This allows responses to be sent to the client even when data is only
partially read, significantly improving the experience for the client
from one with a "connection closed" error to one with a proper response.
The consequence is a lifetime in 'Data'.
Though other non-lifetime-introducing solutions exist, the introduction
of a lifetime to 'Data' is a longstanding desire as it prevents
smuggling 'Data' into a longer-lived context. Use of 'Data' in that
context was unspecified with various runtime consequences. The addition
of a lifetime bound by the request prevents this error statically.
In summary, the changes are:
* Clients receive responses even when data isn't fully read.
* 'Data' becomes 'Data<'r>'. 'FromData' changes accordingly.
* Route 'Outcome's are strictly tied to the request lifetime.
Tangentially, the invalid length form field validation error message has
improved to format length in byte units if it exceeds 1024.
If stars aligned properly, we might imagine writing this:
#[non_exhaustive]
struct Config {
pub field: Foo,
pub other: Bar,
}
...with semantics that would allow the defining crate (here, Rocket), to
construct the structure directly while consumers would need to use
public constructors or struct update syntax:
Config {
field: Foo,
other: Bar,
..Default::default()
}
Alas, this is not the way `non_exhaustive` works on structs. You cannot
use field-update syntax to construct `Config` above. You must use public
constructors. This means builder methods or mutating an already built
struct. This is not what we want.
I don't know why it works this way. I don't see why it must. Something
something Drop.
So we have this hack from the pre-non_exhaustive era.
This resolves syntax ambiguity issues with public typed-stream macros.
Prior to this commit, greedy single-token matching by macro-rules macros
would result in certain tokens at the beginning of the macro input, such
as 'for', inadvertently triggering a '$ty' matching case resulting in
incorrect expansion.
This commit makes the following improvements to core request handling:
* Absolute target URIs are not rejected. Instead, the path and query
parts are passed through the application. This resolves an issue
where certain HTTP/2 requests would be rejected by Rocket.
* Data is never copied from the request. Previously, Rocket would copy
and allocate for incoming headers.
* Non-UTF-8 headers are dropped with a warning instead of being
lossily, and thus perhaps incorrectly, decoded as UTF-8. The final
fix is to properly support non-UTF-8 headers, no matter how in the
minority they are.
Resolves#1498.
This follows the completed graduation of stable contrib features into
core, removing 'rocket_contrib' in its entirety in favor of two new
crates. These crates are versioned independently of Rocket's core
libraries, allowing upgrades to dependencies without consideration for
versions in core libraries.
'rocket_dyn_templates' replaces the contrib 'templates' features. While
largely a 1-to-1 copy, it makes the following changes:
* the 'tera_templates' feature is now 'tera'
* the 'handlebars_templates' feature is now 'handlebars'
* fails to compile if neither 'tera' nor 'handlebars' is enabled
'rocket_sync_db_pools' replaces the contrib 'database' features. It
makes no changes to the replaced features except that the `database`
attribute is properly documented at the crate root.
This changes 'TempFile' doctests so that different file names are used
across them, avoiding race conditions where one test deletes a file
another test just created and thus expects to subsequently exist.
This has the following nice benefits:
* The 'Uuid' wrapper type is gone.
* 'Uuid' implements 'UriDisplay', 'FromUriParam'.
* The 'serialization' example merges in 'uuid'.
Resolves#1299.
The 'SpaceHelmet' fairing is now called 'Shield'. It features the
following changes and improvements:
* Headers which are now ignored by browsers are removed.
* 'XssFilter' is no longer an on-by-default policy.
* A new 'Permission' policy is introduced.
* 'Shield' is attached to all 'Rocket' instances by default.
* Default headers never allocate on 'Clone'.
* Policy headers are rendered once and cached at start-up.
* Improved use of typed URIs in policy types.
A singleton fairing is guaranteed to be the only instance of its type at
launch time. If more than one instance of a singleton fairing is
attached, only the last instance is retained.
Previously, if a panic occurred with an 'Error' on the stack, 'Error'
would panic as usual during unwinding. This resulted in a double panic.
This commit makes 'Error' detect if a panic is already occurring and
omits its own panic if it is.
This has the following nice benefits:
* The 'JsonValue' wrapper type is gone.
* 'Local{Request, Response}' natively support JSON/MessagePack.
* The 'json' and 'msgpack' limits are officially recognized.
* Soon, Rocket application will not require an explicit 'serde' dep.
This marks the beginning of the end of 'rocket_contrib'.
