mini_static/

server.rs

use std::convert::Infallible;
use std::fs;
use std::io::Write;
use std::net::SocketAddr;
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant, SystemTime};

use bytes::Bytes;
use http_body_util::Full;
use hyper::body::Incoming;
use hyper::header::{self, HeaderName, HeaderValue};
use hyper::http::response::Builder;
use hyper::server::conn::http1;
use hyper::service::service_fn;
use hyper::{HeaderMap, Method, Request, Response, StatusCode};
use hyper_util::rt::{TokioIo, TokioTimer};
use tokio::fs::File;
use tokio::io::{AsyncReadExt, AsyncSeekExt};
use tokio::net::{TcpListener, TcpStream};
use tokio::sync::{OwnedSemaphorePermit, Semaphore};
use tokio::time::timeout;

use crate::error::StaticError;
use crate::handler::{FileBody, ResponseBody};
use crate::reload::{self, SseBody};
use crate::resolve;
use crate::resolve::HiddenFiles;
use crate::spa::{self, SpaTransition};
use crate::watcher::{start_watching, Broadcaster};

const ACCEPT_BACKOFF_INITIAL: Duration = Duration::from_millis(10);
const ACCEPT_BACKOFF_MAX: Duration = Duration::from_secs(1);

/// Default maximum concurrent connections, overridable via `Server::with_max_connections`.
const DEFAULT_MAX_CONNECTIONS: usize = 1024;

/// A source of accepted TCP connections. Abstracted so the accept-error backoff below
/// can be exercised against a listener that fails on demand, without needing to provoke
/// real OS-level accept errors (e.g. EMFILE) in tests.
trait TcpAccept {
    async fn accept(&self) -> std::io::Result<(TcpStream, SocketAddr)>;
}

impl TcpAccept for TcpListener {
    async fn accept(&self) -> std::io::Result<(TcpStream, SocketAddr)> {
        TcpListener::accept(self).await
    }
}

/// Accept a connection and reserve it a connection-limit permit.
///
/// `backoff` retries a failed `accept()` after an exponentially growing delay (reset on
/// the next success, capped at `ACCEPT_BACKOFF_MAX`) instead of ending the accept loop,
/// so a sustained failure — the process being out of file descriptors, say — degrades
/// into periodic retries rather than a CPU-bound busy spin or a permanently dead server.
///
/// Returns `None` only if the semaphore itself has been closed (never happens in normal
/// operation, since nothing ever calls `close()` on it — handled so a caller can still
/// fail safely rather than panic).
async fn accept_and_permit<L: TcpAccept>(
    listener: &L,
    backoff: &mut Duration,
    semaphore: &Arc<Semaphore>,
    log: Option<&Server>,
) -> Option<(TcpStream, OwnedSemaphorePermit)> {
    loop {
        let stream = match listener.accept().await {
            Ok((stream, _)) => {
                *backoff = ACCEPT_BACKOFF_INITIAL;
                stream
            }
            Err(error) => {
                // Logged rather than swallowed: sustained accept failure (out of file
                // descriptors, most often) degrades into an ever-slower retry loop that
                // is otherwise indistinguishable from an idle server.
                if let Some(server) = log {
                    server.log(format_args!(
                        "accept error: {error}; retrying in {backoff:?}"
                    ));
                }
                tokio::time::sleep(*backoff).await;
                *backoff = (*backoff * 2).min(ACCEPT_BACKOFF_MAX);
                continue;
            }
        };
        return semaphore
            .clone()
            .acquire_owned()
            .await
            .ok()
            .map(|permit| (stream, permit));
    }
}

/// A predicate deciding whether a resolved file path should get an immutable cache
/// policy; see [`Server::with_immutable_assets`].
type ImmutablePredicate = Arc<dyn Fn(&Path) -> bool + Send + Sync>;

/// A static file server for serving files securely from a root directory.
///
/// `Server` canonicalizes the root directory once at creation time and uses the
/// canonical form for all subsequent requests, avoiding repeated filesystem calls.
///
/// # Security
///
/// The server protects against:
/// - Path traversal attacks (e.g., `../../etc/passwd`)
/// - Accessing files outside the root via symlinks
/// - Disclosing filesystem structure (traversal and missing files both return 404)
///
/// # Cloning
///
/// `Server` is cheap to clone: a `PathBuf`, a couple of primitives, and an `Arc`'d
/// predicate closure. Multiple clones can be used concurrently in async tasks without
/// synchronization overhead.
///
/// # Example
///
/// ```no_run
/// # async fn example() -> Result<(), Box<dyn std::error::Error>> {
/// use mini_static::Server;
/// use std::path::Path;
/// use std::time::Duration;
///
/// let server = Server::new(Path::new("./public"))?;
/// let (port, _handle) = server.run(Duration::from_secs(30)).await?;
/// println!("Server running on port {}", port);
/// # Ok(())
/// # }
/// ```
/// Headers this server derives from the response it is building, and therefore refuses
/// as fixed values via [`Server::with_response_header`]. A fixed value would be either
/// silently overridden or silently duplicated depending on the response — and a wrong
/// `Content-Length` or `ETag` is a correctness bug, not a policy choice.
const SERVER_COMPUTED_HEADERS: [HeaderName; 13] = [
    header::CONTENT_LENGTH,
    header::CONTENT_TYPE,
    header::CONTENT_ENCODING,
    header::CONTENT_RANGE,
    header::ETAG,
    header::CACHE_CONTROL,
    header::VARY,
    header::ACCEPT_RANGES,
    header::ALLOW,
    header::LOCATION,
    header::CONNECTION,
    header::TRANSFER_ENCODING,
    header::X_CONTENT_TYPE_OPTIONS,
];

/// Where request and connection log lines go.
///
/// A `Server` is cloned per request, so the sink is shared rather than duplicated. The
/// mutex serializes writes from concurrent connections — without it, two responses
/// finishing at once would interleave mid-line and produce log entries belonging to
/// neither request.
type RequestLog = Arc<Mutex<Box<dyn Write + Send>>>;

#[derive(Clone)]
pub struct Server {
    root_canon: PathBuf,
    max_connections: usize,
    live_reload: bool,
    broadcaster: Option<Broadcaster>,
    spa_mode: bool,
    spa_root: Option<String>,
    spa_transition: SpaTransition,
    not_found_page: Option<PathBuf>,
    hidden_files: HiddenFiles,
    request_log: Option<RequestLog>,
    extra_headers: Arc<Vec<(HeaderName, HeaderValue)>>,
    immutable_predicate: Option<ImmutablePredicate>,
}

impl Server {
    /// Create a new server with the given root directory.
    ///
    /// Canonicalizes the root once at startup. All subsequent requests use the
    /// canonical root without re-canonicalizing it, making this suitable for long-lived servers.
    ///
    /// # Errors
    ///
    /// Returns `Err(StaticError::Io)` if the root cannot be canonicalized (e.g., doesn't exist,
    /// no read permissions).
    pub fn new(root: &Path) -> Result<Self, StaticError> {
        let root_canon = root.canonicalize().map_err(StaticError::Io)?;
        Ok(Server {
            root_canon,
            max_connections: DEFAULT_MAX_CONNECTIONS,
            live_reload: false,
            broadcaster: None,
            spa_mode: false,
            spa_root: None,
            spa_transition: SpaTransition::default(),
            not_found_page: None,
            hidden_files: HiddenFiles::Deny,
            request_log: None,
            extra_headers: Arc::new(Vec::new()),
            immutable_predicate: None,
        })
    }

    /// Set the maximum number of connections served concurrently (default 1024).
    ///
    /// Once this many connections are in flight, `run()`'s accept loop stops accepting
    /// new ones — without pausing the accept loop, a client that opens a connection and
    /// sends nothing (see the header-read timeout docs on [`Server::run_on`]) could
    /// otherwise be used, in enough parallel copies, to exhaust the process's file
    /// descriptors or memory with no bound at all.
    pub fn with_max_connections(mut self, max: usize) -> Self {
        self.max_connections = max;
        self
    }

