Native hosting is the ability to host the .NET runtime in an arbitrary process, one which didn't start from .NET Core produced binaries.
- "native host" - the code which uses the proposed APIs. Can be any non .NET Core application (.NET Core applications have easier ways to perform these scenarios).
- "hosting components" - shorthand for .NET Core hosting components. Typically refers to
hostfxr
andhostpolicy
. Sometimes also referred to simply as "host". - "host context" - state which
hostfxr
creates and maintains and represents a logical operation on the hosting components.
- Hosting managed components Native host which wants to load managed assembly and call into it for some functionality. Must support loading multiple such components side by side.
- Hosting managed apps
Native host which wants to run a managed app in-proc. Basically a different implementation of the existing .NET Core hosts (
dotnet.exe
orapphost
). The intent is the ability to modify how the runtime starts and how the managed app is executed (and where it starts from). - App using other .NET Core hosting services App (native or .NET Core both) which needs to use some of the other services provided by the .NET Core hosting components. For example the ability to locate available SDKs and so on.
- C-style ABI in
coreclr
coreclr
exposes ABI to host the .NET runtime and run managed code already using C-style APIs. See this header file for the exposed functions. This API requires the native host to locate the runtime and to fully specify all startup parameters for the runtime. There's no inherent interoperability between these APIs and the .NET SDK. - COM-style ABI in
coreclr
coreclr
exposes COM-style ABI to host the .NET runtime and perform a wide range of operations on it. See this header file for more details. Similarly to the C-style ABI the COM-style ABI also requires the native host to locate the runtime and to fully specify all startup parameters. There's no inherent interoperability between these APIs and the .NET SDK. The COM-style ABI is deprecated and should not be used going forward. hostfxr
andhostpolicy
APIs The hosting components of .NET Core already exposes some functionality as C-style ABI on either thehostfxr
orhostpolicy
libraries. These can execute application, determine available SDKs, determine native dependency locations, resolve component dependencies and so on. Unlike the abovecoreclr
based APIs these don't require the caller to fully specify all startup parameters, instead these APIs understand artifacts produced by .NET SDK making it much easier to consume SDK produced apps/libraries. The native host is still required to locate thehostfxr
orhostpolicy
libraries. These APIs are also designed for specific narrow scenarios, any usage outside of these bounds is typically not possible.
This document focuses on hosting which cooperates with the .NET SDK and consumes the artifacts produced by building the managed app/libraries directly. It completely ignores the COM-style ABI as it's hard to use from some programming languages.
As such the document explicitly excludes any hosting based on directly loading coreclr
. Instead it focuses on using the existing .NET Core hosting components in new ways. For details on the .NET Core hosting components see this document.
This section describes how we think the hosting components should evolve. It is expected that this won't happen in any single release, but each change should fit into this picture and hopefully move us closer to this goal. It is also expected that existing APIs won't change (no breaking changes) and thus it can happen that some of the APIs don't fit nicely into the picture.
The hosting component APIs should provide functionality to cover a wide range of applications. At the same time it needs to stay in sync with what .NET Core SDK can produce so that the end-to-end experience is available.
The overall goal of hosting components is to enable loading and executing managed code, this consists of four main steps:
- Locate and load hosting components - How does a native host find and load the hosting components libraries.
- Locate the managed code and all its dependencies - Determining where the managed code comes from (application, component, ...), what frameworks it relies on (framework resolution) and what dependencies it needs (
.deps.json
processing) - Load the managed code into the process - Specify the environment and settings for the runtime, locating or starting the runtime, deciding how to load the managed code into the runtime (which
AssemblyLoadContext
for example) and actually loading the managed code and its dependencies. - Accessing and executing managed code - Locate the managed entry point (for example assembly entry point
Main
, or a specific method to call, or some other means to transition control to the managed code) and exposing it into the native code (function pointer, COM interface, direct execution, ...).
To achieve these we will structure the hosting components APIs into similar (but not exactly) 4 buckets.
This means providing APIs to find and load the hosting components. The end result should be that the native host has the right version of hostfxr
loaded and can call its APIs. The exact solution is dependent on the native host application and its tooling. Ideally we would have a solution which works great for several common environments like C/C++ apps built in VS, C/C++ apps on Mac/Linux built with CMake (or similar) and so on.
First step in this direction is the introduction of nethost
library described below.
The end result of this step should be an initialized host context for a given managed code input which the native host can use to load the managed code and execute it. The step should not actually load any runtime or managed code into the process. Finally this step must handle both processes which don't have any runtime loaded as well as processes which already have runtime loaded.
Functionality performed by this step:
- Locating and understanding the managed code input (app/component)
- Resolving frameworks or handling self-contained
- Resolving dependencies (
.deps.json
processing) - Populating runtime properties
The vision is to expose a set of APIs in the form of hostfxr_initialize_for_...
which would all perform the above functionality and would differ on how the managed code is located and used. So there should be a different API for loading applications and components at the very least. Going forward we can add API which can initialize the context from in-memory configurations, or API which can initialize an "empty" context (no custom code involved, just frameworks) and so on.
