raw_ptr<T> (aka MiraclePtr, aka BackupRefPtr)

raw_ptr<T> is a non-owning smart pointer that has improved memory-safety over raw pointers. It behaves just like a raw pointer on platforms where USE_BACKUP_REF_PTR is off, and almost like one when it's on. The main difference is that when USE_BACKUP_REF_PTR is enabled, raw_ptr<T> is beneficial for security, because it can prevent a significant percentage of Use-after-Free (UaF) bugs from being exploitable. It does this by poisoning the freed memory and quarantining it as long as a dangling raw_ptr<T> exists.

When USE_BACKUP_REF_PTR is on, raw_ptr<T> it is zero-initialized and cleared on destruction and move. However, when USE_BACKUP_REF_PTR is off, this is not the case. You must continue to explicitly initialize raw_ptr members to ensure consistent behavior across all cases. Because the underlying implementation may vary, always use the raw_ptr<T> member_ = nullptr; form of initialization rather than the so-called uniform initialization form (empty braces) raw_ptr<T> member_{}; whose meaning varies with the implementation.

Unlike std::unique_ptr<T>, base::scoped_refptr<T>, etc., raw_ptr<T> does not manage ownership or lifetime of an allocated object - you are still responsible for freeing the object when no longer used, just as you would with a raw C++ pointer.

raw_ptr<T> has limited impact on stability - dereferencing a dangling pointer remains Undefined Behavior (although poisoning may lead to earlier, easier to debug crashes). Note that the security protection is not yet enabled by default.

raw_ptr<T> is a part of the MiraclePtr project and currently implements the BackupRefPtr algorithm. If needed, please reach out to memory-safety-dev@chromium.org or (Google-internal) chrome-memory-safety@google.com with questions or concerns.

When to use |raw_ptr<T>|

The Chromium C++ Style Guide asks to use raw_ptr<T> for class and struct fields in place of a raw C++ pointer T* whenever possible, except in Renderer-only code. This guide offers more details.

The usage guidelines are not enforced currently (the MiraclePtr team will turn on enforcement via Chromium Clang Plugin after confirming performance results via Stable channel experiments). Afterwards we plan to allow exclusions via:

  • manual-paths-to-ignore.txt to exclude at a directory level. Examples:
    • Renderer-only code (i.e. code in paths that contain /renderer/ or third_party/blink/public/web/)
    • Code that cannot depend on //base
    • Code in //ppapi
  • RAW_PTR_EXCLUSION C++ attribute to exclude individual fields. Examples:
  • No explicit exclusions will be needed for:
    • const char*, const wchar_t*, etc.
    • Function pointers
    • ObjC pointers

Examples of using |raw_ptr<T>| instead of raw C++ pointers

Consider an example struct that uses raw C++ pointer fields:

struct Example {
  int* int_ptr;
  void* void_ptr;
  SomeClass* object_ptr;
  const SomeClass* ptr_to_const;
  SomeClass* const const_ptr;
};

When using raw_ptr<T> the struct above would look as follows:

#include "base/memory/raw_ptr.h"

struct Example {
  raw_ptr<int> int_ptr;
  raw_ptr<void> void_ptr;
  raw_ptr<SomeClass> object_ptr;
  raw_ptr<const SomeClass> ptr_to_const;
  const raw_ptr<SomeClass> const_ptr;
};

In most cases, only the type in the field declaration needs to change. In particular, raw_ptr<T> implements operator->, operator* and other operators that one expects from a raw pointer. Cases where other code needs to be modified are described in the “Recoverable compile-time problems” section below.

Performance

Performance impact of using |raw_ptr<T>| instead of |T*|

Compared to a raw C++ pointer, on platforms where USE_BACKUP_REF_PTR is on, raw_ptr<T> incurs additional runtime overhead for initialization, destruction, and assignment (including ptr++ and ptr += ...). There is no overhead when dereferencing or extracting a pointer (including *ptr, ptr->foobar, ptr.get(), or implicit conversions to a raw C++ pointer). Finally, raw_ptr<T> has exactly the same memory footprint as T* (i.e. sizeof(raw_ptr<T>) == sizeof(T*)).

One source of the performance overhead is a check whether a pointer T* points to a protected memory pool. This happens in raw_ptr<T>'s constructor, destructor, and assignment operators. If the pointed memory is unprotected, then raw_ptr<T> behaves just like a T* and the runtime overhead is limited to the extra check. (The security protection incurs additional overhead described in the “Performance impact of enabling Use-after-Free protection” section below.)

