docs/devel/rust.rst

.. |msrv| replace:: 1.63.0

Rust in QEMU
============

Rust in QEMU is a project to enable using the Rust programming language
to add new functionality to QEMU.

Right now, the focus is on making it possible to write devices that inherit
from ``SysBusDevice`` in `*safe*`__ Rust.  Later, it may become possible
to write other kinds of devices (e.g. PCI devices that can do DMA),
complete boards, or backends (e.g. block device formats).

__ https://doc.rust-lang.org/nomicon/meet-safe-and-unsafe.html

Building the Rust in QEMU code
------------------------------

The Rust in QEMU code is included in the emulators via Meson.  Meson
invokes rustc directly, building static libraries that are then linked
together with the C code.  This is completely automatic when you run
``make`` or ``ninja``.

However, QEMU's build system also tries to be easy to use for people who
are accustomed to the more "normal" Cargo-based development workflow.
In particular:

* the set of warnings and lints that are used to build QEMU always
  comes from the ``rust/Cargo.toml`` workspace file

* it is also possible to use ``cargo`` for common Rust-specific coding
  tasks, in particular to invoke ``clippy``, ``rustfmt`` and ``rustdoc``.

To this end, QEMU includes a ``build.rs`` build script that picks up
generated sources from QEMU's build directory and puts it in Cargo's
output directory (typically ``rust/target/``).  A vanilla invocation
of Cargo will complain that it cannot find the generated sources,
which can be fixed in different ways:

* by using Makefile targets, provided by Meson, that run ``clippy`` or
  ``rustdoc``:

    make clippy
    make rustdoc

A target for ``rustfmt`` is also declared in ``rust/meson.build``:

    make rustfmt

* by invoking ``cargo`` through the Meson `development environment`__
  feature::

    pyvenv/bin/meson devenv -w ../rust cargo clippy --tests
    pyvenv/bin/meson devenv -w ../rust cargo fmt

  If you are going to use ``cargo`` repeatedly, ``pyvenv/bin/meson devenv``
  will enter a shell where commands like ``cargo fmt`` just work.

__ https://mesonbuild.com/Commands.html#devenv

* by pointing the ``MESON_BUILD_ROOT`` to the top of your QEMU build
  tree.  This third method is useful if you are using ``rust-analyzer``;
  you can set the environment variable through the
  ``rust-analyzer.cargo.extraEnv`` setting.

As shown above, you can use the ``--tests`` option as usual to operate on test
code.  Note however that you cannot *build* or run tests via ``cargo``, because
they need support C code from QEMU that Cargo does not know about.  Tests can
be run via ``meson test`` or ``make``::

   make check-rust

Note that doctests require all ``.o`` files from the build to be available.

Supported tools
'''''''''''''''

QEMU supports rustc version 1.77.0 and newer.  Notably, the following features
are missing:

* inline const expression (stable in 1.79.0), currently worked around with
  associated constants in the ``FnCall`` trait.

* associated constants have to be explicitly marked ``'static`` (`changed in
  1.81.0`__)

* ``&raw`` (stable in 1.82.0).  Use ``addr_of!`` and ``addr_of_mut!`` instead,
  though hopefully the need for raw pointers will go down over time.

* ``new_uninit`` (stable in 1.82.0).  This is used internally by the ``pinned_init``
  crate, which is planned for inclusion in QEMU, but it can be easily patched
  out.

* referencing statics in constants (stable in 1.83.0).  For now use a const
  function; this is an important limitation for QEMU's migration stream
  architecture (VMState).  Right now, VMState lacks type safety because
  it is hard to place the ``VMStateField`` definitions in traits.

* NUL-terminated file names with ``#[track_caller]`` are scheduled for
  inclusion as ``#![feature(location_file_nul)]``, but it will be a while
  before QEMU can use them.  For now, there is special code in
  ``util/error.c`` to support non-NUL-terminated file names.

* associated const equality would be nice to have for some users of
  ``callbacks::FnCall``, but is still experimental.  ``ASSERT_IS_SOME``
  replaces it.

__ https://github.com/rust-lang/rust/pull/125258

QEMU also supports version 0.60.x of bindgen, which is missing option
``--generate-cstr``.  This option requires version 0.66.x and will
be adopted as soon as supporting these older versions is not necessary
anymore.

Writing Rust code in QEMU
-------------------------

QEMU includes four crates:

* ``qemu_api`` for bindings to C code and useful functionality

* ``qemu_api_macros`` defines several procedural macros that are useful when
  writing C code

* ``pl011`` (under ``rust/hw/char/pl011``) and ``hpet`` (under ``rust/hw/timer/hpet``)
  are sample devices that demonstrate ``qemu_api`` and ``qemu_api_macros``, and are
  used to further develop them.  These two crates are functional\ [#issues]_ replacements
  for the ``hw/char/pl011.c`` and ``hw/timer/hpet.c`` files.