First step in this direction is the introduction of these APIs as described below:
hostfxr_initialize_for_dotnet_command_line
- this is the initialize for an application. The API is a bit more complicated as it allows for command line parsing as well. Eventually we might want to add another "initialize app" API without the command line support.hostfxr_initialize_for_runtime_config
- this is the initialize for a component (naming is not ideal, but can't be changed anymore).
After this step it should be possible for the native host to inspect and potentially modify runtime properties. APIs like hostfxr_get_runtime_property_value
and similar described below are an example of this.
This step might be "virtual" in the sense that it's a way to setup the context for the desired functionality. In several scenarios it's not practical to be able to perform this step in separation from the "execute code" step below. The goal of this step is to determine which managed code to load, where it should be loaded into, and to setup the correct inputs so that runtime can perform the right dependency resolution.
Logically this step finds/loads the runtime and initializes it before it can load any managed code into the process.
The main flexibility this step should provide is to determine which assembly load context will be used for loading which code:
- Load all managed code into the default load context
- Load the specified custom managed code into an isolated load context
Eventually we could also add functionality for example to setup the isolated load context for unloadability and so on.
Note that there are no APIs proposed in this document which would only perform this step for now. All the APIs so far fold this step into the next one and perform both loading and executing of the managed code in one step. We should try to make this step more explicit in the future to allow for greater flexibility.
After this step it is no longer possible to modify runtime properties (inspection is still allowed) since the runtime has been loaded.
The end result of this step is either execution of the desired managed code, or returning a native callable representation of the managed code.
The different options should include at least:
- running an application (where the API takes over the thread and runs the application on behalf of the calling thread) - the
hostfxr_run_app
API described bellow is an example of this. - getting a native function pointer to a managed method - the
hostfxr_get_runtime_delegate
and thehdt_load_assembly_and_get_function_pointer
described below is an example of this. - getting a native callable interface (COM) to a managed instance - the
hostfxr_get_runtime_delegate
and thehdt_com_activation
described below is an example of this. - and more...
Ideally it should be possible to combine the different types of initialization, loading and executing of managed code in any way the native host desires. In reality not all combinations are possible or practical. Few examples of combinations which we should probably NOT support (note that this may change in the future):
- running a component as an application - runtime currently makes too many assumptions on what it means to "Run an application"
- loading application into an isolated load context - for now there are too many limitations in the frameworks around secondary load contexts (several large framework features don't work well in a secondary ALC), so the hosting components should prevent this as a way to prevent users from getting into a bad situation.
- loading multiple copies of the runtime into the process - support for side-by-side loading of different .NET Core runtime in a single process is not something we want to implement (at least yet). Hosting components should not allow this to make the user experience predictable. This means that full support for loading self-contained components at all times is not supported.
Lot of other combinations will not be allowed simply as a result of shipping decisions. There's only so much functionality we can fit into any given release, so many combinations will be explicitly disabled to reduce the work necessary to actually ship this. The document below should describe which combinations are allowed in which release.
In .NET Core 3.0 the hosting components (see here) ships with several hosts. These are binaries which act as the entry points to the .NET Core hosting components/runtime:
- The "muxer" (
dotnet.exe
) - The
apphost
(.exe
which is part of the app) - The
comhost
(.dll
which is part of the app and acts as COM server) - The
ijwhost
(.dll
consumed via.lib
used by IJW assemblies)
Every one of these hosts serve different scenario and expose different APIs. The one thing they have in common is that their main purpose is to find the right hostfxr
, load it and call into it to execute the desired scenario. For the most part all these hosts are basically just wrappers around functionality provided by hostfxr
.
The proposal is to add a new host library nethost
which can be used by native host to easily locate hostfxr
. Going forward the library could also include easy-to-use APIs for common scenarios - basically just a simplification of the hostfxr
API surface.
At the same time add the ability to pass additional runtime properties when starting the runtime through the hosting components APIs (starting app, loading component). This can be used by the native host to:
- Register startup hook without modifying environment variables (which are inherited by child processes)
- Introduce new runtime knobs which are only available for native hosts without the need to update the hosting components APIs every time.
Technical note: All strings in the proposed APIs are using the char_t
in this document for simplicity. In real implementation they are of the type pal::char_t
. In particular:
- On Windows - they are
WCHAR *
usingUTF16
encoding - On Linux/macOS - they are
char *
usingUTF8
encoding
New library nethost
which provides a way to locate the right hostfxr
.
This is a dynamically loaded library (.dll
, .so
, .dylib
). For ease of use there is a header file for C/C++ apps as well as .lib
for easy linking on Windows.
Native hosts ship this library as part of the app. Unlike the apphost
, comhost
and ijwhost
, the nethost
will not be directly supported by the .NET SDK since it's target usage is not from .NET Core apps.
The nethost
is part of the Microsoft.NETCore.DotNetAppHost
package. Users are expected to either download the package directly or rely on .NET SDK to pull it down.
