web_net/

app.rs

//! The application: its site, its compiled routes, its gate, and the
//! pipeline every request runs.

use std::borrow::Cow;
use std::cell::Cell;
use std::fmt;
use std::panic::{AssertUnwindSafe, catch_unwind};
use std::sync::Arc;

use serve_net::http_net::MethodSet;
use serve_net::{
    Envelope, HttpHead, HttpReply, HttpRequest, HttpService, Method, Outcome, Scheme, Verdict,
    WorkClass,
};

use crate::assets::{self, Assets};
use crate::csrf;
use crate::cx::{Out, is_local_target};
use crate::error::{self, BuildError, Problem};
use crate::headers::SecurityHeaders;
use crate::route::{Access, BoxHandler, Group, Route};
use crate::router::{self, Captures, Trie};
use crate::url::{self, TrailingSlash};
use crate::{
    Cx, Denial, DenialKind, Event, EventKind, Events, Fields, Gate, Identity, NoEvents, Open,
    Output, Render, Request, RequestId, Result, Stage, TemplateId, Value,
};

/// The default budget for one rendered page.
pub const DEFAULT_PAGE_LIMIT: usize = 8 << 20;

/// Statuses with an error template lookup (`error/404`, ...).
const ERROR_STATUSES: [u16; 14] = [
    400, 401, 403, 404, 405, 409, 410, 413, 421, 422, 429, 500, 501, 503,
];

/// A named point in the pipeline where registered code runs, only for the
/// routes it registered for. A route with no hooks pays nothing.
#[non_exhaustive]
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Hook {
    /// After the access check passed, before the handler: locale, feature
    /// flags, an A/B variant. An error ends the request.
    Routed,
    /// After the handler succeeded, before the reply leaves: add headers.
    /// An error replaces the reply with the error page.
    Finish,
}

/// Which routes a hook runs for.
#[derive(Clone, Debug)]
pub struct Routes(Selector);

#[derive(Clone, Debug)]
enum Selector {
    All,
    Prefix(String),
    Named(Vec<&'static str>),
}

impl Routes {
    /// Every route.
    #[must_use]
    pub fn all() -> Self {
        Self(Selector::All)
    }

    /// Routes whose pattern starts with `prefix` (`/blog`).
    #[must_use]
    pub fn prefix(prefix: impl Into<String>) -> Self {
        Self(Selector::Prefix(prefix.into()))
    }

    /// Routes with one of these names.
    #[must_use]
    pub fn named(names: &[&'static str]) -> Self {
        Self(Selector::Named(names.to_vec()))
    }

    fn matches(&self, pattern: &str, name: Option<&str>) -> bool {
        match &self.0 {
            Selector::All => true,
            Selector::Prefix(p) => {
                let p = p.trim_end_matches('/');
                pattern == p
                    || pattern
                        .strip_prefix(p)
                        .is_some_and(|rest| rest.starts_with('/'))
            }
            Selector::Named(names) => name.is_some_and(|n| names.contains(&n)),
        }
    }
}

type HookFn<I> = Arc<dyn Fn(&mut Cx<'_, I>) -> Result + Send + Sync>;

/// What [`App::classify`] tells a dispatcher before admission.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Classified {
    /// The route that will run, if one matched.
    pub route: Option<RouteId>,
    /// How its handler behaves (`Quick` when nothing matched).
    pub work: WorkClass,
}

/// A declared route (one method of one pattern), numbered in declaration
/// order.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub struct RouteId(pub u32);

/// The parts of the application every handler can reach.
pub(crate) struct Shared {
    pub(crate) render: Option<Box<dyn Render>>,
    pub(crate) checkpoint: Option<Box<dyn crate::Checkpoint>>,
    pub(crate) events: Box<dyn Events>,
    pub(crate) page_limit: usize,
    pub(crate) base: Box<str>,
    errors: Vec<(u16, TemplateId)>,
    sign_in: Option<Box<str>>,
    challenge: Box<str>,
    wipe: bool,
    security: SecurityHeaders,
    pub(crate) csrf: Option<csrf_sec::Csrf>,
    pub(crate) redirect_hosts: Vec<Box<str>>,
}

/// Where a reply is going: the request's envelope, and whether security
/// headers for HTTPS apply (HTTPS, and not a development host).
#[derive(Clone, Copy)]
pub(crate) struct At<'a> {
    pub(crate) env: &'a Envelope,
    pub(crate) tls: bool,
}

type ReadyFn = Arc<dyn Fn() -> bool + Send + Sync>;

#[derive(Clone)]
enum Health {
    Live,
    Ready(ReadyFn),
}

#[derive(Clone, Debug, Default)]
struct Site {
    /// Lower case; the first is canonical.
    hosts: Vec<Box<str>>,
    aliases: Vec<Box<str>>,
    development: bool,
    slash: TrailingSlash,
    lowercase: bool,
    https_only: bool,
}

struct Entry<G: Gate> {
    handler: BoxHandler<G::Identity>,
    access: Access<G::Rule>,
    names: Box<[Box<str>]>,
    pattern: Box<str>,
    work: WorkClass,
    template: Option<TemplateId>,
    csrf: bool,
    limits: form_net::Limits,
    routed: Box<[HookFn<G::Identity>]>,
    finish: Box<[HookFn<G::Identity>]>,
}

struct Endpoint {
    /// Entry per method, by `Method::index`.
    methods: [Option<u32>; 10],
    allow: Box<str>,
}

