| //! Extensions to [`Target`](super::Target) which add support for various |
| //! subsets of the GDB Remote Serial Protocol. |
| //! |
| //! ### Note: Missing Protocol Extensions |
| //! |
| //! `gdbstub`'s development is guided by the needs of its contributors, with |
| //! new features being added on an "as-needed" basis. |
| //! |
| //! If there's a GDB protocol extensions you're interested in that hasn't been |
| //! implemented in `gdbstub` yet, (e.g: remote filesystem access, tracepoint |
| //! support, etc...), consider opening an issue / filing a PR on the |
| //! [`gdbstub` GitHub repo](https://github.com/daniel5151/gdbstub/). |
| //! |
| //! Check out the [GDB Remote Configuration Docs](https://sourceware.org/gdb/onlinedocs/gdb/Remote-Configuration.html) |
| //! for a table of GDB commands + their corresponding Remote Serial Protocol |
| //! packets. |
| //! |
| //! ## How Protocol Extensions Work - Inlineable Dyn Extension Traits (IDETs) |
| //! |
| //! The GDB protocol is massive, and contains all sorts of optional |
| //! functionality. In the early versions of `gdbstub`, the `Target` trait |
| //! directly implemented a method for _every single protocol extension_. If this |
| //! trend continued, there would've been literally _hundreds_ of associated |
| //! methods - of which only a small subset were ever used at once! |
| //! |
| //! Aside from the cognitive complexity of having so many methods on a single |
| //! trait, this approach had numerous other drawbacks as well: |
| //! |
| //! - Implementations that did not implement all available protocol extensions |
| //! still had to "pay" for the unused packet parsing/handler code, resulting |
| //! in substantial code bloat, even on `no_std` platforms. |
| //! - `GdbStub`'s internal implementation needed to include a large number of |
| //! _runtime_ checks to deal with incorrectly implemented `Target`s. |
| //! - No way to enforce "mutually-dependent" trait methods at compile-time. |
| //! - e.g: When implementing hardware breakpoint extensions, targets |
| //! _must_ implement both the `add_breakpoint` and |
| //! `remove_breakpoints` methods. |
| //! - No way to enforce "mutually-exclusive" trait methods at compile-time. |
| //! - e.g: The `resume` method for single-threaded targets has a much |
| //! simpler API than for multi-threaded targets, but it would be |
| //! incorrect for a target to implement both. |
| //! |
| //! At first blush, it seems the the solution to all these issues is obvious: |
| //! simply tie each protocol extension to a `cargo` feature! And yes, while |
| //! this would indeed work, there would be several serious ergonomic drawbacks: |
| //! |
| //! - There would be _hundreds_ of individual feature flags that would need to |
| //! be toggled by end users. |
| //! - It would be functionally impossible to _test_ all permutations of |
| //! enabled/disabled cargo features. |
| //! - A single binary would need to rely on some [non-trivial `cargo`-fu](https://github.com/rust-lang/cargo/issues/674) |
| //! in order to have multiple `Target` implementations in a single binary. |
| //! |
| //! After much experimentation and iteration, `gdbstub` ended up taking a |
| //! radically different approach to implementing and enumerating available |
| //! features, using a technique called **Inlineable Dyn Extension Traits**. |
| //! |
| //! > _Author's note:_ As far as I can tell, this isn't a very well-known trick, |
| //! or at the very least, I've personally never encountered any library that |
| //! uses this sort of API. As such, I've decided to be a bit cheeky and give it |
| //! a name! At some point, I'm hoping to write a standalone blog post which |
| //! further explores this technique, comparing it to other/existing approaches, |
| //! and diving into details of the how the compiler optimizes this sort of code. |
| //! In fact, I've already got a [very rough github repo](https://github.com/daniel5151/optional-trait-methods) with some of my |
| //! findings. |
| //! |
| //! So, what are "Inlineable Dyn Extension Traits"? Well, let's break it down: |
| //! |
| //! - **Extension Traits** - A common [Rust convention](https://rust-lang.github.io/rfcs/0445-extension-trait-conventions.html#what-is-an-extension-trait) |
| //! to extend the functionality of a Trait, _without_ modifying the original |
