mitigation enforcement

Author: Ariel Ben-Yehuda

text/0000-mitigation-enforcement.md

@@ -0,0 +1,347 @@
+- Feature Name: `mitigation_enforcement`
+- Start Date: 2025-09-13
+- RFC PR: [rust-lang/rfcs#3855](https://github.com/rust-lang/rfcs/pull/3855)
+- Rust Issue: None
+
+# Summary
+[summary]: #summary
+
+Introduce the concept of "mitigation enforcement", so that when compiling
+a crate with mitigations enabled (for example, `-C stack-protector`),
+a compilation error will happen if the produced artifact would contain Rust
+code without the same mitigations enabled.
+
+This in many cases would require use of `-Z build-std`, since the standard
+library only comes with a single set of enabled mitigations per target.
+
+Mitigation enforcement should be disableable by the end-user via a compiler
+flag.
+
+# Motivation
+[motivation]: #motivation
+
+Memory unsafety mitigations are important for reducing the chance that a vulnerability
+ends up being exploitable.
+
+While in Rust, memory unsafety is less of a concern than in C, mitigations are
+still important for several reasons:
+
+1. Some mitigations (for example, straight line speculation mitigation,
+   [`-Z harden-sls`]) mitigate the impact of Spectre-style speculative
+   execution vulnerabilities, that exist in Rust just as well as C.
+2. Many Rust programs also contain large C/C++ components, that can have
+   memory vulnerabilities.
+3. Many Rust programs use `unsafe`, that can introduce memory unsafety
+   and vulnerabilities.
+
+Mitigations are generally enabled by passing a flag to the compiler (for
+example, [`-Z harden-sls`] or [`-Z stack-protector`]). If the compilation
+process of a program is complex, it is very easy to end up accidentally
+not passing the flag to one of the constituent object files.
+
+This can have one of several consequences:
+
+1. In some cases (for example `-Z fixed-x18 -Z sanitizer=shadow-call-stack`),
+   the mitigation changes the ABI, and linking together code with different
+   mitigation settings leads to undefined behavior such as crashes even
+   in the absence of an attack. In these cases, the sanitizer should be a
+   [target modifier] rather than using this RFC.
+2. For "Spectre-type" mitigations (e.g. `harden-sls`), if there is some reachable
+   code in your address space without a retpoline, attackers can execute a
+   Spectre attack, even if there is 0 UB in your code.
+3. For "CFI-type" mitigations (e.g. kcfi), if there is reachable code in your
+   address space that does not have that sanitizer enabled, attackers can use it to
+   leverage an already-existing memory vulnerability into ROP execution, even
+   if the memory vulnerability is in a completely different part of the code than
+   the part that has the mitigation disabled
+4. For "local" mitigations (e.g. stack protector, or C's `-fwrapv` - which I don't think
+   Rust has), the mitigation protects the code when it is in the right place 
+   relative to the bug - a stack protector helps basically when it protects the buffer
+   that overflows, and it does not matter which other functions have a stack protector.
+
+To avoid these consequences, teams that write software with high security needs - for
+example, browsers and the Linux kernel - need to have a way to make sure that the
+programs they produce have the mitigations they want enabled.
+
+On the other hand, for teams that write software in a more messy environment, it
+can be hard to chase down all dependencies, and especially for "local" mitigations,
+being able to enable them on an object-by-object basis is the only thing that allows
+for the mitigations to actually be deployed. Especially important is progressive
+deployment - it's much easier to introduce mitigations 1 crate at a time than
+to introduce mitigations a whole program at a time, even if the end goal is
+to introduce the mitigations to the entire program.
+
+[target modifier]: https://github.com/rust-lang/rfcs/pull/3716
+[`-Z harden-sls`]: https://github.com/rust-lang/compiler-team/issues/869
+[`-Z stack-protector`]: https://github.com/rust-lang/rust/issues/114903
+[example by Alice Ryhl]: https://rust-lang.zulipchat.com/#narrow/channel/131828-t-compiler/topic/Target.20modifiers.20and.20-Cunsafe-allow-abi-mismatch/near/483871803
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+When you use a mitigation, such as `-C stack-protector=strong`, if one of your
+dependencies does not have that mitigation enabled, compilation will fail.