The binary itself should be signed by Microsoft as there will be no support for modifying the binary as part of custom application build (unlike apphost
).
struct get_hostfxr_parameters {
size_t size;
const char_t * assembly_path;
const char_t * dotnet_root;
};
int get_hostfxr_path(
char_t * result_buffer,
size_t * buffer_size,
const get_hostfxr_parameters * parameters);
This API locates the hostfxr
library and returns its path by populating result_buffer
.
result_buffer
- Buffer that will be populated with the hostfxr path, including a null terminator.buffer_size
- On input this points to the size of theresult_buffer
inchar_t
units. On output this points to the number ofchar_t
units used from theresult_buffer
(including the null terminator). Ifresult_buffer
isNULL
the input value is ignored and only the minimum required size inchar_t
units is set on output.parameters
- Optional. Additional parameters that modify the behaviour for locating thehostfxr
library. IfNULL
,hostfxr
is located using the environment variable or global registrationsize
- Size of the structure. This is used for versioning and should be set tosizeof(get_hostfxr_parameters)
.assembly_path
- Optional. Path to the application or to the component's assembly.- If specified,
hostfxr
is located as if theassembly_path
is an application withapphost
- If specified,
dotnet_root
- Optional. Path to the root of a .NET Core installation (i.e. folder containing the dotnet executable).- If specified,
hostfxr
is located as if an application is started usingdotnet app.dll
, which means it will be searched for under thedotnet_root
path and theassembly_path
is ignored.
- If specified,
nethost
library uses the __stdcall
calling convention.
- All hosts should be able to use the new API (whether they will is a separate question as the old API has to be kept for backward compat reasons)
- Hide implementation details as much as possible
- Make the API generally easier to use/understand
- Give the implementation more freedom
- Allow future improvements without breaking the API
- Consider explicitly documenting types of behaviors which nobody should take dependency on (specifically failure scenarios)
- Extensible
- It should allow additional parameters to some of the operations without a need to add new exported APIs
- It should allow additional interactions with the host - for example modifying how the runtime is initialized via some new options, without a need for a completely new set of APIs
The API should allow these scenarios:
- Runtime properties
- Specify additional runtime properties from the native host
- Implement conflict resolution for runtime properties
- Inspect calculated runtime properties (the ones calculated by
hostfxr
/hostpolicy
)
- Load managed component and get native function pointer for managed method
- From native app start the runtime and load an assembly
- The assembly is loaded in isolation and with all its dependencies as directed by
.deps.json
- The native app can get back a native function pointer which calls specified managed method
- Get native function pointer for managed method
- From native code get a native function pointer to already loaded managed method
- Load managed component
- From native code load an assembly
All the proposed APIs will be exports of the hostfxr
library and will use the same calling convention and name mangling as existing hostfxr
exports.
All the "initialize" functions will
- Process the
.runtimeconfig.json
- Resolve framework references and find actual frameworks
- Find the root framework (
Microsoft.NETCore.App
) and load thehostpolicy
from it - The
hostpolicy
will then process all relevant.deps.json
files and produce the list of assemblies, native search paths and other artifacts needed to initialize the runtime.
The functions will NOT load the CoreCLR runtime. They just prepare everything to the point where it can be loaded.
The functions return a handle to a new host context:
- The handle must be closed via
hostfxr_close
. - The handle is not thread safe - the consumer should only call functions on it from one thread at a time.
The hostfxr
will also track active runtime in the process. Due to limitations (and to simplify implementation) this tracking will actually not look at the actual coreclr
module (or try to communicate with the runtime in any way). Instead hostfxr
itself will track the host context initialization. The first host context initialization in the process will represent the "loaded runtime". It is only possible to have one "loaded runtime" in the process. Any subsequent host context initialization will just "attach" to the "loaded runtime" instead of creating a new one.
typedef void* hostfxr_handle;
struct hostfxr_initialize_parameters
{
size_t size;
const char_t * host_path;
const char_t * dotnet_root;
};
The hostfxr_initialize_parameters
structure stores parameters which are common to all forms of initialization.
size
- the size of the structure. This is used for versioning. Should be set tosizeof(hostfxr_initialize_parameters)
.host_path
- path to the native host (typically the.exe
). This value is not used for anything by the hosting components. It's just passed to the CoreCLR as the path to the executable. It can point to a file which is not executable itself, if such file doesn't exist (for example in COM activation scenarios this points to thecomhost.dll
). This is used by PAL to initialize internal command line structures, process name and so on.dotnet_root
- Optional. Path to the root of the .NET Core installation in use. This typically points to the install location from which thehostfxr
has been loaded. For example on Windows this would typically point toC:\Program Files\dotnet
. The path is used to search for shared frameworks and potentially SDKs.- If not specified,
hostfxr
will determine the root based on its own location - the same directory if the runtime (coreclr
) exists next to it (as in self-contained applications), otherwise a relative path per the install layout.
- If not specified,
int hostfxr_initialize_for_dotnet_command_line(
int argc,
const char_t * argv[],
const hostfxr_initialize_parameters * parameters,
hostfxr_handle * host_context_handle
);
Initializes the hosting components for running a managed application.
The command line is parsed to determine the app path. The app path will be used to locate the .runtimeconfig.json
and the .deps.json
which will be used to load the application and its dependent frameworks.
argc
andargv
- the command line for running a managed application. These represent the arguments which would have been passed to the muxer if the app was being run from the command line. These are the parameters which are valid for the runtime installation by itself - SDK/CLI commands are not supported. For example, the arguments could beapp.dll app_argument_1 app_argument_2
. This API specifically doesn't support thedotnet run
SDK command.parameters
- additional parameters - seehostfxr_initialize_parameters
for details. (Could be made optional potentially)host_context_handle
- output parameter. On success receives an opaque value which identifies the initialized host context. The handle should be closed by callinghostfxr_close
.
