Skip to content

Latest commit

 

History

History
1587 lines (1229 loc) · 83.7 KB

RATIONALE.md

File metadata and controls

1587 lines (1229 loc) · 83.7 KB
Raw

Notable rationale of wazero

Zero dependencies

Wazero has zero dependencies to differentiate itself from other runtimes which have heavy impact usually due to CGO. By avoiding CGO, wazero avoids prerequisites such as shared libraries or libc, and lets users keep features like cross compilation.

Avoiding go.mod dependencies reduces interference on Go version support, and size of a statically compiled binary. However, doing so brings some responsibility into the project.

Go's native platform support is good: We don't need platform-specific code to get monotonic time, nor do we need much work to implement certain features needed by our compiler such as mmap. That said, Go does not support all common operating systems to the same degree. For example, Go 1.18 includes Mprotect on Linux and Darwin, but not FreeBSD.

The general tradeoff the project takes from a zero dependency policy is more explicit support of platforms (in the compiler runtime), as well a larger and more technically difficult codebase.

At some point, we may allow extensions to supply their own platform-specific hooks. Until then, one end user impact/tradeoff is some glitches trying untested platforms (with the Compiler runtime).

Why do we use CGO to implement system calls on darwin?

wazero is dependency and CGO free by design. In some cases, we have code that can optionally use CGO, but retain a fallback for when that's disabled. The only operating system (GOOS) we use CGO by default in is darwin.

Unlike other operating systems, regardless of CGO_ENABLED, Go always uses "CGO" mechanisms in the runtime layer of darwin. This is explained in Statically linked binaries on Mac OS X:

Apple does not support statically linked binaries on Mac OS X. A statically linked binary assumes binary compatibility at the kernel system call interface, and we do not make any guarantees on that front. Rather, we strive to ensure binary compatibility in each dynamically linked system library and framework.

This plays to our advantage for system calls that aren't yet exposed in the Go standard library, notably futimens for nanosecond-precision timestamp manipulation.

Why not x/sys

Going beyond Go's SDK limitations can be accomplished with their x/sys library. For example, this includes zsyscall_freebsd_amd64.go missing from the Go SDK.

However, like all dependencies, x/sys is a source of conflict. For example, x/sys had to be in order to upgrade to Go 1.18.

If we depended on x/sys, we could get more precise functionality needed for features such as clocks or more platform support for the compiler runtime.

That said, formally supporting an operating system may still require testing as even use of x/sys can require platform-specifics. For example, mmap-go uses x/sys, but also mentions limitations, some not surmountable with x/sys alone.

Regardless, we may at some point introduce a separate go.mod for users to use x/sys as a platform plugin without forcing all users to maintain that dependency.

Project structure

wazero uses internal packages extensively to balance API compatibility desires for end users with the need to safely share internals between compilers.

End-user packages include wazero, with Config structs, api, with shared types, and the built-in wasi library. Everything else is internal.

We put the main program for wazero into a directory of the same name to match conventions used in go install, notably the name of the folder becomes the binary name. We chose to use cmd/wazero as it is common practice and less surprising than wazero/wazero.

Internal packages

Most code in wazero is internal, and it is acknowledged that this prevents external implementation of facets such as compilers or decoding. It also prevents splitting this code into separate repositories, resulting in a larger monorepo. This also adds work as more code needs to be centrally reviewed.

However, the alternative is neither secure nor viable. To allow external implementation would require exporting symbols public, such as the CodeSection, which can easily create bugs. Moreover, there's a high drift risk for any attempt at external implementations, compounded not just by wazero's code organization, but also the fast moving Wasm and WASI specifications.

For example, implementing a compiler correctly requires expertise in Wasm, Golang and assembly. This requires deep insight into how internals are meant to be structured and the various tiers of testing required for wazero to result in a high quality experience. Even if someone had these skills, supporting external code would introduce variables which are constants in the central one. Supporting an external codebase is harder on the project team, and could starve time from the already large burden on the central codebase.

The tradeoffs of internal packages are a larger codebase and responsibility to implement all standard features. It also implies thinking about extension more as forking is not viable for reasons above also. The primary mitigation of these realities are friendly OSS licensing, high rigor and a collaborative spirit which aim to make contribution in the shared codebase productive.

Avoiding cyclic dependencies

wazero shares constants and interfaces with internal code by a sharing pattern described below:

  • shared interfaces and constants go in one package under root: api.
  • user APIs and structs depend on api and go into the root package wazero.
    • e.g. InstantiateModule -> /wasm.go depends on the type api.Module.
  • implementation code can also depend on api in a corresponding package under /internal.
    • Ex package wasm -> /internal/wasm/*.go and can depend on the type api.Module.

The above guarantees no cyclic dependencies at the cost of having to re-define symbols that exist in both packages. For example, if wasm.Store is a type the user needs access to, it is narrowed by a cover type in the wazero:

type runtime struct {
	s *wasm.Store
}

This is not as bad as it sounds as mutations are only available via configuration. This means exported functions are limited to only a few functions.

Avoiding security bugs

In order to avoid security flaws such as code insertion, nothing in the public API is permitted to write directly to any mutable symbol in the internal package. For example, the package api is shared with internal code. To ensure immutability, the api package cannot contain any mutable public symbol, such as a slice or a struct with an exported field.

In practice, this means shared functionality like memory mutation need to be implemented by interfaces.

Here are some examples:

  • api.Memory protects access by exposing functions like WriteFloat64Le instead of exporting a buffer ([]byte).
  • There is no exported symbol for the []byte representing the CodeSection

Besides security, this practice prevents other bugs and allows centralization of validation logic such as decoding Wasm.

API Design

Why is context.Context inconsistent?

It may seem strange that only certain API have an initial context.Context parameter. We originally had a context.Context for anything that might be traced, but it turned out to be only useful for lifecycle and host functions.

For instruction-scoped aspects like memory updates, a context parameter is too fine-grained and also invisible in practice. For example, most users will use the compiler engine, and its memory, global or table access will never use go's context.

Why does api.ValueType map to uint64?

WebAssembly allows functions to be defined either by the guest or the host, with signatures expressed as WebAssembly types. For example, i32 is a 32-bit type which might be interpreted as signed. Function signatures can have zero or more parameters or results even if WebAssembly 1.0 allows up to one result.

The guest can export functions, so that the host can call it. In the case of wazero, the host is Go and an exported function can be called via api.Function. api.Function allows users to supply parameters and read results as a slice of uint64. For example, if there are no results, an empty slice is returned. The user can learn the signature via FunctionDescription, which returns the api.ValueType corresponding to each parameter or result. api.ValueType defines the mapping of WebAssembly types to uint64 values for reason described in this section. The special case of v128 is also mentioned below.

wazero maps each value type to a uint64 values because it holds the largest type in WebAssembly 1.0 (i64). A slice allows you to express empty (e.g. a nullary signature), for example a start function.

Here's an example of calling a function, noting this syntax works for both a signature (param i32 i32) (result i32) and (param i64 i64) (result i64)

x, y := uint64(1), uint64(2)
results, err := mod.ExportedFunction("add").Call(ctx, x, y)
if err != nil {
	log.Panicln(err)
}
fmt.Printf("%d + %d = %d\n", x, y, results[0])

WebAssembly does not define an encoding strategy for host defined parameters or results. This means the encoding rules above are defined by wazero instead. To address this, we clarified mapping both in api.ValueType and added helper functions like api.EncodeF64. This allows users conversions typical in Go programming, and utilities to avoid ambiguity and edge cases around casting.

Alternatively, we could have defined a byte buffer based approach and a binary encoding of value types in and out. For example, an empty byte slice would mean no values, while a non-empty could use a binary encoding for supported values. This could work, but it is more difficult for the normal case of i32 and i64. It also shares a struggle with the current approach, which is that value types were added after WebAssembly 1.0 and not all of them have an encoding. More on this below.

In summary, wazero chose an approach for signature mapping because there was none, and the one we chose biases towards simplicity with integers and handles the rest with documentation and utilities.

Post 1.0 value types

Value types added after WebAssembly 1.0 stressed the current model, as some have no encoding or are larger than 64 bits. While problematic, these value types are not commonly used in exported (extern) functions. However, some decisions were made and detailed below.