This commit entirely rewrites Rocket's URI parsing routines and
overhauls the 'uri!' macro resolving all known issues and removing any
potential limitations for compile-time URI creation. This commit:
* Introduces a new 'Reference' URI variant for URI-references.
* Modifies 'Redirect' to accept 'TryFrom<Reference>'.
* Introduces a new 'Asterisk' URI variant for parity.
* Allows creation of any URI type from a string literal via 'uri!'.
* Enables dynamic/static prefixing/suffixing of route URIs in 'uri!'.
* Unifies 'Segments' and 'QuerySegments' into one generic 'Segments'.
* Consolidates URI formatting types/traits into a 'uri::fmt' module.
* Makes APIs more symmetric across URI types.
It also includes the following less-relevant changes:
* Implements 'FromParam' for a single-segment 'PathBuf'.
* Adds 'FileName::is_safe()'.
* No longer reparses upstream request URIs.
Resolves#842.
Resolves#853.
Resolves#998.
This has the following positive effects:
1) The lifetime retrieved through 'Deref' is now long-lived.
2) An '&State<T>` can be created via an '&T'.
3) '&State<T>' is shorter to type than 'State<'_, T>'.
The crux of the implementation is as follows:
* Configurable ctrl-c, signals that trigger a graceful shutdown.
* Configurable grace period before forced I/O termination.
* Programatic triggering via an application-wide method.
* A future (`Shutdown`) that resolves only when shutdown is requested.
Resolves#180.
This is a breaking change for many consumers of the 'Response' and all
consumers of the 'Body' API. The summary of breaking changes is:
* 'Response::body()', 'Response::body_mut()' are infallible.
* A 'Body' can represent an empty body in more cases.
* 'ResponseBuilder' is now simply 'Builder'.
* Direct body read methods on 'Response' were removed in favor of
chaining through 'body_mut()': 'r.body_mut().to_string()'.
* Notion of a 'chunked_body' was removed as it was inaccurate.
* Maximum chunk size can be set on any body.
* 'Response' no longer implements 'Responder'.
A few bugs were fixed in the process. Specifically, 'Body' will emit an
accurate size even for bodies that are partially read, and the size of
seek-determined bodies is emitted on HEAD request where it wasn't
before. Specifics on transport were clarified, and 'Body' docs greatly
improved as a result.
This removes the export of each of these macros from the root, limiting
their export-scope to their respective module. This is accomplished
using a new internal macro, 'export!', which does some "magic" to work
around rustdoc deficiencies.
This commit includes changes that improve how and what Rocket logs
automatically. Rocket now logs:
* All guard errors, indicating the failing guard kind and type.
* A warning when a 'TempFile' is used as a data guard for a request
that specifies a 'form' Content-Type.
* Only the top/sub of a request's format.
This commit makes the following breaking changes:
* '<T as FromData>::Error' must implement 'Debug'.
Furthermore, this commit restores the previous behavior of always
logging launch info. It further restores the unspecified behavior of
modifying logging state only when the set logger is Rocket's logger.
Sentinels resolve a long-standing usability and functional correctness
issue in Rocket: starting an application with guards and/or responders
that depend on state that isn't available. The canonical example is the
'State' guard. Prior to this commit, an application with routes that
queried unmanaged state via 'State' would fail at runtime. With this
commit, the application refuses to launch with a detailed error message.
The 'Sentinel' docs explains it as:
A sentinel, automatically run on ignition, can trigger a launch
abort should an instance fail to meet arbitrary conditions. Every
type that appears in a mounted route's type signature is eligible to
be a sentinel. Of these, those that implement 'Sentinel' have their
'abort()' method invoked automatically, immediately after ignition,
once for each unique type. Sentinels inspect the finalized instance
of 'Rocket' and can trigger a launch abort by returning 'true'.
The following types are now sentinels:
* 'contrib::databases::Connection' (any '#[database]' type)
* 'contrib::templates::Metadata'
* 'contrib::templates::Template'
* 'core::State'
The following are "specialized" sentinels, which allow sentinel
discovery even through type aliases:
* 'Option<T>', 'Debug<T>' if 'T: Sentinel'
* 'Result<T, E>', 'Either<T, E>' if 'T: Sentinel', 'E: Sentinel'
Closes#464.