Some additional overhead comes from setting raw_ptr<T> to nullptr when default-constructed, destructed, or moved.

During the “Big Rewrite” most Chromium T* fields have been rewritten to raw_ptr<T> (excluding fields in Renderer-only code). The cumulative performance impact of such rewrite has been measured by earlier A/B binary experiments. There was no measurable impact, except that 32-bit platforms have seen a slight increase in jankiness metrics (for more detailed results see the document here).

Performance impact of enabling Use-after-Free protection

When the Use-after-Free protection is enabled, then raw_ptr<T> has some additional performance overhead. This protection is currently disabled by default. We will enable the protection incrementally, starting with more non-Renderer processes first.

The protection can increase memory usage:

  • For each memory allocation Chromium‘s allocator (PartitionAlloc) allocates extra 16 bytes (4 bytes to store the BackupRefPtr’s ref-count associated with the allocation, the rest to maintain alignment requirements).
  • Freed memory is quarantined and not available for reuse as long as dangling raw_ptr<T> pointers exist.
  • Enabling protection requires additional partitions in PartitionAlloc, which increases memory fragmentation.

The protection can increase runtime costs - raw_ptr<T>‘s constructor, destructor, and assignment operators (including ptr++ and ptr += ...) need to maintain BackupRefPtr’s ref-count.

When it is okay to continue using raw C++ pointers

Unsupported cases leading to compile errors

Continue to use raw C++ pointers in the following cases, which may otherwise result in compile errors:

  • Function pointers
  • Pointers to Objective-C objects
  • Pointer fields in classes/structs that are used as global, static, or thread_local variables (see more details in the Rewrite exclusion statistics )
  • Pointer fields that require non-null, constexpr initialization (see more details in the Rewrite exclusion statistics )
  • Pointer fields in classes/structs that have to be trivially constructible or destructible
  • Code that doesn’t depend on //base (including non-Chromium repositories and third party libraries)
  • Code in //ppapi

Pointers to unprotected memory (performance optimization)

Using raw_ptr<T> offers no security benefits (no UaF protection) for pointers that don’t point to protected memory (only PartitionAlloc-managed heap allocations in non-Renderer processes are protected). Therefore in the following cases raw C++ pointers may be used instead of raw_ptr<T>:

  • Pointer fields that can only point outside PartitionAlloc, including literals, stack allocated memory, shared memory, mmap'ed memory, V8/Oilpan/Java heaps, TLS, etc.
  • const char* (and const wchar_t*) pointer fields, unless you’re convinced they can point to a heap-allocated object, not just a string literal
  • Pointer fields that can only point to aligned allocations (requested via PartitionAlloc’s AlignedAlloc or memalign family of functions, with alignment higher than base::kAlignment)
  • Pointer fields in Renderer-only code. (This might change in the future as we explore expanding raw_ptr<T> usage in https://crbug.com/1273204.)

Other perf optimizations

As a performance optimization, raw C++ pointers may be used instead of raw_ptr<T> if it would have a significant performance impact.

Pointers in locations other than fields

Use raw C++ pointers instead of raw_ptr<T> in the following scenarios:

  • Pointers in local variables and function parameters and return values. This includes pointer fields in classes/structs that are used only on the stack. (We plan to enforce this in the Chromium Clang Plugin. Using raw_ptr<T> here would cumulatively lead to performance regression and the security benefit of UaF protection is lower for such short-lived pointers.)
  • Pointer fields in unions. However, note that a much better, modern alternative is absl::variant + raw_ptr<T>. If use of C++ union is absolutely unavoidable, prefer a regular C++ pointer: incorrect management of a raw_ptr<T> field can easily lead to refcount corruption.
  • Pointers whose addresses are used only as identifiers and which are never dereferenced (e.g. keys in a map). There is a performance gain by not using raw_ptr in this case; prefer to use uintptr_t to emphasize that the entity can dangle and must not be dereferenced.

You don’t have to, but may use raw_ptr<T>, in the following scenarios:

  • Pointers that are used as an element type of collections/wrappers. E.g. std::vector<T*> and std::vector<raw_ptr<T>> are both okay, but prefer the latter if the collection is a class field (note that some of the perf optimizations above might still apply and argue for using a raw C++ pointer).

