You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

229 lines
9.5 KiB

/**
* \mainpage Wasmtime C API
*
* This documentation is an overview and API reference for the C API of
* Wasmtime. The C API is spread between three different header files:
*
* * \ref wasmtime.h
* * \ref wasi.h
* * \ref wasm.h
*
* The \ref wasmtime.h header file includes all the other header files and is
* the main header file you'll likely be using. The \ref wasm.h header file
* comes directly from the
* [WebAssembly/wasm-c-api](https://github.com/WebAssembly/wasm-c-api)
* repository, and at this time the upstream header file does not have
* documentation so Wasmtime provides documentation here. It should be noted
* some semantics may be Wasmtime-specific and may not be portable to other
* engines.
*
* ## Installing the C API
*
* To install the C API from precompiled binaries you can download the
* appropriate binary from the [releases page of
* Wasmtime](https://github.com/bytecodealliance/wasmtime/releases). Artifacts
* for the C API all end in "-c-api" for the filename.
*
* Each archive contains an `include` directory with necessary headers, as well
* as a `lib` directory with both a static archive and a dynamic library of
* Wasmtime. You can link to either of them as you see fit.
*
* ## Installing the C API through CMake
*
* CMake can be used to make the process of linking and compiling easier. An
* example of this if you have wasmtime as a git submodule at
* `third_party/wasmtime`:
* ```
* add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third_party/wasmtime/crates/c-api
* ${CMAKE_CURRENT_BINARY_DIR}/wasmtime)
* ...
* target_include_directories(YourProject PUBLIC wasmtime)
* target_link_libraries(YourProject PUBLIC wasmtime)
* ```
* `BUILD_SHARED_LIBS` is provided as a define if you would like to build a
* shared library instead. You must distribute the appropriate shared library
* for your platform if you do this.
*
* ## Linking against the C API
*
* You'll want to arrange the `include` directory of the C API to be in your
* compiler's header path (e.g. the `-I` flag). If you're compiling for Windows
* and you're using the static library then you'll also need to pass
* `-DWASM_API_EXTERN=` and `-DWASI_API_EXTERN=` to disable dllimport.
*
* Your final artifact can then be linked with `-lwasmtime`. If you're linking
* against the static library you may need to pass other system libraries
* depending on your platform:
*
* * Linux - `-lpthread -ldl -lm`
* * macOS - no extra flags needed
* * Windows - `ws2_32.lib advapi32.lib userenv.lib ntdll.lib shell32.lib ole32.lib bcrypt.lib`
*
* ## Building from Source
*
* The C API is located in the
* [`crates/c-api`](https://github.com/bytecodealliance/wasmtime/tree/main/crates/c-api)
* directory of the [Wasmtime
* repository](https://github.com/bytecodealliance/wasmtime). To build from
* source you'll need a Rust compiler and a checkout of the `wasmtime` project.
* Afterwards you can execute:
*
* ```
* $ cargo build --release -p wasmtime-c-api
* ```
*
* This will place the final artifacts in `target/release`, with names depending
* on what platform you're compiling for.
*
* ## Other resources
*
* Some other handy resources you might find useful when exploring the C API
* documentation are:
*
* * [Rust `wasmtime` crate
* documentation](https://bytecodealliance.github.io/wasmtime/api/wasmtime/) -
* although this documentation is for Rust and not C, you'll find that many
* functions mirror one another and there may be extra documentation in Rust
* you find helpful. If you find yourself having to frequently do this,
* though, please feel free to [file an
* issue](https://github.com/bytecodealliance/wasmtime/issues/new).
*
* * [C embedding
* examples](https://bytecodealliance.github.io/wasmtime/examples-c-embed.html)
* are available online and are tested from the Wasmtime repository itself.
*
* * [Contribution documentation for
* Wasmtime](https://bytecodealliance.github.io/wasmtime/contributing.html) in
* case you're interested in helping out!
*/
/**
* \file wasmtime.h
*
* \brief Wasmtime's C API
*
* This file is the central inclusion point for Wasmtime's C API. There are a
* number of sub-header files but this file includes them all. The C API is
* based on \ref wasm.h but there are many Wasmtime-specific APIs which are
* tailored to Wasmtime's implementation.
*
* The #wasm_config_t and #wasm_engine_t types are used from \ref wasm.h.
* Additionally all type-level information (like #wasm_functype_t) is also
* used from \ref wasm.h. Otherwise, though, all wasm objects (like
* #wasmtime_store_t or #wasmtime_func_t) are used from this header file.
*
* ### Thread Safety
*
* The multithreading story of the C API very closely follows the
* multithreading story of the Rust API for Wasmtime. All objects are safe to
* send to other threads so long as user-specific data is also safe to send to
* other threads. Functions are safe to call from any thread but some functions
* cannot be called concurrently. For example, functions which correspond to
* `&T` in Rust can be called concurrently with any other methods that take
* `&T`. Functions that take `&mut T` in Rust, however, cannot be called
* concurrently with any other function (but can still be invoked on any
* thread).
*
* This generally equates to mutation of internal state. Functions which don't
* mutate anything, such as learning type information through
* #wasmtime_func_type, can be called concurrently. Functions which do require
* mutation, for example #wasmtime_func_call, cannot be called concurrently.
* This is conveyed in the C API with either `const wasmtime_context_t*`
* (concurrency is ok as it's read-only) or `wasmtime_context_t*` (concurrency
* is not ok, mutation may happen).
