Merge pull request #2325 from ehuss/poll-watcher

Add a poll-based file watcher.

9 files changed, 638 insertions, 199 deletions

diff --git a/Cargo.toml b/Cargo.toml
index 2beec26a..db1b1825 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -41,6 +41,7 @@ notify = { version = "6.1.1", optional = true }
 notify-debouncer-mini = { version = "0.4.1", optional = true }
 ignore = { version = "0.4.20", optional = true }
 pathdiff = { version = "0.2.1", optional = true }
+walkdir = { version = "2.3.3", optional = true }
 
 # Serve feature
 futures-util = { version = "0.3.28", optional = true }
@@ -61,7 +62,7 @@ walkdir = "2.3.3"
 
 [features]
 default = ["watch", "serve", "search"]
-watch = ["dep:notify", "dep:notify-debouncer-mini", "dep:ignore", "dep:pathdiff"]
+watch = ["dep:notify", "dep:notify-debouncer-mini", "dep:ignore", "dep:pathdiff", "dep:walkdir"]
 serve = ["dep:futures-util", "dep:tokio", "dep:warp"]
 search = ["dep:elasticlunr-rs", "dep:ammonia"]
 
diff --git a/guide/src/cli/arg-watcher.md b/guide/src/cli/arg-watcher.md
new file mode 100644
index 00000000..bddf2dd2
--- /dev/null
+++ b/guide/src/cli/arg-watcher.md
@@ -0,0 +1,7 @@
+#### `--watcher`
+
+There are different backends used to determine when a file has changed.
+
+* `poll` (default) --- Checks for file modifications by scanning the filesystem every second.
+* `native` --- Uses the native operating system facilities to receive notifications when files change.
+  This can have less constant overhead, but may not be as reliable as the `poll` based watcher. See these issues for more information: [#383](https://github.com/rust-lang/mdBook/issues/383) [#1441](https://github.com/rust-lang/mdBook/issues/1441) [#1707](https://github.com/rust-lang/mdBook/issues/1707) [#2035](https://github.com/rust-lang/mdBook/issues/2035) [#2102](https://github.com/rust-lang/mdBook/issues/2102)
diff --git a/guide/src/cli/serve.md b/guide/src/cli/serve.md
index 5a92d8cb..4603df8e 100644
--- a/guide/src/cli/serve.md
+++ b/guide/src/cli/serve.md
@@ -44,6 +44,8 @@ book. Relative paths are interpreted relative to the book's root directory. If
 not specified it will default to the value of the `build.build-dir` key in
 `book.toml`, or to `./book`.
 
+{{#include arg-watcher.md}}
+
 #### Specify exclude patterns
 
 The `serve` command will not automatically trigger a build for files listed in
diff --git a/guide/src/cli/watch.md b/guide/src/cli/watch.md
index 5e868a6c..be2f5be4 100644
--- a/guide/src/cli/watch.md
+++ b/guide/src/cli/watch.md
@@ -27,6 +27,7 @@ book. Relative paths are interpreted relative to the book's root directory. If
 not specified it will default to the value of the `build.build-dir` key in
 `book.toml`, or to `./book`.
 
+{{#include arg-watcher.md}}
 
 #### Specify exclude patterns
 
diff --git a/src/cmd/command_prelude.rs b/src/cmd/command_prelude.rs
index b6362e60..ea2d3095 100644
--- a/src/cmd/command_prelude.rs
+++ b/src/cmd/command_prelude.rs
@@ -36,6 +36,19 @@ pub trait CommandExt: Sized {
     fn arg_open(self) -> Self {
         self._arg(arg!(-o --open "Opens the compiled book in a web browser"))
     }
+
+    fn arg_watcher(self) -> Self {
+        #[cfg(feature = "watch")]
+        return self._arg(
+            Arg::new("watcher")
+                .long("watcher")
+                .value_parser(["poll", "native"])
+                .default_value("poll")
+                .help("The filesystem watching technique"),
+        );
+        #[cfg(not(feature = "watch"))]
+        return self;
+    }
 }
 