.. [#issues] The ``pl011`` crate is synchronized with ``hw/char/pl011.c``
   as of commit 3e0f118f82.  The ``hpet`` crate is synchronized as of
   commit 1433e38cc8.  Both are lacking tracing functionality.

This section explains how to work with them.

Status
''''''

Modules of ``qemu_api`` can be defined as:

- *complete*: ready for use in new devices; if applicable, the API supports the
  full functionality available in C

- *stable*: ready for production use, the API is safe and should not undergo
  major changes

- *proof of concept*: the API is subject to change but allows working with safe
  Rust

- *initial*: the API is in its initial stages; it requires large amount of
  unsafe code; it might have soundness or type-safety issues

The status of the modules is as follows:

================ ======================
module           status
================ ======================
``assertions``   stable
``bitops``       complete
``callbacks``    complete
``cell``         stable
``errno``        complete
``error``        stable
``irq``          complete
``memory``       stable
``module``       complete
``qdev``         stable
``qom``          stable
``sysbus``       stable
``timer``        stable
``vmstate``      proof of concept
``zeroable``     stable
================ ======================

.. note::
  API stability is not a promise, if anything because the C APIs are not a stable
  interface either.  Also, ``unsafe`` interfaces may be replaced by safe interfaces
  later.

Naming convention
'''''''''''''''''

C function names usually are prefixed according to the data type that they
apply to, for example ``timer_mod`` or ``sysbus_connect_irq``.  Furthermore,
both function and structs sometimes have a ``qemu_`` or ``QEMU`` prefix.
Generally speaking, these are all removed in the corresponding Rust functions:
``QEMUTimer`` becomes ``timer::Timer``, ``timer_mod`` becomes ``Timer::modify``,
``sysbus_connect_irq`` becomes ``SysBusDeviceMethods::connect_irq``.

Sometimes however a name appears multiple times in the QOM class hierarchy,
and the only difference is in the prefix.  An example is ``qdev_realize`` and
``sysbus_realize``.  In such cases, whenever a name is not unique in
the hierarchy, always add the prefix to the classes that are lower in
the hierarchy; for the top class, decide on a case by case basis.

For example:

========================== =========================================
``device_cold_reset()``    ``DeviceMethods::cold_reset()``
``pci_device_reset()``     ``PciDeviceMethods::pci_device_reset()``
``pci_bridge_reset()``     ``PciBridgeMethods::pci_bridge_reset()``
========================== =========================================

Here, the name is not exactly the same, but nevertheless ``PciDeviceMethods``
adds the prefix to avoid confusion, because the functionality of
``device_cold_reset()`` and ``pci_device_reset()`` is subtly different.

In this case, however, no prefix is needed:

========================== =========================================
``device_realize()``       ``DeviceMethods::realize()``
``sysbus_realize()``       ``SysbusDeviceMethods::sysbus_realize()``
``pci_realize()``          ``PciDeviceMethods::pci_realize()``
========================== =========================================

Here, the lower classes do not add any functionality, and mostly
provide extra compile-time checking; the basic *realize* functionality
is the same for all devices.  Therefore, ``DeviceMethods`` does not
add the prefix.

Whenever a name is unique in the hierarchy, instead, you should
always remove the class name prefix.

Common pitfalls
'''''''''''''''

Rust has very strict rules with respect to how you get an exclusive (``&mut``)
reference; failure to respect those rules is a source of undefined behavior.
In particular, even if a value is loaded from a raw mutable pointer (``*mut``),
it *cannot* be casted to ``&mut`` unless the value was stored to the ``*mut``
from a mutable reference.  Furthermore, it is undefined behavior if any
shared reference was created between the store to the ``*mut`` and the load::

    let mut p: u32 = 42;
    let p_mut = &mut p;                              // 1
    let p_raw = p_mut as *mut u32;                   // 2

    // p_raw keeps the mutable reference "alive"

    let p_shared = &p;                               // 3
    println!("access from &u32: {}", *p_shared);

    // Bring back the mutable reference, its lifetime overlaps
    // with that of a shared reference.
    let p_mut = unsafe { &mut *p_raw };              // 4
    println!("access from &mut 32: {}", *p_mut);

    println!("access from &u32: {}", *p_shared);     // 5

These rules can be tested with `MIRI`__, for example.