/// A web application: one site, its routes, and its gate. Built once at
/// start-up with [`App::builder`] or [`App::with_gate`]; immutable after.
///
/// It is a `serve_net::HttpService`, so it runs on serve-net as it is, and
/// it has [`respond`](Self::respond) and
/// [`respond_identified`](Self::respond_identified) for a wrapper that
/// calls it from its own service.
///
/// ```
/// use web_net::{App, get};
///
/// let app = App::builder()
///     .route(get("/", |cx| { cx.html(200, "<h1>Hello</h1>"); Ok(()) }).public())
///     .route(get("/users/:id", |cx| {
///         let id: u64 = cx.param_as("id")?;
///         cx.text(200, &format!("user {id}"));
///         Ok(())
///     }).public())
///     .build()?;
///
/// let reply = serve_net::testing::respond(&app, b"GET /users/7 HTTP/1.1\r\nHost: a\r\n\r\n");
/// assert_eq!((reply.status, reply.text()), (200, "user 7"));
///
/// let reply = serve_net::testing::respond(&app, b"GET /users/7/ HTTP/1.1\r\nHost: a\r\n\r\n");
/// assert_eq!(reply.status, 301);
/// assert_eq!(reply.header("location"), Some(&b"/users/7"[..]));
/// # Ok::<(), web_net::BuildError>(())
/// ```
pub struct App<G: Gate = Open> {
    gate: G,
    shared: Shared,
    site: Site,
    trie: Trie,
    endpoints: Vec<Endpoint>,
    entries: Vec<Entry<G>>,
    allow_all: Box<str>,
    assets: Vec<Assets>,
    health: Vec<(Box<str>, Health)>,
}

impl App<Open> {
    /// A builder with the [`Open`] gate: every visitor is a guest and every
    /// route is declared public.
    #[must_use]
    pub fn builder() -> Builder<Open> {
        Builder::new(Open)
    }
}

impl<G: Gate> App<G> {
    /// A builder with `gate` for authentication and authorization.
    #[must_use]
    pub fn with_gate(gate: G) -> Builder<G> {
        Builder::new(gate)
    }
}

/// Declares an application. Every mistake is reported by
/// [`build`](Self::build) at once, before anything listens.
pub struct Builder<G: Gate> {
    gate: G,
    routes: Vec<Route<G::Identity, G::Rule>>,
    hooks: Vec<(Hook, Routes, HookFn<G::Identity>)>,
    site: Site,
    base: String,
    render: Option<Box<dyn Render>>,
    checkpoint: Option<Box<dyn crate::Checkpoint>>,
    events: Box<dyn Events>,
    page_limit: usize,
    sign_in: Option<String>,
    challenge: String,
    wipe: bool,
    security: SecurityHeaders,
    assets: Vec<Assets>,
    health: Vec<(String, Health)>,
    csrf: Option<csrf_sec::Csrf>,
    redirect_hosts: Vec<Box<str>>,
}

impl<G: Gate> Builder<G> {
    fn new(gate: G) -> Self {
        Self {
            gate,
            routes: Vec::new(),
            hooks: Vec::new(),
            site: Site::default(),
            base: String::new(),
            render: None,
            checkpoint: None,
            events: Box::new(NoEvents),
            page_limit: DEFAULT_PAGE_LIMIT,
            sign_in: None,
            challenge: "Session".into(),
            wipe: false,
            security: SecurityHeaders::default(),
            assets: Vec::new(),
            health: Vec::new(),
            csrf: None,
            redirect_hosts: Vec::new(),
        }
    }

    /// CSRF protection with csrf-sec (re-exported as
    /// [`web_net::csrf_sec`](crate::csrf_sec)), built from a persisted
    /// 32-byte key. Required as soon as one route takes an unsafe method and
    /// is not [exempt](crate::Route::csrf_exempt).
    #[must_use]
    pub fn csrf(mut self, csrf: csrf_sec::Csrf) -> Self {
        self.csrf = Some(csrf);
        self
    }

    /// Hosts that [`Cx::redirect_external`] may send a visitor to
    /// (`accounts.example.com`). Everything else is refused.
    #[must_use]
    pub fn redirect_hosts(mut self, hosts: &[&str]) -> Self {
        self.redirect_hosts.extend(
            hosts
                .iter()
                .map(|h| h.to_ascii_lowercase().into_boxed_str()),
        );
        self
    }

    /// Serves a folder of files under a mount path, before routing and
    /// without the gate (files are public).
    #[must_use]
    pub fn assets(mut self, assets: Assets) -> Self {
        self.assets.push(assets);
        self
    }

    /// A liveness path (`/healthz`): `200 ok` whenever the process answers.
    /// Answered before routing and without the gate.
    #[must_use]
    pub fn health(mut self, path: &str) -> Self {
        self.health.push((path.to_owned(), Health::Live));
        self
    }

    /// A readiness path (`/readyz`): `200 ready` when `check` says so,
    /// `503 not ready` otherwise. Keep `check` to a flag read: it runs on the
    /// request path.
    #[must_use]
    pub fn ready(mut self, path: &str, check: impl Fn() -> bool + Send + Sync + 'static) -> Self {
        self.health
            .push((path.to_owned(), Health::Ready(Arc::new(check))));
        self
    }

    /// The security headers every reply carries (default
    /// [`SecurityHeaders::default`]).
    #[must_use]
    pub fn security_headers(mut self, policy: SecurityHeaders) -> Self {
        self.security = policy;
        self
    }

    /// Adds a route.
    #[must_use]
    pub fn route(mut self, route: Route<G::Identity, G::Rule>) -> Self {
        self.routes.push(route);
        self
    }

    /// Adds a group's routes.
    #[must_use]
    pub fn group(mut self, group: Group<G::Identity, G::Rule>) -> Self {
        self.routes.extend(group.into_routes());
        self
    }

    /// Runs `f` at `hook` for the routes `routes` selects.
    #[must_use]
    pub fn hook<F>(mut self, hook: Hook, routes: Routes, f: F) -> Self
    where
        F: Fn(&mut Cx<'_, G::Identity>) -> Result + Send + Sync + 'static,
    {
        self.hooks.push((hook, routes, Arc::new(f)));
        self
    }