This function only supports arguments for running an application as through the muxer. It does not support SDK commands.
This function can only be called once per-process. It's not supported to run multiple apps in one process (even sequentially).
This function will fail if there already is a CoreCLR running in the process as it's not possible to run two apps in a single process.
This function supports both framework-dependent and self-contained applications.
Note: This is effectively a replacement for hostfxr_main_startupinfo
and hostfxr_main
. Currently it is not a goal to fully replace these APIs because they also support SDK commands which are special in lot of ways and don't fit well with the rest of the native hosting. There's no scenario right now which would require the ability to issue SDK commands from a native host. That said nothing in this proposal should block enabling even SDK commands through these APIs.
int hostfxr_initialize_for_runtime_config(
const char_t * runtime_config_path,
const hostfxr_initialize_parameters * parameters,
hostfxr_handle * host_context_handle
);
This function would load the specified .runtimeconfig.json
, resolve all frameworks, resolve all the assets from those frameworks and then prepare runtime initialization where the TPA contains only frameworks. Note that this case does NOT consume any .deps.json
from the app/component (only processes the framework's .deps.json
). This entry point is intended for comhost
/ijwhost
/nethost
and similar scenarios.
runtime_config_path
- path to the.runtimeconfig.json
file to process. Unlike withhostfxr_initialize_for_dotnet_command_line
, any.deps.json
from the app/component will not be processed by the hosting components during the initialize call.parameters
- additional parameters - seehostfxr_initialize_parameters
for details. (Could be made optional potentially)host_context_handle
- output parameter. On success receives an opaque value which identifies the initialized host context. The handle should be closed by callinghostfxr_close
.
This function can be called multiple times in a process.
- If it's called when no runtime is present, it will run through the steps to "initialize" the runtime (resolving frameworks and so on).
- If it's called when there already is CoreCLR in the process (loaded through the
hostfxr
, direct usage ofcoreclr
is not supported), then the function determines if the specified runtime configuration is compatible with the existing runtime and frameworks. If it is, it returns a valid handle, otherwise it fails.
It needs to be possible to call this function simultaneously from multiple threads at the same time.
It also needs to be possible to call this function while there is an active host context created by hostfxr_initialize_for_dotnet_command_line
and running inside the hostfxr_run_app
.
The function returns specific return code for the first initialized host context, and a different one for any subsequent one. Both return codes are considered "success". If there already was initialized host context in the process then the returned host context has these limitations:
- It won't allow setting runtime properties.
- The initialization will compare the runtime properties from the
.runtimeconfig.json
specified in theruntime_config_path
with those already set to the runtime in the process- If all properties from the new runtime config are already set and have the exact same values (case sensitive string comparison), the initialization succeeds with no additional consequences. (Note that this is the most typical case where the runtime config have no properties in it.)
- If there are either new properties which are not set in the runtime or ones which have different values, the initialization will return a special return code - a "warning". It's not a full on failure as initialized context will be returned.
- In both cases only the properties specified by the new runtime config will be reported on the host context. This is to allow the native host to decide in the "warning" case if it's OK to let the component run or not.
- In both cases the returned host context can still be used to get a runtime delegate, the properties from the new runtime config will be ignored (as there's no way to modify those in the runtime).
The specified .runtimeconfig.json
must be for a framework dependent component. That is it must specify at least one shared framework in its frameworks
section. Self-contained components are not supported.
These functions allow the native host to inspect and modify runtime properties.
- If the
host_context_handle
represents the first initialized context in the process, these functions expose all properties from runtime configurations as well as those computed by the hosting components. These functions will allow modification of the properties viahostfxr_set_runtime_property_value
. - If the
host_context_handle
represents any other context (so not the first one), these functions expose only properties from runtime configuration. These functions won't allow modification of the properties.
It is possible to access runtime properties of the first initialized context in the process at any time (for reading only), by specifying NULL
as the host_context_handle
.
int hostfxr_get_runtime_property_value(
const hostfxr_handle host_context_handle,
const char_t * name,
const char_t ** value);
Returns the value of a runtime property specified by its name.
host_context_handle
- the initialized host context. If set toNULL
the function will operate on runtime properties of the first host context in the process.name
- the name of the runtime property to get. Must not beNULL
.value
- returns a pointer to a buffer with the property value. The buffer is owned by the host context. The caller should make a copy of it if it needs to store it for anything longer than immediate consumption. The lifetime is only guaranteed until any of the below happens:- one of the "run" methods is called on the host context
- the host context is closed via
hostfxr_close
- the value of the property is changed via
hostfxr_set_runtime_property_value
Trying to get a property which doesn't exist is an error and will return an appropriate error code.
We're proposing a fix in hostpolicy
which will make sure that there are no duplicates possible after initialization (see dotnet/runtime#3514). With that hostfxr_get_runtime_property_value
will work always (as there can only be one value).
int hostfxr_set_runtime_property_value(
const hostfxr_handle host_context_handle,
const char_t * name,
const char_t * value);
Sets the value of a property.
host_context_handle
- the initialized host context. (Must not beNULL
)name
- the name of the runtime property to set. Must not beNULL
.value
- the value of the property to set. If the property already has a value in the host context, this function will overwrite it. When set toNULL
and if the property already has a value then the property is "unset" - removed from the runtime property collection.