__ https://github.com/rust-lang/miri

Almost all Rust code in QEMU will involve QOM objects, and pointers to these
objects are *shared*, for example because they are part of the QOM composition
tree.  This creates exactly the above scenario:

1. a QOM object is created

2. a ``*mut`` is created, for example as the opaque value for a ``MemoryRegion``

3. the QOM object is placed in the composition tree

4. a memory access dereferences the opaque value to a ``&mut``

5. but the shared reference is still present in the composition tree

Because of this, QOM objects should almost always use ``&self`` instead
of ``&mut self``; access to internal fields must use *interior mutability*
to go from a shared reference to a ``&mut``.

Whenever C code provides you with an opaque ``void *``, avoid converting it
to a Rust mutable reference, and use a shared reference instead.  The
``qemu_api::cell`` module provides wrappers that can be used to tell the
Rust compiler about interior mutability, and optionally to enforce locking
rules for the "Big QEMU Lock".  In the future, similar cell types might
also be provided for ``AioContext``-based locking as well.

In particular, device code will usually rely on the ``BqlRefCell`` and
``BqlCell`` type to ensure that data is accessed correctly under the
"Big QEMU Lock".  These cell types are also known to the ``vmstate``
crate, which is able to "look inside" them when building an in-memory
representation of a ``struct``'s layout.  Note that the same is not true
of a ``RefCell`` or ``Mutex``.

Bindings code instead will usually use the ``Opaque`` type, which hides
the contents of the underlying struct and can be easily converted to
a raw pointer, for use in calls to C functions.  It can be used for
example as follows::

    #[repr(transparent)]
    #[derive(Debug, qemu_api_macros::Wrapper)]
    pub struct Object(Opaque<bindings::Object>);

where the special ``derive`` macro provides useful methods such as
``from_raw``, ``as_ptr`, ``as_mut_ptr`` and ``raw_get``.  The bindings will
then manually check for the big QEMU lock with assertions, which allows
the wrapper to be declared thread-safe::

    unsafe impl Send for Object {}
    unsafe impl Sync for Object {}

Writing bindings to C code
''''''''''''''''''''''''''

Here are some things to keep in mind when working on the ``qemu_api`` crate.

**Look at existing code**
  Very often, similar idioms in C code correspond to similar tricks in
  Rust bindings.  If the C code uses ``offsetof``, look at qdev properties
  or ``vmstate``.  If the C code has a complex const struct, look at
  ``MemoryRegion``.  Reuse existing patterns for handling lifetimes;
  for example use ``&T`` for QOM objects that do not need a reference
  count (including those that can be embedded in other objects) and
  ``Owned<T>`` for those that need it.

**Use the type system**
  Bindings often will need access information that is specific to a type
  (either a builtin one or a user-defined one) in order to pass it to C
  functions.  Put them in a trait and access it through generic parameters.
  The ``vmstate`` module has examples of how to retrieve type information
  for the fields of a Rust ``struct``.

**Prefer unsafe traits to unsafe functions**
  Unsafe traits are much easier to prove correct than unsafe functions.
  They are an excellent place to store metadata that can later be accessed
  by generic functions.  C code usually places metadata in global variables;
  in Rust, they can be stored in traits and then turned into ``static``
  variables.  Often, unsafe traits can be generated by procedural macros.

**Document limitations due to old Rust versions**
  If you need to settle for an inferior solution because of the currently
  supported set of Rust versions, document it in the source and in this
  file.  This ensures that it can be fixed when the minimum supported
  version is bumped.

**Keep locking in mind**.
  When marking a type ``Sync``, be careful of whether it needs the big
  QEMU lock.  Use ``BqlCell`` and ``BqlRefCell`` for interior data,
  or assert ``bql_locked()``.

**Don't be afraid of complexity, but document and isolate it**
  It's okay to be tricky; device code is written more often than bindings
  code and it's important that it is idiomatic.  However, you should strive
  to isolate any tricks in a place (for example a ``struct``, a trait
  or a macro) where it can be documented and tested.  If needed, include
  toy versions of the code in the documentation.