    /// A host the site answers (`example.com`, or `localhost:8080` with its
    /// port). The first one is canonical. With no host declared the site
    /// answers any host; with hosts declared, any other host gets 421.
    #[must_use]
    pub fn host(mut self, host: &str) -> Self {
        self.site.hosts.push(host.to_ascii_lowercase().into());
        self
    }

    /// A host that is redirected (301, or 308 for unsafe methods) to the
    /// canonical host (`www.example.com` to `example.com`).
    #[must_use]
    pub fn alias(mut self, host: &str) -> Self {
        self.site.aliases.push(host.to_ascii_lowercase().into());
        self
    }

    /// A development instance: `localhost`, `127.0.0.1`, `[::1]`, and
    /// `*.localhost` are answered as the site (with any port), and HTTPS is
    /// not required for them. Never set this in production: it is the only
    /// way a `Host: localhost` is trusted.
    #[must_use]
    pub fn development(mut self) -> Self {
        self.site.development = true;
        self
    }

    /// The path the site is installed under (`/blog`). It is removed before
    /// routing and added back by [`Cx::url`], so routes and links are written
    /// as if the site were at `/`. Requests outside it get 404.
    #[must_use]
    pub fn base(mut self, base: &str) -> Self {
        self.base = base.to_owned();
        self
    }

    /// The trailing-slash policy (default [`TrailingSlash::Strip`]).
    #[must_use]
    pub fn trailing_slash(mut self, policy: TrailingSlash) -> Self {
        self.site.slash = policy;
        self
    }

    /// Paths are lower case: one with an upper-case letter is redirected.
    #[must_use]
    pub fn lowercase_paths(mut self) -> Self {
        self.site.lowercase = true;
        self
    }

    /// Plain HTTP is redirected to HTTPS on the canonical host (behind a
    /// TLS proxy, serve-net must trust the proxy's scheme). Needs a host.
    #[must_use]
    pub fn https_only(mut self) -> Self {
        self.site.https_only = true;
        self
    }

    /// The site-relative sign-in page. A [`Denial::sign_in`] on a `GET` or
    /// `HEAD` is then answered with a 303 to it, the page asked for passed
    /// as `next`; without it, or for other methods, the answer is 401.
    #[must_use]
    pub fn sign_in(mut self, path: &str) -> Self {
        self.sign_in = Some(path.to_owned());
        self
    }

    /// The `WWW-Authenticate` challenge sent with a 401 (default
    /// `Session`).
    #[must_use]
    pub fn challenge(mut self, challenge: &str) -> Self {
        challenge.clone_into(&mut self.challenge);
        self
    }

    /// The template engine.
    #[must_use]
    pub fn render(mut self, engine: impl Render) -> Self {
        self.render = Some(Box::new(engine));
        self
    }

    /// Where refusals, denials, and failures are reported.
    #[must_use]
    pub fn events(mut self, events: impl Events) -> Self {
        self.events = Box::new(events);
        self
    }

    /// The in-flight revocation check at web-net's stages.
    #[must_use]
    pub fn checkpoint(mut self, checkpoint: impl crate::Checkpoint) -> Self {
        self.checkpoint = Some(Box::new(checkpoint));
        self
    }

    /// The budget for one rendered page (default 8 MiB).
    #[must_use]
    pub fn page_limit(mut self, bytes: usize) -> Self {
        self.page_limit = bytes;
        self
    }

    /// Zero the reply buffers web-net reuses after every request, so one
    /// visitor's page never lingers in memory another request reuses. Costs
    /// one pass over the bytes written.
    #[must_use]
    pub fn wipe_buffers(mut self) -> Self {
        self.wipe = true;
        self
    }

    /// Compiles the application.
    ///
    /// # Errors
    ///
    /// A [`BuildError`] listing every problem: a route without declared
    /// access, a malformed pattern, two handlers for one method of one
    /// path, an unknown template, a malformed host or base path.
    pub fn build(self) -> std::result::Result<App<G>, BuildError> {
        let mut problems = Vec::new();
        let mut problem = |subject: &str, message: String| {
            problems.push(Problem {
                subject: subject.to_owned(),
                message,
            });
        };

        // Site.
        let base = self.base.trim_end_matches('/');
        if !base.is_empty() {
            match url::normalize(base, self.site.lowercase) {
                Ok(n) if n == base && !base.contains(['?', '#']) => {}
                _ => problem(
                    "base",
                    format!("'{}' is not a canonical path such as '/blog'", self.base),
                ),
            }
        }
        for h in self.site.hosts.iter().chain(&self.site.aliases) {
            if !valid_host(h) {
                problem("host", format!("'{h}' is not a host name or address"));
            }
        }
        if self.site.https_only && self.site.hosts.is_empty() {
            problem(
                "https_only",
                "redirecting to HTTPS needs a declared host; add .host(\"example.com\")".into(),
            );
        }
        if let Some(s) = &self.sign_in {
            if !is_local_target(s) {
                problem("sign_in", format!("'{s}' is not a site-relative path"));
            }
        }
        if self.challenge.is_empty() || !self.challenge.bytes().all(|b| (0x20..0x7f).contains(&b)) {
            problem("challenge", "the challenge must be visible ASCII".into());
        }

        if !self.security.valid() {
            problem(
                "security_headers",
                "every header value must be visible ASCII and not empty".into(),
            );
        }
        let mut assets = self.assets;
        for a in &mut assets {
            let subject = format!("assets {}", a.mount);
            if a.mount.is_empty()
                || !is_local_target(&a.mount)
                || url::normalize(&a.mount, self.site.lowercase).map_or(true, |n| n != a.mount)
            {
                problem(
                    &subject,
                    "the mount must be a canonical path such as '/static'".into(),
                );
            }
            match std::fs::canonicalize(&a.dir) {
                Ok(root) if root.is_dir() => a.root = root,
                _ => problem(
                    &subject,
                    format!("'{}' is not a folder that exists", a.dir.display()),
                ),
            }
            for r in &self.routes {
                if a.strip(&r.pattern).is_some() {
                    problem(
                        &r.pattern,
                        format!(
                            "the route is under the assets mount '{}' and can never run",
                            a.mount
                        ),
                    );
                }
            }
        }
        for h in &self.redirect_hosts {
            if !valid_host(h) {
                problem(
                    "redirect_hosts",
                    format!("'{h}' is not a host name or address"),
                );
            }
        }
        let mut health = Vec::new();
        for (p, h) in self.health {
            if !is_local_target(&p) || url::normalize(&p, false).map_or(true, |n| n != p) {
                problem(
                    "health",
                    format!("'{p}' is not a canonical site-relative path"),
                );
            }
            health.push((p.into_boxed_str(), h));
        }