For example externref has no guest representation. wazero chose to map references to uint64 as that's the largest value needed to encode a pointer on supported platforms. While there are two reference types, externref and functype, the latter is an internal detail of function tables, and the former is rarely if ever used in function signatures as of the end of 2022.

The only value larger than 64 bits is used for SIMD (v128). Vectorizing via host functions is not used as of the end of 2022. Even if it were, it would be inefficient vs guest vectorization due to host function overhead. In other words, the v128 value type is unlikely to be in an exported function signature. That it requires two uint64 values to encode is an internal detail and not worth changing the exported function interface api.Function, as doing so would break all users.

Interfaces, not structs

All exported types in public packages, regardless of configuration vs runtime, are interfaces. The primary benefits are internal flexibility and avoiding people accidentally mis-initializing by instantiating the types on their own vs using the NewXxx constructor functions. In other words, there's less support load when things can't be done incorrectly.

Here's an example:

rt := &RuntimeConfig{} // not initialized properly (fields are nil which shouldn't be)
rt := RuntimeConfig{} // not initialized properly (should be a pointer)
rt := wazero.NewRuntimeConfig() // initialized properly

There are a few drawbacks to this, notably some work for maintainers.

  • Interfaces are decoupled from the structs implementing them, which means the signature has to be repeated twice.
  • Interfaces have to be documented and guarded at time of use, that 3rd party implementations aren't supported.
  • As of Golang 1.21, interfaces are still not well supported in godoc.

Config

wazero configures scopes such as Runtime and Module using XxxConfig types. For example, RuntimeConfig configures Runtime and ModuleConfig configure Module (instantiation). In all cases, config types begin defaults and can be customized by a user, e.g., selecting features or a module name override.

Why don't we make each configuration setting return an error?

No config types create resources that would need to be closed, nor do they return errors on use. This helps reduce resource leaks, and makes chaining easier. It makes it possible to parse configuration (ex by parsing yaml) independent of validating it.

Instead of:

cfg, err = cfg.WithFS(fs)
if err != nil {
  return err
}
cfg, err = cfg.WithName(name)
if err != nil {
  return err
}
mod, err = rt.InstantiateModuleWithConfig(ctx, code, cfg)
if err != nil {
  return err
}

There's only one call site to handle errors:

cfg = cfg.WithFS(fs).WithName(name)
mod, err = rt.InstantiateModuleWithConfig(ctx, code, cfg)
if err != nil {
  return err
}

This allows users one place to look for errors, and also the benefit that if anything internally opens a resource, but errs, there's nothing they need to close. In other words, users don't need to track which resources need closing on partial error, as that is handled internally by the only code that can read configuration fields.

Why are configuration immutable?

While it seems certain scopes like Runtime won't repeat within a process, they do, possibly in different goroutines. For example, some users create a new runtime for each module, and some re-use the same base module configuration with only small updates (ex the name) for each instantiation. Making configuration immutable allows them to be safely used in any goroutine.

Since config are immutable, changes apply via return val, similar to append in a slice.

For example, both of these are the same sort of error:

append(slice, element) // bug as only the return value has the updated slice.
cfg.WithName(next) // bug as only the return value has the updated name.

Here's an example of correct use: re-assigning explicitly or via chaining.

cfg = cfg.WithName(name) // explicit

mod, err = rt.InstantiateModuleWithConfig(ctx, code, cfg.WithName(name)) // implicit
if err != nil {
  return err
}

Why aren't configuration assigned with option types?

The option pattern is a familiar one in Go. For example, someone defines a type func (x X) err and uses it to update the target. For example, you could imagine wazero could choose to make ModuleConfig from options vs chaining fields.

Ex instead of:

type ModuleConfig interface {
	WithName(string) ModuleConfig
	WithFS(fs.FS) ModuleConfig
}

struct moduleConfig {
	name string
	fs fs.FS
}

func (c *moduleConfig) WithName(name string) ModuleConfig {
    ret := *c // copy
    ret.name = name
    return &ret
}

func (c *moduleConfig) WithFS(fs fs.FS) ModuleConfig {
    ret := *c // copy
    ret.setFS("/", fs)
    return &ret
}

config := r.NewModuleConfig().WithFS(fs)
configDerived := config.WithName("name")

An option function could be defined, then refactor each config method into an name prefixed option function:

type ModuleConfig interface {
}
struct moduleConfig {
    name string
    fs fs.FS
}

type ModuleConfigOption func(c *moduleConfig)

func ModuleConfigName(name string) ModuleConfigOption {
    return func(c *moduleConfig) {
        c.name = name
	}
}

func ModuleConfigFS(fs fs.FS) ModuleConfigOption {
    return func(c *moduleConfig) {
        c.fs = fs
    }
}

func (r *runtime) NewModuleConfig(opts ...ModuleConfigOption) ModuleConfig {
	ret := newModuleConfig() // defaults
    for _, opt := range opts {
        opt(&ret.config)
    }
    return ret
}

func (c *moduleConfig) WithOptions(opts ...ModuleConfigOption) ModuleConfig {
    ret := *c // copy base config
    for _, opt := range opts {
        opt(&ret.config)
    }
    return ret
}

config := r.NewModuleConfig(ModuleConfigFS(fs))
configDerived := config.WithOptions(ModuleConfigName("name"))

wazero took the path of the former design primarily due to:

  • interfaces provide natural namespaces for their methods, which is more direct than functions with name prefixes.
  • parsing config into function callbacks is more direct vs parsing config into a slice of functions to do the same.
  • in either case derived config is needed and the options pattern is more awkward to achieve that.

There are other reasons such as test and debug being simpler without options: the above list is constrained to conserve space. It is accepted that the options pattern is common in Go, which is the main reason for documenting this decision.

Why aren't config types deeply structured?

wazero's configuration types cover the two main scopes of WebAssembly use:

  • RuntimeConfig: This is the broadest scope, so applies also to compilation and instantiation. e.g. This controls the WebAssembly Specification Version.
  • ModuleConfig: This affects modules instantiated after compilation and what resources are allowed. e.g. This defines how or if STDOUT is captured. This also allows sub-configuration of FSConfig.

These default to a flat definition each, with lazy sub-configuration only after proven to be necessary. A flat structure is easier to work with and is also easy to discover. Unlike the option pattern described earlier, more configuration in the interface doesn't taint the package namespace, only ModuleConfig.

We default to a flat structure to encourage simplicity. If we eagerly broke out all possible configurations into sub-types (e.g. ClockConfig), it would be hard to notice configuration sprawl. By keeping the config flat, it is easy to see the cognitive load we may be adding to our users.

In other words, discomfort adding more configuration is a feature, not a bug. We should only add new configuration rarely, and before doing so, ensure it will be used. In fact, this is why we support using context fields for experimental configuration. By letting users practice, we can find out if a configuration was a good idea or not before committing to it, and potentially sprawling our types.

In reflection, this approach worked well for the nearly 1.5 year period leading to version 1.0. We've only had to create a single sub-configuration, FSConfig, and it was well understood why when it occurred.

Why does ModuleConfig.WithStartFunctions default to _start?

We formerly had functions like StartWASICommand that would verify preconditions and start WASI's _start command. However, this caused confusion because both many languages compiled a WASI dependency, and many did so inconsistently.

The conflict is that exported functions need to use features the language runtime provides, such as garbage collection. There's a "chicken-egg problem" where _start needs to complete in order for exported behavior to work.

For example, unlike GOOS=wasip1 in Go 1.21, TinyGo's "wasi" target supports function exports. So, the only way to use FFI style is via the "wasi" target. Not explicitly calling _start before an ABI such as wapc-go, would crash, due to setup not happening (e.g. to implement panic). Other embedders such as Envoy also called _start for the same reason. To avoid a common problem for users unaware of WASI, and also to simplify normal use of WASI (e.g. main), we added _start to ModuleConfig.WithStartFunctions.

In cases of multiple initializers, such as in wapc-go, users can override this to add the others after _start. Users who want to explicitly control _start, such as some of our unit tests, can clear the start functions and remove it.

This decision was made in 2022, and holds true in 2023, even with the introduction of "wasix". It holds because "wasix" is backwards compatible with "wasip1". In the future, there will be other ways to start applications, and may not be backwards compatible with "wasip1".

