly, free the reference held by the used map (these + // also apply for the pool entries). + delete(r.used, ptes) + } + r.pool = r.pool[:0] +} + +// NewPTEs implements Allocator.NewPTEs. +// +// Note that the "physical" address here is actually the virtual address of the +// PTEs structure. The entries are tracked only to avoid garbage collection. +// +// This is guaranteed not to split as long as the pool is sufficiently full. +// +//go:nosplit +func (r *RuntimeAllocator) NewPTEs() *PTEs { + // Pull from the pool if we can. + if len(r.pool) > 0 { + ptes := r.pool[len(r.pool)-1] + r.pool = r.pool[:len(r.pool)-1] + return ptes + } + + // Allocate a new entry. + ptes := newAlignedPTEs() + r.used[ptes] = struct{}{} + return ptes +} + +// PhysicalFor returns the physical address for the given PTEs. +// +//go:nosplit +func (r *RuntimeAllocator) PhysicalFor(ptes *PTEs) uintptr { + return physicalFor(ptes) +} + +// LookupPTEs implements Allocator.LookupPTEs. +// +//go:nosplit +func (r *RuntimeAllocator) LookupPTEs(physical uintptr) *PTEs { + return fromPhysical(physical) +} + +// FreePTEs implements Allocator.FreePTEs. +// +//go:nosplit +func (r *RuntimeAllocator) FreePTEs(ptes *PTEs) { + r.freed = append(r.freed, ptes) +} diff --git a/pkg/ring0/pagetables/allocator_unsafe.go b/pkg/ring0/pagetables/allocator_unsafe.go new file mode 100644 index 000000000..d08bfdeb3 --- /dev/null +++ b/pkg/ring0/pagetables/allocator_unsafe.go @@ -0,0 +1,53 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +import ( + "unsafe" + + "gvisor.dev/gvisor/pkg/usermem" +) + +// newAlignedPTEs returns a set of aligned PTEs. +func newAlignedPTEs() *PTEs { + ptes := new(PTEs) + offset := physicalFor(ptes) & (usermem.PageSize - 1) + if offset == 0 { + // Already aligned. + return ptes + } + + // Need to force an aligned allocation. + unaligned := make([]byte, (2*usermem.PageSize)-1) + offset = uintptr(unsafe.Pointer(&unaligned[0])) & (usermem.PageSize - 1) + if offset != 0 { + offset = usermem.PageSize - offset + } + return (*PTEs)(unsafe.Pointer(&unaligned[offset])) +} + +// physicalFor returns the "physical" address for PTEs. +// +//go:nosplit +func physicalFor(ptes *PTEs) uintptr { + return uintptr(unsafe.Pointer(ptes)) +} + +// fromPhysical returns the PTEs from the "physical" address. +// +//go:nosplit +func fromPhysical(physical uintptr) *PTEs { + return (*PTEs)(unsafe.Pointer(physical)) +} diff --git a/pkg/ring0/pagetables/pagetables.go b/pkg/ring0/pagetables/pagetables.go new file mode 100644 index 000000000..8c0a6aa82 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables.go @@ -0,0 +1,324 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package pagetables provides a generic implementation of pagetables. +// +// The core functions must be safe to call from a nosplit context. Furthermore, +// this pagetables implementation goes to lengths to ensure that all functions +// are free from runtime allocation. Calls to NewPTEs/FreePTEs may be made +// during walks, but these can be cached elsewhere if required. +package pagetables + +import ( + "gvisor.dev/gvisor/pkg/usermem" +) + +// PageTables is a set of page tables. +type PageTables struct { + // Allocator is used to allocate nodes. + Allocator Allocator + + // root is the pagetable root. + // + // For same archs such as amd64, the upper of the PTEs is cloned + // from and owned by upperSharedPageTables which are shared among + // many PageTables if upperSharedPageTables is not nil. + root *PTEs + + // rootPhysical is the cached physical address of the root. + // + // This is saved only to prevent constant translation. + rootPhysical uintptr + + // archPageTables includes architecture-specific features. + archPageTables + + // upperSharedPageTables represents a read-only shared upper + // of the Pagetable. When it is not nil, the upper is not + // allowed to be modified. + upperSharedPageTables *PageTables + + // upperStart is the start address of the upper portion that + // are shared from upperSharedPageTables + upperStart uintptr + + // readOnlyShared indicates the Pagetables are read-only and + // own the ranges that are shared with other Pagetables. + readOnlyShared bool +} + +// Init initializes a set of PageTables. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) Init(allocator Allocator) { + p.Allocator = allocator + p.root = p.Allocator.NewPTEs() + p.rootPhysical = p.Allocator.PhysicalFor(p.root) +} + +// NewWithUpper returns new PageTables. +// +// upperSharedPageTables are used for mapping the upper of addresses, +// starting at upperStart. These pageTables should not be touched (as +// invalidations may be incorrect) after they are passed as an +// upperSharedPageTables. Only when all dependent PageTables are gone +// may they be used. The intenteded use case is for kernel page tables, +// which are static and fixed. +// +// Precondition: upperStart must be between canonical ranges. +// Precondition: upperStart must be pgdSize aligned. +// precondition: upperSharedPageTables must be marked read-only shared. +func NewWithUpper(a Allocator, upperSharedPageTables *PageTables, upperStart uintptr) *PageTables { + p := new(PageTables) + p.Init(a) + + if upperSharedPageTables != nil { + if !upperSharedPageTables.readOnlyShared { + panic("Only read-only shared pagetables can be used as upper") + } + p.upperSharedPageTables = upperSharedPageTables + p.upperStart = upperStart + } + + p.InitArch(a) + return p +} + +// New returns new PageTables. +func New(a Allocator) *PageTables { + return NewWithUpper(a, nil, 0) +} + +// mapVisitor is used for map. +type mapVisitor struct { + target uintptr // Input. + physical uintptr // Input. + opts MapOpts // Input. + prev bool // Output. +} + +// visit is used for map. +// +//go:nosplit +func (v *mapVisitor) visit(start uintptr, pte *PTE, align uintptr) bool { + p := v.physical + (start - uintptr(v.target)) + if pte.Valid() && (pte.Address() != p || pte.Opts() != v.opts) { + v.prev = true + } + if p&align != 0 { + // We will install entries at a smaller granulaity if we don't + // install a valid entry here, however we must zap any existing + // entry to ensure this happens. + pte.Clear() + return true + } + pte.Set(p, v.opts) + return true +} + +//go:nosplit +func (*mapVisitor) requiresAlloc() bool { return true } + +//go:nosplit +func (*mapVisitor) requiresSplit() bool { return true } + +// Map installs a mapping with the given physical address. +// +// True is returned iff there was a previous mapping in the range. +// +// Precondition: addr & length must be page-aligned, their sum must not overflow. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) Map(addr usermem.Addr, length uintptr, opts MapOpts, physical uintptr) bool { + if p.readOnlyShared { + panic("Should not modify read-only shared pagetables.") + } + if uintptr(addr)+length < uintptr(addr) { + panic("addr & length overflow") + } + if p.upperSharedPageTables != nil { + // ignore change to the read-only upper shared portion. + if uintptr(addr) >= p.upperStart { + return false + } + if uintptr(addr)+length > p.upperStart { + length = p.upperStart - uintptr(addr) + } + } + w := mapWalker{ + pageTables: p, + visitor: mapVisitor{ + target: uintptr(addr), + physical: physical, + opts: opts, + }, + } + w.iterateRange(uintptr(addr), uintptr(addr)+length) + return w.visitor.prev +} + +// unmapVisitor is used for unmap. +type unmapVisitor struct { + count int +} + +//go:nosplit +func (*unmapVisitor) requiresAlloc() bool { return false } + +//go:nosplit +func (*unmapVisitor) requiresSplit() bool { return true } + +// visit unmaps the given entry. +// +//go:nosplit +func (v *unmapVisitor) visit(start uintptr, pte *PTE, align uintptr) bool { + pte.Clear() + v.count++ + return true +} + +// Unmap unmaps the given range. +// +// True is returned iff there was a previous mapping in the range. +// +// Precondition: addr & length must be page-aligned, their sum must not overflow. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) Unmap(addr usermem.Addr, length uintptr) bool { + if p.readOnlyShared { + panic("Should not modify read-only shared pagetables.") + } + if uintptr(addr)+length < uintptr(addr) { + panic("addr & length overflow") + } + if p.upperSharedPageTables != nil { + // ignore change to the read-only upper shared portion. + if uintptr(addr) >= p.upperStart { + return false + } + if uintptr(addr)+length > p.upperStart { + length = p.upperStart - uintptr(addr) + } + } + w := unmapWalker{ + pageTables: p, + visitor: unmapVisitor{ + count: 0, + }, + } + w.iterateRange(uintptr(addr), uintptr(addr)+length) + return w.visitor.count > 0 +} + +// emptyVisitor is used for emptiness checks. +type emptyVisitor struct { + count int +} + +//go:nosplit +func (*emptyVisitor) requiresAlloc() bool { return false } + +//go:nosplit +func (*emptyVisitor) requiresSplit() bool { return false } + +// visit unmaps the given entry. +// +//go:nosplit +func (v *emptyVisitor) visit(start uintptr, pte *PTE, align uintptr) bool { + v.count++ + return true +} + +// IsEmpty checks if the given range is empty. +// +// Precondition: addr & length must be page-aligned. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) IsEmpty(addr usermem.Addr, length uintptr) bool { + w := emptyWalker{ + pageTables: p, + } + w.iterateRange(uintptr(addr), uintptr(addr)+length) + return w.visitor.count == 0 +} + +// lookupVisitor is used for lookup. +type lookupVisitor struct { + target uintptr // Input & Output. + findFirst bool // Input. + physical uintptr // Output. + size uintptr // Output. + opts MapOpts // Output. +} + +// visit matches the given address. +// +//go:nosplit +func (v *lookupVisitor) visit(start uintptr, pte *PTE, align uintptr) bool { + if !pte.Valid() { + // If looking for the first, then we just keep iterating until + // we find a valid entry. + return v.findFirst + } + // Is this within the current range? + v.target = start + v.physical = pte.Address() + v.size = (align + 1) + v.opts = pte.Opts() + return false +} + +//go:nosplit +func (*lookupVisitor) requiresAlloc() bool { return false } + +//go:nosplit +func (*lookupVisitor) requiresSplit() bool { return false } + +// Lookup returns the physical address for the given virtual address. +// +// If findFirst is true, then the next valid address after addr is returned. +// If findFirst is false, then only a mapping for addr will be returned. +// +// Note that if size is zero, then no matching entry was found. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) Lookup(addr usermem.Addr, findFirst bool) (virtual usermem.Addr, physical, size uintptr, opts MapOpts) { + mask := uintptr(usermem.PageSize - 1) + addr &^= usermem.Addr(mask) + w := lookupWalker{ + pageTables: p, + visitor: lookupVisitor{ + target: uintptr(addr), + findFirst: findFirst, + }, + } + end := ^usermem.Addr(0) &^ usermem.Addr(mask) + if !findFirst { + end = addr + 1 + } + w.iterateRange(uintptr(addr), uintptr(end)) + return usermem.Addr(w.visitor.target), w.visitor.physical, w.visitor.size, w.visitor.opts +} + +// MarkReadOnlyShared marks the pagetables read-only and can be shared. +// +// It is usually used on the pagetables that are used as the upper +func (p *PageTables) MarkReadOnlyShared() { + p.readOnlyShared = true +} diff --git a/pkg/ring0/pagetables/pagetables_aarch64.go b/pkg/ring0/pagetables/pagetables_aarch64.go new file mode 100644 index 000000000..163a3aea3 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_aarch64.go @@ -0,0 +1,214 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package pagetables + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/usermem" +) + +// archPageTables is architecture-specific data. +type archPageTables struct { + // root is the pagetable root for kernel space. + root *PTEs + + // rootPhysical is the cached physical address of the root. + // + // This is saved only to prevent constant translation. + rootPhysical uintptr + + asid uint16 +} + +// TTBR0_EL1 returns the translation table base register 0. +// +//go:nosplit +func (p *PageTables) TTBR0_EL1(noFlush bool, asid uint16) uint64 { + return uint64(p.rootPhysical) | (uint64(asid)&ttbrASIDMask)<<ttbrASIDOffset +} + +// TTBR1_EL1 returns the translation table base register 1. +// +//go:nosplit +func (p *PageTables) TTBR1_EL1(noFlush bool, asid uint16) uint64 { + return uint64(p.archPageTables.rootPhysical) | (uint64(asid)&ttbrASIDMask)<<ttbrASIDOffset +} + +// Bits in page table entries. +const ( + typeTable = 0x3 << 0 + typeSect = 0x1 << 0 + typePage = 0x3 << 0 + pteValid = 0x1 << 0 + pteTableBit = 0x1 << 1 + pteTypeMask = 0x3 << 0 + present = pteValid | pteTableBit + user = 0x1 << 6 /* AP[1] */ + readOnly = 0x1 << 7 /* AP[2] */ + accessed = 0x1 << 10 + dbm = 0x1 << 51 + writable = dbm + cont = 0x1 << 52 + pxn = 0x1 << 53 + xn = 0x1 << 54 + dirty = 0x1 << 55 + nG = 0x1 << 11 + shared = 0x3 << 8 +) + +const ( + mtDevicenGnRE = 0x1 << 2 + mtNormal = 0x4 << 2 +) + +const ( + executeDisable = xn + optionMask = 0xfff | 0xffff<<48 + protDefault = accessed | shared +) + +// MapOpts are x86 options. +type MapOpts struct { + // AccessType defines permissions. + AccessType usermem.AccessType + + // Global indicates the page is globally accessible. + Global bool + + // User indicates the page is a user page. + User bool +} + +// PTE is a page table entry. +type PTE uintptr + +// Clear clears this PTE, including sect page information. +// +//go:nosplit +func (p *PTE) Clear() { + atomic.StoreUintptr((*uintptr)(p), 0) +} + +// Valid returns true iff this entry is valid. +// +//go:nosplit +func (p *PTE) Valid() bool { + return atomic.LoadUintptr((*uintptr)(p))&present != 0 +} + +// Opts returns the PTE options. +// +// These are all options except Valid and Sect. +// +//go:nosplit +func (p *PTE) Opts() MapOpts { + v := atomic.LoadUintptr((*uintptr)(p)) + + return MapOpts{ + AccessType: usermem.AccessType{ + Read: true, + Write: v&readOnly == 0, + Execute: v&xn == 0, + }, + Global: v&nG == 0, + User: v&user != 0, + } +} + +// SetSect sets this page as a sect page. +// +// The page must not be valid or a panic will result. +// +//go:nosplit +func (p *PTE) SetSect() { + if p.Valid() { + // This is not allowed. + panic("SetSect called on valid page!") + } + atomic.StoreUintptr((*uintptr)(p), typeSect) +} + +// IsSect returns true iff this page is a sect page. +// +//go:nosplit +func (p *PTE) IsSect() bool { + return atomic.LoadUintptr((*uintptr)(p))&pteTypeMask == typeSect +} + +// Set sets this PTE value. +// +// This does not change the sect page property. +// +//go:nosplit +func (p *PTE) Set(addr uintptr, opts MapOpts) { + v := (addr &^ optionMask) | nG | readOnly | protDefault + if p.IsSect() { + // Note that this is inherited from the previous instance. Set + // does not change the value of Sect. See above. + v |= typeSect + } else { + v |= typePage + } + if !opts.AccessType.Any() { + // Leave as