Setting properties is only supported on the first host context in the process. This is really a limitation of the runtime for which the runtime properties are immutable. Once the first host context is initialized and starts a runtime there's no way to change these properties. For now we will not consider the scenario where the host context is initialized but the runtime hasn't started yet, mainly for simplicity of implementation and lack of requirements.
int hostfxr_get_runtime_properties(
const hostfxr_handle host_context_handle,
size_t * count,
const char_t **keys,
const char_t **values);
Returns the full set of all runtime properties for the specified host context.
host_context_handle
- the initialized host context. If set toNULL
the function will operate on runtime properties of the first host context in the process.count
- in/out parameter which must not beNULL
. On input it specifies the size of thekeys
andvalues
buffers. On output it contains the number of entries used fromkeys
andvalues
buffers - the number of properties returned. If the size of the buffers is too small, the function returns a specific error code and fill thecount
with the number of available properties. Ifkeys
orvalues
isNULL
the function ignores the input value ofcount
and just returns the number of properties.keys
- buffer which acts as an array of pointers to buffers with keys for the runtime properties.values
- buffer which acts as an array of pointer to buffers with values for the runtime properties.
keys
and values
store pointers to buffers which are owned by the host context. The caller should make a copy of it if it needs to store it for anything longer than immediate consumption. The lifetime is only guaranteed until any of the below happens:
- one of the "run" methods is called on the host context
- the host context is closed via
hostfxr_close
- the value or existence of any property is changed via
hostfxr_set_runtime_property_value
Note that hostfxr_set_runtime_property_value
can remove or add new properties, so the number of properties returned is only valid as long as no properties were added/removed.
int hostfxr_run_app(const hostfxr_handle host_context_handle);
Runs the application specified by the hostfxr_initialize_for_dotnet_command_line
. It is illegal to try to use this function when the host context was initialized through any other way.
host_context_handle
- handle to the initialized host context.
The function will return only once the managed application exits.
hostfxr_run_app
cannot be used in combination with any other "run" function. It can also only be called once.
int hostfxr_get_runtime_delegate(const hostfxr_handle host_context_handle, hostfxr_delegate_type type, void ** delegate);
Starts the runtime and returns a function pointer to specified functionality of the runtime.
host_context_handle
- handle to the initialized host context.- [.NET 8 and above] If set to
NULL
the function will operate on the active host context in the process.
- [.NET 8 and above] If set to
type
- the type of runtime functionality requestedhdt_load_assembly_and_get_function_pointer
- entry point which loads an assembly (with dependencies) and returns function pointer for a specified static method. See below for details (Loading and calling managed components)hdt_com_activation
,hdt_com_register
,hdt_com_unregister
- COM activation entry-points - see COM activation for more details.hdt_load_in_memory_assembly
- IJW entry-point - see IJW activation for more details.hdt_winrt_activation
[.NET 3.* only] - WinRT activation entry-point - see WinRT activation for more details. The delegate is not supported for .NET 5 and above.hdt_get_function_pointer
[.NET 5 and above] - entry-point which finds a managed method and returns a function pointer to it. See calling managed functions for details.hdt_load_assembly
[.NET 8 and above] - entry-point which loads an assembly by its path. See loading managed components for details.hdt_load_assembly_bytes
[.NET 8 and above] - entry-point which loads an assembly from a byte array. See loading managed components for details.
delegate
- when successful, the native function pointer to the requested runtime functionality.
In .NET Core 3.0 the function only works if hostfxr_initialize_for_runtime_config
was used to initialize the host context.
In .NET 5 the function also works if hostfxr_initialize_for_dotnet_command_line
was used to initialize the host context. Also for .NET 5 it will only be allowed to request hdt_load_assembly_and_get_function_pointer
or hdt_get_function_pointer
on a context initialized via hostfxr_initialize_for_dotnet_command_line
, all other runtime delegates will not be supported in this case.
All returned runtime delegates use the __stdcall
calling convention on x86.
int hostfxr_close(const hostfxr_handle host_context_handle);
Closes a host context.
host_context_handle
- handle to the initialized host context to close.
To load managed components from native app directly (not using COM or WinRT) the hosting components exposes a new runtime helper/delegate hdt_load_assembly_and_get_function_pointer
. Calling the hostfxr_get_runtime_delegate(handle, hdt_load_assembly_and_get_function_pointer, &helper)
returns a function pointer to the runtime helper with this signature:
int load_assembly_and_get_function_pointer_fn(
const char_t *assembly_path,
const char_t *type_name,
const char_t *method_name,
const char_t *delegate_type_name,
void *reserved,
/*out*/ void **delegate)
Calling this function will load the specified assembly in isolation (into its own AssemblyLoadContext
) and it will use AssemblyDependencyResolver
on it to provide dependency resolution. Once loaded it will find the specified type and method and return a native function pointer to that method. The method's signature can be specified via the delegate type name.