Extra pointer rules

raw_ptr<T> requires following some extra rules compared to a raw C++ pointer:

  • Don’t assign invalid, non-null addresses (this includes previously valid and now freed memory, Win32 handles, and more). You can only assign an address of memory that is allocated at the time of assignment. Exceptions:
    • a pointer to the end of a valid allocation (but not even 1 byte further)
    • a pointer to the last page of the address space, e.g. for sentinels like reinterpret_cast<void*>(-1)
  • Don’t initialize or assign raw_ptr<T> memory directly (e.g. reinterpret_cast<ClassWithRawPtr*>(buffer) or memcpy(reinterpret_cast<void*>(&obj_with_raw_ptr), buffer).
  • Don’t assign to a raw_ptr<T> concurrently, even if the same value.
  • Don’t rely on moved-from pointers to keep their old value. Unlike raw pointers, raw_ptr<T> is cleared upon moving.
  • Don't use the pointer after it is destructed. Unlike raw pointers, raw_ptr<T> is cleared upon destruction. This may happen e.g. when fields are ordered such that the pointer field is destructed before the class field whose destructor uses that pointer field (e.g. see Esoteric Issues).
  • Don’t assign to a raw_ptr<T> until its constructor has run. This may happen when a base class’s constructor uses a not-yet-initialized field of a derived class (e.g. see Applying MiraclePtr).

Some of these would result in undefined behavior (UB) even in the world without raw_ptr<T> (e.g. see Field destruction order), but you’d likely get away without any consequences. In the raw_ptr<T> world, an obscure crash may occur. Those crashes often manifest themselves as SEGV or CHECK inside RawPtrBackupRefImpl::AcquireInternal() or RawPtrBackupRefImpl::ReleaseInternal(), but you may also experience memory corruption or a silent drop of UaF protection.

Recoverable compile-time problems

Explicit |raw_ptr.get()| might be needed

If a raw pointer is needed, but an implicit cast from raw_ptr<SomeClass> to SomeClass* doesn't work, then the raw pointer needs to be obtained by explicitly calling .get(). Examples:

  • auto* raw_ptr_var = wrapped_ptr_.get() (auto* requires the initializer to be a raw pointer)
    • Alternatively you can change auto* to auto&. Avoid using auto as it’ll copy the pointer, which incurs a performance overhead.
  • return condition ? raw_ptr : wrapped_ptr_.get(); (ternary operator needs identical types in both branches)
  • base::WrapUniquePtr(wrapped_ptr_.get()); (implicit cast doesn't kick in for arguments in templates)
  • printf("%p", wrapped_ptr_.get()); (can't pass class type arguments to variadic functions)
  • reinterpret_cast<SomeClass*>(wrapped_ptr_.get()) (const_cast and reinterpret_cast sometimes require their argument to be a raw pointer; static_cast should “Just Work”)
  • T2 t2 = t1_wrapped_ptr_.get(); (where there is an implicit conversion constructor T2(T1*) the compiler can handle one implicit conversion, but not two)
  • In general, when type is inferred by a compiler and then used in a context where a pointer is expected.

Out-of-line constructor/destructor might be needed

Out-of-line constructor/destructor may be newly required by the chromium style clang plugin. Error examples:

  • error: [chromium-style] Complex class/struct needs an explicit out-of-line destructor.
  • error: [chromium-style] Complex class/struct needs an explicit out-of-line constructor.

raw_ptr<T> uses a non-trivial constructor/destructor, so classes that used to be POD or have a trivial destructor may require an out-of-line constructor/destructor to satisfy the chromium style clang plugin.

In-out arguments need to be refactored

Due to implementation difficulties, raw_ptr<T> doesn't support an address-of operator. This means that the following code will not compile:

void GetSomeClassPtr(SomeClass** out_arg) {
  *out_arg = ...;
}

struct MyStruct {
  void Example() {
    GetSomeClassPtr(&wrapped_ptr_);  // <- won't compile
  }

  raw_ptr<SomeClass> wrapped_ptr_;
};

The typical fix is to change the type of the out argument (see also an example CL here):

void GetSomeClassPtr(raw_ptr<SomeClass>* out_arg) {
  *out_arg = ...;
}

Similarly this code:

void FillPtr(SomeClass*& out_arg) {
  out_arg = ...;
}

would have to be changed to this:

void FillPtr(raw_ptr<SomeClass>& out_arg) {
  out_arg = ...;
}

Similarly this code:

SomeClass*& GetPtr() {
  return wrapper_ptr_;
}

would have to be changed to this:

raw_ptr<SomeClass>& GetPtr() {
  return wrapper_ptr_;
}

Modern |nullptr| is required

As recommended by the Google C++ Style Guide, use nullptr instead of NULL - the latter might result in compile-time errors when used with raw_ptr<T>.