 impl CommandExt for Command {
diff --git a/src/cmd/serve.rs b/src/cmd/serve.rs
index eeb19cb3..5e637a02 100644
--- a/src/cmd/serve.rs
+++ b/src/cmd/serve.rs
@@ -6,7 +6,6 @@ use clap::builder::NonEmptyStringValueParser;
 use futures_util::sink::SinkExt;
 use futures_util::StreamExt;
 use mdbook::errors::*;
-use mdbook::utils;
 use mdbook::utils::fs::get_404_output_file;
 use mdbook::MDBook;
 use std::net::{SocketAddr, ToSocketAddrs};
@@ -43,12 +42,13 @@ pub fn make_subcommand() -> Command {
                 .help("Port to use for HTTP connections"),
         )
         .arg_open()
+        .arg_watcher()
 }
 
 // Serve command implementation
 pub fn execute(args: &ArgMatches) -> Result<()> {
     let book_dir = get_book_dir(args);
-    let mut book = MDBook::load(book_dir)?;
+    let mut book = MDBook::load(&book_dir)?;
 
     let port = args.get_one::<String>("port").unwrap();
     let hostname = args.get_one::<String>("hostname").unwrap();
@@ -97,23 +97,12 @@ pub fn execute(args: &ArgMatches) -> Result<()> {
     }
 
     #[cfg(feature = "watch")]
-    watch::trigger_on_change(&book, move |paths, book_dir| {
-        info!("Files changed: {:?}", paths);
-        info!("Building book...");
-
-        // FIXME: This area is really ugly because we need to re-set livereload :(
-        let result = MDBook::load(book_dir).and_then(|mut b| {
-            update_config(&mut b);
-            b.build()
-        });
-
-        if let Err(e) = result {
-            error!("Unable to load the book");
-            utils::log_backtrace(&e);
-        } else {
+    {
+        let watcher = watch::WatcherKind::from_str(args.get_one::<String>("watcher").unwrap());
+        watch::rebuild_on_change(watcher, &book_dir, &update_config, &move || {
             let _ = tx.send(Message::text("reload"));
-        }
-    });
+        });
+    }
 
     let _ = thread_handle.join();
 
diff --git a/src/cmd/watch.rs b/src/cmd/watch.rs
index ef3371d8..7adb2bbb 100644
--- a/src/cmd/watch.rs
+++ b/src/cmd/watch.rs
@@ -1,14 +1,11 @@
 use super::command_prelude::*;
 use crate::{get_book_dir, open};
-use ignore::gitignore::Gitignore;
 use mdbook::errors::Result;
-use mdbook::utils;
 use mdbook::MDBook;
-use pathdiff::diff_paths;
 use std::path::{Path, PathBuf};
-use std::sync::mpsc::channel;
-use std::thread::sleep;
-use std::time::Duration;
+
+mod native;
+mod poller;
 
 // Create clap subcommand arguments
 pub fn make_subcommand() -> Command {
@@ -17,12 +14,28 @@ pub fn make_subcommand() -> Command {
         .arg_dest_dir()
         .arg_root_dir()
         .arg_open()
+        .arg_watcher()
+}
+
+pub enum WatcherKind {
+    Poll,
+    Native,
+}
+
+impl WatcherKind {
+    pub fn from_str(s: &str) -> WatcherKind {
+        match s {
+            "poll" => WatcherKind::Poll,
+            "native" => WatcherKind::Native,
+            _ => panic!("unsupported watcher {s}"),
+        }
+    }
 }
 
 // Watch command implementation
 pub fn execute(args: &ArgMatches) -> Result<()> {
     let book_dir = get_book_dir(args);
-    let mut book = MDBook::load(book_dir)?;
+    let mut book = MDBook::load(&book_dir)?;
 
     let update_config = |book: &mut MDBook| {
         if let Some(dest_dir) = args.get_one::<PathBuf>("dest-dir") {
@@ -41,42 +54,21 @@ pub fn execute(args: &ArgMatches) -> Result<()> {
         open(path);
     }
 