        // Error templates.
        let mut errors = Vec::new();
        if let Some(r) = &self.render {
            for s in ERROR_STATUSES {
                if let Some(t) = r.template(&format!("error/{s}")) {
                    errors.push((s, t));
                }
            }
        }

        // Routes.
        let mut trie = Trie::default();
        let mut endpoints: Vec<Endpoint> = Vec::new();
        let mut entries: Vec<Entry<G>> = Vec::new();
        let mut all = MethodSet::EMPTY;
        for r in self.routes {
            let parts = match router::parse(&r.pattern) {
                Ok(p) => p,
                Err(e) => {
                    problem(&r.pattern, e.message());
                    continue;
                }
            };
            if r.handlers.is_empty() {
                problem(&r.pattern, "the route has no handler".into());
                continue;
            }
            if matches!(r.access, Access::Undeclared) {
                problem(
                    &r.pattern,
                    "the route declares no access; add .public() or .allow(rule)".into(),
                );
                continue;
            }
            let template = match (r.template, &self.render) {
                (None, _) => None,
                (Some(t), None) => {
                    problem(
                        &r.pattern,
                        format!("template '{t}' is declared but no render engine is set"),
                    );
                    None
                }
                (Some(t), Some(engine)) => {
                    let id = engine.template(t);
                    if id.is_none() {
                        problem(&r.pattern, format!("the engine has no template '{t}'"));
                    }
                    id
                }
            };
            let ep = trie.insert(&parts, endpoints.len() as u32) as usize;
            if ep == endpoints.len() {
                endpoints.push(Endpoint {
                    methods: [None; 10],
                    allow: Box::from(""),
                });
            }
            let names: Box<[Box<str>]> = router::names(&parts).into();
            if r.csrf_exempt.is_some_and(|reason| reason.trim().is_empty()) {
                problem(&r.pattern, "a CSRF exemption needs its reason".into());
            }
            if let Err(e) = r.form_limits.validate() {
                problem(&r.pattern, format!("the form limits are inconsistent: {e}"));
            }
            for (m, h) in r.handlers {
                let subject = format!("{} {}", m.as_str().unwrap_or("?"), r.pattern);
                if m == Method::Other || m == Method::Connect || m == Method::Trace {
                    problem(&subject, "this method cannot be routed".into());
                    continue;
                }
                let slot = &mut endpoints[ep].methods[m.index() as usize];
                if slot.is_some() {
                    problem(
                        &subject,
                        "a handler for this method and path exists already".into(),
                    );
                    continue;
                }
                *slot = Some(entries.len() as u32);
                let csrf = r.csrf_exempt.is_none() && !m.is_safe();
                if csrf && self.csrf.is_none() {
                    problem(
                        &subject,
                        "an unsafe method needs CSRF protection; set Builder::csrf(..) or mark the route csrf_exempt(\"reason\")".into(),
                    );
                }
                let pick = |at: Hook| -> Box<[HookFn<G::Identity>]> {
                    self.hooks
                        .iter()
                        .filter(|(h, sel, _)| *h == at && sel.matches(&r.pattern, r.name))
                        .map(|(_, _, f)| Arc::clone(f))
                        .collect()
                };
                entries.push(Entry {
                    handler: h,
                    access: r.access.clone(),
                    names: names.clone(),
                    pattern: r.pattern.as_str().into(),
                    work: r.work,
                    template,
                    csrf,
                    limits: r.form_limits,
                    routed: pick(Hook::Routed),
                    finish: pick(Hook::Finish),
                });
            }
        }
        for e in &mut endpoints {
            let mut set = MethodSet::EMPTY;
            for m in [
                Method::Get,
                Method::Head,
                Method::Post,
                Method::Put,
                Method::Delete,
                Method::Options,
                Method::Patch,
            ] {
                if e.methods[m.index() as usize].is_some() {
                    set = set.with(m);
                }
            }
            if set.contains(Method::Get) {
                set = set.with(Method::Head);
            }
            set = set.with(Method::Options);
            all = all.union(set);
            let mut v = Vec::new();
            set.write_allow(&mut v);
            e.allow = String::from_utf8(v).unwrap_or_default().into();
        }
        if !problems.is_empty() {
            return Err(BuildError::new(problems));
        }
        let mut v = Vec::new();
        all.with(Method::Options).write_allow(&mut v);
        Ok(App {
            gate: self.gate,
            shared: Shared {
                render: self.render,
                checkpoint: self.checkpoint,
                events: self.events,
                page_limit: self.page_limit,
                base: base.into(),
                errors,
                sign_in: self.sign_in.map(Into::into),
                challenge: self.challenge.into(),
                wipe: self.wipe,
                security: self.security,
                csrf: self.csrf,
                redirect_hosts: self.redirect_hosts,
            },
            site: self.site,
            trie,
            endpoints,
            entries,
            allow_all: String::from_utf8(v).unwrap_or_default().into(),
            assets,
            health,
        })
    }
}

impl<G: Gate> fmt::Debug for Builder<G> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("Builder")
            .field("routes", &self.routes)
            .field("site", &self.site)
            .field("base", &self.base)
            .finish_non_exhaustive()
    }
}

fn valid_host(h: &str) -> bool {
    let (host, port) = split_port(h);
    let host_ok = if let Some(inner) = host.strip_prefix('[').and_then(|r| r.strip_suffix(']')) {
        inner.parse::<std::net::Ipv6Addr>().is_ok()
    } else {
        !host.is_empty()
            && host.len() <= 253
            && host
                .bytes()
                .all(|b| b.is_ascii_alphanumeric() || b == b'-' || b == b'.')
    };
    host_ok && port.is_none_or(|p| !p.is_empty() && p.parse::<u16>().is_ok())
}