+
+> Error: your program uses the crate `foo`, that is not protected by
+> `-C stack-protector=strong`.
+>
+> Recompile that crate with the mitigation enabled, or use
+> `-C stack-protector=strong-noenforce` to allow creating an artifact
+> that has the mitigation only partially enabled.
+
+# Reference-level explanation
+[reference-level-explanation]: #reference-level-explanation
+
+Every flag value that enables a mitigation for which enforcement is desired is split
+into 2 separate values, "enforcing" and "non-enforcing" mode. The enforcing mode
+is the default, non-enforcing mode is constructed by adding `-noenforce` to the
+name of the value, for example `-C stack-protector=strong-noenforce` or
+`-C sanitizer=shadow-call-stack-noenforce`.
+
+> It is possible to bikeshed the exact naming scheme.
+
+> Every new mitigation would need to decide whether it adopts this scheme,
+> but mitigations are expected to adopt it.
+
+Every crate gets a metadata field that contains the set of mitigations it has enabled.
+
+When compiling a crate, if the current crate has a mitigation with enforcement
+turned on, and one of the dependencies does not have that mitigation turned
+on (whether enforcing or not), a compilation error results.
+
+If a mitigation has multiple "levels", a lower level at a child crate is compatible
+with a higher level at a base crate.
+
+The error happens independent of the target crate type (you get an error
+if you are building an rlib, not just the final executable).
+
+For example, with `-C stack-protector`, the compatibility table will be
+as follows:
+
+|    Base\Child    | none | none-noenforce | strong |   strong-noenforce   |  all  |    all-noenforce     |
+| ---------------- | ---- | -------------- | ------ | -------------------- | ----- | -------------------- |
+| none             |  OK  |      OK        | error  | OK - child noenforce | error | OK - child noenforce |
+| none-noenforce   |  OK  |      OK        | error  | OK - child noenforce | error | OK - child noenforce | 
+| strong           |  OK  |      OK        |   OK   |          OK          | error | OK - child noenforce |
+| strong-noenforce |  OK  |      OK        |   OK   |          OK          | error | OK - child noenforce |
+| all              |  OK  |      OK        |   OK   |          OK          |   OK  |           OK         |
+| all-noenforce    |  OK  |      OK        |   OK   |          OK          |   OK  |           OK         |
+
+If a program has multiple flags of the same kind, the last flag wins, so e.g.
+`-C stack-protector=strong-noenforce -C stack-protector=strong` is the same as
+`-C stack-protector=strong`.
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+The `-noenforce` syntax is ugly, and the
+`-C allow-partial-mitigations=stack-protector` syntax is either order-dependent
+or does not allow for easy appending.
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+## Syntax alternatives
+
+### -C stack-protector=none-noenforce
+
+The option `-C stack-protector=none-noenforce` is the same as
+`-C stack-protector=none`. I am not sure whether we should have both, but
+it feels that orthogonality is in favor of having both.
+
+### -C allow-partial-mitigations
+
+Instead of having `-C stack-protector=strong-noenforce`, we could have the
+syntax be `-C stack-protector=strong -C allow-partial-mitigations=stack-protector`.
+
+Some people feel that syntax is prettier. In that case, we have 2 options:
+
+#### Without order dependency
+
+This is the simplest to implement. With that,
+`-C stack-protector=strong -C allow-partial-mitigations=stack-protector -C stack-protector=strong`
+is the same as `-C stack-protector=strong -C allow-partial-mitigations=stack-protector`.
+
+This is unfortunate, because `-C stack-protector=strong -C allow-partial-mitigations=stack-protector` is
+a pretty good default for distributions to set. If a distribution sets that, and an application
+believes they are turning on enforcing stack protection by using `-C stack-protector=strong`,
+the application will not be getting enforcement due to the distribution setting
+`-C allow-partial-mitigations=stack-protector`.