Most notably WASI "Preview 2" is not implemented in a way compatible with wasip1. Its start function is likely to be different, and defined in the wasi-cli "world". When the design settles, and it is implemented by compilers, wazero will attempt to support "wasip2". However, it won't do so in a way that breaks existing compilers.

In other words, we won't remove _start if "wasip2" continues a path of an alternate function name. If we did, we'd break existing users despite our compatibility promise saying we don't. The most likely case is that when we build-in something incompatible with "wasip1", that start function will be added to the start functions list in addition to _start.

See http://wasix.org See https://github.com/WebAssembly/wasi-cli

Runtime == Engine+Store

wazero defines a single user-type which combines the specification concept of Store with the unspecified Engine which manages them.

Why not multi-store?

Multi-store isn't supported as the extra tier complicates lifecycle and locking. Moreover, in practice it is unusual for there to be an engine that has multiple stores which have multiple modules. More often, it is the case that there is either 1 engine with 1 store and multiple modules, or 1 engine with many stores, each having 1 non-host module. In worst case, a user can use multiple runtimes until "multi-store" is better understood.

If later, we have demand for multiple stores, that can be accomplished by overload. e.g. Runtime.InstantiateInStore or Runtime.Store(name) Store.

Exit

Why do we only return a sys.ExitError on a non-zero exit code?

It is reasonable to think an exit error should be returned, even if the code is success (zero). Even on success, the module is no longer functional. For example, function exports would error later. However, wazero does not. The only time sys.ExitError is on error (non-zero).

This decision was to improve performance and ergonomics for guests that both use WASI (have a _start function), and also allow custom exports. Specifically, Rust, TinyGo and normal wasi-libc, don't exit the module during _start. If they did, it would invalidate their function exports. This means it is unlikely most compilers will change this behavior.

GOOS=waspi1 from Go 1.21 does exit during _start. However, it doesn't support other exports besides _start, and _start is not defined to be called multiple times anyway.

Since sys.ExitError is not always returned, we added Module.IsClosed for defensive checks. This helps integrators avoid calling functions which will always fail.

Why panic with sys.ExitError after a host function exits?

Currently, the only portable way to stop processing code is via panic. For example, WebAssembly "trap" instructions, such as divide by zero, are implemented via panic. This ensures code isn't executed after it.

When code reaches the WASI proc_exit instruction, we need to stop processing. Regardless of the exit code, any code invoked after exit would be in an inconsistent state. This is likely why unreachable instructions are sometimes inserted after exit: emscripten-core/emscripten#12322

WASI

Unfortunately, (WASI Snapshot Preview 1)[https://github.com/WebAssembly/WASI/blob/snapshot-01/phases/snapshot/docs.md] is not formally defined enough, and has APIs with ambiguous semantics. This section describes how Wazero interprets and implements the semantics of several WASI APIs that may be interpreted differently by different wasm runtimes. Those APIs may affect the portability of a WASI application.

Why don't we attempt to pass wasi-testsuite on user-defined fs.FS?

While most cases work fine on an os.File based implementation, we won't promise wasi-testsuite compatibility on user defined wrappers of os.DirFS. The only option for real systems is to use our sysfs.FS.

There are a lot of areas where windows behaves differently, despite the os.File abstraction. This goes well beyond file locking concerns (e.g. EBUSY errors on open files). For example, errors like ACCESS_DENIED aren't properly mapped to EPERM. There are trickier parts too. FileInfo.Sys() doesn't return enough information to build inodes needed for WASI. To rebuild them requires the full path to the underlying file, not just its directory name, and there's no way for us to get that information. At one point we tried, but in practice things became tangled and functionality such as read-only wrappers became untenable. Finally, there are version-specific behaviors which are difficult to maintain even in our own code. For example, go 1.20 opens files in a different way than versions before it.

Why aren't WASI rules enforced?

The snapshot-01 version of WASI has a number of rules for a "command module", but only the memory export rule is enforced. If a "_start" function exists, it is enforced to be the correct signature and succeed, but the export itself isn't enforced. It follows that this means exports are not required to be contained to a "_start" function invocation. Finally, the "__indirect_function_table" export is also not enforced.

The reason for the exceptions are that implementations aren't following the rules. For example, TinyGo doesn't export "__indirect_function_table", so crashing on this would make wazero unable to run TinyGo modules. Similarly, modules loaded by wapc-go don't always define a "_start" function. Since "snapshot-01" is not a proper version, and certainly not a W3C recommendation, there's no sense in breaking users over matters like this.

Why is I/O configuration not coupled to WASI?

WebAssembly System Interfaces (WASI) is a formalization of a practice that can be done anyway: Define a host function to access a system interface, such as writing to STDOUT. WASI stalled at snapshot-01 and as of early 2023, is being rewritten entirely.

This instability implies a need to transition between WASI specs, which places wazero in a position that requires decoupling. For example, if code uses two different functions to call fd_write, the underlying configuration must be centralized and decoupled. Otherwise, calls using the same file descriptor number will end up writing to different places.

In short, wazero defined system configuration in ModuleConfig, not a WASI type. This allows end-users to switch from one spec to another with minimal impact. This has other helpful benefits, as centralized resources are simpler to close coherently (ex via Module.Close).

In reflection, this worked well as more ABI became usable in wazero.

Background on ModuleConfig design

WebAssembly 1.0 (20191205) specifies some aspects to control isolation between modules (sandboxing). For example, wasm.Memory has size constraints and each instance of it is isolated from each other. While wasm.Memory can be shared, by exporting it, it is not exported by default. In fact a WebAssembly Module (Wasm) has no memory by default.

While memory is defined in WebAssembly 1.0 (20191205), many aspects are not. Let's use an example of exec.Cmd as for example, a WebAssembly System Interfaces (WASI) command is implemented as a module with a _start function, and in many ways acts similar to a process with a main function.

To capture "hello world" written to the console (stdout a.k.a. file descriptor 1) in exec.Cmd, you would set the Stdout field accordingly, perhaps to a buffer. In WebAssembly 1.0 (20191205), the only way to perform something like this is via a host function (ex HostModuleFunctionBuilder) and internally copy memory corresponding to that string to a buffer.

WASI implements system interfaces with host functions. Concretely, to write to console, a WASI command Module imports "fd_write" from "wasi_snapshot_preview1" and calls it with the fd parameter set to 1 (STDOUT).

The snapshot-01 version of WASI has no means to declare configuration, although its function definitions imply configuration for example if fd 1 should exist, and if so where should it write. Moreover, snapshot-01 was last updated in late 2020 and the specification is being completely rewritten as of early 2022. This means WASI as defined by "snapshot-01" will not clarify aspects like which file descriptors are required. While it is possible a subsequent version may, it is too early to tell as no version of WASI has reached a stage near W3C recommendation. Even if it did, module authors are not required to only use WASI to write to console, as they can define their own host functions, such as they did before WASI existed.

wazero aims to serve Go developers as a primary function, and help them transition between WASI specifications. In order to do this, we have to allow top-level configuration. To ensure isolation by default, ModuleConfig has WithXXX that override defaults to no-op or empty. One ModuleConfig instance is used regardless of how many times the same WASI functions are imported. The nil defaults allow safe concurrency in these situations, as well lower the cost when they are never used. Finally, a one-to-one mapping with Module allows the module to close the ModuleConfig instead of confusing users with another API to close.

Naming, defaults and validation rules of aspects like STDIN and Environ are intentionally similar to other Go libraries such as exec.Cmd or syscall.SetEnv, and differences called out where helpful. For example, there's no goal to emulate any operating system primitive specific to Windows (such as a 'c:' drive). Moreover, certain defaults working with real system calls are neither relevant nor safe to inherit: For example, exec.Cmd defaults to read STDIN from a real file descriptor ("/dev/null"). Defaulting to this, vs reading io.EOF, would be unsafe as it can exhaust file descriptors if resources aren't managed properly. In other words, blind copying of defaults isn't wise as it can violate isolation or endanger the embedding process. In summary, we try to be similar to normal Go code, but often need act differently and document ModuleConfig is more about emulating, not necessarily performing real system calls.

