Updating Carbon's safety strategy

README.md

@@ -244,9 +244,17 @@ and with a smooth evolutionary path.
 
 Safety, and especially
 [memory safety](https://en.wikipedia.org/wiki/Memory_safety), remains a key
-challenge for C++ and something a successor language needs to address. Our
-initial priority and focus is on immediately addressing important, low-hanging
-fruit in the safety space:
+challenge for C++ and something a successor language needs to address.
+
+We plan to support a two step migration process:
+
+1. Highly automated, minimal supervision migration from C++ to a dialect of
+   Carbon designed for C++ interop and migration.
+2. Incremental refactoring of the Carbon code to adopt memory safe designs,
+   patterns, and APIs.
+
+We also want to address important, low-hanging fruit in the safety space
+immediately when migrating into Carbon:
 
 -   Tracking uninitialized states better, increased enforcement of
     initialization, and systematically providing hardening against
@@ -257,14 +265,7 @@ fruit in the safety space:
     comprehensive than existing C++ build modes even when combined with
     [Address Sanitizer](https://github.com/google/sanitizers/wiki/AddressSanitizer).
 
-Once we can migrate code into Carbon, we will have a simplified language with
-room in the design space to add any necessary annotations or features, and
-infrastructure like [generics](#generics) to support safer design patterns.
-Longer term, we will build on this to introduce **a safe Carbon subset**. This
-will be a large and complex undertaking, and won't be in the 0.1 design.
-Meanwhile, we are closely watching and learning from efforts to add memory safe
-semantics onto C++ such as Rust-inspired
-[lifetime annotations](https://discourse.llvm.org/t/rfc-lifetime-annotations-for-c/61377).
+For more details, see our [safety design](/docs/design/safety).
 
 ## Getting started
 

docs/design/README.md

@@ -372,14 +372,11 @@ required to be the only non-whitespace on the line.
 
 The behavior of the Carbon compiler depends on the _build mode_:
 
--   In a _development build_, the priority is diagnosing problems and fast build
-    time.
--   In a _performance build_, the priority is fastest execution time and lowest
+-   In a _debug build_, the priority is diagnosing problems and fast build time.
+-   In a _release build_, the priority is fastest execution time and lowest
     memory usage.
--   In a _hardened build_, the first priority is safety and second is
-    performance.
 
-> References: [Safety strategy](/docs/project/principles/safety_strategy.md)
+> References: [Safety design](/docs/design/safety#build-modes)
 
 ## Types are values
 
@@ -3746,7 +3743,7 @@ This leads to Carbon's incremental path to safety:
 -   Shift the Carbon code onto an incremental path towards memory safety over
     the next decade.
 
-> References: [Safety strategy](/docs/project/principles/safety_strategy.md)
+> References: [Safety design](/docs/design/safety)
 