/// `example.com:8080` to (`example.com`, `Some("8080")`); `[::1]:80` to
/// (`[::1]`, `Some("80")`).
fn split_port(a: &str) -> (&str, Option<&str>) {
    if a.starts_with('[') {
        return match a.find(']') {
            Some(i) => match a[i + 1..].strip_prefix(':') {
                Some(p) => (&a[..=i], Some(p)),
                None => (a, None),
            },
            None => (a, None),
        };
    }
    match a.rsplit_once(':') {
        Some((h, p)) => (h, Some(p)),
        None => (a, None),
    }
}

fn is_dev_host(host: &str) -> bool {
    matches!(host, "localhost" | "127.0.0.1" | "[::1]") || host.ends_with(".localhost")
}

thread_local! {
    /// Reply buffers reused between requests on one thread. Taken for the
    /// length of one request and put back, so nothing about a request lives
    /// here between calls.
    static SCRATCH: Cell<Option<Box<Out>>> = const { Cell::new(None) };
}

/// Buffers bigger than this are dropped instead of kept.
const KEEP_CAPACITY: usize = 1 << 20;

/// Where the site check sent a request.
enum Placed<'a> {
    Here(Cow<'a, str>),
    Redirect(String),
    Refuse(u16, &'static str, Option<url::UrlError>),
}

impl<G: Gate> App<G> {
    /// Runs the whole pipeline for one request, authentication included.
    pub fn respond(&self, req: &HttpRequest<'_>, env: &Envelope, reply: &mut HttpReply) -> Outcome {
        self.run(req, env, None, reply)
    }

    /// Runs the pipeline for a request whose caller authenticated it
    /// already: web-net takes `who` as given and does not call the gate's
    /// `authenticate`. Authorization still runs for every route that is
    /// not public.
    pub fn respond_identified(
        &self,
        req: &HttpRequest<'_>,
        env: &Envelope,
        who: &G::Identity,
        reply: &mut HttpReply,
    ) -> Outcome {
        self.run(req, env, Some(who), reply)
    }

    /// Looks at a head before admission: which route would run and how it
    /// behaves. Nothing is answered and no gate is called.
    #[must_use]
    pub fn classify(&self, head: &HttpHead<'_>) -> Classified {
        let none = Classified {
            route: None,
            work: WorkClass::Quick,
        };
        let Placed::Here(path) = self.place(head.authority(), head.path(), Scheme::Https, true)
        else {
            return none;
        };
        let Some((ep, _)) = self.trie.find(&path) else {
            return none;
        };
        match self.entry_for(ep as usize, head.method()) {
            Some(i) => Classified {
                route: Some(RouteId(i)),
                work: self.entries[i as usize].work,
            },
            None => none,
        }
    }

    /// The gate.
    #[must_use]
    pub fn gate(&self) -> &G {
        &self.gate
    }

    fn entry_for(&self, ep: usize, m: Method) -> Option<u32> {
        let methods = &self.endpoints[ep].methods;
        match m {
            Method::Other => None,
            Method::Head => {
                methods[Method::Head.index() as usize].or(methods[Method::Get.index() as usize])
            }
            m => methods[m.index() as usize],
        }
    }

    /// Steps 1 and 2: which site, which canonical path. `quiet` (for
    /// `classify`) skips building a redirect.
    fn place<'a>(
        &self,
        authority: Option<&'a str>,
        raw: &'a str,
        scheme: Scheme,
        quiet: bool,
    ) -> Placed<'a> {
        let site = &self.site;
        let authority = authority.unwrap_or("");
        let lower = authority.to_ascii_lowercase();
        let lower = lower.trim_end_matches('.');
        let host_only = split_port(lower).0.trim_end_matches('.');
        let listed = |list: &[Box<str>]| {
            list.iter().any(|h| {
                let h = h.as_ref();
                h == lower || (split_port(h).1.is_none() && h == host_only)
            })
        };
        let dev = site.development && is_dev_host(host_only);
        let mut target_host: Option<&str> = None;
        if !site.hosts.is_empty() && !dev && !authority.is_empty() && !listed(&site.hosts) {
            if listed(&site.aliases) {
                target_host = site.hosts.first().map(AsRef::as_ref);
            } else {
                return Placed::Refuse(421, "host.unknown", None);
            }
        }
        let mut https = scheme == Scheme::Https;
        if site.https_only && !https && !dev {
            https = true;
            target_host = target_host.or(site.hosts.first().map(AsRef::as_ref));
        }