-    trigger_on_change(&book, |paths, book_dir| {
-        info!("Files changed: {:?}\nBuilding book...\n", paths);
-        let result = MDBook::load(book_dir).and_then(|mut b| {
-            update_config(&mut b);
-            b.build()
-        });
-
-        if let Err(e) = result {
-            error!("Unable to build the book");
-            utils::log_backtrace(&e);
-        }
-    });
+    let watcher = WatcherKind::from_str(args.get_one::<String>("watcher").unwrap());
+    rebuild_on_change(watcher, &book_dir, &update_config, &|| {});
 
     Ok(())
 }
 
-fn remove_ignored_files(book_root: &Path, paths: &[PathBuf]) -> Vec<PathBuf> {
-    if paths.is_empty() {
-        return vec![];
-    }
-
-    match find_gitignore(book_root) {
-        Some(gitignore_path) => {
-            let (ignore, err) = Gitignore::new(&gitignore_path);
-            if let Some(err) = err {
-                warn!(
-                    "error reading gitignore `{}`: {err}",
-                    gitignore_path.display()
-                );
-            }
-            filter_ignored_files(ignore, paths)
-        }
-        None => {
-            // There is no .gitignore file.
-            paths.iter().map(|path| path.to_path_buf()).collect()
-        }
+pub fn rebuild_on_change(
+    kind: WatcherKind,
+    book_dir: &Path,
+    update_config: &dyn Fn(&mut MDBook),
+    post_build: &dyn Fn(),
+) {
+    match kind {
+        WatcherKind::Poll => self::poller::rebuild_on_change(book_dir, update_config, post_build),
+        WatcherKind::Native => self::native::rebuild_on_change(book_dir, update_config, post_build),
     }
 }
 