 ### Lifetime and move semantics
 

docs/design/safety/README.md

@@ -0,0 +1,220 @@
+# Safety
+
+<!--
+Part of the Carbon Language project, under the Apache License v2.0 with LLVM
+Exceptions. See /LICENSE for license information.
+SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+-->
+
+<!-- toc -->
+
+## Table of contents
+
+-   [Overview](#overview)
+-   [Safe and unsafe code](#safe-and-unsafe-code)
+-   [Safety modes](#safety-modes)
+-   [Memory safety model](#memory-safety-model)
+    -   [Data races versus unsynchronized temporal safety](#data-races-versus-unsynchronized-temporal-safety)
+-   [Safe library ecosystem](#safe-library-ecosystem)
+-   [Build modes](#build-modes)
+-   [References](#references)
+
+<!-- tocstop -->
+
+## Overview
+
+One of Carbon's core goals is [practical safety]. This is referring to _[code
+safety]_
+as opposed to the larger space of [systems safety]. The largest aspect of code safety
+at the language level is [memory safety], but this also applies to other aspects
+of code safety such as avoiding undefined behavior in other forms.
+
+[practical safety]:
+    /docs/project/goals.md#practical-safety-and-testing-mechanisms
+[code safety]:
+    /docs/design/safety/terminology.md#code-software-or-program-safety
+[systems safety]: /docs/design/safety/terminology.md#safety
+
+However, Carbon also has a goal of fine grained, smooth interop-with and
+migration-from existing C++ code, and C++ is a fundamentally unsafe language. It
+has pervasive sources of undefined behavior and minimal memory safety
+guarantees. Our safety strategy has to address how C++ code fits into it, and
+provide an incremental path from where the code is at today towards increasing
+levels of safety.
+
+Ultimately, Carbon will both provide a [memory safe language], _and_ provide a
+language that can interop with C++ and be targeted for mechanical migration from
+C++.
+
+[memory safe language]: /docs/design/safety/terminology.md#memory-safe-language
+
+## Safe and unsafe code
+
+Carbon will have both _safe_ and _unsafe_ code. Safe code provides limits on the
+potential behavior of the program even in the face of bugs in order to prevent
+[safety bugs] from becoming [vulnerabilities]. Unsafe code is any code or operation
+which may lack limits or guarantees on behavior, and as a result may result in undefined
+behavior and a safety bug.
+
+[safety bugs]: /docs/design/safety/terminology.md#safety-bugs
+[vulnerabilities]:
+    /docs/design/safety/terminology.md#vulnerability-or-security-vulnerability
+
+All things being equal, safe code constructs are preferable to unsafe
+constructs, but many unsafe constructs are necessary. Where unsafe constructs
+are needed, Carbon follows three principles to incorporate them into the
+language:
+
+-   The unsafe capabilities provided should be semantically narrow: only the
+    minimal unsafe operation to achieve the desired result.
+-   The unsafe code should be syntactically narrow and separable from
+    surrounding safe code.
+-   There should be a reasonable way of including the keyword `unsafe` in
+    whatever syntax is used so that this keyword can be used as a common
+    annotation in audits and review.
+
+The result is that Carbon does not have a "safe" mode and an "unsafe" mode, but
+rather narrow and specific unsafe operations and constructs.
+
+## Safety modes
+
+The _safety mode_ of Carbon governs the extent to which unsafe code must include
+the local `unsafe` keyword in its syntax to delineate it from safe code.
+
+_Strict Carbon_ is the mode in which all unsafe code is marked with the `unsafe`
+keyword. This mode is designed to satisfy the requirements of a [memory
+safe language].
+
+_Permissive Carbon_ is a mode optimized for interop with C++ and automated
+migration from C++. In this mode, some unsafe code to not require an `unsafe`
+keyword: specific aspects of C++ interop or pervasive patterns that occur when
+migrating from C++. However, not _all_ unsafe code will omit the keyword, the
+permissive mode is designed to be minimal in the unsafe code allowed.
+
+## Memory safety model
+
+Carbon will use a hybrid of different techniques to achieve memory safety in its
+safe code, largely broken down by the categories of memory safety:
+
+-   [Type safety]: compile-time enforcement, the same as other statically typed languages
+    with generic type systems.
+-   [Initialization safety]: hybrid of run-time and compile-time enforcement.
+-   [Spatial safety]: run-time enforcement.
+-   [Temporal safety]: compile-time enforcement through its type system.
+-   [Data-race safety]: compile-time enforcement through its type system.
+
+[type safety]: /docs/design/safety/terminology.md#type-safety
+[initialization safety]:
+    /docs/design/safety/terminology.md#initialization-safety
+[spatial safety]: /docs/design/safety/terminology.md#spatial-safety
+[temporal safety]: /docs/design/safety/terminology.md#temporal-safety
+[data-race safety]: /docs/design/safety/terminology.md#data-race-safety
+
+**At this high level, this means Carbon's memory safety model will largely match
+Rust's.** The only minor deviation at this level from Rust is the use of
+run-time enforcement for initialization, where Carbon will more heavily leverage
+run-time techniques such as automatic initialization and dynamic "optional"
+semantics to improve the ergonomics in Carbon.
+