    /// Enable live-reload for this server (disabled by default).
    ///
    /// Once enabled, the `run*` methods start a background watcher (mtime polling,
    /// bounded 500ms interval — see [`crate::start_watching`]) the first time the server
    /// actually starts accepting connections. It watches the served root; when a build
    /// pipeline is configured it watches that pipeline's source folders instead, because
    /// the pipeline broadcasts its own outputs once they are written. Then it will:
    ///
    /// - serve a live-reload SSE stream at [`crate::LIVE_RELOAD_PATH`], broadcasting a
    ///   change event (with [`crate::ChangeType`]) whenever a served file is added,
    ///   modified, or removed;
    /// - inject a small `<script>` into every served `text/html` response that connects
    ///   to that stream and reloads the page (or hot-swaps stylesheet `<link>`s, for CSS
    ///   changes) — no manual client wiring required.
    ///
    /// This is meant for local development, not production: leave it disabled (the
    /// default) for any server serving real traffic. A typical call site gates it behind
    /// `#[cfg(debug_assertions)]` so a release build never pays for the watcher or the
    /// injected script.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?;
    /// #[cfg(debug_assertions)]
    /// let server = server.with_live_reload();
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_live_reload(mut self) -> Self {
        self.live_reload = true;
        self
    }

    /// Enable spa-mode navigation for this server, swapping `document.body` on
    /// each navigation (disabled by default).
    ///
    /// Once enabled, every served `text/html` response gets a small `<script>`
    /// injected (see [`Server::with_spa_root`] for what it does) that treats
    /// `document.body` as the swap target. Calling this after
    /// [`Server::with_spa_root`] does not clear a previously configured root
    /// selector — the two methods set independent fields, so
    /// `.with_spa_root(sel).with_spa_mode()` and
    /// `.with_spa_mode().with_spa_root(sel)` both end up with spa-mode on and
    /// root `sel`. Use this one alone when there's no persistent chrome to
    /// preserve across navigations.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?.with_spa_mode();
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_spa_mode(mut self) -> Self {
        self.spa_mode = true;
        self
    }

    /// Enable spa-mode navigation for this server, swapping only the element
    /// matched by the CSS `selector` on each navigation (disabled by default;
    /// also enables spa-mode the same as [`Server::with_spa_mode`]).
    ///
    /// Once enabled, every served `text/html` response gets a small `<script>`
    /// injected that intercepts left-clicks on same-origin `<a href>`
    /// elements (skipping links with a non-`_self` `target`, a `download`
    /// attribute, `rel="external"`, a `data-no-spa` attribute, or a same-page
    /// hash-only href) and, instead of a normal navigation:
    ///
    /// - fetches the target URL;
    /// - on a non-OK or non-`text/html` response (or a fetch error), falls
    ///   back to a real `location.href` navigation — spa-mode never renders a
    ///   broken page;
    /// - otherwise replaces the matched element's `innerHTML` with the
    ///   corresponding content from the fetched document, updates the page
    ///   title, and pushes the new URL via `history.pushState`, animating the
    ///   swap with `document.startViewTransition()` where supported;
    /// - dispatches a `mini-static:navigate` `CustomEvent` on `window` after
    ///   every client-side navigation, so page scripts can re-run any
    ///   per-page initialization that would otherwise only execute once
    ///   (content swapped in via `innerHTML` never executes its own
    ///   `<script>` tags);
    /// - handles browser back/forward by re-fetching and swapping to the new
    ///   `location.href`.
    ///
    /// `selector` is matched against both the current page and the fetched
    /// page; a link click where the selector matches neither falls back to a
    /// real navigation, same as a fetch failure. Choose a `selector` that
    /// wraps only the content that varies between pages, leaving persistent
    /// chrome (nav/header/footer) outside it so it survives navigation
    /// untouched.
    ///
    /// This is meant to be usable in production, not just local development
    /// (unlike [`Server::with_live_reload`]): a click on a link mini-static
    /// doesn't intercept, or on a browser without JS or View Transitions
    /// support, still works as a normal navigation.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?.with_spa_root("#app");
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_spa_root(mut self, selector: &str) -> Self {
        self.spa_mode = true;
        self.spa_root = Some(selector.to_string());
        self
    }

    /// Set how spa-mode animates the swap between pages (also enables
    /// spa-mode the same as [`Server::with_spa_mode`]; default
    /// [`SpaTransition::Fade`] when spa-mode is enabled without calling this).
    ///
    /// [`SpaTransition::Slide`] injects its own `<style>` tag alongside the
    /// spa-mode `<script>` — no site CSS is required. See [`SpaTransition`]
    /// and [`SlideOptions`] for what each variant does and how to
    /// configure the slide's duration, direction, and easing.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::{Server, SlideOptions, SpaTransition};
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?
    ///     .with_spa_root("#app")
    ///     .with_spa_transition(SpaTransition::Slide(
    ///         SlideOptions::default().duration_ms(500),
    ///     ));
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_spa_transition(mut self, transition: SpaTransition) -> Self {
        self.spa_mode = true;
        self.spa_transition = transition;
        self
    }

    /// Serves `path` as the body of every `404`, instead of the default plain-text
    /// `not found`.
    ///
    /// `path` is resolved relative to the served root and must exist when this is
    /// called: a missing 404 page is a deployment mistake, and finding out on the first
    /// broken link — the one moment the page exists to handle — is too late. It is read
    /// from disk per response rather than cached, so editing it during a live-reload
    /// session takes effect without a restart.
    ///
    /// The response keeps its `404` status. Serving a custom page with `200` is a soft
    /// 404: search engines index it, and monitoring stops seeing the failures. It also
    /// carries `Cache-Control: no-store`, so a client never holds this page as though it
    /// were the resource that was actually requested.
    ///
    /// Nothing about the failed request reaches the page — no path, no reason. A
    /// traversal attempt and an ordinary miss are deliberately indistinguishable
    /// (`StaticError::user_message`), and templating the requested path into the
    /// response would undo that and hand back a reflected-content vector besides.
    ///
    /// # Errors
    ///
    /// Returns `Err(StaticError::Io)` if `path` cannot be canonicalized (typically:
    /// it does not exist), or `Err(StaticError::Traversal)` if it lies outside the
    /// served root.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?
    ///     .with_not_found_page(Path::new("404.html"))?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_not_found_page(mut self, path: &Path) -> Result<Self, StaticError> {
        let joined = self.root_canon.join(path);
        let canon = joined.canonicalize().map_err(StaticError::Io)?;

        if !canon.starts_with(&self.root_canon) {
            return Err(StaticError::Traversal(format!(
                "404 page {} lies outside the served root {}",
                canon.display(),
                self.root_canon.display()
            )));
        }

        self.not_found_page = Some(canon);
        Ok(self)
    }

    /// Serve dot-prefixed paths (`.env`, `.git/config`) instead of answering them as a
    /// miss.
    ///
    /// Hidden files are denied by default. A served root is routinely a build output
    /// directory, a repository working copy, or a folder someone dropped a `.env` into,
    /// and the traversal guard cannot help: those files are legitimately *inside* the
    /// root, so anyone who guesses the name gets them. The default trades a rarely-wanted
    /// capability for not leaking credentials by accident.
    ///
    /// `/.well-known/` is served either way — it is where the web puts resources that
    /// are meant to be fetched (ACME challenges for certificate issuance,
    /// `security.txt`), and denying it would break certificate renewal. The exception is
    /// the first segment only: `/.well-known/.hidden` is still denied.
    ///
    /// Call this when the served root is a curated directory whose dotfiles are content
    /// — a static site that publishes a `.htaccess` for a downstream server, say.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?.with_hidden_files();
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_hidden_files(mut self) -> Self {
        self.hidden_files = HiddenFiles::Serve;
        self
    }