non-valid if no access is available. + v &^= pteValid + } + + if opts.Global { + v = v &^ nG + } + + if opts.AccessType.Execute { + v = v &^ executeDisable + } else { + v |= executeDisable + } + if opts.AccessType.Write { + v = v &^ readOnly + } + + if opts.User { + v |= user + v |= mtNormal + } else { + v = v &^ user + v |= mtNormal + } + atomic.StoreUintptr((*uintptr)(p), v) +} + +// setPageTable sets this PTE value and forces the write bit and sect bit to +// be cleared. This is used explicitly for breaking sect pages. +// +//go:nosplit +func (p *PTE) setPageTable(pt *PageTables, ptes *PTEs) { + addr := pt.Allocator.PhysicalFor(ptes) + if addr&^optionMask != addr { + // This should never happen. + panic("unaligned physical address!") + } + v := addr | typeTable | protDefault | mtNormal + atomic.StoreUintptr((*uintptr)(p), v) +} + +// Address extracts the address. This should only be used if Valid returns true. +// +//go:nosplit +func (p *PTE) Address() uintptr { + return atomic.LoadUintptr((*uintptr)(p)) &^ optionMask +} diff --git a/pkg/ring0/pagetables/pagetables_aarch64_state_autogen.go b/pkg/ring0/pagetables/pagetables_aarch64_state_autogen.go new file mode 100644 index 000000000..a24523f87 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_aarch64_state_autogen.go @@ -0,0 +1,6 @@ +// automatically generated by stateify. + +// +build arm64 +// +build arm64 + +package pagetables diff --git a/pkg/ring0/pagetables/pagetables_amd64.go b/pkg/ring0/pagetables/pagetables_amd64.go new file mode 100644 index 000000000..a217f404c --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_amd64.go @@ -0,0 +1,77 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +// Address constraints. +// +// The lowerTop and upperBottom currently apply to four-level pagetables; +// additional refactoring would be necessary to support five-level pagetables. +const ( + lowerTop = 0x00007fffffffffff + upperBottom = 0xffff800000000000 + + pteShift = 12 + pmdShift = 21 + pudShift = 30 + pgdShift = 39 + + pteMask = 0x1ff << pteShift + pmdMask = 0x1ff << pmdShift + pudMask = 0x1ff << pudShift + pgdMask = 0x1ff << pgdShift + + pteSize = 1 << pteShift + pmdSize = 1 << pmdShift + pudSize = 1 << pudShift + pgdSize = 1 << pgdShift + + executeDisable = 1 << 63 + entriesPerPage = 512 +) + +// InitArch does some additional initialization related to the architecture. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) InitArch(allocator Allocator) { + if p.upperSharedPageTables != nil { + p.cloneUpperShared() + } +} + +//go:nosplit +func pgdIndex(upperStart uintptr) uintptr { + if upperStart&(pgdSize-1) != 0 { + panic("upperStart should be pgd size aligned") + } + if upperStart >= upperBottom { + return entriesPerPage/2 + (upperStart-upperBottom)/pgdSize + } + if upperStart < lowerTop { + return upperStart / pgdSize + } + panic("upperStart should be in canonical range") +} + +// cloneUpperShared clone the upper from the upper shared page tables. +// +//go:nosplit +func (p *PageTables) cloneUpperShared() { + start := pgdIndex(p.upperStart) + copy(p.root[start:entriesPerPage], p.upperSharedPageTables.root[start:entriesPerPage]) +} + +// PTEs is a collection of entries. +type PTEs [entriesPerPage]PTE diff --git a/pkg/ring0/pagetables/pagetables_amd64_state_autogen.go b/pkg/ring0/pagetables/pagetables_amd64_state_autogen.go new file mode 100644 index 000000000..f48a8acd1 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_amd64_state_autogen.go @@ -0,0 +1,5 @@ +// automatically generated by stateify. + +// +build amd64 + +package pagetables diff --git a/pkg/ring0/pagetables/pagetables_arm64.go b/pkg/ring0/pagetables/pagetables_arm64.go new file mode 100644 index 000000000..fef7a0fd1 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_arm64.go @@ -0,0 +1,71 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +// Address constraints. +// +// The lowerTop and upperBottom currently apply to four-level pagetables; +// additional refactoring would be necessary to support five-level pagetables. +const ( + lowerTop = 0x0000ffffffffffff + upperBottom = 0xffff000000000000 + pteShift = 12 + pmdShift = 21 + pudShift = 30 + pgdShift = 39 + + pteMask = 0x1ff << pteShift + pmdMask = 0x1ff << pmdShift + pudMask = 0x1ff << pudShift + pgdMask = 0x1ff << pgdShift + + pteSize = 1 << pteShift + pmdSize = 1 << pmdShift + pudSize = 1 << pudShift + pgdSize = 1 << pgdShift + + ttbrASIDOffset = 48 + ttbrASIDMask = 0xff + + entriesPerPage = 512 +) + +// InitArch does some additional initialization related to the architecture. +// +// +checkescape:hard,stack +//go:nosplit +func (p *PageTables) InitArch(allocator Allocator) { + if p.upperSharedPageTables != nil { + p.cloneUpperShared() + } else { + p.archPageTables.root = p.Allocator.NewPTEs() + p.archPageTables.rootPhysical = p.Allocator.PhysicalFor(p.archPageTables.root) + } +} + +// cloneUpperShared clone the upper from the upper shared page tables. +// +//go:nosplit +func (p *PageTables) cloneUpperShared() { + if p.upperStart != upperBottom { + panic("upperStart should be the same as upperBottom") + } + + p.archPageTables.root = p.upperSharedPageTables.archPageTables.root + p.archPageTables.rootPhysical = p.upperSharedPageTables.archPageTables.rootPhysical +} + +// PTEs is a collection of entries. +type PTEs [entriesPerPage]PTE diff --git a/pkg/ring0/pagetables/pagetables_arm64_state_autogen.go b/pkg/ring0/pagetables/pagetables_arm64_state_autogen.go new file mode 100644 index 000000000..ae9d2b272 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_arm64_state_autogen.go @@ -0,0 +1,5 @@ +// automatically generated by stateify. + +// +build arm64 + +package pagetables diff --git a/pkg/ring0/pagetables/pagetables_state_autogen.go b/pkg/ring0/pagetables/pagetables_state_autogen.go new file mode 100644 index 000000000..4c4540603 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_state_autogen.go @@ -0,0 +1,3 @@ +// automatically generated by stateify. + +package pagetables diff --git a/pkg/ring0/pagetables/pagetables_unsafe_state_autogen.go b/pkg/ring0/pagetables/pagetables_unsafe_state_autogen.go new file mode 100644 index 000000000..4c4540603 --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_unsafe_state_autogen.go @@ -0,0 +1,3 @@ +// automatically generated by stateify. + +package pagetables diff --git a/pkg/ring0/pagetables/pagetables_x86.go b/pkg/ring0/pagetables/pagetables_x86.go new file mode 100644 index 000000000..32edd2f0a --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_x86.go @@ -0,0 +1,183 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build 386 amd64 + +package pagetables + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/usermem" +) + +// archPageTables is architecture-specific data. +type archPageTables struct { + // pcid is the value assigned by PCIDs.Assign. + // + // Note that zero is a valid PCID. + pcid uint16 +} + +// CR3 returns the CR3 value for these tables. +// +// This may be called in interrupt contexts. A PCID of zero always implies a +// flush and should be passed when PCIDs are not enabled. See pcids_x86.go for +// more information. +// +//go:nosplit +func (p *PageTables) CR3(noFlush bool, pcid uint16) uint64 { + // Bit 63 is set to avoid flushing the PCID (per SDM 4.10.4.1). + const noFlushBit uint64 = 0x8000000000000000 + if noFlush && pcid != 0 { + return noFlushBit | uint64(p.rootPhysical) | uint64(pcid) + } + return uint64(p.rootPhysical) | uint64(pcid) +} + +// Bits