diff --git a/doc/articles/Errors.md b/doc/articles/Errors.md index 2fa48d8d..4d013c33 100644 --- a/doc/articles/Errors.md +++ b/doc/articles/Errors.md @@ -35,7 +35,7 @@ These behave similarly to the Promise-returning JavaScript APIs. Instead of ther These errors include: -- @ref StructChainingError cases. +- @ref OutStructChainError cases. - [Content-timeline](https://www.w3.org/TR/webgpu/#content-timeline) errors other than those which are surfaced as @ref DeviceError in `webgpu.h`. See specific documentation to determine how each error is exposed. Generally these will return some kind of failure status (like \ref WGPUStatus_Error) or `NULL`, and produce an @ref ImplementationDefinedLogging message. @@ -45,16 +45,3 @@ Generally these will return some kind of failure status (like \ref WGPUStatus_Er Entry points may also specify that they produce "implementation-defined logging". These messages are logged in an implementation defined way (e.g. to an implementation-specific callback, or to a logging runtime). They are intended to be intended to be read by humans, useful primarily for development and crash reporting. - -## Struct-Chaining Error {#StructChainingError} - -A struct-chaining error happens when the @ref WGPUSType of a struct in a struct chain is not valid for that chain. - -Struct chains which are used in device-timeline validation/operations (e.g. @ref WGPUBufferDescriptor in @ref wgpuDeviceCreateBuffer) have their chain errors surfaced asynchronously, like any other validation error. - -### Out-Struct-Chain Error {#OutStructChainError} - -Operations which take out-struct-chains (e.g. @ref WGPULimits, in @ref wgpuAdapterGetLimits and @ref wgpuDeviceGetLimits, but not in @ref WGPUDeviceDescriptor) handle struct-chaining errors as follows: - -- The output struct and struct chain is not modified. -- The operation produces a @ref SynchronousError (return value and log message). diff --git a/doc/articles/Ownership.md b/doc/articles/Ownership.md index f23701f3..cbf64bc2 100644 --- a/doc/articles/Ownership.md +++ b/doc/articles/Ownership.md @@ -14,11 +14,13 @@ The application must `Release` this ref before losing the pointer. Variable-sized outputs returned from the API (e.g. the strings in \ref WGPUAdapterInfo, from \ref wgpuAdapterGetInfo) are application-owned. The application must call the appropriate `FreeMembers` function (e.g. \ref wgpuAdapterInfoFreeMembers) before losing the pointers. -Note that such functions will *not* free any previously-allocated data: overwriting an output structure without first freeing members will leak the allocations; e.g.: +Note that such functions will *not* free any previously-allocated data: overwriting an output structure without first releasing ownership will leak the allocations; e.g.: - Overwriting the strings in \ref WGPUAdapterInfo with \ref wgpuAdapterGetInfo without first calling \ref wgpuAdapterInfoFreeMembers. - Overwriting the `texture` in \ref WGPUSurfaceTexture with \ref wgpuSurfaceGetCurrentTexture without first calling \ref wgpuTextureRelease. +Note also that some structs with `FreeMembers` functions may be used as both inputs and outputs. In this case `FreeMembers` must only be used if the member allocations were made by the `webgpu.h` implementation (as an output), not if they were made by the applcation (as an input). + ## Callback Arguments {#CallbackArgs} Output arguments passed from the API to application callbacks include objects and message strings, which are passed to most callbacks. diff --git a/doc/articles/StructChaining.md b/doc/articles/StructChaining.md new file mode 100644 index 00000000..d75d8f28 --- /dev/null +++ b/doc/articles/StructChaining.md @@ -0,0 +1,83 @@ +# Struct-Chaining {#StructChaining} + +Struct-chaining is a C API pattern using linked lists and dynamic typing to extend existing structs with new members, while maintaining API and ABI compatibility. For example: + +An extensible struct is statically typed. It is the root of a linked list, containing a pointer to the next struct in the chain: + +```c +typedef struct WGPUMyStructBase { + WGPUChainedStruct * nextInChain; + uint32_t x; +} WGPUMyBaseStruct; +``` + +Each extension struct is a "subtype" of @ref WGPUChainedStruct; that is, its first member is a @ref WGPUChainedStruct, so that it can be safely cast to @ref WGPUChainedStruct. This allows the implementation to read its @ref WGPUChainedStruct::sType, which is some value of @ref WGPUSType ("Struct Type") that dynamically identifies the struct's type. (The `sType` may come from an implementation-specific extension; @ref WGPUSType is an "open" enum.) + +Once the implementation identifies the struct by its `sType`, it casts the pointer back to the appropriate struct type in order to access its contents. Setting `sType` incorrectly (or pointing to any type that isn't a subtype of @ref WGPUChainedStruct) causes undefined behavior. + +```c +typedef enum WGPUSType { + // ... + WGPUSType_MyStructExtension1 = /* ... */, + WGPUSType_MyStructExtension2 = /* ... */, + // ... +} WGPUStype; + +typedef struct WGPUMyStructExtension1 { + WGPUChainedStruct chain; // .chain.sType must be WGPUSType_MyStructExtension1 + uint32_t y; +} WGPUMyStructExtension1; + +typedef struct WGPUMyStructExtension2 { + WGPUChainedStruct chain; // .chain.sType must be WGPUSType_MyStructExtension2 + uint32_t z; +} WGPUMyStructExtension2; +``` + +This is used like so: + +```c +WGPUMyStructExtension2 ext2 = WGPU_MY_STRUCT_EXTENSION2_INIT; +// .chain.sType is already set correctly by the INIT macro. +// .chain.next is set to NULL indicating the end of the chain. +ext2.z = 2; + +WGPUMyStructExtension1 ext1 = WGPU_MY_STRUCT_EXTENSION1_INIT; +// .chain.sType is already set correctly by the INIT macro. +// .chain.next may be set in either of two ways, equivalently: +ext1.chain.next = &ext2.chain; +ext1.chain.next = (WGPUChainedStruct*) &ext2; +ext1.y = 1; + +WGPUMyStructBase base = WGPU_MY_STRUCT_BASE_INIT; +// Note: base structs do not have STypes (they are statically typed). +base.nextInChain = &ext1.chain; +base.x = 0; +``` + +The pointer links in a struct chain are all mutable pointers. Whether the structs in the chain are actually mutated depends on the function they are passed to (whether the struct chain is passed as an input or an output). Some structs (e.g. @ref WGPULimits) may be either an input or an output depending on the function being called. + +## Struct-Chaining Error {#StructChainingError} + +A struct-chaining error occurs if a struct chain is incorrectly constructed (in a detectable way). +They occur if and only if: + +- The `sType` of a struct in the chain is not valid _in the context of the chain root's static type_. + - If this happens, the implementation must not downcast the pointer to access the rest of the struct (even if it would know how to downcast it in other contexts). +- Multiple instances of the same `sType` value are seen in the same struct chain. (Note this also detects and disallows cycles.) + - Implementation-specific extensions also _should_ avoid designs that use unbounded recursion (such as linked lists) in favor of iterative designs (arrays or arrays-of-pointers). This is to avoid stack overflows in struct handling and serialization/deserialization. + +Struct chains which are used in device-timeline validation/operations (e.g. @ref WGPUBufferDescriptor in @ref wgpuDeviceCreateBuffer) have their chain errors surfaced asynchronously, like any other validation error. + +Struct chains which are used in content-timeline operations (e.g. @ref OutStructChainError) have their chain errors surfaced synchronously, like other content-timeline validation errors. + +### Out-Struct-Chain Error {#OutStructChainError} + +Operations which take out-struct-chains (e.g. @ref WGPULimits, in @ref wgpuAdapterGetLimits and @ref wgpuDeviceGetLimits, but not in @ref WGPUDeviceDescriptor) handle struct-chaining errors as follows: + +- The output struct and struct chain is not modified. +- The operation produces a @ref SynchronousError (return value and log message). + +## Ownership + +TODO(#264): Document whether `FreeMembers` functions traverse the chain (see @ref ReturnedWithOwnership). diff --git a/doc/articles/index.md b/doc/articles/index.md index e7e8bdd6..d15af669 100644 --- a/doc/articles/index.md +++ b/doc/articles/index.md @@ -6,3 +6,4 @@ - \subpage Surfaces - \subpage SentinelValues - \subpage Strings +- \subpage StructChaining diff --git a/tests/compile/main.inl b/tests/compile/main.inl index d017dd4f..d870aeb3 100644 --- a/tests/compile/main.inl +++ b/tests/compile/main.inl @@ -38,6 +38,31 @@ int main(void) { a.endOfPassWriteIndex = WGPU_QUERY_SET_INDEX_UNDEFINED; } + // Simple chaining smoke test + { + // It's not valid to use both WGSL and SPIRV but this test doesn't care. + WGPUShaderSourceSPIRV descSPIRV = WGPU_SHADER_SOURCE_SPIRV_INIT; + WGPUShaderSourceWGSL descWGSL = WGPU_SHADER_SOURCE_WGSL_INIT; + WGPUShaderModuleDescriptor desc = WGPU_SHADER_MODULE_DESCRIPTOR_INIT; + + // Test of linking one extension to another. + descWGSL.chain.next = &descSPIRV.chain; + // Test of linking base struct to an extension. + desc.nextInChain = &descWGSL.chain; + + // Also test the alternate linking style using a cast. +#ifdef __cplusplus +# if __cplusplus >= 201103L + static_assert(offsetof(WGPUShaderSourceWGSL, chain) == 0, ""); +# endif + descWGSL.chain.next = reinterpret_cast(&descSPIRV); + desc.nextInChain = reinterpret_cast(&descWGSL); +#else + descWGSL.chain.next = (WGPUChainedStruct*) &descSPIRV; + desc.nextInChain = (WGPUChainedStruct*) &descWGSL; +#endif + } + // Check that generated initializers are valid // TODO: Would be nice to autogenerate this so we don't forget to test any new structs. { WGPUBufferMapCallbackInfo x = WGPU_BUFFER_MAP_CALLBACK_INFO_INIT; }