Writing procedural macros
'''''''''''''''''''''''''

By conventions, procedural macros are split in two functions, one
returning ``Result<proc_macro2::TokenStream, MacroError>`` with the body of
the procedural macro, and the second returning ``proc_macro::TokenStream``
which is the actual procedural macro.  The former's name is the same as
the latter with the ``_or_error`` suffix.  The code for the latter is more
or less fixed; it follows the following template, which is fixed apart
from the type after ``as`` in the invocation of ``parse_macro_input!``::

    #[proc_macro_derive(Object)]
    pub fn derive_object(input: TokenStream) -> TokenStream {
        let input = parse_macro_input!(input as DeriveInput);
        let expanded = derive_object_or_error(input).unwrap_or_else(Into::into);

        TokenStream::from(expanded)
    }

The ``qemu_api_macros`` crate has utility functions to examine a
``DeriveInput`` and perform common checks (e.g. looking for a struct
with named fields).  These functions return ``Result<..., MacroError>``
and can be used easily in the procedural macro function::

    fn derive_object_or_error(input: DeriveInput) ->
        Result<proc_macro2::TokenStream, MacroError>
    {
        is_c_repr(&input, "#[derive(Object)]")?;

        let name = &input.ident;
        let parent = &get_fields(&input, "#[derive(Object)]")?[0].ident;
        ...
    }

Use procedural macros with care.  They are mostly useful for two purposes:

* Performing consistency checks; for example ``#[derive(Object)]`` checks
  that the structure has ``#[repr[C])`` and that the type of the first field
  is consistent with the ``ObjectType`` declaration.

* Extracting information from Rust source code into traits, typically based
  on types and attributes.  For example, ``#[derive(TryInto)]`` builds an
  implementation of ``TryFrom``, and it uses the ``#[repr(...)]`` attribute
  as the ``TryFrom`` source and error types.

Procedural macros can be hard to debug and test; if the code generation
exceeds a few lines of code, it may be worthwhile to delegate work to
"regular" declarative (``macro_rules!``) macros and write unit tests for
those instead.


Coding style
''''''''''''

Code should pass clippy and be formatted with rustfmt.

Right now, only the nightly version of ``rustfmt`` is supported.  This
might change in the future.  While CI checks for correct formatting via
``cargo fmt --check``, maintainers can fix this for you when applying patches.

It is expected that ``qemu_api`` provides full ``rustdoc`` documentation for
bindings that are in their final shape or close.

Adding dependencies
-------------------

Generally, the set of dependent crates is kept small.  Think twice before
adding a new external crate, especially if it comes with a large set of
dependencies itself.  Sometimes QEMU only needs a small subset of the
functionality; see for example QEMU's ``assertions`` module.

On top of this recommendation, adding external crates to QEMU is a
slightly complicated process, mostly due to the need to teach Meson how
to build them.  While Meson has initial support for parsing ``Cargo.lock``
files, it is still highly experimental and is therefore not used.

Therefore, external crates must be added as subprojects for Meson to
learn how to build them, as well as to the relevant ``Cargo.toml`` files.
The versions specified in ``rust/Cargo.lock`` must be the same as the
subprojects; note that the ``rust/`` directory forms a Cargo `workspace`__,
and therefore there is a single lock file for the whole build.

__ https://doc.rust-lang.org/cargo/reference/workspaces.html#virtual-workspace

Choose a version of the crate that works with QEMU's minimum supported
Rust version (|msrv|).

Second, a new ``wrap`` file must be added to teach Meson how to download the
crate.  The wrap file must be named ``NAME-SEMVER-rs.wrap``, where ``NAME``
is the name of the crate and ``SEMVER`` is the version up to and including the
first non-zero number.  For example, a crate with version ``0.2.3`` will use
``0.2`` for its ``SEMVER``, while a crate with version ``1.0.84`` will use ``1``.

Third, the Meson rules to build the crate must be added at
``subprojects/NAME-SEMVER-rs/meson.build``.  Generally this includes:

* ``subproject`` and ``dependency`` lines for all dependent crates

* a ``static_library`` or ``rust.proc_macro`` line to perform the actual build

* ``declare_dependency`` and a ``meson.override_dependency`` lines to expose
  the result to QEMU and to other subprojects

Remember to add ``native: true`` to ``dependency``, ``static_library`` and
``meson.override_dependency`` for dependencies of procedural macros.
If a crate is needed in both procedural macros and QEMU binaries, everything
apart from ``subproject`` must be duplicated to build both native and
non-native versions of the crate.

It's important to specify the right compiler options.  These include:

* the language edition (which can be found in the ``Cargo.toml`` file)

* the ``--cfg`` (which have to be "reverse engineered" from the ``build.rs``
  file of the crate).

* usually, a ``--cap-lints allow`` argument to hide warnings from rustc
  or clippy.

After every change to the ``meson.build`` file you have to update the patched
version with ``meson subprojects update --reset ``NAME-SEMVER-rs``.  This might
be automated in the future.

Also, after every change to the ``meson.build`` file it is strongly suggested to
do a dummy change to the ``.wrap`` file (for example adding a comment like
``# version 2``), which will help Meson notice that the subproject is out of date.

As a last step, add the new subproject to ``scripts/archive-source.sh``,
``scripts/make-release`` and ``subprojects/.gitignore``.