@@ -86,144 +78,3 @@ fn find_gitignore(book_root: &Path) -> Option<PathBuf> {
         .map(|p| p.join(".gitignore"))
         .find(|p| p.exists())
 }
-
-// Note: The usage of `canonicalize` may encounter occasional failures on the Windows platform, presenting a potential risk.
-// For more details, refer to [Pull Request #2229](https://github.com/rust-lang/mdBook/pull/2229#discussion_r1408665981).
-fn filter_ignored_files(ignore: Gitignore, paths: &[PathBuf]) -> Vec<PathBuf> {
-    let ignore_root = ignore
-        .path()
-        .canonicalize()
-        .expect("ignore root canonicalize error");
-
-    paths
-        .iter()
-        .filter(|path| {
-            let relative_path =
-                diff_paths(path, &ignore_root).expect("One of the paths should be an absolute");
-            !ignore
-                .matched_path_or_any_parents(&relative_path, relative_path.is_dir())
-                .is_ignore()
-        })
-        .map(|path| path.to_path_buf())
-        .collect()
-}
-
-/// Calls the closure when a book source file is changed, blocking indefinitely.
-pub fn trigger_on_change<F>(book: &MDBook, closure: F)
-where
-    F: Fn(Vec<PathBuf>, &Path),
-{
-    use notify::RecursiveMode::*;
-
-    // Create a channel to receive the events.
-    let (tx, rx) = channel();
-
-    let mut debouncer = match notify_debouncer_mini::new_debouncer(Duration::from_secs(1), tx) {
-        Ok(d) => d,
-        Err(e) => {
-            error!("Error while trying to watch the files:\n\n\t{:?}", e);
-            std::process::exit(1)
-        }
-    };
-    let watcher = debouncer.watcher();
-
-    // Add the source directory to the watcher
-    if let Err(e) = watcher.watch(&book.source_dir(), Recursive) {
-        error!("Error while watching {:?}:\n    {:?}", book.source_dir(), e);
-        std::process::exit(1);
-    };
-
-    let _ = watcher.watch(&book.theme_dir(), Recursive);
-
-    // Add the book.toml file to the watcher if it exists
-    let _ = watcher.watch(&book.root.join("book.toml"), NonRecursive);
-
-    for dir in &book.config.build.extra_watch_dirs {
-        let path = book.root.join(dir);
-        let canonical_path = path.canonicalize().unwrap_or_else(|e| {
-            error!("Error while watching extra directory {path:?}:\n    {e}");
-            std::process::exit(1);
-        });
-
-        if let Err(e) = watcher.watch(&canonical_path, Recursive) {
-            error!(
-                "Error while watching extra directory {:?}:\n    {:?}",
-                canonical_path, e
-            );
-            std::process::exit(1);
-        }
-    }
-
-    info!("Listening for changes...");
-
-    loop {
-        let first_event = rx.recv().unwrap();
-        sleep(Duration::from_millis(50));
-        let other_events = rx.try_iter();
-
-        let all_events = std::iter::once(first_event).chain(other_events);
-
-        let paths: Vec<_> = all_events
-            .filter_map(|event| match event {
-                Ok(events) => Some(events),
-                Err(error) => {
-                    log::warn!("error while watching for changes: {error}");
-                    None
-                }
-            })
-            .flatten()
-            .map(|event| event.path)
-            .collect();
-
-        // If we are watching files outside the current repository (via extra-watch-dirs), then they are definitionally
-        // ignored by gitignore. So we handle this case by including such files into the watched paths list.
-        let any_external_paths = paths.iter().filter(|p| !p.starts_with(&book.root)).cloned();
-        let mut paths = remove_ignored_files(&book.root, &paths[..]);
-        paths.extend(any_external_paths);
-
-        if !paths.is_empty() {
-            closure(paths, &book.root);
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use ignore::gitignore::GitignoreBuilder;
-    use std::env;
-
-    #[test]
-    fn test_filter_ignored_files() {
-        let current_dir = env::current_dir().unwrap();
-
-        let ignore = GitignoreBuilder::new(&current_dir)
-            .add_line(None, "*.html")
-            .unwrap()
-            .build()
-            .unwrap();
-        let should_remain = current_dir.join("record.text");
-        let should_filter = current_dir.join("index.html");
-
-        let remain = filter_ignored_files(ignore, &[should_remain.clone(), should_filter]);
-        assert_eq!(remain, vec![should_remain])
-    }
-
-    #[test]
-    fn filter_ignored_files_should_handle_parent_dir() {
-        let current_dir = env::current_dir().unwrap();
-
-        let ignore = GitignoreBuilder::new(&current_dir)
-            .add_line(None, "*.html")
-            .unwrap()
-            .build()
-            .unwrap();
-
-        let parent_dir = current_dir.join("..");
-        let should_remain = parent_dir.join("record.text");
