#attributes #error #error-handling #debug-mode #bubble-up

macro oofs_derive

Error handling library that generates and injects context for you

13 releases

0.2.3 Oct 13, 2022
0.2.2 Oct 12, 2022
0.1.15 Oct 9, 2022
0.1.2 Sep 28, 2022

#32 in #debug-mode


Used in 2 crates (via oofs)

MIT license

105KB
2K SLoC

oofs

Crates.io MIT licensed

Error handling library that generates and injects context for you.

This library provides three main features:

  • #[oofs] attribute that generates and injects context to function calls with ? operators.
  • Tagging an error for categorized error handling.
  • Attaching custom contexts.

Table of Content:

Basic Example 1

Below shows a minimal example of context injection by #[oofs] attribute.

use oofs::{oofs, Oof};

#[oofs]
fn outer_fn() -> Result<(), Oof> {
    let x = 123;
    let y = "hello world";

    inner_fn(x, y)?;

    Ok(())
}

#[oofs]
fn inner_fn(x: usize, y: &str) -> Result<(), Oof> {
    let _ = y.parse::<usize>()?;

    Ok(())
}

Running outer_fn() outputs:

inner_fn($0, $1) failed at `oofs/tests/basic.rs:6:5`

Parameters:
    $0: usize = 123
    $1: &str = "hello world"

Caused by:
    0: y.parse() failed at `oofs/tests/basic.rs:17:13`

    1: invalid digit found in string

The error displays chain of methods that fail, their locations in code, the parameters' types and their debug values. This is what gets displayed when you format the error as Debug (i.e. {:?}).

There should be almost no overhead to performance, as all injected code are either const evaluated (i.e. type_name, call name, etc), or lazily loaded only when an error is encountered (debug string of each argument).

Note: Actually, above statement is semi-true. Let me explain:

Only the arguments that are references or copyable objects like primitives (i.e. bool, usize, etc.) can have their debug strings lazy loaded. For non-copyable objects (i.e. passing an owned object to function args, like String), their debug strings are instantly loaded before the call for debug mode, and disabled for release mode.

Default behavior for debugging non-copyable values (String, custom objects, etc.) are:

  • For debug mode, load debug formatted values before calling each function, incurring overhead at every call.
  • For release mode, skip debugging non-copyable values.

You can change this default behavior with attribute arguments or by enabling either features debug_non_copyable_disabled or debug_non_copyable_full. See more details on them in features.

Basic Example 2

Now, let's look at a slightly longer example. Below is an example from oofs/tests/basic.rs.

The example showcases context-generation, tagging, and attaching custom contexts.

use oofs::{oofs, Oof, OofExt};

// Marker type used for tagging.
struct RetryTag;

#[oofs]
fn application() -> Result<(), Oof> {
    if let Err(e) = middlelayer("hello world") {
        // Check if any of internal errors is tagged as `RetryTag`; if so, try again.
        if e.tagged_nested::<RetryTag>() {
            println!("Retrying middlelayer!\n");

            // If the call fails again, return it.
            // Since `?` operator is used, context is generated and injected to the call.
            middlelayer("hello world")?;
        } else {
            return Err(e);
        }
    }

    Ok(())
}

#[oofs]
fn middlelayer(text: &str) -> Result<u64, Oof> {
    let my_struct = MyStruct {
        field: text.to_owned(),
    };

    // Passing an expression as arg is also fine.
    // All args are evaluated before being debugged in the error.
    // Context is generated and injected to both `?`s in this statement.
    let ret = my_struct.failing_method(get_value()?)?;

    Ok(ret)
}

fn get_value() -> Result<usize, std::io::Error> {
    Ok(123)
}

#[derive(Debug)]
struct MyStruct {
    field: String,
}

// #[oofs] can also be used to `impl` blocks.
// Context will be injected to all methods that return a `Result`.
#[oofs]
impl MyStruct {
    fn failing_method(&self, x: usize) -> Result<u64, Oof> {
        let ret = self
            .field
            .parse::<u64>()
            ._tag::<RetryTag>()                 // tags the error with the type `RetryTag`.
            ._attach(x)                         // attach anything that implements `Debug` as custom context.
            ._attach(&self.field)               // attach the receiver as attachment to debug.
            ._attach_lazy(|| "extra context")?; // lazily evaluate context; useful for something like `|| serde_json::to_string(&x)`.

        Ok(ret)
    }
}

Running application() outputs:

Retrying middlelayer!

middlelayer($0) failed at `oofs/tests/basic.rs:11:13`

Parameters:
    $0: &str = "hello world"

Caused by:
    0: my_struct.failing_method($0) failed at `oofs/tests/basic.rs:26:15`

       Parameters:
           $0: usize = 123

    1: self.field.parse() failed at `oofs/tests/basic.rs:46:14`

       Attachments:
           0: 123
           1: "hello world"
           2: "extra context"

    2: invalid digit found in string

Nice looking error is not all; we also get categorized error handling with tags.