    /// Log one line per request to stderr, plus connection-level errors.
    ///
    /// Off by default: a library that writes to a process's stderr uninvited is a
    /// surprise, and an embedder with its own logging wants the lines somewhere else.
    /// See [`Server::with_request_logging_to`] to choose the destination.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?.with_request_logging();
    /// # Ok(())
    /// # }
    /// ```
    /// Send `name: value` on every response.
    ///
    /// Call repeatedly to add several. Intended for the policy headers a static site
    /// wants applied uniformly — `Strict-Transport-Security`, `Content-Security-Policy`,
    /// `Referrer-Policy` — which this crate has no business choosing on an embedder's
    /// behalf but every business making expressible.
    ///
    /// Both name and value are validated here, at configuration time, so a malformed
    /// header fails when the server is built rather than on a request months later.
    ///
    /// # Errors
    ///
    /// - `StaticError::Config` if `name` or `value` is not a valid HTTP header.
    /// - `StaticError::Config` if `name` is one this server computes per response
    ///   (`Content-Length`, `Content-Type`, `Content-Encoding`, `Content-Range`, `ETag`,
    ///   `Cache-Control`, `Vary`, `Accept-Ranges`, `Allow`, `Location`, `Connection`,
    ///   `Transfer-Encoding`, `X-Content-Type-Options`). A fixed value would either be
    ///   silently overridden or silently duplicated depending on the response — a
    ///   configuration mistake worth surfacing at startup rather than a behavior worth
    ///   supporting. Use [`Server::with_immutable_assets`] for cache policy.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?
    ///     .with_response_header("Strict-Transport-Security", "max-age=63072000")?
    ///     .with_response_header("Referrer-Policy", "strict-origin-when-cross-origin")?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_response_header(mut self, name: &str, value: &str) -> Result<Self, StaticError> {
        let name = HeaderName::from_bytes(name.as_bytes())
            .map_err(|_| StaticError::Config(format!("invalid header name: {name}")))?;
        let value = HeaderValue::from_str(value).map_err(|_| {
            StaticError::Config(format!("invalid value for header {name}: {value}"))
        })?;

        if SERVER_COMPUTED_HEADERS.contains(&name) {
            return Err(StaticError::Config(format!(
                "{name} is computed per response and cannot be set as a fixed header"
            )));
        }

        Arc::make_mut(&mut self.extra_headers).push((name, value));
        Ok(self)
    }

    pub fn with_request_logging(self) -> Self {
        self.with_request_logging_to(Box::new(std::io::stderr()))
    }

    /// Log one line per request to `writer`, plus connection-level errors.
    ///
    /// Each served request writes one line:
    ///
    /// ```text
    /// GET /index.html 200 512 0.421ms
    /// ```
    ///
    /// — method, requested path exactly as received, status, response body bytes (`-`
    /// when the length isn't known, as on a live-reload SSE stream), and how long
    /// handling took. Connection-level failures — a malformed request, a client
    /// vanishing mid-response — write `connection error: <cause>`; before this they were
    /// discarded entirely, so a server that was refusing every request looked exactly
    /// like one nobody was talking to.
    ///
    /// The path is logged as received, *not* decoded: it is attacker-controlled input,
    /// and a log reader deserves to see the bytes that actually arrived rather than a
    /// normalized rendering of them.
    ///
    /// Writes are serialized across connections and write errors are ignored — a
    /// failing log sink must not take down request serving.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::fs::File;
    /// use std::path::Path;
    ///
    /// let log = File::create("access.log")?;
    /// let server = Server::new(Path::new("./public"))?.with_request_logging_to(Box::new(log));
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_request_logging_to(mut self, writer: Box<dyn Write + Send>) -> Self {
        self.request_log = Some(Arc::new(Mutex::new(writer)));
        self
    }

    /// Start a response carrying the baseline security header plus every header the
    /// embedder configured via [`Server::with_response_header`].
    ///
    /// Every response this server builds for a request goes through here, so a
    /// configured policy header cannot be missing from one status and present on
    /// another. The sole exception is the `400` that `finish` falls back to when a
    /// builder produced an invalid header — no `Server` is in scope there, and a
    /// response that exists only because header construction already failed is the wrong
    /// place to add more headers.
    fn response(&self, status: StatusCode) -> Builder {
        let mut builder = response(status);
        for (name, value) in self.extra_headers.iter() {
            builder = builder.header(name, value);
        }
        builder
    }

    /// Write `line` to the configured log sink, if there is one.
    ///
    /// A poisoned mutex (some earlier writer panicked mid-write) and a failed write are
    /// both ignored: neither is a reason to fail a request that was otherwise served
    /// correctly.
    fn log(&self, line: std::fmt::Arguments<'_>) {
        let Some(log) = &self.request_log else {
            return;
        };
        if let Ok(mut sink) = log.lock() {
            let _ = writeln!(sink, "{line}");
            let _ = sink.flush();
        }
    }

    /// Serve files matching `predicate` with a long-lived, immutable cache policy
    /// instead of the default `Cache-Control: no-cache`.
    ///
    /// `predicate` is evaluated against each resolved file's path; a match sends
    /// `Cache-Control: public, max-age=31536000, immutable` on that file's 200 and 304
    /// responses. This is correct only for fingerprinted assets (e.g.
    /// `main.a1b2c3.js`) where a content change always produces a new filename —
    /// caching a mutable filename indefinitely would serve stale content to every
    /// client that already has it cached.
    ///
    /// # Example
    ///
    /// ```no_run
    /// # fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?
    ///     .with_immutable_assets(|path| {
    ///         path.file_name()
    ///             .and_then(|name| name.to_str())
    ///             .is_some_and(|name| name.contains(".fingerprint."))
    ///     });
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_immutable_assets<F>(mut self, predicate: F) -> Self
    where
        F: Fn(&Path) -> bool + Send + Sync + 'static,
    {
        self.immutable_predicate = Some(Arc::new(predicate));
        self
    }

    /// The `Cache-Control` header value for a resolved file path: the immutable policy
    /// if `with_immutable_assets`'s predicate matches, `no-cache` otherwise.
    fn cache_control_for(&self, path: &Path) -> &'static str {
        match &self.immutable_predicate {
            Some(predicate) if predicate(path) => "public, max-age=31536000, immutable",
            _ => "no-cache",
        }
    }

    /// The directories the live-reload watcher polls: the served root, and only that.
    ///
    /// Until 0.29.0 this also returned the build pipeline's source folders, and the
    /// served root was excluded whenever a pipeline was configured — watching the output
    /// dir would have fed each pipeline its own writes back into its own trigger. The
    /// pipeline now lives in `mini-build`, in a separate process, so nothing this server
    /// watches is written by this server and the exclusion has nothing left to prevent.
    fn watch_targets(&self) -> Vec<PathBuf> {
        vec![self.root_canon.clone()]
    }

    /// Resolve a request path under the server's root.
    ///
    /// This is a lower-level API for resolving paths without generating HTTP responses.
    /// For most use cases, prefer [`Server::handle_request`] or the `run*` methods.
    ///
    /// # Returns
    ///
    /// - `Ok(PathBuf)` if the path resolves to a file within root.
    /// - `Err(StaticError)` if the path is invalid, missing, or attempts traversal.
    pub fn resolve(&self, request_path: &str) -> Result<PathBuf, StaticError> {
        resolve::resolve_with_policy(&self.root_canon, request_path, self.hidden_files)
    }