File systems

Motivation on sys.FS

The sys.FS abstraction in wazero was created because of limitations in fs.FS, and fs.File in Go. Compilers targeting wasip1 may access functionality that writes new files. The ability to overcome this was requested even before wazero was named this, via issue #21 in March 2021.

A month later, golang/go#45757 was raised by someone else on the same topic. As of July 2023, this has not resolved to a writeable file system abstraction.

Over the next year more use cases accumulated, consolidated in March 2022 into #390. This closed in January 2023 with a milestone of providing more functionality, limited to users giving a real directory. This didn't yet expose a file abstraction for general purpose use. Internally, this used os.File. However, a wasm module instance is a virtual machine. Only supporting os.File breaks sand-boxing use cases. Moreover, os.File is not an interface. Even though this abstracts functionality, it does allow interception use cases.

Hence, a few days later in January 2023, we had more issues asking to expose an abstraction, #1013 and later #1532, on use cases like masking access to files. In other words, the use case requests never stopped, and aren't solved by exposing only real files.

In summary, the primary motivation for exposing a replacement for fs.FS and fs.File was around repetitive use case requests for years, around interception and the ability to create new files, both virtual and real files. While some use cases are solved with real files, not all are. Regardless, an interface approach is necessary to ensure users can intercept I/O operations.

Why doesn't sys.File have a Fd() method?

There are many features we could expose. We could make File expose underlying file descriptors in case they are supported, for integration of system calls that accept multiple ones, namely poll for multiplexing. This special case is described in a subsequent section.

As noted above, users have been asking for a file abstraction for over two years, and a common answer was to wait. Making users wait is a problem, especially so long. Good reasons to make people wait are stabilization. Edge case features are not a great reason to hold abstractions from users.

Another reason is implementation difficulty. Go did not attempt to abstract file descriptors. For example, unlike fs.ReadFile there is no fs.FdFile interface. Most likely, this is because file descriptors are an implementation detail of common features. Programming languages, including Go, do not require end users to know about file descriptors. Types such as fs.File can be used without any knowledge of them. Implementations may or may not have file descriptors. For example, in Go, os.DirFS has underlying file descriptors while embed.FS does not.

Despite this, some may want to expose a non-standard interface because os.File has Fd() uintptr to return a file descriptor. Mainly, this is handy to integrate with syscall package functions (on GOOS values that declare them). Notice, though that uintptr is unsafe and not an abstraction. Close inspection will find some os.File types internally use poll.FD instead, yet this is not possible to use abstractly because that type is not exposed. For example, plan9 uses a different type than poll.FD. In other words, even in real files, Fd() is not wholly portable, despite it being useful on many operating systems with the syscall package.

The reasons above, why Go doesn't abstract FdFile interface are a subset of reasons why sys.File does not. If we exposed File.Fd() we not only would have to declare all the edge cases that Go describes including impact of finalizers, we would have to describe these in terms of virtualized files. Then, we would have to reason with this value vs our existing virtualized sys.FileTable, mapping whatever type we return to keys in that table, also in consideration of garbage collection impact. The combination of issues like this could lead down a path of not implementing a file system abstraction at all, and instead a weak key mapped abstraction of the syscall package. Once we finished with all the edge cases, we would have lost context of the original reason why we started.. simply to allow file write access!

When wazero attempts to do more than what the Go programming language team, it has to be carefully evaluated, to:

  • Be possible to implement at least for os.File backed files
  • Not be confusing or cognitively hard for virtual file systems and normal use.
  • Affordable: custom code is solely the responsible by the core team, a much smaller group of individuals than who maintain the Go programming language.

Due to problems well known in Go, consideration of the end users who constantly ask for basic file system functionality, and the difficulty virtualizing file descriptors at multiple levels, we don't expose Fd() and likely won't ever expose Fd() on sys.File.

Why does sys.File have a Poll() method, while sys.FS does not?

wazero exposes File.Poll which allows one-at-a-time poll use cases, requested by multiple users. This not only includes abstract tests such as Go 1.21 GOOS=wasip1, but real use cases including python and container2wasm repls, as well listen sockets. The main use cases is non-blocking poll on a single file. Being a single file, this has no risk of problems such as head-of-line blocking, even when emulated.

The main use case of multi-poll are bidirectional network services, something not used in GOOS=wasip1 standard libraries, but could be in the future. Moving forward without a multi-poller allows wazero to expose its file system abstraction instead of continuing to hold back it back for edge cases. We'll continue discussion below regardless, as rationale was requested.

You can loop through multiple sys.File, using File.Poll to see if an event is ready, but there is a head-of-line blocking problem. If a long timeout is used, bad luck could have a file that has nothing to read or write before one that does. This could cause more blocking than necessary, even if you could poll the others just after with a zero timeout. What's worse than this is if unlimited blocking was used (timeout=-1). The host implementations could use goroutines to avoid this, but interrupting a "forever" poll is problematic. All of these are reasons to consider a multi-poll API, but do not require exporting File.Fd().

Should multi-poll becomes critical, sys.FS could expose a Poll function like below, despite it being the non-portable, complicated if possible to implement on all platforms and virtual file systems.

ready, errno := fs.Poll([]sys.PollFile{{f1, sys.POLLIN}, {f2, sys.POLLOUT}}, timeoutMillis)

A real filesystem could handle this by using an approach like the internal unix.Poll function in Go, passing file descriptors on unix platforms, or returning sys.ENOSYS for unsupported operating systems. Implementation for virtual files could have a strategy around timeout to avoid the worst case of head-of-line blocking (unlimited timeout).

Let's remember that when designing abstractions, it is not best to add an interface for everything. Certainly, Go doesn't, as evidenced by them not exposing poll.FD in os.File! Such a multi-poll could be limited to built-in filesystems in the wazero repository, avoiding complexity of trying to support and test this abstractly. This would still permit multiplexing for CLI users, and also permit single file polling as exists now.

Why doesn't wazero implement the working directory?

An early design of wazero's API included a WithWorkDirFS which allowed control over which file a relative path such as "./config.yml" resolved to, independent of the root file system. This intended to help separate concerns like mutability of files, but it didn't work and was removed.

Compilers that target wasm act differently with regard to the working directory. For example, wasi-libc, used by TinyGo, tracks working directory changes in compiled wasm instead: initially "/" until code calls chdir. Zig assumes the first pre-opened file descriptor is the working directory.

The only place wazero can standardize a layered concern is via a host function. Since WASI doesn't use host functions to track the working directory, we can't standardize the storage and initial value of it.

Meanwhile, code may be able to affect the working directory by compiling chdir into their main function, using an argument or ENV for the initial value (possibly PWD). Those unable to control the compiled code should only use absolute paths in configuration.

See

Why ignore the error returned by io.Reader when n > 1?

Per https://pkg.go.dev/io#Reader, if we receive an error, any bytes read should be processed first. At the syscall abstraction (fd_read), the caller is the processor, so we can't process the bytes inline and also return the error (as EIO).

Let's assume we want to return the bytes read on error to the caller. This implies we at least temporarily ignore the error alongside them. The choice remaining is whether to persist the error returned with the read until a possible next call, or ignore the error.

If we persist an error returned, it would be coupled to a file descriptor, but effectively it is boolean as this case coerces to EIO. If we track a "last error" on a file descriptor, it could be complicated for a couple reasons including whether the error is transient or permanent, or if the error would apply to any FD operation, or just read. Finally, there may never be a subsequent read as perhaps the bytes leading up to the error are enough to satisfy the processor.

This decision boils down to whether or not to track an error bit per file descriptor or not. If not, the assumption is that a subsequent operation would also error, this time without reading any bytes.

The current opinion is to go with the simplest path, which is to return the bytes read and ignore the error the there were any. Assume a subsequent operation will err if it needs to. This helps reduce the complexity of the code in wazero and also accommodates the scenario where the bytes read are enough to satisfy its processor.

File descriptor allocation strategy

File descriptor allocation currently uses a strategy similar the one implemented by unix systems: when opening a file, the lowest unused number is picked.

The WASI standard documents that programs cannot expect that file descriptor numbers will be allocated with a lowest-first strategy, and they should instead assume the values will be random. Since random is a very imprecise concept in computers, we technically satisfying the implementation with the descriptor allocation strategy we use in Wazero. We could imagine adding more randomness to the descriptor selection process, however this should never be used as a security measure to prevent applications from guessing the next file number so there are no strong incentives to complicate the logic.