        // The path: canonical, inside the base, with the slash policy.
        let canon = match url::normalize(raw, site.lowercase) {
            Ok(c) => c,
            Err(e) => return Placed::Refuse(400, e.code(), Some(e)),
        };
        let base = self.shared.base.as_ref();
        let c: &str = &canon;
        let rel_start = if base.is_empty() {
            0
        } else if c == base {
            c.len()
        } else if c.strip_prefix(base).is_some_and(|r| r.starts_with('/')) {
            base.len()
        } else {
            return Placed::Refuse(404, "site.outside_base", None);
        };
        let rel = &c[rel_start..];
        let mut fixed: Option<String> = None;
        if rel.is_empty() {
            fixed = Some("/".into());
        } else if rel.len() > 1 {
            match site.slash {
                TrailingSlash::Strip => {
                    if let Some(s) = rel.strip_suffix('/') {
                        fixed = Some(s.to_owned());
                    }
                }
                TrailingSlash::Add => {
                    let last = rel.rsplit('/').next().unwrap_or("");
                    if !rel.ends_with('/') && !last.contains('.') {
                        fixed = Some(format!("{rel}/"));
                    }
                }
                TrailingSlash::Ignore => {}
            }
        }
        if target_host.is_none() && fixed.is_none() {
            if let Cow::Borrowed(c) = canon {
                let mut path = &c[rel_start..];
                if site.slash == TrailingSlash::Ignore && path.len() > 1 {
                    path = path.strip_suffix('/').unwrap_or(path);
                }
                return Placed::Here(Cow::Borrowed(path));
            }
        }
        if quiet {
            return Placed::Refuse(0, "", None);
        }
        let rel_out = fixed.as_deref().unwrap_or(rel);
        let mut loc = String::with_capacity(raw.len() + 32);
        if let Some(h) = target_host {
            loc.push_str(if https { "https://" } else { "http://" });
            loc.push_str(h);
        }
        loc.push_str(base);
        loc.push_str(rel_out);
        Placed::Redirect(loc)
    }

    fn run(
        &self,
        http: &HttpRequest<'_>,
        env: &Envelope,
        who: Option<&G::Identity>,
        reply: &mut HttpReply,
    ) -> Outcome {
        let method = http.method();
        let raw_path = http.path();
        let authority = http.head().authority();
        let at = At {
            env,
            tls: self.tls(env, authority),
        };
        if raw_path == "*" {
            if method == Method::Options {
                reply.status(204).header("allow", self.allow_all.as_bytes());
                self.secure(reply, at.tls, |_| false);
            } else {
                self.refuse(
                    400,
                    "url.asterisk",
                    EventKind::Refused,
                    None,
                    None,
                    at,
                    reply,
                );
            }
            return Outcome::Done;
        }
        let host_raw = authority.unwrap_or("");
        let placed = self.place(authority, raw_path, env.client().scheme(), false);
        let path = match placed {
            Placed::Here(path) => path,
            Placed::Redirect(mut loc) => {
                if let Some(q) = http.query() {
                    loc.push('?');
                    loc.push_str(q);
                }
                let status = if matches!(method, Method::Get | Method::Head) {
                    301
                } else {
                    308
                };
                reply.status(status).header("location", loc.as_bytes());
                self.secure(reply, at.tls, |_| false);
                return Outcome::Done;
            }
            Placed::Refuse(status, code, src) => {
                let src = src
                    .as_ref()
                    .map(|e| e as &(dyn std::error::Error + 'static));
                self.refuse(status, code, EventKind::Refused, None, src, at, reply);
                return Outcome::Done;
            }
        };
        // Early: health paths and files, before the gate and the router.
        if let Some(outcome) = self.early(http, at, &path, reply) {
            return outcome;
        }

        let mut req = Request {
            http,
            env,
            path: &path,
            host: host_raw,
            route: None,
            tls: at.tls,
        };

        // Authentication, unless the caller did it.
        let owned;
        let who: &G::Identity = match who {
            Some(w) => w,
            None => match self.gate.authenticate(&req) {
                Ok(w) => {
                    owned = w;
                    &owned
                }
                Err(d) => {
                    self.deny(d, &req, reply);
                    return Outcome::Done;
                }
            },
        };

        // Route and method.
        let Some((ep, caps)) = self.trie.find(&path) else {
            self.refuse(404, "route.none", EventKind::Refused, None, None, at, reply);
            return Outcome::Done;
        };
        let endpoint = &self.endpoints[ep as usize];
        let Some(idx) = self.entry_for(ep as usize, method) else {
            if method == Method::Options {
                reply.status(204).header("allow", endpoint.allow.as_bytes());
                self.secure(reply, at.tls, |_| false);
                return Outcome::Done;
            }
            let status = if method == Method::Other { 501 } else { 405 };
            let code = if status == 501 {
                "method.unknown"
            } else {
                "method.not_allowed"
            };
            self.refuse(status, code, EventKind::Refused, None, None, at, reply);
            if status == 405 {
                reply.header("allow", endpoint.allow.as_bytes());
            }
            return Outcome::Done;
        };
        let entry = &self.entries[idx as usize];
        req.route = Some(&entry.pattern);

        // Authorization.
        if let Access::Rule(rule) = &entry.access {
            if let Err(d) = self.gate.authorize(who, rule, &req) {
                self.deny(d, &req, reply);
                return Outcome::Done;
            }
        }
        if let Some(cp) = &self.shared.checkpoint {
            match cp.check(env, Stage::Authorized) {
                Verdict::Pass => {}
                Verdict::Deny { status } => {
                    self.refuse(
                        status,
                        "checkpoint.denied",
                        EventKind::Revoked,
                        req.route,
                        None,
                        at,
                        reply,
                    );
                    return Outcome::Done;
                }
                _ => {
                    self.refuse(
                        403,
                        "checkpoint.killed",
                        EventKind::Revoked,
                        req.route,
                        None,
                        at,
                        reply,
                    );
                    return Outcome::Close;
                }
            }
        }

        // CSRF on unsafe methods.
        let binding = who.csrf_binding();
        let csrf_state = match (&self.shared.csrf, entry.csrf) {
            (Some(engine), true) => {
                let scheme = if env.client().scheme() == Scheme::Https {
                    "https://"
                } else {
                    "http://"
                };
                let own = format!("{scheme}{}", host_raw.to_ascii_lowercase());
                match csrf::check(engine, http, binding, &own, &entry.limits) {
                    Ok(s) => s,
                    Err(code) => {
                        self.refuse(403, code, EventKind::Denied, req.route, None, at, reply);
                        return Outcome::Done;
                    }
                }
            }
            // Built only with an engine when a route needs one.
            (None, true) => {
                self.refuse(
                    500,
                    "csrf.no_engine",
                    EventKind::Failed,
                    req.route,
                    None,
                    at,
                    reply,
                );
                return Outcome::Done;
            }
            (_, false) => csrf::State::Off,
        };