At the source method failing_method, we tag the parse method with RetryTag type. At the very top level function application, we call e.tagged_nested::<RetryTag> to check any interal calls were tagged with RetryTag. When the tag is found, we handle the case by calling middlelayer again.

With tagging, we no longer have to go through every error variant at every level. We just look for the tag we want to handle for, and we handle the tagged error accordingly. In the above example, we retry calling middlelayer again if RetryTag tag is found.

#[oofs] attribute

Default Behaviors

There are some default behaviors this attribute chooses to make:

  1. for impl blocks, methods that return Result<_, _> will have context injected.

    • override behavior by specifying #[oofs(skip)] above fn to have that specific method skipped.
  2. for impl blocks, methods that do not return Result<_, _> will be skipped.

    • override behavior by specifying #[oofs] above fn to apply injection regardless.
  3. ? operators inside closures (i.e. || { ... }) will not have context injected.

    • override behavior by specifying #[oofs(closures)] above fn to apply injections to inside closures.
  4. ? operators inside async blocks (i.e. async { ... }) will not have context injected.

    • override behavior by specifying #[oofs(async_blocks)] above fn to apply injections to inside async blocks.
  5. return ... statements and last expression without semicolon will not have context injected.

These default behaviors can be changed by attribute arguments.

Attribute Arguments

Possible attributes arguments are: tag, attach, attach_lazy, closures, async_blocks, skip, debug_skip, debug_with, and debug_non_copyable.

For details on how to use them, see docs.

Tagging Errors

As shown in the example above, you can tag an error with _tag and detect the tag with tagged and tagged_nested.

struct MyTag;

#[oofs]
fn application_level() -> Result<(), Oof> {
    if let Err(e) = source() {
        if e.tagged_nested::<MyTag>() {
            ...handle for this tag
        } else if e.tagged_nested::<OtherTag>() {
            ...handle for this tag
        } else {
            ...
        }
    }
}

...

#[oofs]
fn source() -> Result<(), Oof> {
    some_fn()._tag::<MyTag>()?;

    Ok(())
}

This allows you to categorize errors into different tag groups, and handle for them accordingly. This gives a much better experience when handling errors compared to matching every enum variant in every nested function calls.

Note that you can also tag an error with multiple different tags.

I chose type as tag because types are small, readable and unique. String or usize can lead to having duplicate values by accident.

Attaching Custom Contexts

At some point, you may find the generated context is not enough. After all, it just shows the call that failed, and parameters that were passed to it. It will not capture all the other possibe context information.

You can attach your own context information to the error with _attach and _attach_lazy methods.

#[oofs]
fn outer_fn() -> Result<(), Oof> {
    let x = 123usize;
    let y = std::time::Instant::now();

    "hello world"
        .parse::<usize>()
        ._attach(&x)
        ._attach(&y)?;

    Ok(())
}

Above will print the following error:

$0.parse() failed at `oofs/tests/basic.rs:10:10`

Parameters:
    $0: &str = "hello world"

Attachments:
    0: 123
    1: Instant { t: 11234993365176 }

Caused by:
    invalid digit found in string

_attach takes any type that implements std::fmt::Debug.

_attach_lazy, on the other hand, takes any closure that returns a type that implements ToString.

It can be something &str like ._attach_lazy(|| "some context"), String like ._attach_lazy(|| format!("some context {:?}", x)), or some function that requires some work to display like ._attach_lazy(|| serde_json::to_string(&x)).

Returning Custom Errors

At some point, you also want to return your custom error.

For these cases, you have some options: oof!(...), wrap_err(_), ensure!(...) and ensure_eq!(...).

  • oof!(...): this is a lot like anyhow! or eyre!; you input to macro like you do for println!. This returns Oof struct, and you can call methods on the returned Oof like

    return oof!("my custom error").tag::<MyTag>().attach(&x).into_res();
    

    into_res() wraps Oof into Result::Err(_).

  • wrap_err(_): function that wraps a custom error with Oof.

    return wrap_err(std::io::Error::new(std::io::ErrorKind::Other, "Some Error")).tag::<MyTag>().into_res();
    

    into_res() wraps Oof into Result::Err(_).

  • ensure!(...): this is similar to a lot of other libraries with slight differences.

    Optionally, you can input custom context message like for format!(...).

    Also, you can optionally provide tags and attributes wrapped in braces.

    ensure!(false, "custom context with value {:?}", x, {
      tag: [MyTag, OtherTag],
      attach: [&y, "attachment", Instant::now()],
      attach_lazy: [|| serde_json::to_string(&y), || format!("lazy attachment {}", &z)]
    });
    
  • ensure_eq!(...): this is similar to a lot of other libraries with slight differences.