Why does FSConfig.WithDirMount not match behaviour with os.DirFS?

It may seem that we should require any feature that seems like a standard library in Go, to behave the same way as the standard library. Doing so would present least surprise to Go developers. In the case of how we handle filesystems, we break from that as it is incompatible with the expectations of WASI, the most commonly implemented filesystem ABI.

The main reason is that os.DirFS is a virtual filesystem abstraction while WASI is an abstraction over syscalls. For example, the signature of fs.Open does not permit use of flags. This creates conflict on what default behaviors to take when Go implemented os.DirFS. On the other hand, path_open can pass flags, and in fact tests require them to be honored in specific ways.

This conflict requires us to choose what to be more compatible with, and which type of user to surprise the least. We assume there will be more developers compiling code to wasm than developers of custom filesystem plugins, and those compiling code to wasm will be better served if we are compatible with WASI. Hence on conflict, we prefer WASI behavior vs the behavior of os.DirFS.

See https://github.com/WebAssembly/wasi-testsuite See golang/go#58141

Why is our Readdir function more like Go's os.File than POSIX readdir?

At one point we attempted to move from a bulk Readdir function to something more like the POSIX DIR struct, exposing functions like telldir, seekdir and readdir. However, we chose the design more like os.File.Readdir, because it performs and fits wasip1 better.

wasip1/wasix

fd_readdir in wasip1 (and so also wasix) is like getdents in Linux, not readdir in POSIX. getdents is more like Go's os.File.Readdir.

We currently have an internal type sys.DirentCache which only is used by wasip1 or wasix. When HostModuleBuilder adds support for instantiation state, we could move this to the wasi_snapshot_preview1 package. Meanwhile, all filesystem code is internal anyway, so this special-case is acceptable.

wasip2

directory-entry-stream in wasi-filesystem preview2 is defined in component model, not an ABI, but in wasmtime it is a consuming iterator. A consuming iterator is easy to support with anything (like Readdir(1)), even if it is inefficient as you can neither bulk read nor skip. The implementation of the preview1 adapter (uses preview2) confirms this. They use a dirent cache similar in some ways to our sysfs.DirentCache. As there is no seek concept in preview2, they interpret the cookie as numeric and read on repeat entries when a cache wasn't available. Note: we currently do not skip-read like this as it risks buffering large directories, and no user has requested entries before the cache, yet.

Regardless, wasip2 is not complete until the end of 2023. We can defer design discussion until after it is stable and after the reference impl wasmtime implements it.

See

wasip3

directory-entry-stream is documented to change significantly in wasip3 moving from synchronous to synchronous streams. This is dramatically different than POSIX readdir which is synchronous.

Regardless, wasip3 is not complete until after wasip2, which means 2024 or later. We can defer design discussion until after it is stable and after the reference impl wasmtime implements it.

See

How do we implement Pread with an fs.File?

ReadAt is the Go equivalent to pread: it does not affect, and is not affected by, the underlying file offset. Unfortunately, io.ReaderAt is not implemented by all fs.File. For example, as of Go 1.19, embed.openFile does not.

The initial implementation of fd_pread instead used Seek. To avoid a regression, we fall back to io.Seeker when io.ReaderAt is not supported.

This requires obtaining the initial file offset, seeking to the intended read offset, and resetting the file offset the initial state. If this final seek fails, the file offset is left in an undefined state. This is not thread-safe.

While seeking per read seems expensive, the common case of embed.openFile is only accessing a single int64 field, which is cheap.

Pre-opened files

WASI includes fd_prestat_get and fd_prestat_dir_name functions used to learn any directory paths for file descriptors open at initialization time.

For example, __wasilibc_register_preopened_fd scans any file descriptors past STDERR (1) and invokes fd_prestat_dir_name to learn any path prefixes they correspond to. Zig's preopensAlloc does similar. These pre-open functions are not used again after initialization.

wazero supports stdio pre-opens followed by any mounts e.g .:/. The guest path is a directory and its name, e.g. "/" is returned by fd_prestat_dir_name for file descriptor 3 (STDERR+1). The first longest match wins on multiple pre-opens, which allows a path like "/tmp" to match regardless of order vs "/".

See

fd_prestat_dir_name

fd_prestat_dir_name is a WASI function to return the path of the pre-opened directory of a file descriptor. It has the following three parameters, and the third path_len has ambiguous semantics.

  • fd: a file descriptor
  • path: the offset for the result path
  • path_len: In wazero, FdPrestatDirName writes the result path string to path offset for the exact length of path_len.

Wasmer considers path_len to be the maximum length instead of the exact length that should be written. See https://github.com/wasmerio/wasmer/blob/3463c51268ed551933392a4063bd4f8e7498b0f6/lib/wasi/src/syscalls/mod.rs#L764

The semantics in wazero follows that of wasmtime. See https://github.com/bytecodealliance/wasmtime/blob/2ca01ae9478f199337cf743a6ab543e8c3f3b238/crates/wasi-common/src/snapshots/preview_1.rs#L578-L582

Their semantics match when path_len == the length of path, so in practice this difference won't matter match.

fd_readdir

Why does "wasi_snapshot_preview1" require dot entries when POSIX does not?

In October 2019, WASI project knew requiring dot entries ("." and "..") was not documented in preview1, not required by POSIX and problematic to synthesize. For example, Windows runtimes backed by FindNextFileW could not return these. A year later, the tag representing WASI preview 1 (snapshot-01) was made. This did not include the requested change of making dot entries optional.

The phases/snapshot/docs.md document was altered in subsequent years in significant ways, often in lock-step with wasmtime or wasi-libc. In January 2022, sock_accept was added to phases/snapshot/docs.md, a document later renamed to later renamed to legacy/preview1/docs.md.

As a result, the ABI and behavior remained unstable: The snapshot-01 tag was not an effective basis of portability. A test suite was requested well before this tag, in April 2019. Meanwhile, compliance had no meaning. Developers had to track changes to the latest doc, while clarifying with wasi-libc or wasmtime behavior. This lack of stability could have permitted a fix to the dot entries problem, just as it permitted changes desired by other users.

In November 2022, the wasi-testsuite project began and started solidifying expectations. This quickly led to changes in runtimes and the spec doc. WASI began importing tests from wasmtime as required behaviors for all runtimes. Some changes implied changes to wasi-libc. For example, readdir began to imply inode fan-outs, which caused performance regressions. Most notably a test merged in January required dot entries. Tests were merged without running against any runtime, and even when run ad-hoc only against Linux. Hence, portability issues mentioned over three years earlier did not trigger any failure until wazero (which tests Windows) noticed.

In the same month, wazero requested to revert this change primarily because Go does not return them from os.ReadDir, and materializing them is complicated due to tests also requiring inodes. Moreover, they are discarded by not just Go, but other common programming languages. This was rejected by the WASI lead for preview1, but considered for the completely different ABI named preview2.

In February 2023, the WASI chair declared that new rule requiring preview1 to return dot entries "was decided by the subgroup as a whole", citing meeting notes. According to these notes, the WASI lead stated incorrectly that POSIX conformance required returning dot entries, something it explicitly says are optional. In other words, he said filtering them out would make Preview1 non-conforming, and asked if anyone objects to this. The co-chair was noted to say "Because there are existing P1 programs, we shouldn’t make changes like this." No other were recorded to say anything.

In summary, preview1 was changed retrospectively to require dot entries and preview2 was changed to require their absence. This rule was reverse engineered from wasmtime tests, and affirmed on two false premises:

  • POSIX compliance requires dot entries
    • POSIX literally says these are optional
  • WASI cannot make changes because there are existing P1 programs.
    • Changes to Preview 1 happened before and after this topic.

As of June 2023, wasi-testsuite still only runs on Linux, so compliance of this rule on Windows is left to runtimes to decide to validate. The preview2 adapter uses fake cookies zero and one to refer to dot dirents, uses a real inode for the dot(".") entry and zero inode for dot-dot("..").