+However, Carbon and Rust will have substantial differences in the _details_ of
+how they approach both temporal and data-race safety.
+
+### Data races versus unsynchronized temporal safety
+
+Carbon has the option of distinguishing between two similar but importantly
+different classes of bugs: data races and unsynchronized temporal safety
+violations. Specifically, there evidence from security teams that the second of
+these has been a greater source of exploitation than first. As a consequence,
+Carbon has some flexibility while still being a [memory safe language] according
+to our definition:
+
+-   Carbon might choose to _not_ prevent data race bugs that are not
+    _themselves_ also temporal safety bugs, even though the data race may lead
+    to corruption and cause the program to later violate various other forms of
+    memory safety. So far, evidence has not shown this to be as significant and
+    prevalent source of _vulnerabilities_ as other forms of memory safety bugs.
+-   However, Carbon _must_ detect and prevent cases where a lack of
+    synchronization directly allows temporal safety bugs, such as use after
+    free.
+
+However, preventing these data race bugs remains _highly valuable_ for
+correctness, debugging, and achieving [fearless concurrency]. But it is one aspect
+where Carbon can in theory afford a compromise based on the current security information.
+
+[fearless concurrency]: https://doc.rust-lang.org/book/ch16-00-concurrency.html
+
+## Safe library ecosystem
+
+Carbon will need access to a memory safe library ecosystem. The industry is
+currently investing a massive amount of effort to build out a sufficient
+ecosystem of such software using Rust, and it is critical that Carbon does not
+impede, slow down, or require duplicating that ecosystem.
+
+**Carbon's strategy for safe and generally reusable libraries is to leverage
+Rust libraries through interop.** This is a major motivating reason for seamless
+and safe Rust interop. The Carbon project will work to avoid creating
+duplication between the growing Rust library ecosystem and any future Carbon
+library ecosystem. Carbon's ecosystem will be focused instead on libraries and
+functionality that would either be missing or only available in C++.
+
+## Build modes
+
+Where Carbon's safety mode governs the language rules applied to unsafe code,
+Carbon's build modes will change the _behavior_ of unsafe code.
+
+There are two primary build modes:
+
+-   **Release**: the primary build mode for programs in production, focuses on
+    giving the best possible performance with a practical baseline of safety.
+-   **Debug**: the primary build mode during development, focuses on
+    high-quality developer experience.
+
+The release build will include a baseline of hardening necessary to uphold the
+run-time enforcement components of our
+[memory safety model](#memory-safety-model) above. This means, for example, that
+bounds checking is enabled in the release build. There is [evidence] that the
+cost of these hardening steps is low. Following the specific guidance of our top
+priority for [performance _control_], Carbon will provide ways to write unsafe code
+that disables the run-time enforcement, enabling the control of any overhead incurred.
+
+[evidence]: https://chandlerc.blog/posts/2024/11/story-time-bounds-checking/
+[performance control]: /docs/project/goals.md#performance-critical-software
+
+The debug build will work to cause bugs and any detectable undefined behavior to
+have [fail-stop] behavior and even detailed diagnostics to enable better
+debugging. This mode will at least provide similar bug [detection] capabilities
+to [AddressSanitizer] and [MemorySanitizer].
+
+[fail-stop]: /docs/design/safety/terminology.md#fail-stop
+[detection]: /docs/design/safety/terminology.md#detecting
+[AddressSanitizer]: https://clang.llvm.org/docs/AddressSanitizer.html
+[MemorySanitizer]: https://clang.llvm.org/docs/MemorySanitizer.html
+
+Carbon will also have additional build modes to provide specific, narrow
+capabilities that cannot be reasonably combined into either of the above build
+modes. Each of these is expected to be an extension of either the debug or
+release build for that specific purpose. For example:
+
+-   Debug + [ThreadSanitizer]: detection of [data-races][data-race safety].
+-   Release + specialized [hardening]: some users can afford significant
+    run-time overhead in order to use additional hardening. Carbon will have
+    several specialized build modes in this space. Hardening techniques in this
+    space include [Control-Flow Integrity (CFI)][cfi] and hardware-accelerated
+    address sanitizing ([HWASAN]).
+
+[ThreadSanitizer]: https://clang.llvm.org/docs/ThreadSanitizer.html
+[hardening]: /docs/design/safety/terminology.md#hardening
+[cfi]: https://clang.llvm.org/docs/ControlFlowIntegrity.html
+[hwasan]:
+    https://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html
+
+Carbon will provide a way to disable the default hardening in release builds,
+but not in a supported way. It's use is expected to be strictly for benchmarking
+purposes.
+
+## References
+
+-   Proposal
+    [#196: Language-level safety strategy](https://github.com/carbon-language/carbon-lang/pull/196).
+-   Proposal
+    [#5914: Updating Carbon's safety strategy](https://github.com/carbon-language/carbon-lang/pull/5914).