*
* When in doubt assume that functions cannot be called concurrently with
* aliasing objects.
*
* ### Aliasing
*
* The C API for Wasmtime is intended to be a relatively thin layer over the
* Rust API for Wasmtime. Rust has much more strict rules about aliasing than C
* does, and the Rust API for Wasmtime is designed around these rules to be
* used safely. These same rules must be upheld when using the C API of
* Wasmtime.
*
* The main consequence of this is that the #wasmtime_context_t pointer into
* the #wasmtime_store_t must be carefully used. Since the context is an
* internal pointer into the store it must be used carefully to ensure you're
* not doing something that Rust would otherwise forbid at compile time. A
* #wasmtime_context_t can only be used when you would otherwise have been
* provided access to it. For example in a host function created with
* #wasmtime_func_new you can use #wasmtime_context_t in the host function
* callback. This is because an argument, a #wasmtime_caller_t, provides access
* to #wasmtime_context_t. On the other hand a destructor passed to
* #wasmtime_externref_new, however, cannot use a #wasmtime_context_t because
* it was not provided access to one. Doing so may lead to memory unsafety.
*
* ### Stores
*
* A foundational construct in this API is the #wasmtime_store_t. A store is a
* collection of host-provided objects and instantiated wasm modules. Stores are
* often treated as a "single unit" and items within a store are all allowed to
* reference one another. References across stores cannot currently be created.
* For example you cannot pass a function from one store into another store.
*
* A store is not intended to be a global long-lived object. Stores provide no
* means of internal garbage collections of wasm objects (such as instances),
* meaning that no memory from a store will be deallocated until you call
* #wasmtime_store_delete. If you're working with a web server, for example,
* then it's recommended to think of a store as a "one per request" sort of
* construct. Globally you'd have one #wasm_engine_t and a cache of
* #wasmtime_module_t instances compiled into that engine. Each request would
* create a new #wasmtime_store_t and then instantiate a #wasmtime_module_t
* into the store. This process of creating a store and instantiating a module
* is expected to be quite fast. When the request is finished you'd delete the
* #wasmtime_store_t keeping memory usage reasonable for the lifetime of the
* server.
*/
#ifndef WASMTIME_API_H
#define WASMTIME_API_H
#include <wasi.h>
#include <wasmtime/config.h>
Delete historical interruptable support in Wasmtime (#3925) * Delete historical interruptable support in Wasmtime This commit removes the `Config::interruptable` configuration along with the `InterruptHandle` type from the `wasmtime` crate. The original support for adding interruption to WebAssembly was added pretty early on in the history of Wasmtime when there was no other method to prevent an infinite loop from the host. Nowadays, however, there are alternative methods for interruption such as fuel or epoch-based interruption. One of the major downsides of `Config::interruptable` is that even when it&#39;s not enabled it forces an atomic swap to happen when entering WebAssembly code. This technically could be a non-atomic swap if the configuration option isn&#39;t enabled but that produces even more branch-y code on entry into WebAssembly which is already something we try to optimize. Calling into WebAssembly is on the order of a dozens of nanoseconds at this time and an atomic swap, even uncontended, can add up to 5ns on some platforms. The main goal of this PR is to remove this atomic swap on entry into WebAssembly. This is done by removing the `Config::interruptable` field entirely, moving all existing consumers to epochs instead which are suitable for the same purposes. This means that the stack overflow check is no longer entangled with the interruption check and perhaps one day we could continue to optimize that further as well. Some consequences of this change are: * Epochs are now the only method of remote-thread interruption. * There are no more Wasmtime traps that produces the `Interrupted` trap code, although we may wish to move future traps to this so I left it in place. * The C API support for interrupt handles was also removed and bindings for epoch methods were added. * Function-entry checks for interruption are a tiny bit less efficient since one check is performed for the stack limit and a second is performed for the epoch as opposed to the `Config::interruptable` style of bundling the stack limit and the interrupt check in one. It&#39;s expected though that this is likely to not really be measurable. * The old `VMInterrupts` structure is renamed to `VMRuntimeLimits`.
3 years ago
#include <wasmtime/engine.h>
#include <wasmtime/error.h>
#include <wasmtime/extern.h>
#include <wasmtime/func.h>
#include <wasmtime/global.h>
#include <wasmtime/instance.h>
#include <wasmtime/linker.h>
#include <wasmtime/memory.h>
#include <wasmtime/module.h>
#include <wasmtime/store.h>
#include <wasmtime/table.h>
#include <wasmtime/trap.h>
#include <wasmtime/val.h>
#ifdef __cplusplus
extern "C" {
#endif
/**
* \brief Converts from the text format of WebAssembly to to the binary format.
*
* \param wat this it the input pointer with the WebAssembly Text Format inside of
* it. This will be parsed and converted to the binary format.
* \param wat_len this it the length of `wat`, in bytes.
* \param ret if the conversion is successful, this byte vector is filled in with
* the WebAssembly binary format.
*
* \return a non-null error if parsing fails, or returns `NULL`. If parsing
* fails then `ret` isn't touched.
*
* This function does not take ownership of `wat`, and the caller is expected to
* deallocate the returned #wasmtime_error_t and #wasm_byte_vec_t.
*/
WASM_API_EXTERN wasmtime_error_t* wasmtime_wat2wasm(
const char *wat,
size_t wat_len,
wasm_byte_vec_t *ret
);
#ifdef __cplusplus
} // extern "C"
#endif
#endif // WASMTIME_API_H