-        let should_filter = parent_dir.join("index.html");
-
-        let remain = filter_ignored_files(ignore, &[should_remain.clone(), should_filter]);
-        assert_eq!(remain, vec![should_remain])
-    }
-}
diff --git a/src/cmd/watch/native.rs b/src/cmd/watch/native.rs
new file mode 100644
index 00000000..fad8d7ce
--- /dev/null
+++ b/src/cmd/watch/native.rs
@@ -0,0 +1,189 @@
+//! A filesystem watcher using native operating system facilities.
+
+use ignore::gitignore::Gitignore;
+use mdbook::MDBook;
+use std::path::{Path, PathBuf};
+use std::sync::mpsc::channel;
+use std::thread::sleep;
+use std::time::Duration;
+
+pub fn rebuild_on_change(
+    book_dir: &Path,
+    update_config: &dyn Fn(&mut MDBook),
+    post_build: &dyn Fn(),
+) {
+    use notify::RecursiveMode::*;
+
+    let mut book = MDBook::load(book_dir).unwrap_or_else(|e| {
+        error!("failed to load book: {e}");
+        std::process::exit(1);
+    });
+
+    // Create a channel to receive the events.
+    let (tx, rx) = channel();
+
+    let mut debouncer = match notify_debouncer_mini::new_debouncer(Duration::from_secs(1), tx) {
+        Ok(d) => d,
+        Err(e) => {
+            error!("Error while trying to watch the files:\n\n\t{:?}", e);
+            std::process::exit(1)
+        }
+    };
+    let watcher = debouncer.watcher();
+
+    // Add the source directory to the watcher
+    if let Err(e) = watcher.watch(&book.source_dir(), Recursive) {
+        error!("Error while watching {:?}:\n    {:?}", book.source_dir(), e);
+        std::process::exit(1);
+    };
+
+    let _ = watcher.watch(&book.theme_dir(), Recursive);
+
+    // Add the book.toml file to the watcher if it exists
+    let _ = watcher.watch(&book.root.join("book.toml"), NonRecursive);
+
+    for dir in &book.config.build.extra_watch_dirs {
+        let path = book.root.join(dir);
+        let canonical_path = path.canonicalize().unwrap_or_else(|e| {
+            error!("Error while watching extra directory {path:?}:\n    {e}");
+            std::process::exit(1);
+        });
+
+        if let Err(e) = watcher.watch(&canonical_path, Recursive) {
+            error!(
+                "Error while watching extra directory {:?}:\n    {:?}",
+                canonical_path, e
+            );
+            std::process::exit(1);
+        }
+    }
+
+    info!("Listening for changes...");
+
+    loop {
+        let first_event = rx.recv().unwrap();
+        sleep(Duration::from_millis(50));
+        let other_events = rx.try_iter();
+
+        let all_events = std::iter::once(first_event).chain(other_events);
+
+        let paths: Vec<_> = all_events
+            .filter_map(|event| match event {
+                Ok(events) => Some(events),
+                Err(error) => {
+                    log::warn!("error while watching for changes: {error}");
+                    None
+                }
+            })
+            .flatten()
+            .map(|event| event.path)
+            .collect();
+
+        // If we are watching files outside the current repository (via extra-watch-dirs), then they are definitionally
+        // ignored by gitignore. So we handle this case by including such files into the watched paths list.
+        let any_external_paths = paths.iter().filter(|p| !p.starts_with(&book.root)).cloned();
+        let mut paths = remove_ignored_files(&book.root, &paths[..]);
+        paths.extend(any_external_paths);
+
+        if !paths.is_empty() {
+            info!("Files changed: {paths:?}");
+            match MDBook::load(book_dir) {
+                Ok(mut b) => {
+                    update_config(&mut b);
+                    if let Err(e) = b.build() {
+                        error!("failed to build the book: {e:?}");
+                    } else {
+                        post_build();
+                    }
+                    book = b;
+                }
+                Err(e) => error!("failed to load book config: {e:?}"),
+            }
+        }
+    }
+}
+
+fn remove_ignored_files(book_root: &Path, paths: &[PathBuf]) -> Vec<PathBuf> {
+    if paths.is_empty() {
+        return vec![];
+    }
+
+    match super::find_gitignore(book_root) {
+        Some(gitignore_path) => {
+            let (ignore, err) = Gitignore::new(&gitignore_path);
+            if let Some(err) = err {
+                warn!(
+                    "error reading gitignore `{}`: {err}",
+                    gitignore_path.display()
+                );
+            }
+            filter_ignored_files(ignore, paths)
+        }
+        None => {
+            // There is no .gitignore file.
+            paths.iter().map(|path| path.to_path_buf()).collect()
+        }
+    }
+}
+
+// Note: The usage of `canonicalize` may encounter occasional failures on the Windows platform, presenting a potential risk.