        // The handler, with its hooks, on a buffer web-net holds.
        let mut out = SCRATCH.with(Cell::take).unwrap_or_default();
        out.reset();
        let outcome = self.handle(
            entry,
            req,
            who,
            caps,
            (csrf_state, binding),
            &mut out,
            reply,
        );
        if self.shared.wipe {
            out.wipe();
        }
        if out.body.capacity() <= KEEP_CAPACITY && out.values.capacity() <= KEEP_CAPACITY {
            SCRATCH.with(|s| s.set(Some(out)));
        }
        outcome
    }

    /// Health paths and static files.
    fn early(
        &self,
        http: &HttpRequest<'_>,
        at: At<'_>,
        path: &str,
        reply: &mut HttpReply,
    ) -> Option<Outcome> {
        let method = http.method();
        let get_like = matches!(method, Method::Get | Method::Head);
        if let Some((_, h)) = self.health.iter().find(|(p, _)| p.as_ref() == path) {
            if !get_like {
                self.not_get(method, at, reply);
                return Some(Outcome::Done);
            }
            let (status, text) = match h {
                Health::Live => (200, "ok"),
                Health::Ready(f) if f() => (200, "ready"),
                Health::Ready(_) => (503, "not ready"),
            };
            reply.text(status, text).header("cache-control", "no-store");
            self.secure(reply, at.tls, |_| false);
            return Some(Outcome::Done);
        }
        let (assets, rel) = self
            .assets
            .iter()
            .find_map(|a| a.strip(path).map(|rel| (a, rel)))?;
        if !get_like {
            self.not_get(method, at, reply);
            return Some(Outcome::Done);
        }
        let Some(found) = assets::find(assets, rel, http.header("accept-encoding")) else {
            self.refuse(
                404,
                "assets.none",
                EventKind::Refused,
                None,
                None,
                at,
                reply,
            );
            return Some(Outcome::Done);
        };
        let outcome = assets::serve(assets, &found, http, reply);
        self.secure(reply, at.tls, |_| false);
        Some(outcome)
    }

    /// 405 (or 204 for `OPTIONS`) on a path that takes `GET` and `HEAD` only.
    fn not_get(&self, method: Method, at: At<'_>, reply: &mut HttpReply) {
        if method == Method::Options {
            reply.status(204).header("allow", "GET, HEAD, OPTIONS");
            self.secure(reply, at.tls, |_| false);
        } else {
            let (status, code) = if method == Method::Other {
                (501, "method.unknown")
            } else {
                (405, "method.not_allowed")
            };
            self.refuse(status, code, EventKind::Refused, None, None, at, reply);
            if status == 405 {
                reply.header("allow", "GET, HEAD, OPTIONS");
            }
        }
    }

    #[allow(clippy::too_many_arguments)]
    fn handle(
        &self,
        entry: &Entry<G>,
        req: Request<'_>,
        who: &G::Identity,
        caps: Captures,
        (csrf_state, binding): (csrf::State, Option<&[u8]>),
        out: &mut Out,
        reply: &mut HttpReply,
    ) -> Outcome {
        let at = At {
            env: req.env,
            tls: req.tls,
        };
        let route = req.route;
        let (result, unverified) = {
            let mut cx = Cx {
                req,
                who,
                caps,
                names: &entry.names,
                out: &mut *out,
                shared: &self.shared,
                template: entry.template,
                csrf: csrf_state,
                binding,
                limits: &entry.limits,
            };
            let result = catch_unwind(AssertUnwindSafe(|| -> Result {
                for h in entry.routed.iter() {
                    h(&mut cx)?;
                }
                entry.handler.call(&mut cx)?;
                for h in entry.finish.iter() {
                    h(&mut cx)?;
                }
                Ok(())
            }));
            (result, cx.csrf == csrf::State::Pending)
        };
        match result {
            Ok(Ok(())) if unverified => {
                // A multipart route whose handler never read its form: the
                // token was never checked, so nothing it wrote may leave.
                out.reset();
                self.refuse(
                    403,
                    "csrf.unverified",
                    EventKind::Failed,
                    route,
                    None,
                    at,
                    reply,
                );
                Outcome::Done
            }
            Ok(Ok(())) if !out.invalid => {
                reply.status(out.status);
                for &(ns, ne, vs, ve) in &out.spans {
                    let name = core::str::from_utf8(&out.names[ns as usize..ne as usize])
                        .unwrap_or("x-invalid");
                    reply.header(name, &out.values[vs as usize..ve as usize]);
                }
                if !out.has("x-content-type-options") {
                    reply.header("x-content-type-options", "nosniff");
                }
                self.secure(reply, at.tls, |n| out.has(n));
                reply.body(&out.body);
                Outcome::Done
            }
            Ok(Ok(())) => {
                out.reset();
                self.refuse(
                    500,
                    "reply.invalid_field",
                    EventKind::Failed,
                    route,
                    None,
                    at,
                    reply,
                );
                Outcome::Done
            }
            Ok(Err(e)) => {
                out.reset();
                let src = std::error::Error::source(&e);
                self.refuse(
                    e.status(),
                    e.code(),
                    EventKind::Failed,
                    route,
                    src,
                    at,
                    reply,
                );
                Outcome::Done
            }
            Err(_) => {
                out.reset();
                self.refuse(
                    500,
                    "handler.panic",
                    EventKind::Panicked,
                    route,
                    None,
                    at,
                    reply,
                );
                Outcome::Done
            }
        }
    }