    /// Builds the `404` response: the configured page when there is one and it can be
    /// read, and `fallback` as plain text otherwise.
    ///
    /// `fallback` is the caller's already-sanitized message (see
    /// `StaticError::user_message`) — never the requested path, so an ordinary miss and
    /// a rejected traversal stay indistinguishable to whoever is probing.
    ///
    /// A page that vanished after `with_not_found_page` validated it degrades to that
    /// text rather than to a `500`: the request was still a miss, and answering a
    /// missing page with the wrong status would be a second bug wearing the first one's
    /// clothes.
    async fn not_found_response(&self, fallback: &'static str) -> Response<ResponseBody> {
        let builder = self
            .response(StatusCode::NOT_FOUND)
            .header("Cache-Control", "no-store");

        let Some(page) = &self.not_found_page else {
            return text(builder, format!("{fallback}\n"));
        };
        let Ok(body) = tokio::fs::read(page).await else {
            return text(builder, format!("{fallback}\n"));
        };

        text(
            builder.header("Content-Type", "text/html; charset=utf-8"),
            body,
        )
    }

    /// Run the server on a specific address with a configurable header-read timeout.
    ///
    /// Spawns the server in a background Tokio task and returns immediately with the
    /// assigned port number and a [`ServerHandle`]. Call `handle.shutdown().await` to
    /// stop accepting new connections and wait for in-flight connections to finish.
    /// Dropping the handle instead leaves the server running for the life of the process.
    ///
    /// # Header-Read Timeout
    ///
    /// Connections that don't send complete HTTP headers within `header_timeout` are closed.
    /// This prevents slowloris attacks and resource exhaustion from incomplete requests. The
    /// timeout applies only to the header-read phase — once a complete header block has been
    /// read, the connection is handed off with no further time bound, so long-lived response
    /// bodies (e.g. the live-reload SSE stream from [`Server::with_live_reload`]) are not cut
    /// off mid-stream.
    ///
    /// # Precompressed Sidecars
    ///
    /// If a request's `Accept-Encoding` allows `br` or `gzip` (preferring `br`) and a
    /// sibling `<path>.br`/`<path>.gz` exists next to the resolved file, its bytes are
    /// served instead with a matching `Content-Encoding`. Every file response carries
    /// `Vary: Accept-Encoding` so intermediate caches don't serve the wrong variant to a
    /// differently-capable client.
    ///
    /// # Arguments
    ///
    /// * `addr` - Socket address to bind to (e.g., `127.0.0.1:0` for loopback ephemeral,
    ///   or `0.0.0.0:8080` to bind all interfaces on a fixed port).
    /// * `header_timeout` - Maximum time to wait for complete HTTP headers on each connection.
    ///
    /// # Returns
    ///
    /// - `Ok((u16, ServerHandle))` with the assigned port number and a handle for graceful shutdown.
    /// - `Err(StaticError::Io)` if binding to the socket fails.
    /// - `Err(StaticError::PipelineSetup)` if a configured [`CssTool`]/[`JsTool`]'s binary
    ///   is not found on `PATH`. Checked before the listener binds: a deployment whose
    ///   configured pipeline can never run should fail visibly at boot, not be discovered
    ///   later as a missing/stale asset.
    pub async fn run_on(
        &self,
        addr: SocketAddr,
        header_timeout: Duration,
    ) -> Result<(u16, ServerHandle), StaticError> {
        let listener = TcpListener::bind(addr).await.map_err(StaticError::Io)?;
        let port = listener.local_addr().map_err(StaticError::Io)?.port();

        let mut server = self.clone();
        if server.live_reload {
            // The served root is the only watch target now that nothing writes into it.
            // While the build pipeline lived here the output dir was deliberately never
            // watched, because watching it fed each pipeline its own writes back into its
            // trigger; with the builder in a separate process that loop cannot happen.
            let broadcaster = Broadcaster::new();
            for dir in server.watch_targets() {
                start_watching(Arc::new(dir), broadcaster.clone());
            }
            server.broadcaster = Some(broadcaster);
        }

        let semaphore = Arc::new(Semaphore::new(server.max_connections));
        let (shutdown_tx, shutdown_rx) = tokio::sync::oneshot::channel();

        let accept_task = tokio::spawn(async move {
            let mut backoff = ACCEPT_BACKOFF_INITIAL;
            let mut join_set: tokio::task::JoinSet<()> = tokio::task::JoinSet::new();
            let mut shutdown_pin = std::pin::pin!(shutdown_rx);
            let mut shutting_down = false;

            loop {
                if !shutting_down {
                    // The accept-and-permit step and the shutdown signal race in a single
                    // `select!` so shutdown can preempt a pending accept or a permit wait
                    // cleanly, at any point — not just between loop iterations.
                    tokio::select! {
                        accepted = accept_and_permit(&listener, &mut backoff, &semaphore, Some(&server)) => {
                            match accepted {
                                Some((stream, permit)) => {
                                    let server = server.clone();
                                    join_set.spawn(async move {
                                        let _permit = permit;
                                        serve_connection(stream, server, header_timeout).await;
                                    });
                                }
                                None => shutting_down = true,
                            }
                        }
                        _ = shutdown_pin.as_mut() => {
                            shutting_down = true;
                        }
                    }
                    continue;
                }

                // Stop accepting; drain already-spawned connections before returning.
                match join_set.join_next().await {
                    Some(_) => continue,
                    None => break,
                }
            }
        });

        Ok((
            port,
            ServerHandle {
                shutdown_tx: Some(shutdown_tx),
                accept_task,
            },
        ))
    }

    /// Run the server on loopback (127.0.0.1), binding an ephemeral port.
    ///
    /// Thin wrapper around [`Server::run_on`] — see it for the header-read timeout and
    /// sidecar semantics, and for what the returned [`ServerHandle`] does.
    pub async fn run(&self, header_timeout: Duration) -> Result<(u16, ServerHandle), StaticError> {
        self.run_on(([127, 0, 0, 1], 0).into(), header_timeout)
            .await
    }

    /// Run the server on all interfaces (0.0.0.0) at `port` (0 for an ephemeral port).
    ///
    /// Useful for containerized deployments and reverse-proxy setups. Thin wrapper
    /// around [`Server::run_on`] — see it for the header-read timeout and sidecar
    /// semantics, and for what the returned [`ServerHandle`] does.
    pub async fn run_all(
        &self,
        port: u16,
        header_timeout: Duration,
    ) -> Result<(u16, ServerHandle), StaticError> {
        self.run_on(([0, 0, 0, 0], port).into(), header_timeout)
            .await
    }

    /// Run the server on loopback with the default 30-second header-read timeout.
    ///
    /// The recommended entry point for tests and lightweight services that don't need a
    /// custom timeout. Thin wrapper around [`Server::run`].
    ///
    /// # Example
    ///
    /// ```no_run
    /// # async fn example() -> Result<(), Box<dyn std::error::Error>> {
    /// use mini_static::Server;
    /// use std::path::Path;
    ///
    /// let server = Server::new(Path::new("./public"))?;
    /// let (port, handle) = server.run_ephemeral().await?;
    /// println!("Server ready on http://127.0.0.1:{}", port);
    /// handle.shutdown().await;
    /// # Ok(())
    /// # }
    /// ```
    pub async fn run_ephemeral(&self) -> Result<(u16, ServerHandle), StaticError> {
        self.run(DEFAULT_HEADER_TIMEOUT).await
    }