in page table entries. +const ( + present = 0x001 + writable = 0x002 + user = 0x004 + writeThrough = 0x008 + cacheDisable = 0x010 + accessed = 0x020 + dirty = 0x040 + super = 0x080 + global = 0x100 + optionMask = executeDisable | 0xfff +) + +// MapOpts are x86 options. +type MapOpts struct { + // AccessType defines permissions. + AccessType usermem.AccessType + + // Global indicates the page is globally accessible. + Global bool + + // User indicates the page is a user page. + User bool +} + +// PTE is a page table entry. +type PTE uintptr + +// Clear clears this PTE, including super page information. +// +//go:nosplit +func (p *PTE) Clear() { + atomic.StoreUintptr((*uintptr)(p), 0) +} + +// Valid returns true iff this entry is valid. +// +//go:nosplit +func (p *PTE) Valid() bool { + return atomic.LoadUintptr((*uintptr)(p))&present != 0 +} + +// Opts returns the PTE options. +// +// These are all options except Valid and Super. +// +//go:nosplit +func (p *PTE) Opts() MapOpts { + v := atomic.LoadUintptr((*uintptr)(p)) + return MapOpts{ + AccessType: usermem.AccessType{ + Read: v&present != 0, + Write: v&writable != 0, + Execute: v&executeDisable == 0, + }, + Global: v&global != 0, + User: v&user != 0, + } +} + +// SetSuper sets this page as a super page. +// +// The page must not be valid or a panic will result. +// +//go:nosplit +func (p *PTE) SetSuper() { + if p.Valid() { + // This is not allowed. + panic("SetSuper called on valid page!") + } + atomic.StoreUintptr((*uintptr)(p), super) +} + +// IsSuper returns true iff this page is a super page. +// +//go:nosplit +func (p *PTE) IsSuper() bool { + return atomic.LoadUintptr((*uintptr)(p))&super != 0 +} + +// Set sets this PTE value. +// +// This does not change the super page property. +// +//go:nosplit +func (p *PTE) Set(addr uintptr, opts MapOpts) { + if !opts.AccessType.Any() { + p.Clear() + return + } + v := (addr &^ optionMask) + if opts.AccessType.Any() { + v |= present | accessed + } + if opts.User { + v |= user + } + if opts.Global { + v |= global + } + if !opts.AccessType.Execute { + v |= executeDisable + } + if opts.AccessType.Write { + v |= writable | dirty + } + if p.IsSuper() { + // Note that this is inherited from the previous instance. Set + // does not change the value of Super. See above. + v |= super + } + atomic.StoreUintptr((*uintptr)(p), v) +} + +// setPageTable sets this PTE value and forces the write bit and super bit to +// be cleared. This is used explicitly for breaking super pages. +// +//go:nosplit +func (p *PTE) setPageTable(pt *PageTables, ptes *PTEs) { + addr := pt.Allocator.PhysicalFor(ptes) + if addr&^optionMask != addr { + // This should never happen. + panic("unaligned physical address!") + } + v := addr | present | user | writable | accessed | dirty + atomic.StoreUintptr((*uintptr)(p), v) +} + +// Address extracts the address. This should only be used if Valid returns true. +// +//go:nosplit +func (p *PTE) Address() uintptr { + return atomic.LoadUintptr((*uintptr)(p)) &^ optionMask +} diff --git a/pkg/ring0/pagetables/pagetables_x86_state_autogen.go b/pkg/ring0/pagetables/pagetables_x86_state_autogen.go new file mode 100644 index 000000000..6fe78c51c --- /dev/null +++ b/pkg/ring0/pagetables/pagetables_x86_state_autogen.go @@ -0,0 +1,6 @@ +// automatically generated by stateify. + +// +build 386 amd64 +// +build i386 amd64 + +package pagetables diff --git a/pkg/ring0/pagetables/pcids.go b/pkg/ring0/pagetables/pcids.go new file mode 100644 index 000000000..964496aac --- /dev/null +++ b/pkg/ring0/pagetables/pcids.go @@ -0,0 +1,104 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +import ( + "gvisor.dev/gvisor/pkg/sync" +) + +// PCIDs is a simple PCID database. +// +// This is not protected by locks and is thus suitable for use only with a +// single CPU at a time. +type PCIDs struct { + // mu protects below. + mu sync.Mutex + + // cache are the assigned page tables. + cache map[*PageTables]uint16 + + // avail are available PCIDs. + avail []uint16 +} + +// NewPCIDs returns a new PCID database. +// +// start is the first index to assign. Typically this will be one, as the zero +// pcid will always be flushed on transition (see pagetables_x86.go). This may +// be more than one if specific PCIDs are reserved. +// +// Nil is returned iff the start and size are out of range. +func NewPCIDs(start, size uint16) *PCIDs { + if start+uint16(size) > limitPCID { + return nil // See comment. + } + p := &PCIDs{ + cache: make(map[*PageTables]uint16), + } + for pcid := start; pcid < start+size; pcid++ { + p.avail = append(p.avail, pcid) + } + return p +} + +// Assign assigns a PCID to the given PageTables. +// +// This may overwrite any previous assignment provided. If this in the case, +// true is returned to indicate that the PCID should be flushed. +func (p *PCIDs) Assign(pt *PageTables) (uint16, bool) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + p.mu.Unlock() + return pcid, false // No flush. + } + + // Is there something available? + if len(p.avail) > 0 { + pcid := p.avail[len(p.avail)-1] + p.avail = p.avail[:len(p.avail)-1] + p.cache[pt] = pcid + + // We need to flush because while this is in the available + // pool, it may have been used previously. + p.mu.Unlock() + return pcid, true + } + + // Evict an existing table. + for old, pcid := range p.cache { + delete(p.cache, old) + p.cache[pt] = pcid + + // A flush is definitely required in this case, these page + // tables may still be active. (They will just be assigned some + // other PCID if and when they hit the given CPU again.) + p.mu.Unlock() + return pcid, true + } + + // No PCID. + p.mu.Unlock() + return 0, false +} + +// Drop drops references to a set of page tables. +func (p *PCIDs) Drop(pt *PageTables) { + p.mu.Lock() + if pcid, ok := p.cache[pt]; ok { + delete(p.cache, pt) + p.avail = append(p.avail, pcid) + } + p.mu.Unlock() +} diff --git a/pkg/ring0/pagetables/pcids_aarch64.go b/pkg/ring0/pagetables/pcids_aarch64.go new file mode 100644 index 000000000..fbfd41d83 --- /dev/null +++ b/pkg/ring0/pagetables/pcids_aarch64.go @@ -0,0 +1,32 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package pagetables + +// limitPCID is the maximum value of PCIDs. +// +// In VMSAv8-64, the PCID(ASID) size is an IMPLEMENTATION DEFINED choice +// of 8 bits or 16 bits, and ID_AA64MMFR0_EL1.ASIDBits identifies the +// supported size. When an implementation supports a 16-bit ASID, TCR_ELx.AS +// selects whether the top 8 bits of the ASID are used. +var limitPCID uint16 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +func GetASIDBits() uint8 + +func init() { + limitPCID = uint16(1)<<GetASIDBits() - 1 +} diff --git a/pkg/ring0/pagetables/pcids_aarch64.s b/pkg/ring0/pagetables/pcids_aarch64.s new file mode 100644 index 000000000..e9d62d768 --- /dev/null +++ b/pkg/ring0/pagetables/pcids_aarch64.s @@ -0,0 +1,45 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +#include "funcdata.h" +#include "textflag.h" + +#define ID_AA64MMFR0_ASIDBITS_SHIFT 4 +#define ID_AA64MMFR0_ASIDBITS_16 2 +#define TCR_EL1_AS_BIT 36 + +// GetASIDBits return the system ASID bits, 8 or 16 bits. +// +// func GetASIDBits() uint8 +TEXT ·GetASIDBits(SB),NOSPLIT,$0-1 + // First, check whether 16bits ASID is supported. + // ID_AA64MMFR0_EL1.ASIDBITS[7:4] == 0010. + WORD $0xd5380700 // MRS ID_AA64MMFR0_EL1, R0 + UBFX $ID_AA64MMFR0_ASIDBITS_SHIFT, R0, $4, R0 + CMPW $ID_AA64MMFR0_ASIDBITS_16, R0 + BNE bits_8 + + // Second, check whether 16bits ASID is enabled. + // TCR_EL1.AS[36] == 1. + WORD $0xd5382040 // MRS TCR_EL1, R0 + TBZ $TCR_EL1_AS_BIT, R0, bits_8 + MOVD $16, R0 + B done +bits_8: + MOVD $8, R0 +done: + MOVB R0, ret+0(FP) + RET diff --git a/pkg/ring0/pagetables/pcids_x86.go b/pkg/ring0/pagetables/pcids_x86.go new file mode 100644 index 000000000..91fc5e8dd --- /dev/null +++ b/pkg/ring0/pagetables/pcids_x86.go @@ -0,0 +1,20 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build i386 amd64 + +package pagetables + +// limitPCID is the maximum value of valid PCIDs. +const limitPCID = 4095 diff --git a/pkg/ring0/pagetables/walker_amd64.go b/pkg/ring0/pagetables/walker_amd64.go new file mode 100644 index 000000000..eb4fbcc31 --- /dev/null +++ b/pkg/ring0/pagetables/walker_amd64.go @@ -0,0 +1,221 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build amd64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *Walker) iterateRangeCanonical(start, end uintptr) bool { + for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &w.pageTables.root[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + start = next(start, pgdSize) + continue + } + + // Allocate a new pgd. + pudEntries = w.pageTables.Allocator.NewPTEs() // escapes: depends on allocator. + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) // escapes: see above. + } + + // Map the next level. + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + clearPUDEntries++ + start = next(start, pudSize) + continue + } + + // This level has 1-GB super pages. Is this + // entire region at least as large as a single + // PUD entry? If so, we can skip allocating a + // new page for the pmd. + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = next(start, pudSize) + continue + } + } + + // Allocate a new pud. + pmdEntries = w.pageTables.Allocator.NewPTEs() // escapes: see above. + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSuper() { + // Does this page need to be split? + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < next(start, pudSize)) { + // Install the relevant entries. + pmdEntries = w.pageTables.Allocator.NewPTEs() // escapes: see above. + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSuper() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + // A super page to be checked directly. + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + + // Might have been cleared. + if !pudEntry.Valid() { + clearPUDEntries++ + } + + // Note that the super page was changed. + start = next(start, pudSize) + continue + } + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) // escapes: see above. + } + + // Map the next level, since this is valid. + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + clearPMDEntries++ + start = next(start, pmdSize) + continue + } + + // This level has 2-MB huge pages. If this + // region is contined in a single PMD entry? + // As above, we can skip allocating a new page. + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = next(start, pmdSize) + continue + } + } + + // Allocate a new pmd. + pteEntries = w.pageTables.Allocator.NewPTEs() // escapes: see above. + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSuper() { + // Does this page need to be split? + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < next(start, pmdSize)) { + // Install the relevant entries. + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + // A huge page to be checked directly. + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + + // Might have been cleared. + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + // Note that the huge page was changed. + start = next(start, pmdSize) + continue + } + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) // escapes: see above. + } + + // Map the next level, since this is valid. + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + // At this point, we are guaranteed that start%pteSize == 0. + if !w.visitor.visit(uintptr(start&^(pteSize-1)), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + } + + // Note that the pte was changed. + start += pteSize + continue + } + + // Check if we no longer need this page. + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) // escapes: see above. + clearPMDEntries++ + } + } + + // Check if we no longer need this page. + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) // escapes: see above. + clearPUDEntries++ + } + } + + // Check if we no longer need this page. + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) // escapes: see above. + } + } + return true +} diff --git a/pkg/ring0/pagetables/walker_arm64.go b/pkg/ring0/pagetables/walker_arm64.go new file mode 100644 index 000000000..5ed881c7a --- /dev/null +++ b/pkg/ring0/pagetables/walker_arm64.go @@ -0,0 +1,231 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// +build arm64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *Walker) iterateRangeCanonical(start, end uintptr) bool { + pgdEntryIndex := w.pageTables.root + if start >= upperBottom { + pgdEntryIndex = w.pageTables.archPageTables.root + } + + for pgdIndex := (uint16((start & pgdMask) >> pgdShift)); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &pgdEntryIndex[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + start = next(start, pgdSize) + continue + } + + // Allocate a new pgd. + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + // Map the next level. + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + clearPUDEntries++ + start = next(start, pudSize) + continue + } + + // This level has 1-GB sect pages. Is this + // entire region at least as large as a single + // PUD entry? If so, we can skip allocating a + // new page for the pmd. + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSect() + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = next(start, pudSize) + continue + } + } + + // Allocate a new pud. + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSect() { + // Does this page need to be split? + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < next(start, pudSize)) { + // Install the relevant entries. + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSect() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + // A sect page to be checked directly. + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + + // Might have been cleared. + if !pudEntry.Valid() { + clearPUDEntries++ + } + + // Note that the sect page was changed. + start = next(start, pudSize) + continue + } + + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + // Map the next level, since this is valid. + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + // Skip over this entry. + clearPMDEntries++ + start = next(start, pmdSize) + continue + } + + // This level has 2-MB huge pages. If this + // region is contined in a single PMD entry? + // As above, we can skip allocating a new page. + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSect() + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = next(start, pmdSize) + continue + } + } + + // Allocate a new pmd. + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSect() { + // Does this page need to be split? + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < next(start, pmdSize)) { + // Install the relevant entries. + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + // A huge page to be checked directly. + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + + // Might have been