+// For more details, refer to [Pull Request #2229](https://github.com/rust-lang/mdBook/pull/2229#discussion_r1408665981).
+fn filter_ignored_files(ignore: Gitignore, paths: &[PathBuf]) -> Vec<PathBuf> {
+    let ignore_root = ignore
+        .path()
+        .canonicalize()
+        .expect("ignore root canonicalize error");
+
+    paths
+        .iter()
+        .filter(|path| {
+            let relative_path = pathdiff::diff_paths(&path, &ignore_root)
+                .expect("One of the paths should be an absolute");
+            !ignore
+                .matched_path_or_any_parents(&relative_path, relative_path.is_dir())
+                .is_ignore()
+        })
+        .map(|path| path.to_path_buf())
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use ignore::gitignore::GitignoreBuilder;
+    use std::env;
+
+    #[test]
+    fn test_filter_ignored_files() {
+        let current_dir = env::current_dir().unwrap();
+
+        let ignore = GitignoreBuilder::new(&current_dir)
+            .add_line(None, "*.html")
+            .unwrap()
+            .build()
+            .unwrap();
+        let should_remain = current_dir.join("record.text");
+        let should_filter = current_dir.join("index.html");
+
+        let remain = filter_ignored_files(ignore, &[should_remain.clone(), should_filter]);
+        assert_eq!(remain, vec![should_remain])
+    }
+
+    #[test]
+    fn filter_ignored_files_should_handle_parent_dir() {
+        let current_dir = env::current_dir().unwrap();
+
+        let ignore = GitignoreBuilder::new(&current_dir)
+            .add_line(None, "*.html")
+            .unwrap()
+            .build()
+            .unwrap();
+
+        let parent_dir = current_dir.join("..");
+        let should_remain = parent_dir.join("record.text");
+        let should_filter = parent_dir.join("index.html");
+
+        let remain = filter_ignored_files(ignore, &[should_remain.clone(), should_filter]);
+        assert_eq!(remain, vec![should_remain])
+    }
+}
diff --git a/src/cmd/watch/poller.rs b/src/cmd/watch/poller.rs
new file mode 100644
index 00000000..5e1d1497
--- /dev/null
+++ b/src/cmd/watch/poller.rs
@@ -0,0 +1,386 @@
+//! A simple poll-based filesystem watcher.
+//!
+//! This exists because the native change notifications have historically had
+//! lots of problems. Various operating systems and different filesystems have
+//! had problems correctly reporting changes.
+
+use ignore::gitignore::Gitignore;
+use mdbook::MDBook;
+use pathdiff::diff_paths;
+use std::collections::HashMap;
+use std::fs::FileType;
+use std::path::{Path, PathBuf};
+use std::time::{Duration, Instant, SystemTime};
+use walkdir::WalkDir;
+
+/// Calls the closure when a book source file is changed, blocking indefinitely.
+pub fn rebuild_on_change(
+    book_dir: &Path,
+    update_config: &dyn Fn(&mut MDBook),
+    post_build: &dyn Fn(),
+) {
+    let mut book = MDBook::load(book_dir).unwrap_or_else(|e| {
+        error!("failed to load book: {e}");
+        std::process::exit(1);
+    });
+
+    let mut watcher = Watcher::new(book_dir);
+
+    info!("Watching for changes...");
+    // Scan once to initialize the starting point.
+    watcher.set_roots(&book);
+    watcher.scan();
+
+    // Track average scan time, to help investigate if the poller is taking
+    // undesirably long. This is not a rigorous benchmark, just a rough
+    // estimate.
+    const AVG_SIZE: usize = 60;
+    let mut avgs = vec![0.0; AVG_SIZE];
+    let mut avg_i = 0;
+
+    loop {
+        std::thread::sleep(Duration::new(1, 0));
+        watcher.set_roots(&book);
+        let start = Instant::now();
+        let paths = watcher.scan();
+        let elapsed = start.elapsed().as_secs_f64();
+        avgs[avg_i] = elapsed;
+        avg_i += 1;
+        if avg_i >= AVG_SIZE {
+            avg_i = 0;
+            let avg = avgs.iter().sum::<f64>() / (avgs.len() as f64);
+            trace!(
+                "scan average time: {avg:.2}s, scan size is {}",
+                watcher.path_data.len()
+            );
+        }
+
+        if !paths.is_empty() {
+            info!("Files changed: {paths:?}");
+            match MDBook::load(book_dir) {
+                Ok(mut b) => {
+                    update_config(&mut b);
+                    if let Err(e) = b.build() {
+                        error!("failed to build the book: {e:?}");
+                    } else {
+                        post_build();
+                    }
+                    book = b;
+                }
+                Err(e) => error!("failed to load book config: {e:?}"),
+            }
+        }
+    }
+}
+
+#[derive(PartialEq)]
+struct PathData {
+    file_type: FileType,
+    mtime: SystemTime,
+    size: u64,
+}
+
+/// A very simple poll-watcher that scans for modified files.
+#[derive(Default)]
+struct Watcher {
+    /// The root paths where it will recursively scan for changes.