    /// Produce the HTTP response for a request, streaming file bodies to the client.
    ///
    /// This is the crate's single request-handling path: the `run*` accept loop calls it,
    /// and so should any async server embedding `mini-static` as a fallback route (e.g.
    /// `mini-unified`).
    ///
    /// Filesystem metadata work (path resolution, `open`, `stat`) runs *inline* on the
    /// calling task, deliberately. Until 0.30.0 it was dispatched to Tokio's blocking
    /// pool so a slow filesystem could not stall co-scheduled tasks — measured under
    /// load, that dispatch cost roughly three times the syscalls it sheltered, and a
    /// one-worker server burned nearly four cores on pool handoff. On the local-disk
    /// deployments this crate targets these calls are single-digit microseconds; an
    /// embedder serving from a filesystem with unbounded latency (a network mount)
    /// should use a multi-threaded runtime, which bounds the blast radius of a stall
    /// to one worker.
    ///
    /// File responses are backed by `FileBody`, which hands hyper one 64 KB chunk at a
    /// time as `poll_frame` is driven: memory use stays bounded to one chunk per in-flight
    /// response regardless of file size.
    ///
    /// `headers` are the request's headers; `If-None-Match` (304 on a matching ETag) and
    /// `Accept-Encoding` (precompressed sidecar selection, see [`Server::run_on`]) are the
    /// ones read today. Only `GET` and `HEAD` are allowed; anything else gets a 405 with an
    /// `Allow` header. Missing files and traversal attempts both get an identical 404, so a
    /// response never discloses whether a path exists outside the root.
    pub async fn handle_request(
        &self,
        method: &Method,
        request_path: &str,
        headers: &HeaderMap,
    ) -> Response<ResponseBody> {
        if method != Method::GET && method != Method::HEAD {
            return text(
                self.response(StatusCode::METHOD_NOT_ALLOWED)
                    .header("Allow", "GET, HEAD"),
                "method not allowed\n",
            );
        }

        // Live-reload SSE stream — only reachable when `with_live_reload()` was called
        // and the server was started via a `run*` method (those are the only paths that
        // populate `broadcaster`).
        if *method == Method::GET && request_path == reload::LIVE_RELOAD_PATH {
            if let Some(broadcaster) = &self.broadcaster {
                return finish(
                    self.response(StatusCode::OK)
                        .header("Content-Type", "text/event-stream")
                        .header("Cache-Control", "no-cache")
                        .header("Connection", "keep-alive")
                        .body(ResponseBody::Sse(SseBody::new(broadcaster.subscribe()))),
                );
            }
        }

        // Inline on purpose — see this function's doc comment for the measured case
        // against the old `spawn_blocking` dispatch. One call opens the file and proves
        // containment on the opened fd, so there is no separate open to fail later and
        // no window between the check and the handle that gets served.
        let resolved =
            match resolve::open_with_policy(&self.root_canon, request_path, self.hidden_files) {
                Err(e) => return self.not_found_response(e.user_message()).await,
                Ok(resolved) => resolved,
            };
        let (std_file, metadata, path) = (resolved.file, resolved.metadata, resolved.path);

        // A directory served via its `index.html` needs a trailing slash to establish the
        // correct base for the page's relative links. Compare against the *decoded*
        // request path so a percent-encoded explicit request for index.html (e.g.
        // `/docs/index.htm%6c`) is recognized as such instead of producing a redirect to a
        // still-encoded, broken Location.
        let decoded_request_path = resolve::decode_request_path(request_path);
        if path.file_name().is_some_and(|name| name == "index.html")
            && !decoded_request_path.ends_with('/')
            && !decoded_request_path.ends_with("index.html")
        {
            // `location` is built from the (attacker-controlled) request path; `finish()`
            // degrades to 400 instead of panicking if it ever contains bytes invalid in a
            // header value.
            let location = format!("{}/", request_path.trim_end_matches('/'));
            return text(
                self.response(StatusCode::MOVED_PERMANENTLY)
                    .header("Location", location),
                "moved\n",
            );
        }

        let content_type = mime_type_for_path(&path);
        // Live-reload and spa-mode HTML injection both need the original, uncompressed
        // bytes to splice their script into — never substitute a precompressed sidecar on
        // this path. `broadcaster` is only `Some` when live-reload is enabled (see
        // `Server::with_live_reload`); `spa_mode` is independent of it (see
        // `Server::with_spa_mode`/`with_spa_root`) — either alone is enough to trigger
        // injection.
        // Injection reads the whole file into memory, so it is also gated on size. Every
        // decision keyed off `html_injection` — the sidecar skip below, the range skip,
        // the full read itself — inherits the cap from this one boolean, so an over-cap
        // page takes the ordinary streamed path with no second decision point.
        let wants_injection =
            (self.broadcaster.is_some() || self.spa_mode) && content_type.starts_with("text/html");
        let html_injection = wants_injection && metadata.len() <= MAX_INJECTABLE_HTML_BYTES;

        if wants_injection && !html_injection {
            self.log(format_args!(
                "html injection skipped for {request_path}: {} bytes exceeds the \
                 {MAX_INJECTABLE_HTML_BYTES}-byte limit; serving unmodified",
                metadata.len(),
            ));
        }

        let range_header = header_str(headers, "range");
        let if_range_header = header_str(headers, "if-range");

        let accept_encoding = header_str(headers, "accept-encoding");
        // Skip precompressed sidecars when Range is requested (serve original file instead).
        let sidecar = if html_injection || range_header.is_some() {
            None
        } else {
            select_precompressed_sidecar(&path, accept_encoding)
        };
        let (std_file, metadata, content_encoding) = match sidecar {
            Some((sidecar_file, sidecar_metadata, encoding)) => {
                (sidecar_file, sidecar_metadata, Some(encoding))
            }
            None => (std_file, metadata, None),
        };
        // Async from here only where streaming actually needs it; the buffered path
        // below reads from this same open handle, so what was probed is what is served.
        let mut file = File::from_std(std_file);

        // HTML injection is skipped for a served precompressed sidecar (already final
        // bytes from a build step) — see `html_injection`'s definition above.
        let etag = generate_etag(&metadata);
        let cache_control = self.cache_control_for(&path);

        if header_str(headers, "if-none-match").is_some_and(|value| is_etag_match(value, &etag)) {
            return finish(
                self.response(StatusCode::NOT_MODIFIED)
                    .header("Cache-Control", cache_control)
                    .header("Vary", "Accept-Encoding")
                    .header("ETag", etag)
                    .header("Accept-Ranges", "bytes")
                    .body(ResponseBody::Buffered(Full::new(Bytes::new()))),
            );
        }

        // `Some` when the served representation differs from the file's raw bytes and had
        // to be built in memory; `None` means stream the open file as-is. Computed before
        // the HEAD check below because RFC 9110 requires a HEAD response's headers —
        // `Content-Length` included — to match what a GET would send, even though the body
        // itself is dropped.
        let transformed: Option<Bytes> = if html_injection {
            let mut html = Vec::with_capacity(metadata.len() as usize);
            if file.read_to_end(&mut html).await.is_err() {
                return internal_error_response();
            }
            if self.broadcaster.is_some() {
                reload::inject_reload_script(&mut html);
            }
            if self.spa_mode {
                spa::inject_spa_script(&mut html, self.spa_root.as_deref(), &self.spa_transition);
            }
            Some(Bytes::from(html))
        } else {
            None
        };

        let file_size = transformed
            .as_ref()
            .map_or(metadata.len(), |bytes| bytes.len() as u64);

        // Handle Range requests.
        let range_outcome = range_header.map(|h| parse_range_header(h, file_size));
        let range_check = if let Some(outcome) = &range_outcome {
            match outcome {
                RangeOutcome::Satisfiable(start, end) => {
                    // If-Range validation: stale If-Range ignores Range, serves full 200.
                    if let Some(if_range) = if_range_header {
                        if !if_range_valid(if_range, &etag) {
                            RangeCheck::IgnoreRange
                        } else {
                            RangeCheck::Satisfiable(*start, *end)
                        }
                    } else {
                        RangeCheck::Satisfiable(*start, *end)
                    }
                }
                RangeOutcome::MultiRangeIgnored => RangeCheck::IgnoreRange,
                RangeOutcome::Unsatisfiable => RangeCheck::Unsatisfiable,
                RangeOutcome::NoRange => RangeCheck::IgnoreRange,
            }
        } else {
            RangeCheck::IgnoreRange
        };

        match &range_check {
            RangeCheck::Unsatisfiable => {
                return finish(
                    Response::builder()
                        .status(StatusCode::RANGE_NOT_SATISFIABLE)
                        .header("Content-Range", format!("bytes */{}", file_size))
                        .header("Accept-Ranges", "bytes")
                        .body(ResponseBody::Buffered(Full::new(Bytes::new()))),
                );
            }
            RangeCheck::Satisfiable(start, end) => {
                let range_len = end - start + 1;