assembly_path
- Path to the assembly to load. In case of complex component, this should be the main assembly of the component (the one with the.deps.json
next to it). Note that this does not have to be the assembly from which thetype_name
andmethod_name
are.type_name
- Assembly qualified type name to findmethod_name
- Name of the method on thetype_name
to find. The method must bestatic
and must match the signature ofdelegate_type_name
.delegate_type_name
- Assembly qualified delegate type name for the method signature, or null. If this is null, the method signature is assumed to be:This maps to native signature:public delegate int ComponentEntryPoint(IntPtr args, int sizeBytes);
[.NET 5 and above] Theint component_entry_point_fn(void *arg, int32_t arg_size_in_bytes);
delegate_type_name
can be also specified asUNMANAGEDCALLERSONLY_METHOD
(defined as(const char_t*)-1
) which means that the managed method is marked withUnmanagedCallersOnlyAttribute
.reserved
- parameter reserved for future extensibility, currently unused and must beNULL
.delegate
- out parameter which receives the native function pointer to the requested managed method.
The helper will always load the assembly into an isolated load context. This is the case regardless if the requested assembly is also available in the default load context or not.
[.NET 5 and above] It is allowed to call this helper on a host context which came from hostfxr_initialize_for_dotnet_command_line
which will make all the application assemblies available in the default load context. In such case it is recommended to only use this helper for loading plugins external to the application. Using the helper to load assembly from the application itself will lead to duplicate copies of assemblies and duplicate types.
It is allowed to call the returned runtime helper many times for different assemblies or different methods from the same assembly. It is not required to get the helper every time. The implementation of the helper will cache loaded assemblies, so requests to load the same assembly twice will load it only once and reuse it from that point onward. Ideally components should not take a dependency on this behavior, which means components should not have global state. Global state in components is typically just cause for problems. For example it may create ordering issues or unintended side effects and so on.
The returned native function pointer to managed method has the lifetime of the process and can be used to call the method many times over. Currently there's no way to unload the managed component or otherwise free the native function pointer. Such support may come in future releases.
In .NET 5 the hosting components add a new functionality which allows just getting a native function pointer for already available managed method. This is implemented in a runtime helper hdt_get_function_pointer
. Calling the hostfxr_get_runtime_delegate(handle, hdt_get_function_pointer, &helper)
returns a function pointer to the runtime helper with this signature:
int get_function_pointer_fn(
const char_t *type_name,
const char_t *method_name,
const char_t *delegate_type_name,
void *load_context,
void *reserved,
/*out*/ void **delegate)
Calling this function will find the specified type in the default load context, locate the required method on it and return a native function pointer to that method. The method's signature can be specified via the delegate type name.
type_name
- Assembly qualified type name to findmethod_name
- Name of the method on thetype_name
to find. The method must bestatic
and must match the signature ofdelegate_type_name
.delegate_type_name
- Assembly qualified delegate type name for the method signature, or null. If this is null, the method signature is assumed to be:This maps to native signature:public delegate int ComponentEntryPoint(IntPtr args, int sizeBytes);
Theint component_entry_point_fn(void *arg, int32_t arg_size_in_bytes);
delegate_type_name
can be also specified asUNMANAGEDCALLERSONLY_METHOD
(defined as(const char_t*)-1
) which means that the managed method is marked withUnmanagedCallersOnlyAttribute
.load_context
- eventually this parameter should support specifying which load context should be used to locate the type/method specified in previous parameters. For .NET 5 this parameter must beNULL
and the API will only locate the type/method in the default load context.reserved
- parameter reserved for future extensibility, currently unused and must beNULL
.delegate
- out parameter which receives the native function pointer to the requested managed method.
The helper will lookup the type_name
from the default load context (AssemblyLoadContext.Default
) and then return method on it. If the type and method lookup requires assemblies which have not been loaded by the Default ALC yet, this process will resolve them against the Default ALC and load them there (most likely from TPA). This helper will not register any additional assembly resolution logic onto the Default ALC, it will solely rely on the existing functionality of the Default ALC.
It is allowed to ask for this helper on any valid host context. Because the helper operates on default load context only it should mostly be used with context initialized via hostfxr_initialize_for_dotnet_command_line
as in that case the default load context will have the application code available in it. Contexts initialized via hostfxr_initialize_for_runtime_config
have only the framework assemblies available in the default load context. The type_name
must resolve within the default load context, so in the case where only framework assemblies are loaded into the default load context it would have to come from one of the framework assemblies only.
It is allowed to call the returned runtime helper many times for different types or methods. It is not required to get the helper every time.
The returned native function pointer to managed method has the lifetime of the process and can be used to call the method many times over. Currently there's no way to "release" the native function pointer (and the respective managed delegate), this functionality may be added in a future release.
The runtime delegate type hdt_load_assembly
allows loading a managed assembly by its path. Calling hostfxr_get_runtime_delegate(handle, hdt_load_assembly, &helper)
returns a function pointer to the runtime helper with this signature:
int load_assembly(
const char_t *assembly_path,
void *load_context,
void *reserved);
Calling this function will load the specified assembly in the default load context. It uses AssemblyDependencyResolver
to register additional dependency resolution for the load context.
assembly_path
- Path to the assembly to load - requirements match theassemblyPath
parameter of AssemblyLoadContext.LoadFromAssemblyPath. This path will also be used for dependency resolution via any.deps.json
corresponding to the assembly.load_context
- the load context that will be used to load the assembly. For .NET 8 this parameter must beNULL
and the API will only load the assembly in the default load context.reserved
- parameter reserved for future extensibility, currently unused and must beNULL
.
