Support #[serde(validate = "function")] for container, compatible with validator crate

[zao111222333]

This PR aims to sovle #939, #642

The basic idea is only specify validate function(s) in container attribute, since it is the most general way.
One can directly use validator::Validate::validate as validation function, so it is not necessary to have field validator in serde.
The Error needs to be impl Display, to convert it by serde::de::Error::custom.

For general usage, one can use arbitrary fn (&T) -> Result<(), impl Display> function to validate.
And for validator user, there is no extra overhead to do validate.

#[derive(validator::Validate, Deserialize)]
#[serde(validate = "validate_struct")]
#[serde(validate = "validator::Validate::validate")]
struct ValidatorDemo {
    #[validate(email)]
    mail: String,
    a: u16,
}

fn validate_struct(deserialized: &ValidatorDemo) -> Result<(), impl Display> {
    if deserialized.a == 0 {
        return Err("field `a` can not be zero");
    }
    Ok(())
}

The #[serde(validator)] can be seen as the short of #[serde(validate = "validator::Validate::validate")], so the above code is equivalent to

#[derive(validator::Validate, Deserialize)]
#[serde(validate = "validate_struct")]
#[serde(validator)]
struct ValidatorDemo {
    #[validate(email)]
    mail: String,
    a: u16,
}

The full unit-test & demo see here.
I am glad to write documents if this idea looks good to you :)

[zao111222333]

And I notice that the fn (&T) -> Result<(), impl Display> way is also compatible with https://github.com/ohkami-rs/serdev

[oli-obk]

I'll need to read up on the issue discussions and the extent to which we may want to simplify validator crate support

[zao111222333]

I'll need to read up on the issue discussions and the extent to which we may want to simplify validator crate support

The most user friendly API should like that

#[derive(Deserialize)]
#[serde(validator)]
struct ValidatorDemo {
    #[validate(email)]
    mail: String,
}

However it needs to expand into follows, assume the validator is a block box to us, and we only wrap it up

#[derive(validator::Validate, Deserialize)]
#[serde(validate = "validator::Validate::validate")]
struct ValidatorDemo {
    #[validate(email)]
    mail: String,
}

And I think it is impossible to make the Deserialize proc_macro_derive to introduce another proc_macro_derive for the same container ...

[oli-obk]

I think it's ok for Validate to be required in the list of derives, but I do like the idea of an emtpy serde(validator) to just expand to validator::Validate::validate and hope that it's in scope.

[zao111222333]

Great, I will do that

[zao111222333]

Now we have two ways to use validator:

#[derive(validator::Validate, Deserialize)]
#[serde(validate = "validator::Validate::validate")]
struct ValidateStruct {
    #[validate(email)]
    mail: String,
}

#[derive(validator::Validate, Deserialize)]
#[serde(validator)]
struct ValidatorStruct {
    #[validate(email)]
    mail: String,
}

And it will send error message when both serde(validator) and serde(validate = "...") are defined.
I add complie failed tests in test_suite/tests/ui/validate.

[oli-obk]

it's not really wrong to have two validation schemes running at the same time. serde(validate = "..") could be specified twice, too (add a test?)

[zao111222333]

You mean user can define a list of validation schemes?
I think that's a good idea.

[zao111222333]

Now there can be multiple validations, I add two test cases

1. Multiple serde(validate = "..."), see here
2. Both serde(validate = "...") and serde(validator), see here

[Mingun]

What the advantage of using #[serde(validator)] over #[serde(validate = "...")]? It seems to me that this is unnecessary introduction of implicit dependency.

Because Rust already supports using paths in attributes, why you choosed using string for validate values instead of a path?

[zao111222333]

Hi @Mingun, thank you for your review.

What the advantage of using #[serde(validator)] over #[serde(validate = "...")]? It seems to me that this is unnecessary introduction of implicit dependency.

#[serde(validator)] is the short of #[serde(validate = "validator::Validate::validate")]
#[serde(validate = "...")] is the most general way, and I am open for whether or not use #[serde(validator)]

Because Rust already supports using paths in attributes, why you choosed using string for validate values instead of a path?

Considering the user interface consistency, I just reuse the serialize_with's internal APIs.
If you think we can use a path to do #[serde(validate = ...)], I can change to it.

[zao111222333]

Hi, I am wondering is there any suggestions for this PR? So that I can modify under that.
Take your time and thank you!

[zao111222333]

Hi @dtolnay , I summary the modifications and discussions under this PR, and update them to #2891 (comment). Pls have a look when you are free.
Take your time and thank you :)