                // Seek to start position; if sidecar, we already skipped it above.
                if transformed.is_none()
                    && file.seek(std::io::SeekFrom::Start(*start)).await.is_err()
                {
                    return internal_error_response();
                }

                // HEAD must not return a body (RFC 9110).
                let body = if *method == Method::HEAD {
                    ResponseBody::Buffered(Full::new(Bytes::new()))
                } else {
                    match transformed {
                        Some(ref bytes) => ResponseBody::Buffered(Full::new(
                            bytes.slice(*start as usize..(*end as usize + 1)),
                        )),
                        None => ResponseBody::Streamed(FileBody::new_ranged(file, range_len)),
                    }
                };

                let mut builder = Response::builder()
                    .status(StatusCode::PARTIAL_CONTENT)
                    .header("Content-Type", content_type)
                    .header("Content-Length", range_len.to_string())
                    .header(
                        "Content-Range",
                        format!("bytes {}-{}/{}", start, end, file_size),
                    )
                    .header("Cache-Control", cache_control)
                    .header("Vary", "Accept-Encoding")
                    .header("ETag", etag)
                    .header("Accept-Ranges", "bytes");
                if let Some(encoding) = content_encoding {
                    builder = builder.header("Content-Encoding", encoding);
                }
                return finish(builder.body(body));
            }
            RangeCheck::IgnoreRange => {}
        }

        // HEAD must not return a body (RFC 9110).
        let body = if *method == Method::HEAD {
            ResponseBody::Buffered(Full::new(Bytes::new()))
        } else {
            match transformed {
                Some(bytes) => ResponseBody::Buffered(Full::new(bytes)),
                None if metadata.len() <= INLINE_BODY_BYTES => {
                    // From the handle opened above — never by re-opening the path — so
                    // the bytes served are provably the file that was probed and
                    // stat'd, sidecars included, with no reopen window in between.
                    let mut std_file = file.into_std().await;
                    let mut bytes = Vec::with_capacity(metadata.len() as usize);
                    use std::io::Read as _;
                    if std_file.read_to_end(&mut bytes).is_err() {
                        return internal_error_response();
                    }
                    ResponseBody::Buffered(Full::new(Bytes::from(bytes)))
                }
                None => ResponseBody::Streamed(FileBody::new(file)),
            }
        };

        let mut builder = self
            .response(StatusCode::OK)
            .header("Content-Type", content_type)
            .header("Content-Length", file_size.to_string())
            .header("Cache-Control", cache_control)
            .header("Vary", "Accept-Encoding")
            .header("ETag", etag)
            .header("Accept-Ranges", "bytes");
        if let Some(encoding) = content_encoding {
            builder = builder.header("Content-Encoding", encoding);
        }
        finish(builder.body(body))
    }
}

/// Ceiling on how many bytes hyper buffers for a single request's header block before
/// rejecting it. Without this, a client that trickles bytes forever without ever sending
/// the terminating blank line could grow the buffer without limit — the header-read
/// timeout alone doesn't bound memory, only wall-clock time, and a sufficiently patient
/// sender could still send unbounded data before the deadline fires.
const MAX_HEADER_BYTES: usize = 64 * 1024;

/// Ceiling on the size of an HTML file this server will buffer in memory to splice a
/// live-reload or spa-mode `<script>` into.
///
/// Injection is the one code path that reads a whole file into memory rather than
/// streaming it in bounded chunks, and it does so *per request* — so without a cap, a
/// single large HTML file turns every concurrent request for it into another full copy
/// in memory, and spa-mode is a production feature, not a development-only one. An
/// over-cap page is served unmodified (and streamed) instead: losing a client-side
/// navigation enhancement on an 8 MiB document is a far smaller failure than an
/// allocation proportional to file size times concurrency.
///
/// 8 MiB is comfortably above any hand-written HTML page and any realistic
/// static-site-generator output, so the cap should never fire on content this feature
/// was designed for.
const MAX_INJECTABLE_HTML_BYTES: u64 = 8 * 1024 * 1024;

/// Bodies at or below this size are read synchronously and served from memory; larger
/// ones stream through `FileBody`. Equal to `FileBody`'s chunk size on purpose: at or
/// under one chunk the streaming path performed exactly one read anyway, so buffering
/// changes only *where* that read runs (inline, instead of a blocking-pool round trip
/// per chunk) — never how much memory a response can hold.
const INLINE_BODY_BYTES: u64 = 64 * 1024;

/// Wires an accepted connection up to the hyper HTTP/1.1 service.
///
/// HTTP/1.1 is the whole surface, deliberately. An earlier implementation used
/// `hyper_util`'s auto builder, whose `server-auto` feature transitively enables
/// `hyper/http2` — so a prior-knowledge h2c client negotiated HTTP/2 against a server
/// that documented, tested, and tuned only HTTP/1: no stream-concurrency limit, no
/// frame-size bound, and a header pre-read whose `\r\n\r\n` scan the HTTP/2 preface
/// satisfies without being an HTTP/1 request at all. Serving a protocol nobody
/// configured is worse than not serving it; browsers reach h2 over TLS only, which this
/// crate does not terminate.
///
/// `header_timeout` and `MAX_HEADER_BYTES` are enforced by hyper itself, per request.
/// They were previously enforced by a hand-rolled pre-read that ran once, before the
/// connection was handed to hyper — which meant the first request on a connection was
/// bounded and every subsequent keep-alive request on that same connection was not:
/// a client could complete one cheap request and then trickle headers forever, or send
/// an unbounded header block, with neither the timeout nor the size ceiling in play.
/// Delegating to hyper applies both bounds to every request, and deletes ~100 lines of
/// socket plumbing (`read_header_prefix`, `PrefixedIo`) whose only job was to hand the
/// already-read bytes back to hyper.
///
/// Only the *header* phase is bounded. A response body may legitimately outlive
/// `header_timeout` by design — the live-reload SSE stream stays open until a watched
/// file changes, possibly hours later — and `header_read_timeout` does not apply once a
/// request's headers are complete. The connection-count ceiling
/// (`Server::with_max_connections`) is what bounds resource use from connections held
/// open indefinitely.
///
/// `.timer(TokioTimer::new())` is load-bearing, not boilerplate: hyper resolves
/// `header_read_timeout` against an installed timer and panics with "timeout set, but no
/// timer set" if there isn't one (`hyper::common::time`). The failure is loud rather than
/// silent, but it happens per connection inside a spawned task — where a panic takes out
/// the connection, not the server — so it is the keep-alive tests, not a startup check,
/// that hold this wiring in place.
async fn serve_connection(stream: TcpStream, server: Server, header_timeout: Duration) {
    let io = TokioIo::new(stream);
    let log_server = server.clone();
    let svc = service_fn(move |req: Request<Incoming>| {
        let server = server.clone();
        async move {
            let started = Instant::now();
            let method = req.method().clone();
            let path = req.uri().path().to_string();

            let resp = server
                .handle_request(req.method(), req.uri().path(), req.headers())
                .await;

            let bytes = header_str(resp.headers(), "content-length").unwrap_or("-");
            server.log(format_args!(
                "{method} {path} {} {bytes} {:.3}ms",
                resp.status().as_u16(),
                started.elapsed().as_secs_f64() * 1000.0,
            ));
            Ok::<_, Infallible>(resp)
        }
    });
    if let Err(error) = http1::Builder::new()
        .timer(TokioTimer::new())
        .header_read_timeout(header_timeout)
        .max_buf_size(MAX_HEADER_BYTES)
        .serve_connection(io, svc)
        .await
    {
        log_server.log(format_args!("connection error: {error}"));
    }
}