cleared. + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + // Note that the huge page was changed. + start = next(start, pmdSize) + continue + } + + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + // Map the next level, since this is valid. + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + // At this point, we are guaranteed that start%pteSize == 0. + if !w.visitor.visit(uintptr(start), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() { + if w.visitor.requiresAlloc() { + panic("PTE not set after iteration with requiresAlloc!") + } + clearPTEEntries++ + } + + // Note that the pte was changed. + start += pteSize + continue + } + + // Check if we no longer need this page. + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + // Check if we no longer need this page. + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + // Check if we no longer need this page. + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} diff --git a/pkg/ring0/pagetables/walker_empty_amd64.go b/pkg/ring0/pagetables/walker_empty_amd64.go new file mode 100644 index 000000000..a3cd7a1a2 --- /dev/null +++ b/pkg/ring0/pagetables/walker_empty_amd64.go @@ -0,0 +1,265 @@ +// +build amd64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *emptyWalker) iterateRangeCanonical(start, end uintptr) bool { + for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &w.pageTables.root[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = emptynext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = emptynext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = emptynext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < emptynext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSuper() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = emptynext(start, pudSize) + continue + } + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = emptynext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = emptynext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < emptynext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = emptynext(start, pmdSize) + continue + } + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start&^(pteSize-1)), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type emptyWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor emptyVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *emptyWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func emptynext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_empty_arm64.go b/pkg/ring0/pagetables/walker_empty_arm64.go new file mode 100644 index 000000000..d61b44b65 --- /dev/null +++ b/pkg/ring0/pagetables/walker_empty_arm64.go @@ -0,0 +1,275 @@ +// +build arm64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *emptyWalker) iterateRangeCanonical(start, end uintptr) bool { + pgdEntryIndex := w.pageTables.root + if start >= upperBottom { + pgdEntryIndex = w.pageTables.archPageTables.root + } + + for pgdIndex := (uint16((start & pgdMask) >> pgdShift)); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &pgdEntryIndex[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = emptynext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = emptynext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSect() + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = emptynext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < emptynext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSect() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = emptynext(start, pudSize) + continue + } + + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = emptynext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSect() + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = emptynext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < emptynext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = emptynext(start, pmdSize) + continue + } + + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() { + if w.visitor.requiresAlloc() { + panic("PTE not set after iteration with requiresAlloc!") + } + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type emptyWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor emptyVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *emptyWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func emptynext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_generic.go b/pkg/ring0/pagetables/walker_generic.go new file mode 100644 index 000000000..34fba7b84 --- /dev/null +++ b/pkg/ring0/pagetables/walker_generic.go @@ -0,0 +1,110 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package pagetables + +// Visitor is a generic type. +type Visitor interface { + // visit is called on each PTE. The returned boolean indicates whether + // the walk should continue. + visit(start uintptr, pte *PTE, align uintptr) bool + + // requiresAlloc indicates that new entries should be allocated within + // the walked range. + requiresAlloc() bool + + // requiresSplit indicates that entries in the given range should be + // split if they are huge or jumbo pages. + requiresSplit() bool +} + +// Walker walks page tables. +type Walker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor Visitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *Walker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func next(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_lookup_amd64.go b/pkg/ring0/pagetables/walker_lookup_amd64.go new file mode 100644 index 000000000..c92c1cb44 --- /dev/null +++ b/pkg/ring0/pagetables/walker_lookup_amd64.go @@ -0,0 +1,265 @@ +// +build amd64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *lookupWalker) iterateRangeCanonical(start, end uintptr) bool { + for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &w.pageTables.root[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = lookupnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = lookupnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = lookupnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < lookupnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSuper() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = lookupnext(start, pudSize) + continue + } + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = lookupnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = lookupnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < lookupnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = lookupnext(start, pmdSize) + continue + } + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start&^(pteSize-1)), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type lookupWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor lookupVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *lookupWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func lookupnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_lookup_arm64.go b/pkg/ring0/pagetables/walker_lookup_arm64.go new file mode 100644 index 000000000..74062a00a --- /dev/null +++ b/pkg/ring0/pagetables/walker_lookup_arm64.go @@ -0,0 +1,275 @@ +// +build arm64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *lookupWalker) iterateRangeCanonical(start, end uintptr) bool { + pgdEntryIndex := w.pageTables.root + if start >= upperBottom { + pgdEntryIndex = w.pageTables.archPageTables.root + } + + for pgdIndex := (uint16((start & pgdMask) >> pgdShift)); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &pgdEntryIndex[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = lookupnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = lookupnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSect() + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = lookupnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < lookupnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSect() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = lookupnext(start, pudSize) + continue + } + + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = lookupnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSect() + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = lookupnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < lookupnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = lookupnext(start, pmdSize) + continue + } + + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() { + if w.visitor.requiresAlloc() { + panic("PTE not set after iteration with requiresAlloc!") + } + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type lookupWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor lookupVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *lookupWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func lookupnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_map_amd64.go b/pkg/ring0/pagetables/walker_map_amd64.go new file mode 100644 index 000000000..1c6c1a032 --- /dev/null +++ b/pkg/ring0/pagetables/walker_map_amd64.go @@ -0,0 +1,265 @@ +// +build amd64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *mapWalker) iterateRangeCanonical(start, end uintptr) bool { + for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &w.pageTables.root[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = mapnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = mapnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = mapnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < mapnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSuper() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = mapnext(start, pudSize) + continue + } + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = mapnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = mapnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < mapnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = mapnext(start, pmdSize) + continue + } + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start&^(pteSize-1)), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type mapWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor mapVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *mapWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func mapnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_map_arm64.go b/pkg/ring0/pagetables/walker_map_arm64.go new file mode 100644 index 000000000..8223de306 --- /dev/null +++ b/pkg/ring0/pagetables/walker_map_arm64.go @@ -0,0 +1,275 @@ +// +build arm64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *mapWalker) iterateRangeCanonical(start, end uintptr) bool { + pgdEntryIndex := w.pageTables.root + if start >= upperBottom { + pgdEntryIndex = w.pageTables.archPageTables.root + } + + for pgdIndex := (uint16((start & pgdMask) >> pgdShift)); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &pgdEntryIndex[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = mapnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = mapnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSect() + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = mapnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < mapnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSect() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = mapnext(start, pudSize) + continue + } + + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = mapnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSect() + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = mapnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < mapnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = mapnext(start, pmdSize) + continue + } + + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() { + if w.visitor.requiresAlloc() { + panic("PTE not set after iteration with requiresAlloc!") + } + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type mapWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor mapVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *mapWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func mapnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_unmap_amd64.go b/pkg/ring0/pagetables/walker_unmap_amd64.go new file mode 100644 index 000000000..82b27ab64 --- /dev/null +++ b/pkg/ring0/pagetables/walker_unmap_amd64.go @@ -0,0 +1,265 @@ +// +build amd64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *unmapWalker) iterateRangeCanonical(start, end uintptr) bool { + for pgdIndex := uint16((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &w.pageTables.root[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = unmapnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = unmapnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = unmapnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < unmapnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSuper() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pudSize-1)), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = unmapnext(start, pudSize) + continue + } + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = unmapnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSuper() + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = unmapnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSuper() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < unmapnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start&^(pmdSize-1)), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = unmapnext(start, pmdSize) + continue + } + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start&^(pteSize-1)), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type unmapWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor unmapVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *unmapWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func unmapnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/pagetables/walker_unmap_arm64.go b/pkg/ring0/pagetables/walker_unmap_arm64.go new file mode 100644 index 000000000..1ecccbf27 --- /dev/null +++ b/pkg/ring0/pagetables/walker_unmap_arm64.go @@ -0,0 +1,275 @@ +// +build arm64 + +package pagetables + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *unmapWalker) iterateRangeCanonical(start, end uintptr) bool { + pgdEntryIndex := w.pageTables.root + if start >= upperBottom { + pgdEntryIndex = w.pageTables.archPageTables.root + } + + for pgdIndex := (uint16((start & pgdMask) >> pgdShift)); start < end && pgdIndex < entriesPerPage; pgdIndex++ { + var ( + pgdEntry = &pgdEntryIndex[pgdIndex] + pudEntries *PTEs + ) + if !pgdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + start = unmapnext(start, pgdSize) + continue + } + + pudEntries = w.pageTables.Allocator.NewPTEs() + pgdEntry.setPageTable(w.pageTables, pudEntries) + } else { + pudEntries = w.pageTables.Allocator.LookupPTEs(pgdEntry.Address()) + } + + clearPUDEntries := uint16(0) + + for pudIndex := uint16((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ { + var ( + pudEntry = &pudEntries[pudIndex] + pmdEntries *PTEs + ) + if !pudEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPUDEntries++ + start = unmapnext(start, pudSize) + continue + } + + if start&(pudSize-1) == 0 && end-start >= pudSize { + pudEntry.SetSect() + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + if pudEntry.Valid() { + start = unmapnext(start, pudSize) + continue + } + } + + pmdEntries = w.pageTables.Allocator.NewPTEs() + pudEntry.setPageTable(w.pageTables, pmdEntries) + + } else if pudEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pudSize-1) != 0 || end < unmapnext(start, pudSize)) { + + pmdEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pmdEntries[index].SetSect() + pmdEntries[index].Set( + pudEntry.Address()+(pmdSize*uintptr(index)), + pudEntry.Opts()) + } + pudEntry.setPageTable(w.pageTables, pmdEntries) + } else { + + if !w.visitor.visit(uintptr(start), pudEntry, pudSize-1) { + return false + } + + if !pudEntry.Valid() { + clearPUDEntries++ + } + + start = unmapnext(start, pudSize) + continue + } + + } else { + pmdEntries = w.pageTables.Allocator.LookupPTEs(pudEntry.Address()) + } + + clearPMDEntries := uint16(0) + + for pmdIndex := uint16((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ { + var ( + pmdEntry = &pmdEntries[pmdIndex] + pteEntries *PTEs + ) + if !pmdEntry.Valid() { + if !w.visitor.requiresAlloc() { + + clearPMDEntries++ + start = unmapnext(start, pmdSize) + continue + } + + if start&(pmdSize-1) == 0 && end-start >= pmdSize { + pmdEntry.SetSect() + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + if pmdEntry.Valid() { + start = unmapnext(start, pmdSize) + continue + } + } + + pteEntries = w.pageTables.Allocator.NewPTEs() + pmdEntry.setPageTable(w.pageTables, pteEntries) + + } else if pmdEntry.IsSect() { + + if w.visitor.requiresSplit() && (start&(pmdSize-1) != 0 || end < unmapnext(start, pmdSize)) { + + pteEntries = w.pageTables.Allocator.NewPTEs() + for index := uint16(0); index < entriesPerPage; index++ { + pteEntries[index].Set( + pmdEntry.Address()+(pteSize*uintptr(index)), + pmdEntry.Opts()) + } + pmdEntry.setPageTable(w.pageTables, pteEntries) + } else { + + if !w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) { + return false + } + + if !pmdEntry.Valid() { + clearPMDEntries++ + } + + start = unmapnext(start, pmdSize) + continue + } + + } else { + pteEntries = w.pageTables.Allocator.LookupPTEs(pmdEntry.Address()) + } + + clearPTEEntries := uint16(0) + + for pteIndex := uint16((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ { + var ( + pteEntry = &pteEntries[pteIndex] + ) + if !pteEntry.Valid() && !w.visitor.requiresAlloc() { + clearPTEEntries++ + start += pteSize + continue + } + + if !w.visitor.visit(uintptr(start), pteEntry, pteSize-1) { + return false + } + if !pteEntry.Valid() { + if w.visitor.requiresAlloc() { + panic("PTE not set after iteration with requiresAlloc!") + } + clearPTEEntries++ + } + + start += pteSize + continue + } + + if clearPTEEntries == entriesPerPage { + pmdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pteEntries) + clearPMDEntries++ + } + } + + if clearPMDEntries == entriesPerPage { + pudEntry.Clear() + w.pageTables.Allocator.FreePTEs(pmdEntries) + clearPUDEntries++ + } + } + + if clearPUDEntries == entriesPerPage { + pgdEntry.Clear() + w.pageTables.Allocator.FreePTEs(pudEntries) + } + } + return true +} + +// Walker walks page tables. +type unmapWalker struct { + // pageTables are the tables to walk. + pageTables *PageTables + + // Visitor is the set of arguments. + visitor unmapVisitor +} + +// iterateRange iterates over all appropriate levels of page tables for the given range. +// +// If requiresAlloc is true, then Set _must_ be called on all given PTEs. The +// exception is super pages. If a valid super page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of super/sect pages whenever +// possible. Whether a super page is provided will be clear through the range +// provided in the callback. +// +// Note that if requiresAlloc is true, then no gaps will be present. However, +// if alloc is not set, then the iteration will likely be full of gaps. +// +// Note that this function should generally be avoided in favor of Map, Unmap, +// etc. when not necessary. +// +// Precondition: start must be page-aligned. +// Precondition: start must be less than end. +// Precondition: If requiresAlloc is true, then start and end should not span +// non-canonical ranges. If they do, a panic will result. +// +//go:nosplit +func (w *unmapWalker) iterateRange(start, end uintptr) { + if start%pteSize != 0 { + panic("unaligned start") + } + if end < start { + panic("start > end") + } + if start < lowerTop { + if end <= lowerTop { + w.iterateRangeCanonical(start, end) + } else if end > lowerTop && end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(start, lowerTop) + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + if !w.iterateRangeCanonical(start, lowerTop) { + return + } + w.iterateRangeCanonical(upperBottom, end) + } + } else if start < upperBottom { + if end <= upperBottom { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + } else { + if w.visitor.requiresAlloc() { + panic("alloc spans non-canonical range") + } + w.iterateRangeCanonical(upperBottom, end) + } + } else { + w.iterateRangeCanonical(start, end) + } +} + +// next returns the next address quantized by the given size. +// +//go:nosplit +func unmapnext(start uintptr, size uintptr) uintptr { + start &= ^(size - 1) + start += size + return start +} diff --git a/pkg/ring0/ring0.go b/pkg/ring0/ring0.go new file mode 100644 index 000000000..cdeb1b43a --- /dev/null +++ b/pkg/ring0/ring0.go @@ -0,0 +1,16 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package ring0 provides basic operating system-level stubs. +package ring0 diff --git a/pkg/ring0/ring0_amd64_state_autogen.go b/pkg/ring0/ring0_amd64_state_autogen.go new file mode 100644 index 000000000..96cf5d331 --- /dev/null +++ b/pkg/ring0/ring0_amd64_state_autogen.go @@ -0,0 +1,7 @@ +// automatically generated by stateify. + +// +build amd64 +// +build amd64 +// +build amd64 + +package ring0 diff --git a/pkg/ring0/ring0_arm64_state_autogen.go b/pkg/ring0/ring0_arm64_state_autogen.go new file mode 100644 index 000000000..7f2ab3537 --- /dev/null +++ b/pkg/ring0/ring0_arm64_state_autogen.go @@ -0,0 +1,7 @@ +// automatically generated by stateify. + +// +build arm64 +// +build arm64 +// +build arm64 + +package ring0 diff --git a/pkg/ring0/ring0_impl_amd64_state_autogen.go b/pkg/ring0/ring0_impl_amd64_state_autogen.go new file mode 100644 index 000000000..770a13dfa --- /dev/null +++ b/pkg/ring0/ring0_impl_amd64_state_autogen.go @@ -0,0 +1,7 @@ +// automatically generated by stateify. + +// +build amd64 +// +build amd64 +// +build 386 amd64 + +package ring0 diff --git a/pkg/ring0/ring0_impl_arm64_state_autogen.go b/pkg/ring0/ring0_impl_arm64_state_autogen.go new file mode 100644 index 000000000..7f2ab3537 --- /dev/null +++ b/pkg/ring0/ring0_impl_arm64_state_autogen.go @@ -0,0 +1,7 @@ +// automatically generated by stateify. + +// +build arm64 +// +build arm64 +// +build arm64 + +package ring0 diff --git a/pkg/ring0/ring0_state_autogen.go b/pkg/ring0/ring0_state_autogen.go new file mode 100644 index 000000000..327aba163 --- /dev/null +++ b/pkg/ring0/ring0_state_autogen.go @@ -0,0 +1,3 @@ +// automatically generated by stateify. + +package ring0 diff --git a/pkg/ring0/ring0_unsafe_state_autogen.go b/pkg/ring0/ring0_unsafe_state_autogen.go new file mode 100644 index 000000000..327aba163 --- /dev/null +++ b/pkg/ring0/ring0_unsafe_state_autogen.go @@ -0,0 +1,3 @@ +// automatically generated by stateify. + +package ring0 |