See WebAssembly/wasi-filesystem#3 See https://github.com/WebAssembly/WASI/tree/snapshot-01 See WebAssembly/WASI#9 See WebAssembly/WASI#458 See WebAssembly/wasi-testsuite#32 See WebAssembly/wasi-libc#345 See WebAssembly/wasi-testsuite#52 See WebAssembly/WASI#516 See https://github.com/WebAssembly/meetings/blob/main/wasi/2023/WASI-02-09.md#should-preview1-fd_readdir-filter-out--and- See https://github.com/bytecodealliance/preview2-prototyping/blob/e4c04bcfbd11c42c27c28984948d501a3e168121/crates/wasi-preview1-component-adapter/src/lib.rs#L1026-L1041

Why are dot (".") and dot-dot ("..") entries problematic?

When reading a directory, dot (".") and dot-dot ("..") entries are problematic. For example, Go does not return them from os.ReadDir, and materializing them is complicated (at least dot-dot is).

A directory entry has stat information in it. The stat information includes inode which is used for comparing file equivalence. In the simple case of dot, we could materialize a special entry to expose the same info as stat on the fd would return. However, doing this and not doing dot-dot would cause confusion, and dot-dot is far more tricky. To back-fill inode information about a parent directory would be costly and subtle. For example, the pre-open (mount) of the directory may be different than its logical parent. This is easy to understand when considering the common case of mounting "/" and "/tmp" as pre-opens. To implement ".." from "/tmp" requires information from a separate pre-open, this includes state to even know the difference. There are easier edge cases as well, such as the decision to not return ".." from a root path. In any case, this should start to explain that faking entries when underlying stdlib doesn't return them is tricky and requires quite a lot of state.

Another issue is around the Dirent.Off value of a directory entry, sometimes called a "cookie" in Linux man pagers. When the host operating system or library function does not return dot entries, to support functions such as seekdir, you still need a value for Dirent.Off. Naively, you can synthesize these by choosing sequential offsets zero and one. However, POSIX strictly says offsets should be treated opaquely. The backing filesystem could use these to represent real entries. For example, a directory with one entry could use zero as the Dirent.Off value. If you also used zero for the "." dirent, there would be a clash. This means if you synthesize Dirent.Off for any entry, you need to synthesize this value for all entries. In practice, the simplest way is using an incrementing number, such as done in the WASI preview2 adapter.

Working around these issues causes expense to all users of wazero, so we'd then look to see if that would be justified or not. However, the most common compilers involved in end user questions, as of early 2023 are TinyGo, Rust and Zig. All of these compile code which ignores dot and dot-dot entries. In other words, faking these entries would not only cost our codebase with complexity, but it would also add unnecessary overhead as the values aren't commonly used.

The final reason why we might do this, is an end users or a specification requiring us to. As of early 2023, no end user has raised concern over Go and by extension wazero not returning dot and dot-dot. The snapshot-01 spec of WASI does not mention anything on this point. Also, POSIX has the following to say, which summarizes to "these are optional"

The readdir() function shall not return directory entries containing empty names. If entries for dot or dot-dot exist, one entry shall be returned for dot and one entry shall be returned for dot-dot; otherwise, they shall not be returned.

Unfortunately, as described above, the WASI project decided in early 2023 to require dot entries in both the spec and the wasi-testsuite. For only this reason, wazero adds overhead to synthesize dot entries despite it being unnecessary for most users.

See https://pubs.opengroup.org/onlinepubs/9699919799/functions/readdir.html See https://github.com/golang/go/blob/go1.20/src/os/dir_unix.go#L108-L110 See https://github.com/bytecodealliance/preview2-prototyping/blob/e4c04bcfbd11c42c27c28984948d501a3e168121/crates/wasi-preview1-component-adapter/src/lib.rs#L1026-L1041

Why don't we pre-populate an inode for the dot-dot ("..") entry?

We only populate an inode for dot (".") because wasi-testsuite requires it, and we likely already have it (because we cache it). We could attempt to populate one for dot-dot (".."), but chose not to.

Firstly, wasi-testsuite does not require the inode of dot-dot, possibly because the wasip2 adapter doesn't populate it (but we don't really know why).

The only other reason to populate it would be to avoid wasi-libc's stat fanout when it is missing. However, wasi-libc explicitly doesn't fan-out to lstat on the ".." entry on a zero ino.

Fetching dot-dot's inode despite the above not only doesn't help wasi-libc, but it also hurts languages that don't use it, such as Go. These languages would pay a stat syscall penalty even if they don't need the inode. In fact, Go discards both dot entries!

In summary, there are no significant upsides in attempting to pre-fetch dot-dot's inode, and there are downsides to doing it anyway.

See

Why don't we require inodes to be non-zero?

We don't require a non-zero value for Dirent.Ino because doing so can prevent a real one from resolving later via Stat_t.Ino.

We define Ino like d_ino in POSIX which doesn't special-case zero. It can be zero for a few reasons:

  • The file is not a regular file or directory.
  • The underlying filesystem does not support inodes. e.g. embed:fs
  • A directory doesn't include inodes, but a later stat can. e.g. Windows
  • The backend is based on wasi-filesystem (a.k.a wasip2), which has directory_entry.inode optional, and might remove it entirely.

There are other downsides to returning a zero inode in widely used compilers:

  • File equivalence utilities, like os.SameFile will not work.
  • wasi-libc's wasip1 mode will call lstat and attempt to retrieve a non-zero value (unless the entry is named "..").

A new compiler may accidentally skip a Dirent with a zero Ino if emulating a non-POSIX function and re-using Dirent.Ino for d_fileno.

  • Linux getdents doesn't define d_fileno must be non-zero
  • BSD getdirentries is implementation specific. For example, OpenBSD will return dirents with a zero d_fileno, but Darwin will skip them.

The above shouldn't be a problem, even in the case of BSD, because wasip1 is defined more in terms of getdents than getdirentries. The bottom half of either should treat wasip1 (or any similar ABI such as wasix or wasip2) as a different operating system and either use different logic that doesn't skip, or synthesize a fake non-zero d_fileno when d_ino is zero.

However, this has been a problem. Go's syscall.ParseDirent utility is shared for all GOOS=unix. For simplicity, this abstracts direntIno with data from d_fileno or d_ino, and drops if either are zero, even if d_fileno is the only field with zero explicitly defined. This led to a change to special case GOOS=wasip1 as otherwise virtual files would be unconditionally skipped.

In practice, this problem is rather unique due to so many compilers relying on wasi-libc, which tolerates a zero inode. For example, while issues were reported about the performance regression when wasi-libc began doing a fan-out on zero Dirent.Ino, no issues were reported about dirents being dropped as a result.

In summary, rather than complicating implementation and forcing non-zero inodes for a rare case, we permit zero. We instead document this topic thoroughly, so that emerging compilers can re-use the research and reference it on conflict. We also document that Ino should be non-zero, so that users implementing that field will attempt to get it.

See

sys.Walltime and Nanotime

The sys package has two function types, Walltime and Nanotime for real and monotonic clock exports. The naming matches conventions used in Go.

func time_now() (sec int64, nsec int32, mono int64) {
	sec, nsec = walltime()
	return sec, nsec, nanotime()
}

Splitting functions for wall and clock time allow implementations to choose whether to implement the clock once (as in Go), or split them out.

Each can be configured with a ClockResolution, although is it usually incorrect as detailed in a sub-heading below. The only reason for exposing this is to satisfy WASI:

See https://github.com/WebAssembly/wasi-clocks

Why default to fake time?

WebAssembly has an implicit design pattern of capabilities based security. By defaulting to a fake time, we reduce the chance of timing attacks, at the cost of requiring configuration to opt-into real clocks.

See https://gruss.cc/files/fantastictimers.pdf for an example attacks.

Why does fake time increase on reading?

Both the fake nanotime and walltime increase by 1ms on reading. Particularly in the case of nanotime, this prevents spinning.

Why not time.Clock?

wazero can't use time.Clock as a plugin for clock implementation as it is only substitutable with build flags (faketime) and conflates wall and monotonic time in the same call.

Go's time.Clock was added monotonic time after the fact. For portability with prior APIs, a decision was made to combine readings into the same API call.

See https://go.googlesource.com/proposal/+/master/design/12914-monotonic.md

WebAssembly time imports do not have the same concern. In fact even Go's imports for clocks split walltime from nanotime readings.