Example:

struct SomeStruct {
  raw_ptr<int> ptr_field;
};

void bar() {
  SomeStruct some_struct;
  some_struct.ptr_field = NULL;
}

Error:

../../base/memory/checked_ptr_unittest.cc:139:25: error: use of overloaded
operator '=' is ambiguous (with operand types raw_ptr<int>' and 'long')
  some_struct.ptr_field = NULL;
  ~~~~~~~~~~~~~~~~~~~~~ ^ ~~~~
../../base/memory/raw_ptr.h:369:29: note: candidate function
  ALWAYS_INLINE raw_ptr& operator=(std::nullptr_t) noexcept {
                         ^
../../base/memory/raw_ptr.h:374:29: note: candidate function
  ALWAYS_INLINE raw_ptr& operator=(T* p)
                         noexcept {

[rare] Explicit overload or template specialization for |raw_ptr<T>|

In rare cases, the default template code won’t compile when raw_ptr<...> is substituted for a template argument. In such cases, it might be necessary to provide an explicit overload or template specialization for raw_ptr<T>.

Example (more details in Applying MiraclePtr and the Add CheckedPtr support for cbor_extract::Element CL):

// An explicit overload (taking raw_ptr<T> as an argument)
// was needed below:
template <typename S>
constexpr StepOrByte<S> Element(
    const Is required,
    raw_ptr<const std::string> S::*member,  // <- HERE
    uintptr_t offset) {
  return ElementImpl<S>(required, offset, internal::Type::kString);
}

AddressSanitizer support

For years, AddressSanitizer has been the main tool for diagnosing memory corruption issues in Chromium. MiraclePtr alters the security properties of some of some such issues, so ideally it should be integrated with ASan. That way an engineer would be able to check whether a given use-after-free vulnerability is covered by the protection without having to switch between ASan and non-ASan builds.

Unfortunately, MiraclePtr relies heavily on PartitionAlloc, and ASan needs its own allocator to work. As a result, the default implementation of raw_ptr<T> can't be used with ASan builds. Instead, a special version of raw_ptr<T> has been implemented, which is based on the ASan quarantine and acts as a sufficiently close approximation for diagnostic purposes. At crash time, the tool will tell the user if the dangling pointer access would have been protected by MiraclePtr in a regular build.

You can configure the diagnostic tool by modifying the parameters of the feature flag PartitionAllocBackupRefPtr. For example, launching Chromium as follows:

path/to/chrome --enable-features=PartitionAllocBackupRefPtr:enabled-processes/browser-only/asan-enable-dereference-check/true/asan-enable-extraction-check/true/asan-enable-instantiation-check/true

activates all available checks in the browser process.

Available checks

MiraclePtr provides ASan users with three kinds of security checks, which differ in when a particular check occurs:

Dereference

This is the basic check type that helps diagnose regular heap-use-after-free bugs. It's enabled by default.

Extraction

The user will be warned if a dangling pointer is extracted from a raw_ptr<T> variable. If the pointer is then dereferenced, an ASan error report will follow. In some cases, extra work on the reproduction case is required to reach the faulty memory access. However, even without memory corruption, relying on the value of a dangling pointer may lead to problems. For example, it's a common (anti-)pattern in Chromium to use a raw pointer as a key in a container. Consider the following example:

std::map<T*, std::unique_ptr<Ext>> g_map;

struct A {
  A() {
    g_map[this] = std::make_unique<Ext>(this);
  }

  ~A() {
    g_map.erase(this);
  }
};

raw_ptr<A> dangling = new A;
// ...
delete dangling.get();
A* replacement = new A;
// ...
auto it = g_map.find(dangling);
if (it == g_map.end())
  return 0;
it->second.DoStuff();

Depending on whether the allocator reuses the same memory region for the second A object, the program may inadvertently call DoStuff() on the wrong Ext instance. This, in turn, may corrupt the state of the program or bypass security controls if the two A objects belong to different security contexts.

Given the proportion of false positives reported in the mode, it is disabled by default. It's mainly intended to be used by security researchers who are willing to spend a significant amount of time investigating these early warnings.

Instantiation

This check detects violations of the rule that when instantiating a raw_ptr<T> from a T* , it is only allowed if the T* is a valid (i.e. not dangling) pointer. This rule exists to help avoid an issue called “pointer laundering” which can result in unsafe raw_ptr<T> instances that point to memory that is no longer in quarantine. This is important, since subsequent use of these raw_ptr<T> might appear to be safe.

