node/deps/v8/include/v8-sandbox.h

174 lines
7.4 KiB
C++

// Copyright 2024 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#ifndef INCLUDE_V8_SANDBOX_H_
#define INCLUDE_V8_SANDBOX_H_
#include <cstdint>
#include "v8-internal.h" // NOLINT(build/include_directory)
#include "v8config.h" // NOLINT(build/include_directory)
namespace v8 {
/**
* A pointer tag used for wrapping and unwrapping `CppHeap` pointers as used
* with JS API wrapper objects that rely on `v8::Object::Wrap()` and
* `v8::Object::Unwrap()`.
*
* The CppHeapPointers use a range-based type checking scheme, where on access
* to a pointer, the actual type of the pointer is checked to be within a
* specified range of types. This allows supporting type hierarchies, where a
* type check for a supertype must succeed for any subtype.
*
* The tag is currently in practice limited to 15 bits since it needs to fit
* together with a marking bit into the unused parts of a pointer (the top 16
* bits).
*/
enum class CppHeapPointerTag : uint16_t {
kFirstTag = 0,
kNullTag = 0,
/**
* The lower type ids are reserved for the embedder to assign. For that, the
* main requirement is that all (transitive) child classes of a given parent
* class have type ids in the same range, and that there are no unrelated
* types in that range. For example, given the following type hierarchy:
*
* A F
* / \
* B E
* / \
* C D
*
* a potential type id assignment that satistifes these requirements is
* {C: 0, D: 1, B: 2, A: 3, E: 4, F: 5}. With that, the type check for type A
* would check for the range [0, 4], while the check for B would check range
* [0, 2], and for F it would simply check [5, 5].
*
* In addition, there is an option for performance tweaks: if the size of the
* type range corresponding to a supertype is a power of two and starts at a
* power of two (e.g. [0x100, 0x13f]), then the compiler can often optimize
* the type check to use even fewer instructions (essentially replace a AND +
* SUB with a single AND).
*/
kDefaultTag = 0x7000,
kZappedEntryTag = 0x7ffd,
kEvacuationEntryTag = 0x7ffe,
kFreeEntryTag = 0x7fff,
// The tags are limited to 15 bits, so the last tag is 0x7fff.
kLastTag = 0x7fff,
};
// Convenience struct to represent tag ranges. This is used for type checks
// against supertypes, which cover a range of types (their subtypes).
// Both the lower- and the upper bound are inclusive. In other words, this
// struct represents the range [lower_bound, upper_bound].
struct CppHeapPointerTagRange {
constexpr CppHeapPointerTagRange(CppHeapPointerTag lower,
CppHeapPointerTag upper)
: lower_bound(lower), upper_bound(upper) {}
CppHeapPointerTag lower_bound;
CppHeapPointerTag upper_bound;
// Check whether the tag of the given CppHeapPointerTable entry is within
// this range. This method encodes implementation details of the
// CppHeapPointerTable, which is necessary as it is used by
// ReadCppHeapPointerField below.
// Returns true if the check is successful and the tag of the given entry is
// within this range, false otherwise.
bool CheckTagOf(uint64_t entry) {
// Note: the cast to uint32_t is important here. Otherwise, the uint16_t's
// would be promoted to int in the range check below, which would result in
// undefined behavior (signed integer undeflow) if the actual value is less
// than the lower bound. Then, the compiler would take advantage of the
// undefined behavior and turn the range check into a simple
// `actual_tag <= last_tag` comparison, which is incorrect.
uint32_t actual_tag = static_cast<uint16_t>(entry);
// The actual_tag is shifted to the left by one and contains the marking
// bit in the LSB. To ignore that during the type check, simply add one to
// the (shifted) range.
constexpr int kTagShift = internal::kCppHeapPointerTagShift;
uint32_t first_tag = static_cast<uint32_t>(lower_bound) << kTagShift;
uint32_t last_tag = (static_cast<uint32_t>(upper_bound) << kTagShift) + 1;
return actual_tag >= first_tag && actual_tag <= last_tag;
}
};
constexpr CppHeapPointerTagRange kAnyCppHeapPointer(
CppHeapPointerTag::kFirstTag, CppHeapPointerTag::kLastTag);
class SandboxHardwareSupport {
public:
/**
* Initialize sandbox hardware support. This needs to be called before
* creating any thread that might access sandbox memory since it sets up
* hardware permissions to the memory that will be inherited on clone.
*/
V8_EXPORT static void InitializeBeforeThreadCreation();
};
namespace internal {
#ifdef V8_COMPRESS_POINTERS
V8_INLINE static Address* GetCppHeapPointerTableBase(v8::Isolate* isolate) {
Address addr = reinterpret_cast<Address>(isolate) +
Internals::kIsolateCppHeapPointerTableOffset +
Internals::kExternalPointerTableBasePointerOffset;
return *reinterpret_cast<Address**>(addr);
}
#endif // V8_COMPRESS_POINTERS
template <typename T>
V8_INLINE static T* ReadCppHeapPointerField(v8::Isolate* isolate,
Address heap_object_ptr, int offset,
CppHeapPointerTagRange tag_range) {
#ifdef V8_COMPRESS_POINTERS
// See src/sandbox/cppheap-pointer-table-inl.h. Logic duplicated here so
// it can be inlined and doesn't require an additional call.
const CppHeapPointerHandle handle =
Internals::ReadRawField<CppHeapPointerHandle>(heap_object_ptr, offset);
const uint32_t index = handle >> kExternalPointerIndexShift;
const Address* table = GetCppHeapPointerTableBase(isolate);
const std::atomic<Address>* ptr =
reinterpret_cast<const std::atomic<Address>*>(&table[index]);
Address entry = std::atomic_load_explicit(ptr, std::memory_order_relaxed);
Address pointer = entry;
if (V8_LIKELY(tag_range.CheckTagOf(entry))) {
pointer = entry >> kCppHeapPointerPayloadShift;
} else {
// If the type check failed, we simply return nullptr here. That way:
// 1. The null handle always results in nullptr being returned here, which
// is a desired property. Otherwise, we would need an explicit check for
// the null handle above, and therefore an additional branch. This
// works because the 0th entry of the table always contains nullptr
// tagged with the null tag (i.e. an all-zeros entry). As such,
// regardless of whether the type check succeeds, the result will
// always be nullptr.
// 2. The returned pointer is guaranteed to crash even on platforms with
// top byte ignore (TBI), such as Arm64. The alternative would be to
// simply return the original entry with the left-shifted payload.
// However, due to TBI, an access to that may not always result in a
// crash (specifically, if the second most significant byte happens to
// be zero). In addition, there shouldn't be a difference on Arm64
// between returning nullptr or the original entry, since it will
// simply compile to a `csel x0, x8, xzr, lo` instead of a
// `csel x0, x10, x8, lo` instruction.
pointer = 0;
}
return reinterpret_cast<T*>(pointer);
#else // !V8_COMPRESS_POINTERS
return reinterpret_cast<T*>(
Internals::ReadRawField<Address>(heap_object_ptr, offset));
#endif // !V8_COMPRESS_POINTERS
}
} // namespace internal
} // namespace v8
#endif // INCLUDE_V8_SANDBOX_H_