The runtime delegate type hdt_load_assembly_bytes
allows loading a managed assembly from a byte array. Calling hostfxr_get_runtime_delegate(handle, hdt_load_assembly_bytes, &helper)
returns a function pointer to the runtime helper with this signature:
int load_assembly_bytes(
const void *assembly_bytes,
size_t assembly_bytes_len,
const void *symbols_bytes,
size_t symbols_bytes_len,
void *load_context,
void *reserved);
Calling this function will load the specified assembly in the default load context. It does not provide a mechanism for registering additional dependency resolution, as mechanisms like .deps.json
and AssemblyDependencyResolver
are file-based. Dependencies can be pre-loaded (for example, via a previous call to this function) or the specified assembly can explicitly register its own resolution logic (for example, via the AssemblyLodContext.Resolving
event).
assembly_bytes
- Bytes of the assembly to load.assembly_bytes_len
- Byte length of the assembly to load.symbols_bytes
- Bytes of the symbols for the assembly to load.symbols_bytes_len
- Byte length of the symbols for the assembly to load.load_context
- the load context that will be used to load the assembly. For .NET 8 this parameter must beNULL
and the API will only load the assembly in the default load context.reserved
- parameter reserved for future extensibility, currently unused and must beNULL
.
These runtime delegates simply load the assembly. They do not return any representation of the loaded assembly and do not execute code in the assembly. To run code from the assembly, the delegate for calling a managed function can be used to get a function pointer to a method in a loaded assembly by specifying the assembly-qualified type name containing the method.
It is allowed to ask for this helper on any valid host context. The returned runtime helper can be called multiple times for different assemblies. It is not required to get the helper every time.
It is important to correctly synchronize some of these operations to achieve the desired API behavior as well as thread safety requirements. The following behaviors will be used to achieve this.
first host context
is the one which is used to load and initialize the CoreCLR runtime in the process. At any given time there can only be onefirst host context
.secondary host context
is any other initialized host context whenfirst host context
already exists in the process.
- If there's no
first host context
in the process the first call tohostfxr_initialize_...
will create a newfirst host context
. There can only be onefirst host context
in existence at any point in time. - Calling
hostfxr_initialize...
whenfirst host context
already exists will always return asecondary host context
. - The
first host context
blocks creation of any other host context until it is used to load and initialize the CoreCLR runtime. This means thathostfxr_initialize...
and subsequently one of the "run" methods must be called on thefirst host context
to unblock creation ofsecondary host contexts
. - Calling
hostfxr_initialize...
will block until thefirst host context
is initialized, a "run" method is called on it and the CoreCLR is loaded and initialized. Thehostfxr_initialize...
will block potentially indefinitely. The method will block very early on. All of the operations done by the initialize will only happen once it's unblocked. first host context
can fail to initialize the runtime (or anywhere up to that point). If this happens, it's marked as failed and is not considered afirst host context
anymore. This unblocks the potentially waitinghostfxr_initialize...
calls. In this case the firsthostfxr_initialize...
after the failure will create a newfirst host context
.first host context
can be closed usinghostfxr_close
before it is used to initialize the CoreCLR runtime. This is similar to the failure above, the host context is marked as "closed/failed" and is not consideredfirst host context
anymore. This unblocks any waitinghostfxr_initialize...
calls.- Once the
first host context
successfully initialized the CoreCLR runtime it is permanently marked as "successful" and will remain thefirst host context
for the lifetime of the process. Such host context should still be closed once not needed viahostfxr_close
.
- It is invalid to initialize a host context via
hostfxr_initialize...
and then never callhostfxr_close
on it. An initialized but not closed host context is considered abandoned. Abandonedfirst host context
will cause infinite blocking of any futurehostfxr_initialize...
calls.
The above behaviors should make sure that some important scenarios are possible and work reliably.
One such scenario is a COM host on multiple threads. The app is not running any .NET Core yet (no CoreCLR loaded). On two threads in parallel COM activation is invoked which leads to two invocations into the comhost
to active .NET Core objects. The comhost
will use the hostfxr_initialize...
and hostfxr_get_runtime_delegate
APIs on two threads in parallel then. Only one of them can load and initialize the runtime (and also perform full framework resolution and determine the framework versions and assemblies to load). The other has to become a secondary host context
and try to conform to the first one. The above behavior of hostfxr_initialize...
blocking until the first host context
is done initializing the runtime will make sure of the correct behavior in this case.
At the same time it gives the native app (comhost
in this case) the ability to query and modify runtime properties in between the hostfxr_initialize...
and hostfxr_get_runtime_delegate
calls on the first host context
.
The hostfxr
exports are defined in the hostfxr.h header file.
The runtime helper and method signatures for loading managed components are defined in coreclr_delegates.h header file.
Currently we don't plan to ship these files, but it's possible to take them from the repo and use it.