/// Default header-read timeout used by [`Server::run_ephemeral`].
const DEFAULT_HEADER_TIMEOUT: Duration = Duration::from_secs(30);

/// Default grace period `ServerHandle::shutdown()` waits for in-flight connections to
/// finish on their own before aborting whatever is left. A connection with no
/// self-imposed end — an open live-reload SSE stream, or any keep-alive connection whose
/// peer simply never closes it — would otherwise let `shutdown()` hang forever waiting
/// for it to finish naturally. Every wait in this crate has a stated upper bound;
/// shutdown is no exception.
const DEFAULT_SHUTDOWN_DRAIN_TIMEOUT: Duration = Duration::from_secs(5);

/// A handle to a server started by one of the `Server::run*` methods.
///
/// Dropping this handle without calling `shutdown()` leaves the server running in the
/// background for the life of the process. Call `shutdown()` to stop accepting new
/// connections and wait for already-accepted connections to finish before returning.
pub struct ServerHandle {
    shutdown_tx: Option<tokio::sync::oneshot::Sender<()>>,
    accept_task: tokio::task::JoinHandle<()>,
}

impl ServerHandle {
    /// Stop accepting new connections and wait up to `DEFAULT_SHUTDOWN_DRAIN_TIMEOUT`
    /// (5s) for in-flight connections to finish on their own. Equivalent to
    /// `shutdown_with_timeout(DEFAULT_SHUTDOWN_DRAIN_TIMEOUT)` — see that method for what
    /// happens to connections still open once the grace period elapses.
    pub async fn shutdown(self) {
        self.shutdown_with_timeout(DEFAULT_SHUTDOWN_DRAIN_TIMEOUT)
            .await;
    }

    /// Stop accepting new connections and wait up to `drain_timeout` for in-flight
    /// connections to finish on their own.
    ///
    /// Connections still open once `drain_timeout` elapses are aborted rather than
    /// waited on further: dropping the accept task drops its `JoinSet`, which aborts
    /// every task still tracked in it (see `tokio::task::JoinSet`'s own drop behavior) —
    /// which in turn drops each connection's socket, closing it. This is what bounds
    /// shutdown when a connection has no natural end of its own (the live-reload SSE
    /// stream is the motivating case: it stays open until a watched file changes, which
    /// may never happen before the process needs to exit).
    pub async fn shutdown_with_timeout(mut self, drain_timeout: Duration) {
        if let Some(tx) = self.shutdown_tx.take() {
            let _ = tx.send(());
        }
        if timeout(drain_timeout, &mut self.accept_task).await.is_err() {
            self.accept_task.abort();
        }
    }
}

/// Read a request header as a `&str`, or `None` if it's absent or not valid ASCII.
fn header_str<'h>(headers: &'h HeaderMap, name: &str) -> Option<&'h str> {
    headers.get(name).and_then(|value| value.to_str().ok())
}

/// Start a response carrying the baseline security header every response in this crate
/// sends — 304s included. A 304 otherwise repeats only the caching validators, which is
/// why it once built its own builder and was the single response able to arrive without
/// `nosniff`; a client that caches the header set alongside the representation would
/// then hold a copy missing it.
///
/// Prefer [`Server::response`], which also applies the embedder's configured headers.
/// This bare form exists for `bad_request_response`, which is reachable from `finish`
/// where no `Server` is in scope.
fn response(status: StatusCode) -> Builder {
    Response::builder()
        .status(status)
        .header("X-Content-Type-Options", "nosniff")
}

/// Finish `builder` with a plain-text body. `&'static str` bodies borrow rather than
/// allocate; owned bodies are moved in.
fn text(builder: Builder, body: impl Into<Bytes>) -> Response<ResponseBody> {
    finish(builder.body(ResponseBody::Buffered(Full::new(body.into()))))
}

/// Finishes building a response, degrading to a generic 400 instead of panicking if any
/// header value turns out to be invalid for use as an HTTP header value.
///
/// Every header value that reaches `Response::builder()` in this module is either a
/// static string or formatted from internal, already-validated data (a byte count, an
/// mtime, a fixed method list) — none of it can actually fail today. But `.unwrap()`
/// on that assumption is exactly the kind of thing that turns "can't happen" into a
/// production panic the day someone adds a header built from new input without
/// re-deriving that guarantee. Routing every response through this one fallible path
/// means that mistake fails safe instead of panicking.
fn finish(built: Result<Response<ResponseBody>, hyper::http::Error>) -> Response<ResponseBody> {
    built.unwrap_or_else(|_| bad_request_response())
}

// `internal_error_response()` and `bad_request_response()` are the fallback responses
// `finish()` itself degrades to — every header and body here is a fixed string with no
// external input, so `.body(...)` cannot fail. They can't be routed through `finish()`
// without it degrading to itself on failure.
fn internal_error_response() -> Response<ResponseBody> {
    response(StatusCode::INTERNAL_SERVER_ERROR)
        .body(ResponseBody::Buffered(Full::new(Bytes::from_static(
            b"internal server error\n",
        ))))
        .unwrap()
}

fn bad_request_response() -> Response<ResponseBody> {
    response(StatusCode::BAD_REQUEST)
        .body(ResponseBody::Buffered(Full::new(Bytes::from_static(
            b"bad request\n",
        ))))
        .unwrap()
}

/// `Content-Encoding` name and sidecar file extension for each supported precompressed
/// variant, in preference order — brotli wins when a client accepts both and both
/// sidecars exist.
const SIDECAR_ENCODINGS: [(&str, &str); 2] = [("br", ".br"), ("gzip", ".gz")];

/// `q`-values are carried in thousandths — RFC 9110 allows at most three decimal places
/// — so weights compare exactly as integers instead of through float equality.
const QVALUE_SCALE: f32 = 1000.0;

/// An `Accept-Encoding` entry with no explicit `q` parameter has weight 1.
const DEFAULT_QVALUE: u16 = 1000;

/// The weight `accept_encoding` gives `encoding`, or `None` if it does not list it.
///
/// Entries are matched as whole tokens, case-insensitively, per RFC 9110 — not by
/// substring. The substring form this replaces got two things wrong that a client can
/// trigger: `Accept-Encoding: gzip;q=0` selected gzip, because the header *contains*
/// "gzip" while explicitly refusing it, and a token like `brotli` matched `br`.
///
/// `*` is deliberately not honored: treating the wildcard as matching nothing can only
/// cost a bandwidth optimization, while treating it as matching everything risks sending
/// an encoding the client did not ask for. The conservative reading is the safe one when
/// the payoff is choosing between two static files.
fn encoding_quality(accept_encoding: &str, encoding: &str) -> Option<u16> {
    accept_encoding.split(',').find_map(|entry| {
        let mut parts = entry.split(';');
        if !parts.next()?.trim().eq_ignore_ascii_case(encoding) {
            return None;
        }

        let quality = parts
            .find_map(|parameter| {
                let (key, value) = parameter.split_once('=')?;
                key.trim().eq_ignore_ascii_case("q").then(|| value.trim())
            })
            .and_then(|value| value.parse::<f32>().ok())
            .map(|value| (value.clamp(0.0, 1.0) * QVALUE_SCALE).round() as u16)
            .unwrap_or(DEFAULT_QVALUE);