+
+On the other hand, maybe there is not actually desire to add
+`-C stack-protector=strong -C allow-partial-mitigations=stack-protector` as a default?
+
+Maybe it is actually possible to ship a `-C stack-protector=strong` standard library and
+add a `-C stack-protector=strong` default, since the enforcement check only works
+"towards roots"?
+
+#### With order dependency
+
+With a small amount of implementation effort, we could have `-C stack-protector=strong` reset the
+`-C allow-partial-mitigations=stack-protector` state, so that
+`-C stack-protector=strong -C allow-partial-mitigations=stack-protector -C stack-protector=strong`
+is equivalent to `-C stack-protector=strong`.
+
+This would work quite well, but I am not sure that rustc wants to have order between different
+kinds of CLI arguments.
+
+## Interaction with `-C unsafe-allow-abi-mismatch` / `-C pretend-mitigation-enabled`
+
+The proposed rules do not interact with `-C unsafe-allow-abi-mismatch` at all, so if
+you have a "sanitizer runtime" crate that is compiled with the following options:
+
+> -C no-fixed-x18 -C sanitizer=shadow-call-stack=off -C unsafe-allow-abi-mismatch=fixed-x18 -C unsafe-allow-abi-mismatch=shadow-call-stack
+
+Then dependencies will need to use it via `-C sanitizer=shadow-call-stack-noenforce`
+rather than `-C sanitizer=shadow-call-stack`, otherwise they will get an error.
+
+As far as I can need, there is no current demand for that sort of sanitizer runtime,
+but if that is desired, it might be a good idea to add a
+`-C pretend-mitigation-enabled=shadow-call-stack`, and possibly to make
+`-C unsafe-allow-abi-mismatch` do that for crates that are target modifiers.
+
+## Defaults
+
+We want that the most obvious way to enable mitigations (e.g.
+`-C stack-protector=strong` or `-C sanitizer=shadow-call-stack`) to turn on
+enforcement, since that will set people up to a pit of success where mitigations
+are enabled throughout.
+
+However, we do want an easy way for distribution owners (for example,
+Ubuntu) to turn on mitigations in a non-enforcing way, as is done
+today e.g. [by Ubuntu with `-fstack-protector-strong`]. Distributions can't easily
+add a new mitigation in an enforcing way, as that will cause widespread
+breakage, but they can fairly easily turn a mitigation on in a non-enforcing way.
+
+We do want the combination of defaults to combine in a nice way - if the
+distributioon sets `-C stack-protector=strong-noenforce`, and the user adds
+`-C stack-protector=strong`, we want the result to be stack-protector set
+to strong and enforcing.
+
+On the other hand, maybe there is not actually desire to add
+`-C stack-protector=strong -C allow-partial-mitigations=stack-protector` as a default,
+which would make this less interesting?
+
+Maybe it is actually possible to ship a `-C stack-protector=strong` standard library and
+add a `-C stack-protector=strong` default, since the enforcement check only works
+"towards roots"?
+
+[by Ubuntu with `-fstack-protector-strong`]: https://wiki.ubuntu.com/ToolChain/CompilerFlags
+
+## The standard library
+
+One big place where it's very easy to end up with mixed mitigations is the
+standard library. The standard library comes compiled with just a single
+set of mitigations enabled (as of Rust 1.88: none), and without `-Z build-std`,
+it is only possible to use the mitigation settings in the shipped standard
+library.
+
+If we find out that some mitigations have a positive cost-benefit ratio
+for the standard library (probably at least [`-Z stack-protector`]), we
+probably want to ship a standard library supporting them by default, but
+in a way that still allows people to compile code without mitigations,
+if that fulfills their cost/benefit ratios better.
+
+## Why not target modifiers?
+
+The [target modifier] feature provides a similar goal of preventing mismatches in compiler
+settings.