    Optionally, you can input custom context message like for format!(...).

    Also, you can optionally provide tags and attributes wrapped in braces.

    ensure_eq!(1u8, 2u8, "custom context with value {:?}", x, {
      tag: [MyTag, OtherTag],
      attach: [&y, "attachment", Instant::now()],
      attach_lazy: [|| serde_json::to_string(&y), || format!("lazy attachment {}", &z)]
    });
    

Features

  • location (default: true): enables printing location of code that fails.

  • debug_non_copyable_disabled (default: false): Disables debugging non-copy-able function arguments.

    Default behavior is to instantly load debug strings of non-copyable arguments before each call for debug mode, but disabling them for release mode.

  • debug_non_copyable_full (default: false): Enables instant loading debug strings of non-copy-able arguments even for release mode.

Notes/Limitations About the Library

About #[oofs] Attribute

  • #[oofs] generates and injects contexts into all statements and expressions that have ? operator.

  • return Err(...) or last expression without semicolon do not get injected with context.

  • If the receiver of a method is a variable (i.e. x.some_method()), or a field of a variable (i.e. x.field.some_method()), values of x or x.field are not displayed. This is because there is no way to determine in the macro whether this receiver is a reference, mutable reference, or an owned variable.

    • For these cases, you can attach the variable like x.some_method()._attach(&x) to display the value of x in the error.

About Oof Error Struct

  • Oof does not implement From<E> where E: std::error::Error, and so must be built by attribute macro. So, if you don't include #[oofs], it will throw a comiler error; this is intentional because it will catch the user's eyes and force them to include the attribute.

  • Unlike anyhow::Error or eyre::Report, Oof does implement std::error::Error. This is nice because it makes it compatible with these boxed error types. For example, this works:

    #[oofs]
    fn outer_fn() -> Result<(), anyhow::Error> {
        inner_fn()?;
        Ok(())
    }
    

    It works since ? operator will implicitly convert Oof into anyhow::Error.

About Underscore Methods like ._tag() and ._attach(_)

In the basic examples above, you may have noticed that all the methods used for oof starts with an underscore; you could call them 'meta-methods' as they do not affect the logic, but only the result that is returned.

The reason for this is that there has to be a way for the macro to differentiate between functional methods and meta methods. This is because macro would also try to include these meta methods as part of the displayed method chain, and things like _attach(x) would be displayed twice in Parameters section and Attachments section.

This may seem disturbing and unnatural at first; it was for me, too. But after trying it out, I got used to it; and now I think I like it because I can easily differentiate between functional methods and meta methods.

I apologize for the inconvenience, and please let me know if there was a better way to do this.

Debugging Non-Copyable Arguments

One of the pain points while creating the library was to lazy-load values of copyable arguments and instantly load values of non-copyable arguments at compile time. I figured out how to do this with a cool rust hack.

Now, should the default behavior be to always instantly load values of non-copyable arguments? this could incur unwanted performance costs, as it would load them for non-error cases.

As a compromise, I made it so that, for debug mode, it will instantly load values of non-copyable arguments; and, for release mode, it will not load values of non-copyable arguments.

You can change this behavior with features debug_non_copyable_disabled and debug_non_copyable_full.

debug_non_copyable_disabled will disable loading values of non-copyable arguments even for debug mode. debug_non_copyable_full will enable loading values of non-copyable arguments even for releaes mode.

Compatibility with #[async_trait]

#[async_trait] parses and converts async fn in traits into fn -> Box<Future<Output = Result<...>>>. Since #[oofs] by default only applies context injection to methods that returns Result<_, _>, it will not apply injection once #[async_trait] is applied.

There are two ways to deal with this:

  • Place #[oofs] above #[async_trait], so that oofs is applied first, then #[async_trait].

    #[oofs]
    #[async_trait]
    impl Trait for Struct {
      ...
    }
    
  • In the impl block, place #[oofs(closures, async_blocks)] above fn ..., and oofs attr will tell the macro to apply injection regardless, and closures and async_blocks will tell the macro to apply injection for closures and async blocks, which are disabled by default.

    #[async_trait]
    #[oofs]
    impl Trait for Struct {
      #[oofs(closures, async_blocks)]
      async fn do_something() -> Result<(), _> {
          ...
      }
    }
    

Future Plans

This library is still very much WIP.

I plan to test the error handling for performance, optimize memory footprints of errors, and implement attribute arguments like #[oofs(tag(MyTag))], #[oofs(skip)], etc.

Also, it does not inject context into closures and async blocks. I plan to add attribute args like #[oofs(closures)] and #[oofs(async_blocks)] to enable injecting context to closures and async blocks.

Dependencies

~1.5MB
~38K SLoC