        Some(quality)
    })
}

/// Look for a precompressed sidecar (`<path>.br` / `<path>.gz`) matching the client's
/// `Accept-Encoding`, and return its open file, metadata, and encoding name if found.
///
/// `path` must already be the fully resolved, canonicalized path `resolve()` produced.
/// The sidecar path is built by appending an extension to it — never by re-resolving a
/// modified request path — so this lookup can't become a second traversal surface: any
/// path this function reads is provably a sibling of a path `resolve()` already cleared.
fn select_precompressed_sidecar(
    path: &Path,
    accept_encoding: Option<&str>,
) -> Option<(std::fs::File, fs::Metadata, &'static str)> {
    // Highest `q` first; ties keep `SIDECAR_ENCODINGS` order (brotli over gzip) because
    // the sort is stable. Without this, `br;q=0.5, gzip` would serve brotli purely
    // because it is listed first here, ignoring the preference the client stated.
    let mut candidates: Vec<(&'static str, &'static str, u16)> = SIDECAR_ENCODINGS
        .iter()
        .filter_map(|(encoding, ext)| {
            let quality = accept_encoding.and_then(|header| encoding_quality(header, encoding))?;
            (quality > 0).then_some((*encoding, *ext, quality))
        })
        .collect();
    candidates.sort_by_key(|(_, _, quality)| std::cmp::Reverse(*quality));

    for (encoding, ext, _) in candidates {
        let mut sidecar = path.as_os_str().to_os_string();
        sidecar.push(ext);
        let sidecar_path = PathBuf::from(sidecar);

        // Tripwire for the traversal boundary: a sidecar path built by appending a suffix
        // must stay in the same directory as `path` (which `resolve()` already proved is
        // inside root). `ext` is always one of the two static literals in
        // `SIDECAR_ENCODINGS`, never derived from request input, so this can only fire if
        // a future change starts deriving `sidecar` some other way.
        debug_assert_eq!(
            sidecar_path.parent(),
            path.parent(),
            "sidecar path must stay in the same directory as the already-resolved path"
        );

        // Synchronous, and it stays an *open* rather than a cheaper `stat`: handing back
        // an already-open file is what keeps there being no gap between probing the
        // sidecar and serving it. Browsers send `Accept-Encoding` on every request, so
        // this probe is the common path — as `tokio::fs` opens, two misses per request
        // kept the blocking pool hot for files that do not exist.
        if let Ok(sidecar_std) = std::fs::File::open(&sidecar_path) {
            if let Ok(sidecar_metadata) = sidecar_std.metadata() {
                return Some((sidecar_std, sidecar_metadata, encoding));
            }
        }
    }
    None
}

/// Generate an ETag for a file based on modification time and size.
///
/// Format: `"<size>-<mtime_secs>.<mtime_nanos>"`.
///
/// The sub-second component is what makes this crate's choice to serve an ETag *instead*
/// of `Last-Modified`/`If-Modified-Since` sound. That choice rests on an ETag being able
/// to distinguish representations a whole-second timestamp cannot — two writes inside the
/// same second — which a whole-second ETag plainly cannot do either: rewriting a file
/// within a second of its last write, without changing its length, reproduced the
/// previous ETag exactly and every revalidating client was told `304 Not Modified` while
/// holding stale bytes. Build pipelines that rewrite generated assets are the realistic
/// way to hit that, and this crate ships one.
///
/// A filesystem whose timestamps are only second-granular gives `subsec_nanos() == 0`
/// and the same behavior as before — no worse, and no false confidence beyond what the
/// filesystem actually provides.
fn generate_etag(metadata: &fs::Metadata) -> String {
    let mtime = metadata
        .modified()
        .ok()
        .and_then(|t| t.duration_since(SystemTime::UNIX_EPOCH).ok())
        .unwrap_or_default();
    format!(
        "\"{}-{}.{}\"",
        metadata.len(),
        mtime.as_secs(),
        mtime.subsec_nanos()
    )
}

/// Determine MIME type from file path extension.
fn mime_type_for_path(path: &Path) -> &'static str {
    let ext = path
        .extension()
        .and_then(|ext| ext.to_str())
        .unwrap_or_default()
        .to_lowercase();

    match ext.as_str() {
        "html" | "htm" => "text/html; charset=utf-8",
        "css" => "text/css; charset=utf-8",
        "js" => "application/javascript; charset=utf-8",
        "json" => "application/json; charset=utf-8",
        "svg" => "image/svg+xml",
        "png" => "image/png",
        "jpg" | "jpeg" => "image/jpeg",
        "gif" => "image/gif",
        "webp" => "image/webp",
        "ico" => "image/x-icon",
        "woff" => "font/woff",
        "woff2" => "font/woff2",
        "ttf" => "font/ttf",
        "md" | "markdown" => "text/markdown; charset=utf-8",
        "txt" => "text/plain; charset=utf-8",
        "xml" => "application/xml",
        "pdf" => "application/pdf",
        "zip" => "application/zip",
        _ => "application/octet-stream",
    }
}

/// Check if the If-None-Match header matches the current ETag.
/// Handles both exact match and wildcard (*) comparison per RFC 9110.
fn is_etag_match(if_none_match: &str, etag: &str) -> bool {
    if if_none_match == "*" {
        return true;
    }
    if_none_match.split(',').any(|tag| tag.trim() == etag)
}

#[derive(Debug)]
enum RangeOutcome {
    NoRange,
    Satisfiable(u64, u64),
    Unsatisfiable,
    MultiRangeIgnored,
}

enum RangeCheck {
    IgnoreRange,
    Satisfiable(u64, u64),
    Unsatisfiable,
}

fn parse_range_header(header: &str, file_size: u64) -> RangeOutcome {
    let header = header.trim();
    if !header.starts_with("bytes=") {
        return RangeOutcome::NoRange;
    }

    let range_spec = &header[6..];

    if range_spec.contains(',') {
        return RangeOutcome::MultiRangeIgnored;
    }

    if let Some(suffix_pos) = range_spec.find('-') {
        if suffix_pos == 0 {
            let suffix_len_str = &range_spec[1..];
            if let Ok(suffix_len) = suffix_len_str.parse::<u64>() {
                if suffix_len == 0 {
                    return RangeOutcome::Unsatisfiable;
                }
                if suffix_len >= file_size {
                    return RangeOutcome::Satisfiable(0, file_size - 1);
                }
                return RangeOutcome::Satisfiable(file_size - suffix_len, file_size - 1);
            }
            return RangeOutcome::Unsatisfiable;
        }

        let start_str = &range_spec[..suffix_pos];
        let end_str = &range_spec[suffix_pos + 1..];

        if let Ok(start) = start_str.parse::<u64>() {
            if start >= file_size {
                return RangeOutcome::Unsatisfiable;
            }

            if end_str.is_empty() {
                return RangeOutcome::Satisfiable(start, file_size - 1);
            }

            if let Ok(end) = end_str.parse::<u64>() {
                if end < start {
                    return RangeOutcome::Unsatisfiable;
                }
                let clamped_end = (end + 1).min(file_size) - 1;
                if start > clamped_end {
                    return RangeOutcome::Unsatisfiable;
                }
                return RangeOutcome::Satisfiable(start, clamped_end);
            }
        }
    }

    RangeOutcome::Unsatisfiable
}

fn if_range_valid(if_range_header: &str, current_etag: &str) -> bool {
    if_range_header.trim() == current_etag
}

#[cfg(test)]
#[path = "../tests/unit/server/precompressed_sidecar.rs"]
mod precompressed_sidecar_tests;

#[cfg(test)]
#[path = "../tests/unit/server/file_body.rs"]
mod file_body_tests;

#[cfg(test)]
#[path = "../tests/unit/server/accept.rs"]
mod accept_tests;

#[cfg(test)]
#[path = "../tests/unit/server/finish.rs"]
mod finish_tests;

#[cfg(test)]
#[path = "../tests/unit/server/etag.rs"]
mod etag_tests;

#[cfg(test)]
#[path = "../tests/unit/server/range_header.rs"]
mod range_header_tests;