See https://github.com/golang/go/blob/go1.20/misc/wasm/wasm_exec.js#L243-L255

Finally, Go's clock is not an interface. WebAssembly users who want determinism or security need to be able to substitute an alternative clock implementation from the host process one.

ClockResolution

A clock's resolution is hardware and OS dependent so requires a system call to retrieve an accurate value. Go does not provide a function for getting resolution, so without CGO we don't have an easy way to get an actual value. For now, we return fixed values of 1us for realtime and 1ns for monotonic, assuming that realtime clocks are often lower precision than monotonic clocks. In the future, this could be improved by having OS+arch specific assembly to make syscalls.

For example, Go implements time.Now for linux-amd64 with this assembly. Because retrieving resolution is not generally called often, unlike getting time, it could be appropriate to only implement the fallback logic that does not use VDSO (executing syscalls in user mode). The syscall for clock_getres is 229 and should be usable. https://pkg.go.dev/syscall#pkg-constants.

If implementing similar for Windows, mingw is often a good source to find the Windows API calls that correspond to a POSIX method.

Writing assembly would allow making syscalls without CGO, but comes with the cost that it will require implementations across many combinations of OS and architecture.

sys.Nanosleep

All major programming languages have a sleep mechanism to block for a duration. Sleep is typically implemented by a WASI poll_oneoff relative clock subscription.

For example, the below ends up calling wasi_snapshot_preview1.poll_oneoff:

const std = @import("std");
pub fn main() !void {
    std.time.sleep(std.time.ns_per_s * 5);
}

Besides Zig, this is also the case with TinyGo (-target=wasi) and Rust (--target wasm32-wasi).

We decided to expose sys.Nanosleep to allow overriding the implementation used in the common case, even if it isn't used by Go, because this gives an easy and efficient closure over a common program function. We also documented sys.Nanotime to warn users that some compilers don't optimize sleep.

sys.Osyield

We expose sys.Osyield, to allow users to control the behavior of WASI's sched_yield without a new build of wazero. This is mainly for parity with all other related features which we allow users to implement, including sys.Nanosleep. Unlike others, we don't provide an out-of-box implementation primarily because it will cause performance problems when accessed.

For example, the below implementation uses CGO, which might result in a 1us delay per invocation depending on the platform.

See golang/go#19409 (comment)

//go:noescape
//go:linkname osyield runtime.osyield
func osyield()

In practice, a request to customize this is unlikely to happen until other thread based functions are implemented. That said, as of early 2023, there are a few signs of implementation interest and cross-referencing:

See WebAssembly/stack-switching#38 See https://github.com/WebAssembly/wasi-threads#what-can-be-skipped See https://slinkydeveloper.com/Kubernetes-controllers-A-New-Hope/

sys.Stat_t

We expose stat information as sys.Stat_t, like syscall.Stat_t except defined without build constraints. For example, you can use sys.Stat_t on GOOS=windows which doesn't define syscall.Stat_t.

The first use case of this is to return inodes from fs.FileInfo without relying on platform-specifics. For example, a user could return *sys.Stat_t from info.Sys() and define a non-zero inode for a virtual file, or map a real inode to a virtual one.

Notable choices per field are listed below, where sys.Stat_t is unlike syscall.Stat_t on GOOS=linux, or needs clarification. One common issue not repeated below is that numeric fields are 64-bit when at least one platform defines it that large. Also, zero values are equivalent to nil or absent.

  • Dev and Ino (Inode) are both defined unsigned as they are defined opaque, and most syscall.Stat_t also defined them unsigned. There are separate sections in this document discussing the impact of zero in Ino.
  • Mode is defined as a fs.FileMode even though that is not defined in POSIX and will not map to all possible values. This is because the current use is WASI, which doesn't define any types or features not already supported. By using fs.FileMode, we can re-use routine experience in Go.
  • NLink is unsigned because it is defined that way in syscall.Stat_t: there can never be less than zero links to a file. We suggest defaulting to 1 in conversions when information is not knowable because at least that many links exist.
  • Size is signed because it is defined that way in syscall.Stat_t: while regular files and directories will always be non-negative, irregular files are possibly negative or not defined. Notably sparse files are known to return negative values.
  • Atim, Mtim and Ctim are signed because they are defined that way in syscall.Stat_t: Negative values are time before 1970. The resolution is nanosecond because that's the maximum resolution currently supported in Go.

Why do we use sys.EpochNanos instead of time.Time or similar?

To simplify documentation, we defined a type alias sys.EpochNanos for int64. time.Time is a data structure, and we could have used this for syscall.Stat_t time values. The most important reason we do not is conversion penalty deriving time from common types.

The most common ABI used in wasip2. This, and compatible ABI such as wasix, encode timestamps in memory as a 64-bit number. If we used time.Time, we would have to convert an underlying type like syscall.Timespec to time.Time only to later have to call .UnixNano() to convert it back to a 64-bit number.

In the future, the component model module "wasi-filesystem" may represent stat timestamps with a type shared with "wasi-clocks", abstractly structured similar to time.Time. However, component model intentionally does not define an ABI. It is likely that the canonical ABI for timestamp will be in two parts, but it is not required for it to be intermediately represented this way. A utility like syscall.NsecToTimespec could split an int64 so that it could be written to memory as 96 bytes (int64, int32), without allocating a struct.

Finally, some may confuse epoch nanoseconds with 32-bit epoch seconds. While 32-bit epoch seconds has "The year 2038" problem, epoch nanoseconds has "The Year 2262" problem, which is even less concerning for this library. If the Go programming language and wazero exist in the 2200's, we can make a major version increment to adjust the sys.EpochNanos approach. Meanwhile, we have faster code.

poll_oneoff

poll_oneoff is a WASI API for waiting for I/O events on multiple handles. It is conceptually similar to the POSIX poll(2) syscall. The name is not poll, because it references “the fact that this function is not efficient when used repeatedly with the same large set of handles”.

We chose to support this API in a handful of cases that work for regular files and standard input. We currently do not support other types of file descriptors such as socket handles.

Clock Subscriptions

As detailed above in sys.Nanosleep, poll_oneoff handles relative clock subscriptions. In our implementation we use sys.Nanosleep() for this purpose in most cases, except when polling for interactive input from os.Stdin (see more details below).

FdRead and FdWrite Subscriptions

When subscribing a file descriptor (except Stdin) for reads or writes, the implementation will generally return immediately with success, unless the file descriptor is unknown. The file descriptor is not checked further for new incoming data. Any timeout is cancelled, and the API call is able to return, unless there are subscriptions to Stdin: these are handled separately.

FdRead and FdWrite Subscription to Stdin

Subscribing Stdin for reads (writes make no sense and cause an error), requires extra care: wazero allows to configure a custom reader for Stdin.

In general, if a custom reader is found, the behavior will be the same as for regular file descriptors: data is assumed to be present and a success is written back to the result buffer.

However, if the reader is detected to read from os.Stdin, a special code path is followed, invoking sysfs.poll().

sysfs.poll() is a wrapper for poll(2) on POSIX systems, and it is emulated on Windows.

Poll on POSIX

On POSIX systems, poll(2) allows to wait for incoming data on a file descriptor, and block until either data becomes available or the timeout expires.

Usage of syfs.poll() is currently only reserved for standard input, because

  1. it is really only necessary to handle interactive input: otherwise, there is no way in Go to peek from Standard Input without actually reading (and thus consuming) from it;

  2. if Stdin is connected to a pipe, it is ok in most cases to return with success immediately;

  3. syfs.poll() is currently a blocking call, irrespective of goroutines, because the underlying syscall is; thus, it is better to limit its usage.

So, if the subscription is for os.Stdin and the handle is detected to correspond to an interactive session, then sysfs.poll() will be invoked with a the Stdin handle and the timeout.

This also means that in this specific case, the timeout is uninterruptible, unless data becomes available on Stdin itself.

Select on Windows

On Windows sysfs.poll() cannot be delegated to a single syscall, because there is no single syscall to handle sockets, pipes and regular files.

Instead, we emulate its behavior for the cases that are currently of interest.

  • For regular files, we always report them as ready, as most operating systems do anyway.