    /// A gate's denial as a response.
    fn deny(&self, d: Denial, req: &Request<'_>, reply: &mut HttpReply) {
        let env = req.env;
        let at = At { env, tls: req.tls };
        if d.kind() == DenialKind::SignIn {
            if let Some(page) = &self.shared.sign_in {
                if matches!(req.method(), Method::Get | Method::Head) {
                    let mut loc = String::with_capacity(64);
                    loc.push_str(&self.shared.base);
                    loc.push_str(page);
                    loc.push_str(if page.contains('?') {
                        "&next="
                    } else {
                        "?next="
                    });
                    let mut next = String::new();
                    next.push_str(&self.shared.base);
                    next.push_str(req.path);
                    if let Some(q) = req.query() {
                        next.push('?');
                        next.push_str(q);
                    }
                    encode_component(&next, &mut loc);
                    self.event(EventKind::Denied, env, req.route, 303, d.code(), None);
                    reply.status(303).header("location", loc.as_bytes());
                    self.secure(reply, at.tls, |_| false);
                    return;
                }
            }
        }
        let status = d.status_code();
        self.refuse(
            status,
            d.code(),
            EventKind::Denied,
            req.route,
            None,
            at,
            reply,
        );
        if status == 401 {
            reply.header("www-authenticate", self.shared.challenge.as_bytes());
        }
    }

    fn event(
        &self,
        kind: EventKind,
        env: &Envelope,
        route: Option<&str>,
        status: u16,
        code: &'static str,
        source: Option<&(dyn std::error::Error + 'static)>,
    ) {
        self.shared.events.event(&Event {
            kind,
            id: RequestId::of(env),
            route,
            status,
            code,
            source,
        });
    }

    /// The one function every refusal and error goes through: the status,
    /// the error template for it or the built-in page, the reference; the
    /// code and source to the events only.
    #[allow(clippy::too_many_arguments)]
    fn refuse(
        &self,
        status: u16,
        code: &'static str,
        kind: EventKind,
        route: Option<&str>,
        source: Option<&(dyn std::error::Error + 'static)>,
        at: At<'_>,
        reply: &mut HttpReply,
    ) {
        self.event(kind, at.env, route, status, code, source);
        let reference = RequestId::of(at.env).to_string();
        let mut body = Vec::with_capacity(512);
        let mut done = false;
        if let (Some(engine), Some(&(_, t))) = (
            self.shared.render.as_deref(),
            self.shared.errors.iter().find(|(s, _)| *s == status),
        ) {
            let fields = [
                ("status", Value::Int(i64::from(status))),
                ("title", Value::Text(error::title(status))),
                ("message", Value::Text(error::message(status))),
                ("reference", Value::Text(&reference)),
            ];
            let view = Fields(&fields);
            let mut out = Output::new(&mut body, self.shared.page_limit);
            if engine.render(t, &view, &mut out).is_ok() {
                done = true;
            } else {
                body.clear();
            }
        }
        if !done {
            error::builtin_page(status, &reference, &mut body);
        }
        reply
            .status(status)
            .header("content-type", "text/html; charset=utf-8")
            .header("x-content-type-options", "nosniff")
            .header("cache-control", "no-store")
            .body(&body);
        self.secure(reply, at.tls, |_| false);
    }

    /// Adds the policy's security headers, except those `has` says the reply
    /// set already.
    fn secure(&self, reply: &mut HttpReply, tls: bool, has: impl Fn(&str) -> bool) {
        for (name, value) in self.shared.security.fields(tls) {
            if !has(name) {
                reply.header(name, value.as_bytes());
            }
        }
    }

    /// Whether HTTPS-only headers apply: the request came over HTTPS and is
    /// not for a development host.
    fn tls(&self, env: &Envelope, authority: Option<&str>) -> bool {
        if env.client().scheme() != Scheme::Https {
            return false;
        }
        let a = authority.unwrap_or("").to_ascii_lowercase();
        !(self.site.development && is_dev_host(split_port(a.trim_end_matches('.')).0))
    }
}

impl<G: Gate> HttpService for App<G> {
    fn call(&self, request: &HttpRequest<'_>, env: &Envelope, reply: &mut HttpReply) -> Outcome {
        self.respond(request, env, reply)
    }
}

impl<G: Gate> fmt::Debug for App<G> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("App")
            .field("site", &self.site)
            .field("base", &self.shared.base)
            .field("routes", &self.entries.len())
            .finish_non_exhaustive()
    }
}

/// Percent-encodes a query component: everything but unreserved and `/`.
fn encode_component(s: &str, out: &mut String) {
    const HEX: &[u8; 16] = b"0123456789ABCDEF";
    for b in s.bytes() {
        if b.is_ascii_alphanumeric() || matches!(b, b'-' | b'.' | b'_' | b'~' | b'/') {
            out.push(b as char);
        } else {
            out.push('%');
            out.push(HEX[(b >> 4) as usize] as char);
            out.push(HEX[(b & 15) as usize] as char);
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_split_port() {
        assert_eq!(split_port("a.test:80"), ("a.test", Some("80")));
        assert_eq!(split_port("a.test"), ("a.test", None));
        assert_eq!(split_port("[::1]:8080"), ("[::1]", Some("8080")));
        assert_eq!(split_port("[::1]"), ("[::1]", None));
    }

    #[test]
    fn test_valid_host() {
        for ok in ["example.com", "localhost:8080", "[::1]", "127.0.0.1:80"] {
            assert!(valid_host(ok), "{ok}");
        }
        for bad in ["", "a b", "a:", "a:x", "[zz]", "a/b"] {
            assert!(!valid_host(bad), "{bad}");
        }
    }

    #[test]
    fn test_selector() {
        assert!(Routes::prefix("/blog").matches("/blog", None));
        assert!(Routes::prefix("/blog/").matches("/blog/:slug", None));
        assert!(!Routes::prefix("/blog").matches("/blogger", None));
        assert!(Routes::named(&["a"]).matches("/x", Some("a")));
        assert!(!Routes::named(&["a"]).matches("/x", None));
    }
}