+
+There are several issues with using target modifiers for mitigations:
+
+### The name unsafe-allow-abi-mismatch
+
+The name of the flag that allows mixing target modifiers, `-C unsafe-allow-abi-mismatch`,
+does not make sense for cases that are not "unsafe ABI mismatches". It also uses the
+word "unsafe", which we prefer not to use except in cases that can result in actual
+unsoundness.
+
+### The behavior of unsafe-allow-abi-mismatch
+
+The behavior of `-C unsafe-allow-abi-mismatch` is also not ideal for mitigations.
+
+The flag marks a crate as basically having a "wildcard target modifier", which allows it
+to compile with crates with any value of the target modifier.
+
+This is quite good for the original use case - it's an "I know what I am doing" flag
+that allows "runtime" crates to be mixed in even if they subtly play with the
+ABI rules - for example, in a kernel, where floating point execution is mostly
+forbidden, there are a few compilation units using real floats. To call into them, first
+you call some special functions that make the floating point registers usable, and then
+you call into the CU safely by not having any floats in the signature of the function on
+the boundary ([example by Alice Ryhl]).
+
+However, for mitigations, the expected case for disabling mitigations is less people
+knowing what they are doing, and more people that don't agree with the performance/security
+tradeoff they bring. In that case, we should allow the executable-writer to be aware
+of the tradeoff being made, rather than letting libraries in the middle decide it
+for them.
+
+## Why not an external tool?
+
+This is somewhat hard to do with an external tool, since there is
+no way of looking at a binary and telling what mitigations its components
+have (for example [`hardening-check(1)`], exists, but its check for
+stack smashing protection only checks that at least 1 function has stack
+cookies, rather than checking that every interesting function has it
+enabled).
+
+[`hardening-check(1)`]: https://manpages.debian.org/testing/devscripts/hardening-check.1.en.html
+
+## .note.gnu.property
+
+The `.note.gnu.property` field contains a number of properties
+(for example, [`GNU_PROPERTY_AARCH64_FEATURE_1_BTI`]) that are used to indicate
+that the compiled code contains certain mitigations, for example BTI
+(`-Zbranch-protection=bti`).
+
+When linking multiple objects, the linker sets the resulting property to be the
+logical AND of the properties of the constituent objects.
+
+For protections such as BTI, the mitigation can only be turned on if all code
+within the compiled binary supports it - if one of the object files doesn't,
+the loader has to leave the mitigation turned off entirely. The ELF loader uses
+the value of the property within the loaded executable to decide whether
+to turn on the mitigation.
+
+If it could be arranged, using `.note.gnu.property` could allow mitigation tracking
+to propagate across languages - with the final compilation step intentionally erroring
+out if the property is not enabled. However, this is also a disadvantage - adding a
+new property to `.note.gnu.property` requires cooperation from the target owners.
+
+Therefore, it might be useful as a future step with cooperation from the target owners,
+but is not good if we want to be able to add new enforced mitigations without requiring
+cooperation from all platforms.
+
+[`GNU_PROPERTY_AARCH64_FEATURE_1_BTI`]: https://docs.rs/object/0.37/object/elf/constant.GNU_PROPERTY_AARCH64_FEATURE_1_BTI.html
+
+# Prior art
+[prior-art]: #prior-art
+
+## The panic strategy
+
+The Rust compiler already *has* infrastructure to detect flag mismatches: the
+flags `-Cpanic` and `-Zpanic-in-drop`. The prebuilt stdlib comes with different
+pieces depending on which strategy is used, although panic landing flags are
+not entirely removed when using `-Cpanic=abort`, as only part of the prebuilt
+stdlib is switched out.
+
+## Target modifiers
+
+## .note.gnu.property
+
+The `.note.gnu.property` section discussed previously is an example of C code
+detecting mismatches of a flag at link time.
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+# Future possibilities
+[future-possibilities]: #future-possibilities
+
+A possible future extension could be to provide a mechanism to enforce
+mitigations across C code and Rust code. This would be an interesting
+extension, but it would require cross-language effort that will
+take a long period of time to finish. Similarly, another possible
+future extension could be to catch mitigation mismatches when using
+dynamic linking.