The core 'Rocket' type is parameterized: 'Rocket<P: Phase>', where
'Phase' is a newly introduced, sealed marker trait. The trait is
implemented by three new marker types representing the three launch
phases: 'Build', 'Ignite', and 'Orbit'. Progression through these three
phases, in order, is enforced, as are the invariants guaranteed by each
phase. In particular, an instance of 'Rocket' is guaranteed to be in its
final configuration after the 'Build' phase and represent a running
local or public server in the 'Orbit' phase. The 'Ignite' phase serves
as an intermediate, enabling inspection of a finalized but stationary
instance. Transition between phases validates the invariants required
by the transition.
All APIs have been adjusted appropriately, requiring either an instance
of 'Rocket' in a particular phase ('Rocket<Build>', 'Rocket<Ignite>', or
'Rocket<Orbit>') or operating generically on a 'Rocket<P>'.
Documentation is also updated and substantially improved to mention
required and guaranteed invariants.
Additionally, this commit makes the following relevant changes:
* 'Rocket::ignite()' is now a public interface.
* 'Rocket::{build,custom}' methods can no longer panic.
* 'Launch' fairings are now 'ignite' fairings.
* 'Liftoff' fairings are always run, even in local mode.
* All 'ignite' fairings run concurrently at ignition.
* Launch logging occurs on launch, not any point prior.
* Launch log messages have improved formatting.
* A new launch error kind, 'Config', was added.
* A 'fairing::Result' type alias was introduced.
* 'Shutdown::shutdown()' is now 'Shutdown::notify()'.
Some internal changes were also introduced:
* Fairing 'Info' name for 'Templates' is now 'Templating'.
* Shutdown is implemented using 'tokio::sync::Notify'.
* 'Client::debug()' is used nearly universally in tests.
Resolves#1154.
Resolves#1136.
...because loading up a Rocket while it's ignited is a bad idea.
More seriously, because 'Rocket.ignite()' will become an "execute
everything up to here" method.
This is now consistent with the serialization of 'Flash' and better
illustrates the purpose of the late 'name' property.
Additionally, take 'Into<String>' instead of 'AsRef<str>' so we only
allocate when necessary.
Launch fairings are now fallible and take the place of attach fairings,
but they are only run, as the name implies, at launch time.
This is is a fundamental shift from eager execution of set-up routines,
including the now defunct attach fairings, to lazy execution,
precipitated by the transition to `async`. The previous functionality,
while simple, caused grave issues:
1. A instance of 'Rocket' with async attach fairings requires an async
runtime to be constructed.
2. The instance is accessible in non-async contexts.
3. The async attach fairings have no runtime in which to be run.
Here's an example:
```rust
let rocket = rocket::ignite()
.attach(AttachFairing::from(|rocket| async {
Ok(rocket.manage(load_from_network::<T>().await))
}));
let state = rocket.state::<T>();
```
This had no real meaning previously yet was accepted by running the
attach fairing future in an isolated runtime. In isolation, this causes
no issue, but when attach fairing futures share reactor state with other
futures in Rocket, panics ensue.
The new Rocket application lifecycle is this:
* Build - A Rocket instance is constructed. No fairings are run.
* Ignition - All launch fairings are run.
* Liftoff - If all launch fairings succeeded, the server is started.
New 'liftoff' fairings are run in this third phase.
Catchers can now be scoped to paths, with preference given to the
longest-prefix, then the status code. This a breaking change for all
applications that register catchers:
* `Rocket::register()` takes a base path to scope catchers under.
- The previous behavior is recovered with `::register("/", ...)`.
* Catchers now fallibly, instead of silently, collide.
* `ErrorKind::Collision` is now `ErrorKind::Collisions`.
Related changes:
* `Origin` implements `TryFrom<String>`, `TryFrom<&str>`.
* All URI variants implement `TryFrom<Uri>`.
* Added `Segments::prefix_of()`.
* `Rocket::mount()` takes a `TryInto<Origin<'_>>` instead of `&str`
for the base mount point.
* Extended `errors` example with scoped catchers.
* Added scoped sections to catchers guide.
Internal changes:
* Moved router code to `router/router.rs`.
This surfaced a dormant concurrency related issue. Prior to this commit,
the router used `routed_segments()` to retrieve the path segments of the
request. This was okay as there was no route in the request, and matched
segments were retrieved eagerly.
This commit makes segment matching lazy, so no matching occurs if
unnecessary. Between two matches, a `route` is atomically set of
`Request`. This is now visible in `routed_segments()`, which should not
have considered the current route in the first place. This was fixed.
Sergio Benitez
ThouCheese
toshokan
Jeb Rosen
Jonah Brüchert
Ben Sully