Since hostfxr
and the other hosting components are versioned independently there are several interesting cases of version mismatches:
For muxer it should almost always match, but apphost
can be different. That is, it's perfectly valid to use older 2.* apphost
with a new 3.0 hostfxr
. The opposite should be rare, but in theory can happen as well. To keep the code simple both muxer and apphost
will keep using the existing 2.* APIs on hostfxr
even in situation where both are 3.0 and thus could start using the new APIs.
hostfxr
must be backward compatible and support 2.* APIs.
Potentially we could switch just apphost
to use the new APIs (since it doesn't have the same compatibility burden as the muxer), but it's probably safer to not do that.
It should really only happen that hostfxr
is equal or newer than hostpolicy
. The opposite should be very rare. In any case hostpolicy
should support existing 2.* APIs and thus the rare case will keep working anyway.
The interesting case is 3.0 hostfxr
using 2.* hostpolicy
. This will be very common, basically any 2.* app running on a machine with 3.0 installed will be in that situation. This case has two sub-cases:
hostfxr
is invoked using one of the 2.* APIs. In this case the simple solution is to keep using the 2.*hostpolicy
APIs always.hostfxr
is invoked using one of the new 3.0 APIs (likehostfxr_initialize...
). In this case it's not possible to completely support the new APIs, since they require new functionality fromhostpolicy
. For now thehostfxr
should simply fail. It is in theory possible to support some kind of emulation mode where for some scenarios the new APIs would work even with oldhostpolicy
, but for simplicity it's better to start with just failing.
The existing 2.* APIs in hostfxr
could switch to internally use the new functionality and in turn use the new 3.0 hostpolicy
APIs. The tricky bit here is scenario like this:
- 3.0 App is started via
apphost
or muxer which as mentioned above will keep on using the 2.*hostfxr
APIs. This will load CoreCLR into the process. - COM component is activated in the same process. This will go through the new 3.0
hostfxr
APIs, and to work correctly will require the internal representation of thefirst host context
.
If the 2.* hostfxr
APIs would continue to use the old 2.* hostpolicy
APIs even if hostpolicy
is new, then the above scenario will be hard to achieve as there will be no first host context
. hostpolicy
could somehow "emulate" the first host context
, but without hostpolicy
cooperation this would be hard.
On the other hand switching to use the new hostpolicy
APIs even in 2.* hostfxr
APIs is risky for backward compatibility. This will have to be decided during implementation.
All samples assume that the native host has found the hostfxr
, loaded it and got the exports (possibly by using the nethost
).
Samples in general ignore error handling.
hostfxr_initialize_parameters params;
params.size = sizeof(params);
params.host_path = get_path_to_the_host_exe(); // Path to the current executable
params.dotnet_root = get_directory(get_directory(get_directory(hostfxr_path))); // Three levels up from hostfxr typically
hostfxr_handle host_context_handle;
hostfxr_initialize_for_dotnet_command_line(
_argc_,
_argv_, // For example, 'app.dll app_argument_1 app_argument_2'
¶ms,
&host_context_handle);
size_t buffer_used = 0;
if (hostfxr_get_runtime_property(host_context_handle, "TEST_PROPERTY", NULL, 0, &buffer_used) == HostApiMissingProperty)
{
hostfxr_set_runtime_property(host_context_handle, "TEST_PROPERTY", "TRUE");
}
hostfxr_run_app(host_context_handle);
hostfxr_close(host_context_handle);
using load_assembly_and_get_function_pointer_fn = int (STDMETHODCALLTYPE *)(
const char_t *assembly_path,
const char_t *type_name,
const char_t *method_name,
const char_t *delegate_type_name,
void *reserved,
void **delegate);
hostfxr_handle host_context_handle;
hostfxr_initialize_for_runtime_config(config_path, NULL, &host_context_handle);
load_assembly_and_get_function_pointer_fn runtime_delegate = NULL;
hostfxr_get_runtime_delegate(
host_context_handle,
hostfxr_delegate_type::load_assembly_and_get_function_pointer,
(void **)&runtime_delegate);
using managed_entry_point_fn = int (STDMETHODCALLTYPE *)(void *arg, int argSize);
managed_entry_point_fn entry_point = NULL;
runtime_delegate(assembly_path,
type_name,
method_name,
NULL,
NULL,
(void **)&entry_point);
ArgStruct arg;
entry_point(&arg, sizeof(ArgStruct));
hostfxr_close(host_context_handle);
The exact impact on the hostfxr
/hostpolicy
interface needs to be investigated. The assumption is that new APIs will have to be added to hostpolicy
to implement the proposed functionality.
Part if this investigation will also be compatibility behavior. Currently "any" version of hostfxr
needs to be able to use "any" version of hostpolicy
. But the proposed functionality will need both new hostfxr
and new hostpolicy
to work. It is likely the proposed APIs will fail if the app resolves to a framework with old hostpolicy
without the necessary new APIs. Part of the investigation will be if it's feasible to use the new hostpolicy
APIs to implement existing old hostfxr
APIs.
Native hosting support on managed side is disabled by default on trimmed apps. Native hosting and trimming are incompatible since the trimmer cannot analyze methods that are called by native hosts. Native hosting support for trimming can be managed through the feature switch settings specific to each native host.