  • For pipes, we invoke PeekNamedPipe for each file handle we detect is a pipe open for reading. We currently ignore pipes open for writing.

  • Notably, we include also support for sockets using the WinSock implementation of poll, but instead of relying on the timeout argument of the WSAPoll function, we set a 0-duration timeout so that it behaves like a peek.

This way, we can check for regular files all at once, at the beginning of the function, then we poll pipes and sockets periodically using a cancellable time.Tick, which plays nicely with the rest of the Go runtime.

Impact of blocking

Because this is a blocking syscall, it will also block the carrier thread of the goroutine, preventing any means to support context cancellation directly.

There are ways to obviate this issue. We outline here one idea, that is however not currently implemented. A common approach to support context cancellation is to add a signal file descriptor to the set, e.g. the read-end of a pipe or an eventfd on Linux. When the context is canceled, we may unblock a Select call by writing to the fd, causing it to return immediately. This however requires to do a bit of housekeeping to hide the "special" FD from the end-user.

Signed encoding of integer global constant initializers

wazero treats integer global constant initializers signed as their interpretation is not known at declaration time. For example, there is no signed integer value type.

To get at the problem, let's use an example.

(global (export "start_epoch") i64 (i64.const 1620216263544))

In both signed and unsigned LEB128 encoding, this value is the same bit pattern. The problem is that some numbers are not. For example, 16256 is 807f encoded as unsigned, but 80ff00 encoded as signed.

While the specification mentions uninterpreted integers are in abstract unsigned values, the binary encoding is clear that they are encoded signed.

For consistency, we go with signed encoding in the special case of global constant initializers.

Implementation limitations

WebAssembly 1.0 (20191205) specification allows runtimes to limit certain aspects of Wasm module or execution.

wazero limitations are imposed pragmatically and described below.

Number of functions in a module

The possible number of function instances in a module is not specified in the WebAssembly specifications since funcaddr corresponding to a function instance in a store can be arbitrary number. wazero limits the maximum function instances to 2^27 as even that number would occupy 1GB in function pointers.

That is because not only we believe that all use cases are fine with the limitation, but also we have no way to test wazero runtimes under these unusual circumstances.

Number of function types in a store

There's no limitation on the number of function types in a store according to the spec. In wazero implementation, we assign each function type to a unique ID, and choose to use uint32 to represent the IDs. Therefore the maximum number of function types a store can have is limited to 2^27 as even that number would occupy 512MB just to reference the function types.

This is due to the same reason for the limitation on the number of functions above.

Number of values on the stack in a function

While the the spec does not clarify a limitation of function stack values, wazero limits this to 2^27 = 134,217,728. The reason is that we internally represent all the values as 64-bit integers regardless of its types (including f32, f64), and 2^27 values means 1 GiB = (2^30). 1 GiB is the reasonable for most applications as we see a Goroutine has 250 MB as a limit on the stack for 32-bit arch, considering that WebAssembly is (currently) 32-bit environment.

All the functions are statically analyzed at module instantiation phase, and if a function can potentially reach this limit, an error is returned.

Number of globals in a module

Theoretically, a module can declare globals (including imports) up to 2^32 times. However, wazero limits this to 2^27(134,217,728) per module. That is because internally we store globals in a slice with pointer types (meaning 8 bytes on 64-bit platforms), and therefore 2^27 globals means that we have 1 GiB size of slice which seems large enough for most applications.

Number of tables in a module

While the the spec says that a module can have up to 2^32 tables, wazero limits this to 2^27 = 134,217,728. One of the reasons is even that number would occupy 1GB in the pointers tables alone. Not only that, we access tables slice by table index by using 32-bit signed offset in the compiler implementation, which means that the table index of 2^27 can reach 2^27 * 8 (pointer size on 64-bit machines) = 2^30 offsets in bytes.

We believe that all use cases are fine with the limitation, but also note that we have no way to test wazero runtimes under these unusual circumstances.

If a module reaches this limit, an error is returned at the compilation phase.

Compiler engine implementation

Why it's safe to execute runtime-generated machine codes against async Goroutine preemption

Goroutine preemption is the mechanism of the Go runtime to switch goroutines contexts on an OS thread. There are two types of preemption: cooperative preemption and async preemption. The former happens, for example, when making a function call, and it is not an issue for our runtime-generated functions as they do not make direct function calls to Go-implemented functions. On the other hand, the latter, async preemption, can be problematic since it tries to interrupt the execution of Goroutine at any point of function, and manipulates CPU register states.

Fortunately, our runtime-generated machine codes do not need to take the async preemption into account. All the assembly codes are entered via the trampoline implemented as Go Assembler Function (e.g. arch_amd64.s), and as of Go 1.20, these assembler functions are considered as unsafe for async preemption:

From the Go runtime point of view, the execution of runtime-generated machine codes is considered as a part of that trampoline function. Therefore, runtime-generated machine code is also correctly considered unsafe for async preemption.

Why context cancellation is handled in Go code rather than native code

Since wazero v1.0.0-pre.9, the runtime supports integration with Go contexts to interrupt execution after a timeout, or in response to explicit cancellation. This support is internally implemented as a special opcode builtinFunctionCheckExitCode that triggers the execution of a Go function (ModuleInstance.FailIfClosed) that atomically checks a sentinel value at strategic points in the code.

It is indeed possible to check the sentinel value directly, without leaving the native world, thus sparing some cycles; however, because native code never preempts (see section above), this may lead to a state where the other goroutines never get the chance to run, and thus never get the chance to set the sentinel value; effectively preventing cancellation from taking place.

Golang patterns

Hammer tests

Code that uses concurrency primitives, such as locks or atomics, should include "hammer tests", which run large loops inside a bounded amount of goroutines, run by half that many GOMAXPROCS. These are named consistently "hammer", so they are easy to find. The name inherits from some existing tests in golang/go.

Here is an annotated description of the key pieces of a hammer test:

  1. P declares the count of goroutines to use, defaulting to 8 or 4 if testing.Short.
    • Half this amount are the cores used, and 4 is less than a modern laptop's CPU. This allows multiple "hammer" tests to run in parallel.
  2. N declares the scale of work (loop) per goroutine, defaulting to value that finishes in ~0.1s on a modern laptop.
    • When in doubt, try 1000 or 100 if testing.Short
    • Remember, there are multiple hammer tests and CI nodes are slow. Slower tests hurt feedback loops.
  3. defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(P/2)) makes goroutines switch cores, testing visibility of shared data.
  4. To ensure goroutines execute at the same time, block them with sync.WaitGroup, initialized to Add(P).
    • sync.WaitGroup internally uses runtime_Semacquire not available in any other library.
    • sync.WaitGroup.Add with a negative value can unblock many goroutines at the same time, e.g. without a for loop.
  5. Track goroutines progress via finished := make(chan int) where each goroutine in P defers finished <- 1.
    1. Tests use require.XXX, so recover() into t.Fail in a defer function before finished <- 1.
      • This makes it easier to spot larger concurrency problems as you see each failure, not just the first.
    2. After the defer function, await unblocked, then run the stateful function N times in a normal loop.
      • This loop should trigger shared state problems as locks or atomics are contended by P goroutines.
  6. After all P goroutines launch, atomically release all of them with WaitGroup.Add(-P).
  7. Block the runner on goroutine completion, by (<-finished) for each P.
  8. When all goroutines complete, return if t.Failed(), otherwise perform follow-up state checks.

This is implemented in wazero in hammer.go

Lock-free, cross-goroutine observations of updates

How to achieve cross-goroutine reads of a variable are not explicitly defined in https://go.dev/ref/mem. wazero uses atomics to implement this following unofficial practice. For example, a Close operation can be guarded to happen only once via compare-and-swap (CAS) against a zero value. When we use this pattern, we consistently use atomics to both read and update the same numeric field.

In lieu of formal documentation, we infer this pattern works from other sources (besides tests):

  • sync.WaitGroup by definition must support calling Add from other goroutines. Internally, it uses atomics.
  • rsc in golang/go#5045 writes "atomics guarantee sequential consistency among the atomic variables".

See https://github.com/golang/go/blob/go1.20/src/sync/waitgroup.go#L64 See golang/go#5045 (comment) See https://www.youtube.com/watch?v=VmrEG-3bWyM