| //! trait. |
| //! - **Dyn** - Alludes to the use of Dynamic Dispatch via [Trait Objects](https://doc.rust-lang.org/book/ch17-02-trait-objects.html). |
| //! - **Inlineable** - Alludes to the fact that this approach can be easily |
| //! inlined, making it a truly zero-cost abstraction. |
| //! |
| //! In a nutshell, Inlineable Dyn Extension Traits (or IDETs) are an abuse of |
| //! the Rust trait system + modern compiler optimizations to emulate zero-cost, |
| //! runtime-enumerable optional trait methods! |
| //! |
| //! #### Technical overview |
| //! |
| //! The basic principles behind Inlineable Dyn Extension Traits are best |
| //! explained though example: |
| //! |
| //! Lets say we want to add an optional protocol extension described by an |
| //! `ProtocolExt` trait to a base `Protocol` trait. How would we do that using |
| //! IDETs? |
| //! |
| //! - (library) Define a `trait ProtocolExt: Protocol { ... }` which includes |
| //! all the methods required by the protocol extension: |
| //! - _Note:_ Making `ProtocolExt` a subtrait of `Protocol` is not strictly |
| //! required, but it does enable transparently using `Protocol`'s |
| //! associated types as part of `ProtocolExt`'s method definitions. |
| //! |
| //! ```rust,ignore |
| //! /// `foo` and `bar` are mutually-dependent methods. |
| //! trait ProtocolExt: Protocol { |
| //! fn foo(&self); |
| //! // can use associated types in method signature! |
| //! fn bar(&mut self) -> Result<(), Self::Error>; |
| //! } |
| //! ``` |
| //! |
| //! - (library) "Associate" the `ProtocolExt` extension trait to the original |
| //! `Protocol` trait by adding a new `Protocol` method that "downcasts" `self` |
| //! into a `&mut dyn ProtocolExt`. |
| //! |
| //! ```rust,ignore |
| //! trait Protocol { |
| //! // ... other methods ... |
| //! |
| //! // Optional extension |
| //! #[inline(always)] |
| //! fn support_protocol_ext(&mut self) -> Option<ProtocolExtOps<Self>> { |
| //! // disabled by default |
| //! None |
| //! } |
| //! |
| //! // Mutually-exclusive extensions |
| //! fn get_ext_a_or_b(&mut self) -> EitherOrExt<Self::Arch, Self::Error>; |
| //! } |
| //! |
| //! // Using a typedef for readability |
| //! type ProtocolExtOps<T> = |
| //! &'a mut dyn ProtocolExt<Arch = <T as Protocol>::Arch, Error = <T as Protocol>::Error>; |
| //! |
| //! enum EitherOrExt<A, E> { |
| //! ProtocolExtA(&'a mut dyn ProtocolExtA<Arch = A, Error = E>), |
| //! ProtocolExtB(&'a mut dyn ProtocolExtB<Arch = A, Error = E>), |
| //! } |
| //! ``` |
| //! |
| //! - (user) Implements the `ProtocolExt` extension for their target (just like |
| //! a normal trait). |
| //! |
| //! ```rust,ignore |
| //! impl ProtocolExt for MyTarget { |
| //! fn foo(&self) { ... } |
| //! fn bar(&mut self) -> Result<(), Self::Error> { ... } |
| //! } |
| //! ``` |
| //! |
| //! - (user) Implements the base `Protocol` trait, overriding the |
| //! `support_protocol_ext` method to return `Some(self)`, which will |
| //! effectively "enable" the extension. |
| //! |
| //! ```rust,ignore |
| //! impl Protocol for MyTarget { |
| //! // Optional extension |
| //! #[inline(always)] |
| //! fn support_protocol_ext(&mut self) -> Option<ProtocolExtOps<Self>> { |
| //! Some(self) // will not compile unless `MyTarget` also implements `ProtocolExt` |
| //! } |
| //! |
| //! // Mutually-exclusive extensions |
| //! #[inline(always)] |
| //! fn get_ext_a_or_b(&mut self) -> EitherOrExt<Self::Arch, Self::Error> { |
| //! EitherOrExt::ProtocolExtA(self) |
| //! } |
| //! } |
| //! ``` |
| //! |
| //! > Please note the use of `#[inline(always)]` when enabling IDET methods. |
| //! While LLVM is usually smart enough to inline single-level IDETs (such as in |
| //! the example above), nested IDETs will often require a bit of "help" from the |
| //! `inline` directive to be correctly optimized. |
| //! |
| //! Now, here's where IDETs really shine: If the user didn't implement |
| //! `ProtocolExt`, but _did_ try to enable the feature by overriding |
| //! `support_protocol_ext` to return `Some(self)`, they'll get a compile-time |
| //! error that looks something like this: |
| //! |
| //! ```text |
| //! error[E0277]: the trait bound `MyTarget: ProtocolExt` is not satisfied |
| //! --> path/to/implementation.rs:44:14 |
| //! | |
| //! 44 | Some(self) |
| //! | ^^^^ the trait `ProtocolExt` is not implemented for `MyTarget` |