In order for “pointer laundering” to occur, we need (1) a dangling T* (pointing to memory that has been freed) to be assigned to a raw_ptr<T>, while (2) there is no other raw_ptr<T> pointing to the same object/allocation at the time of assignment.

The check only detects (1), a dangling T* being assigned to a raw_ptr<T>, so in order to determine whether “pointer laundering” has occurred, we need to determine whether (2) could plausibly occur, not just in the specific reproduction testcase, but in the more general case.

In the absence of thorough reasoning about (2), the assumption here should be that any failure of this check is a security issue of the same severity as an unprotected use-after-free.

Protection status

When ASan generates a heap-use-after-free report, it will include a new section near the bottom, which starts with the line MiraclePtr Status: <status>. At the moment, it has three possible options:

Protected

The system is sufficiently confident that MiraclePtr makes the discovered issue unexploitable. In the future, the security severity of such bugs will be reduced.

Manual analysis required

Dangling pointer extraction was detected before the crash, but there might be extra code between the extraction and dereference. Most of the time, the code in question will look similar to the following:

struct A {
  raw_ptr<T> dangling_;
};

void trigger(A* a) {
  // ...
  T* local = a->dangling_;
  DoStuff();
  local->DoOtherStuff();
  // ...
}

In this scenario, even though dangling_ points to freed memory, that memory is protected and will stay in quarantine until dangling_ (and all other raw_ptr<T> variables pointing to the same region) changes its value or gets destroyed. Therefore, the expression a_->dangling->DoOtherStuff() wouldn't trigger an exploitable use-after-free.

You will need to make sure that DoStuff() is sufficiently trivial and can‘t (not only for the particular reproduction case, but even in principle) make dangling_ change its value or get destroyed. If that’s the case, the DoOtherStuff() call may be considered protected. The tool will provide you with the stack trace for both the extraction and dereference events.

Not protected

The dangling T* doesn‘t appear to originate from a raw_ptr<T> variable, which means MiraclePtr can’t prevent this issue from being exploited. In practice, there may still be a raw_ptr<T> in a different part of the code that protects the same allocation indirectly, but such protection won't be considered robust enough to impact security-related decisions.

Limitations

The main limitation of MiraclePtr in ASan builds is the main limitation of ASan itself: the capacity of the quarantine is limited. Eventually, every allocation in quarantine will get reused regardless of whether there are still references to it.

In the context of MiraclePtr combined with ASan, it's a problem when:

  1. A heap allocation that isn't supported by MiraclePtr is made. At the moment, the only such class is allocations made early during the process startup before MiraclePtr can be activated.
  2. Its address is assigned to a raw_ptr<T> variable.
  3. The allocation gets freed.
  4. A new allocation is made in the same memory region as the first one, but this time it is supported.
  5. The second allocation gets freed.
  6. The raw_ptr<T> variable is accessed.

In this case, MiraclePtr will incorrectly assume the memory access is protected. Luckily, considering the small number of unprotected allocations in Chromium, the size of the quarantine, and the fact that most reproduction cases take relatively short time to run, the odds of this happening are very low.

The problem is relatively easy to spot if you look at the ASan report: the allocation and deallocation stack traces won‘t be consistent across runs and the allocation type won’t match the use stack trace.

If you encounter a suspicious ASan report, it may be helpful to re-run Chromium with an increased quarantine capacity as follows:

ASAN_OPTIONS=quarantine_size_mb=1024 path/to/chrome

Appendix: Is raw_ptr Live?

Diagram showing how both code support and feature flag must be presentfor raw_ptr to be BRP.

Note that

  • RawPtrNoOpImpl is thought to have no overhead. However, this has yet to be verified.

  • “Inert BackupRefPtr” has overhead - once BRP support is compiled in, every raw_ptr will (at assignment) perform the check that asks, “is BRP protection active?”

As for general BRP enablement,

  • BRP is live in most browser tests and Chromium targets.

    • This is nuanced by platform type and process type.
  • In unit tests,

    • raw_ptr is the no-op impl when the build is ASan.

    • raw_ptr is live BRP on bots.

    • raw_ptr is inert BRP otherwise (see https://crbug.com/1440658).