This addresses #1051. There is one more pull request that needs to be completed/merged (for `CompositeTagHelperDescriptorResolver` and friends). After that, runtime should work!
When an action contained an attribute derived from HttpMethodAttribute,
doesn't specify an attribute route and there is also another attribute
extending HttpMethodAttribute that has a route defined on it; we ignored
the HttpMethodAttribute attribute without a defined route when building
the set of action selectors for the method.
This caused the resulting action to be unbounded and to accept requests
for other verbs not associated with it. The root cause of the problem was
that attributes override equality and do a field by field comparison but
ignore fields in the base classes of the type, so if an attribute is part
of a class hierarchy (like Http*Attributes) there might be two different
attributes that get considered equal.
The fix for the problem has been to change using Contains on a couple of
collections (that uses the equals method on the underlying object) and
check for the existence of the attribute on the collection directly by
using reference equality.
- #4690
- move `ModelBindingMessageProvider` init from `DefaultBindingMetadataProvider` to `DefaultModelMetadata`
- in addition to avoiding error cases, this removes some boilerplate
- add specific errors to `BodyModelBinderProvider`, `CompilerCache`, `CompositeViewEngine`, `ModelBinderFactory`,
and `ObjectResultExecutor`
- `DefaultRazorViewEngineFileProviderAccessor.FileProvider` now a `NullFileProvider` in empty case
'ValidationExcludeFilter' -> 'SuppressChildValidationMetadataProvider'
Also moved to .ModelBinding for improved discoverability. There aren't
many reasons user code would have a using for .Validation.
Moves IControllerArgumentBinder and IControllerPropertyActivator into
.Internal. Also renames ControllerArgumentBinder ->
DefaultControllerArgumentBinder for consistency with other controller
extensibility types.
We don't think these are 100% baked for our long term maintenance of the
product, and want to reserve the ability to make changes in the future.
This change splits up the conventional routing path from the attribute
routing path *inside* routing, instead of inside `MvcRouteHandler`. Each
attribute route group now gets its own instance of
`MvcAttributeRouteHandler` which just knows about the actions it can
reach.
This removes the concept of a route-group-token and removes the lookup
table entirely for attribute routing. This also means that the
`DefaultHandler` on `IRouteBuilder` will not be used for attribute routes,
which we are OK with for 1.0.0.
The action selector's functionality is now split into two methods. We
think this is OK for 1.0.0 because any customization of `IActionSelector`
up to now had to implement virtually the same policy as ours in order to
work with attribute routing. It should now be possible to customize the
selector in a meaningful way without interfering with attribute routing.
This change to ModelBinderFactory makes the caching much more aggressive,
by caching all non-root binders. There's some trickiness here around
making sure we have the right behavior when all providers return null. See
the tests and comments.
I also kept the change I made for a temporary workaround to use a
dictionary rather than a "stack" for cycle breaking. This seems like an
overall improvement in clarity.
- #4652
- previously ignored for top-level models
- `ModelBinderProviderContext.BindingInfo` is now never `null`
- similarly, use type metadata (as well as parameter info) for `ModelBindingContext.BinderModelName`
- previously ignored when overridden in `ControllerArgumentBinder`
This change simplifies a bunch of code and fits more in line with the
current design of model binding.
Now, a model binder only has to do anything if it was successful.
'return' is enough to indicate failure.
This will help future generations maintain this class. Notice that the
protected methods that are going away all just call into another
extensibility point (other than the executor). If we need to we could make
that extensible in the future and then we have the same support with fewer
hooks and less complexity.
This change corrects and ordering bug between the creation of the
'context' and the diagnostic source event that occurs before a synchronous
filter's 'after' stage.
Also made some simple changes to avoid allocating Task<T> in many common
cases. Now we'll only create the Task<T> when we really need it (async
filters).
This change just rearranges some code in the argument binder with a mind
towards performance and clarity. We're removing a few Task<T>'s here as
well in certain cases, but not yet all of them. We additionally save a
dictionary in the case where you have bound properties.
Hopefully these changes break the code into more discrete and sensible
units without multiple levels of indirection without abstraction.
- Main 'driver' code
- BindModel
- ActivateProperty
We want this change to avoid MVC eagerly reading the form. This is good
for general perf and also for scenarios where you want read the body
yourself (large file uploads).
We DO have scenarios where you want to configure the value providers
per-request or also to change the limits on the value providers (form) so
it's worth keeping these around on the context.
When the service receives a model (say, via a POST message) MVC validates it to ensure the model is in a correct state. Validation currently incurs in many allocations that can be avoided. This tackles two of them:
1. We're now caching the generic `GetEnumerator<T>` method infos generated on the fly during collection validation, and
2. We're now only initializing `ModelErrorCollection` on demand.
The first one incurs in the additional allocation of 1 long-lived dictionary object, which will grow only to the amount of `Collection<T>` types used by the model being validated. This is expected to be a small to medium number.
The second change assumes that class `ModelStateEntry` isn't thread safe, as model validation isn't multithreaded.
This resolves#4434 and #4435.