| //! | |
| //! = note: required for the cast to the object type `dyn ProtocolExt<Arch = ..., Error = ...>` |
| //! ``` |
| //! |
| //! The Rust compiler is preventing you from enabling a feature you haven't |
| //! implemented _at compile time!_ |
| //! |
| //! - (library) Is able to _query_ whether or not an extension is available, |
| //! _without_ having to actually invoke any method on the target! |
| //! |
| //! ```rust,ignore |
| //! fn execute_protocol(mut target: impl Target) { |
| //! match target.support_protocol_ext() { |
| //! Some(ops) => ops.foo(), |
| //! None => { /* fallback when not enabled */ } |
| //! } |
| //! } |
| //! ``` |
| //! |
| //! This is already pretty cool, but what's _even cooler_ is that if you take a |
| //! look at the generated assembly of a monomorphized `execute_protocol` method |
| //! (e.g: using godbolt.org), you'll find that the compiler is able to |
| //! efficiently inline and devirtualize _all_ the calls to |
| //! `support_protocol_ext` method, which in-turn allows the dead-code-eliminator |
| //! to work its magic, and remove the unused branches from the generated code! |
| //! i.e: If a target implemention didn't implement the `ProtocolExt` extension, |
| //! then that `match` statement in `execute_protocol` would simply turn into a |
| //! noop! |
| //! |
| //! If IDETs are something you're interested in, consider checking out |
| //! [daniel5151/optional-trait-methods](https://github.com/daniel5151/optional-trait-methods) |
| //! for some sample code that shows off the power of IDETs. It's not |
| //! particularly polished, but it does includes code snippets which can be |
| //! pasted into godbolt.org directly to confirm the optimizations described |
| //! above, and a brief writeup which compares / contrasts alternatives to IDETs. |
| //! |
| //! Long story short: Optimizing compilers really are magic! |
| //! |
| //! #### Summary: The Benefits of IDETs |
| //! |
| //! IDETs solve the numerous issues and shortcomings that arise from the |
| //! traditional single trait + "optional" methods approach: |
| //! |
| //! - **Compile-time enforcement of mutually-dependent methods** |
| //! - By grouping mutually-dependent methods behind a single extension trait |
| //! and marking them all as required methods, the Rust compiler is able to |
| //! catch missing mutually-dependent methods at compile time, with no need |
| //! for any runtime checks! |
| //! - **Compile-time enforcement of mutually-exclusive methods** |
| //! - By grouping mutually-exclusive methods behind two extension traits, and |
| //! wrapping those in an `enum`, the API is able to document |
| //! mutually-exclusive functions _at the type-level_, in-turn enabling the |
| //! library to omit any runtime checks! |
| //! - _Note:_ Strictly speaking, this isn't really compile time |
| //! "enforcement", as there's nothing stopping an "adversarial" |
| //! implementation from implementing both sets of methods, and then |
| //! "flipping" between the two at runtime. Nonetheless, it serves as a good |
| //! guardrail. |
| //! - **Enforce dead-code-elimination _without_ `cargo` feature flags** |
| //! - This is a really awesome trick: by wrapping code in an `if |
| //! target.support_protocol_ext().is_some()` block, it's possible to |
| //! specify _arbitrary_ blocks of code to be feature-dependent! |
| //! - This is used to great effect in `gdbstub` to optimize-out any packet |
| //! parsing / handler code for unimplemented protocol extensions. |
| |
| macro_rules! doc_comment { |
| ($x:expr, $($tt:tt)*) => { |
| #[doc = $x] |
| $($tt)* |
| }; |
| } |
| |
| macro_rules! define_ext { |
| ($extname:ident, $exttrait:ident) => { |
| doc_comment! { |
| concat!("See [`", stringify!($exttrait), "`](trait.", stringify!($exttrait), ".html)."), |
| pub type $extname<'a, T> = |
| &'a mut dyn $exttrait<Arch = <T as Target>::Arch, Error = <T as Target>::Error>; |
| } |
| }; |
| } |
| |
| pub mod auxv; |
| pub mod base; |
| pub mod breakpoints; |
| pub mod catch_syscalls; |
| pub mod exec_file; |
| pub mod extended_mode; |
| pub mod host_io; |
| pub mod libraries; |
| pub mod lldb_register_info_override; |
| pub mod memory_map; |
| pub mod monitor_cmd; |
| pub mod section_offsets; |
| pub mod target_description_xml_override; |
| pub mod thread_extra_info; |