diff options
author | gVisor bot <gvisor-bot@google.com> | 2021-01-12 06:42:30 +0000 |
---|---|---|
committer | gVisor bot <gvisor-bot@google.com> | 2021-01-12 06:42:30 +0000 |
commit | 1ea6658d24215e9fbcdb693b00884e1bdbdcc95d (patch) | |
tree | 12d40a4d89ff89b33b6828fe0a772f8539e2ca25 /pkg/sentry/platform | |
parent | 3f46e58c75a67dbab7422b1f2063cb633c135838 (diff) | |
parent | a20da708291e2e5bdece5176dce61c1b4b10b7d9 (diff) |
Merge release-20201216.0-86-ga20da7082 (automated)
Diffstat (limited to 'pkg/sentry/platform')
13 files changed, 2178 insertions, 59 deletions
diff --git a/pkg/sentry/platform/ring0/defs_impl_amd64.go b/pkg/sentry/platform/ring0/defs_impl_amd64.go index c0d7a8d7e..3282254a7 100644 --- a/pkg/sentry/platform/ring0/defs_impl_amd64.go +++ b/pkg/sentry/platform/ring0/defs_impl_amd64.go @@ -1,3 +1,7 @@ +// +build amd64 +// +build amd64 +// +build 386 amd64 + package ring0 import ( diff --git a/pkg/sentry/platform/ring0/defs_impl_arm64.go b/pkg/sentry/platform/ring0/defs_impl_arm64.go index 614aba4ad..92c64fe1f 100644 --- a/pkg/sentry/platform/ring0/defs_impl_arm64.go +++ b/pkg/sentry/platform/ring0/defs_impl_arm64.go @@ -1,3 +1,7 @@ +// +build arm64 +// +build arm64 +// +build arm64 + package ring0 import ( diff --git a/pkg/sentry/platform/ring0/entry_impl_arm64.s b/pkg/sentry/platform/ring0/entry_impl_arm64.s index dbc1adf8f..b3d300fb7 100644 --- a/pkg/sentry/platform/ring0/entry_impl_arm64.s +++ b/pkg/sentry/platform/ring0/entry_impl_arm64.s @@ -3,71 +3,90 @@ // 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 +#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 _RFLAGS_IF 0x200 -#define _RFLAGS_IOPL0 0x1000 -#define _KERNEL_FLAGS 0x02 +#define _KERNEL_FLAGS 0x3c5 // 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 +#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_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 +#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"); diff --git a/pkg/sentry/platform/ring0/pagetables/walker_empty_amd64.go b/pkg/sentry/platform/ring0/pagetables/walker_empty_amd64.go new file mode 100644 index 000000000..d940a9299 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_empty_amd64.go @@ -0,0 +1,257 @@ +// +build amd64 + +package pagetables + +// 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 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *emptyWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_empty_arm64.go b/pkg/sentry/platform/ring0/pagetables/walker_empty_arm64.go new file mode 100644 index 000000000..2ec9675e5 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_empty_arm64.go @@ -0,0 +1,264 @@ +// +build arm64 + +package pagetables + +// 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 sect pages. If a valid sect page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of sect pages whenever +// possible. Whether a sect 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *emptyWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_lookup_amd64.go b/pkg/sentry/platform/ring0/pagetables/walker_lookup_amd64.go new file mode 100644 index 000000000..8f2daa57e --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_lookup_amd64.go @@ -0,0 +1,257 @@ +// +build amd64 + +package pagetables + +// 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 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *lookupWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_lookup_arm64.go b/pkg/sentry/platform/ring0/pagetables/walker_lookup_arm64.go new file mode 100644 index 000000000..81bf6a384 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_lookup_arm64.go @@ -0,0 +1,264 @@ +// +build arm64 + +package pagetables + +// 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 sect pages. If a valid sect page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of sect pages whenever +// possible. Whether a sect 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *lookupWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_map_amd64.go b/pkg/sentry/platform/ring0/pagetables/walker_map_amd64.go new file mode 100644 index 000000000..755c9afc3 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_map_amd64.go @@ -0,0 +1,257 @@ +// +build amd64 + +package pagetables + +// 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 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *mapWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_map_arm64.go b/pkg/sentry/platform/ring0/pagetables/walker_map_arm64.go new file mode 100644 index 000000000..9a8f0b1d6 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_map_arm64.go @@ -0,0 +1,264 @@ +// +build arm64 + +package pagetables + +// 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 sect pages. If a valid sect page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of sect pages whenever +// possible. Whether a sect 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *mapWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_unmap_amd64.go b/pkg/sentry/platform/ring0/pagetables/walker_unmap_amd64.go new file mode 100644 index 000000000..16ab12d73 --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_unmap_amd64.go @@ -0,0 +1,257 @@ +// +build amd64 + +package pagetables + +// 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 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *unmapWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/pagetables/walker_unmap_arm64.go b/pkg/sentry/platform/ring0/pagetables/walker_unmap_arm64.go new file mode 100644 index 000000000..695aecb6d --- /dev/null +++ b/pkg/sentry/platform/ring0/pagetables/walker_unmap_arm64.go @@ -0,0 +1,264 @@ +// +build arm64 + +package pagetables + +// 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 sect pages. If a valid sect page (huge or jumbo) cannot be +// installed, then the walk will continue to individual entries. +// +// This algorithm will attempt to maximize the use of sect pages whenever +// possible. Whether a sect 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") + } + w.iterateRangeCanonical(start, lowerTop) + 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 +} + +// iterateRangeCanonical walks a canonical range. +// +//go:nosplit +func (w *unmapWalker) iterateRangeCanonical(start, end uintptr) { + 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() + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + 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 { + + w.visitor.visit(uintptr(start), pudEntry, pudSize-1) + + 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() + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + 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 { + + w.visitor.visit(uintptr(start), pmdEntry, pmdSize-1) + + 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 + } + + w.visitor.visit(uintptr(start), pteEntry, pteSize-1) + 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) + } + } +} diff --git a/pkg/sentry/platform/ring0/ring0_impl_amd64_state_autogen.go b/pkg/sentry/platform/ring0/ring0_impl_amd64_state_autogen.go index 327aba163..770a13dfa 100644 --- a/pkg/sentry/platform/ring0/ring0_impl_amd64_state_autogen.go +++ b/pkg/sentry/platform/ring0/ring0_impl_amd64_state_autogen.go @@ -1,3 +1,7 @@ // automatically generated by stateify. +// +build amd64 +// +build amd64 +// +build 386 amd64 + package ring0 diff --git a/pkg/sentry/platform/ring0/ring0_impl_arm64_state_autogen.go b/pkg/sentry/platform/ring0/ring0_impl_arm64_state_autogen.go index 327aba163..7f2ab3537 100644 --- a/pkg/sentry/platform/ring0/ring0_impl_arm64_state_autogen.go +++ b/pkg/sentry/platform/ring0/ring0_impl_arm64_state_autogen.go @@ -1,3 +1,7 @@ // automatically generated by stateify. +// +build arm64 +// +build arm64 +// +build arm64 + package ring0 |