+    root_paths: Vec<PathBuf>,
+    /// Data about files on disk.
+    path_data: HashMap<PathBuf, PathData>,
+    /// Filters paths that will be watched.
+    ignore: Option<(PathBuf, Gitignore)>,
+}
+
+impl Watcher {
+    fn new(book_root: &Path) -> Watcher {
+        // FIXME: ignore should be reloaded when it changes.
+        let ignore = super::find_gitignore(book_root).map(|gitignore_path| {
+            let (ignore, err) = Gitignore::new(&gitignore_path);
+            if let Some(err) = err {
+                warn!(
+                    "error reading gitignore `{}`: {err}",
+                    gitignore_path.display()
+                );
+            }
+            // Note: The usage of `canonicalize` may encounter occasional
+            // failures on the Windows platform, presenting a potential risk.
+            // For more details, refer to [Pull Request
+            // #2229](https://github.com/rust-lang/mdBook/pull/2229#discussion_r1408665981).
+            let ignore_path = ignore
+                .path()
+                .canonicalize()
+                .expect("ignore root canonicalize error");
+            (ignore_path, ignore)
+        });
+
+        Watcher {
+            ignore,
+            ..Default::default()
+        }
+    }
+
+    /// Sets the root directories where scanning will start.
+    fn set_roots(&mut self, book: &MDBook) {
+        let mut root_paths = vec![
+            book.source_dir(),
+            book.theme_dir(),
+            book.root.join("book.toml"),
+        ];
+        root_paths.extend(
+            book.config
+                .build
+                .extra_watch_dirs
+                .iter()
+                .map(|path| book.root.join(path)),
+        );
+        if let Some(html_config) = book.config.html_config() {
+            root_paths.extend(
+                html_config
+                    .additional_css
+                    .iter()
+                    .chain(html_config.additional_js.iter())
+                    .map(|path| book.root.join(path)),
+            );
+        }
+
+        self.root_paths = root_paths;
+    }
+
+    /// Scans for changes.
+    ///
+    /// Returns the paths that have changed.
+    fn scan(&mut self) -> Vec<PathBuf> {
+        let ignore = &self.ignore;
+        let new_path_data: HashMap<_, _> = self
+            .root_paths
+            .iter()
+            .filter(|root| root.exists())
+            .flat_map(|root| {
+                WalkDir::new(root)
+                    .follow_links(true)
+                    .into_iter()
+                    .filter_entry(|entry| {
+                        if let Some((ignore_path, ignore)) = ignore {
+                            let path = entry.path();
+                            // Canonicalization helps with removing `..` and
+                            // `.` entries, which can cause issues with
+                            // diff_paths.
+                            let path = path.canonicalize().unwrap_or_else(|_| path.to_path_buf());
+                            let relative_path = diff_paths(&path, &ignore_path)
+                                .expect("One of the paths should be an absolute");
+                            if ignore
+                                .matched_path_or_any_parents(&relative_path, relative_path.is_dir())
+                                .is_ignore()
+                            {
+                                trace!("ignoring {path:?}");
+                                return false;
+                            }
+                        }
+                        true
+                    })
+                    .filter_map(move |entry| {
+                        let entry = match entry {
+                            Ok(e) => e,
+                            Err(e) => {
+                                debug!("failed to scan {root:?}: {e}");
+                                return None;
+                            }
+                        };
+                        if entry.file_type().is_dir() {
+                            // Changes to directories themselves aren't
+                            // particularly interesting.
+                            return None;
+                        }
+                        let path = entry.path().to_path_buf();
+
+                        let meta = match entry.metadata() {
+                            Ok(meta) => meta,
+                            Err(e) => {
+                                debug!("failed to scan {path:?}: {e}");
+                                return None;
+                            }
+                        };
+                        let mtime = meta.modified().unwrap_or(SystemTime::UNIX_EPOCH);
+                        let pd = PathData {
+                            file_type: meta.file_type(),
+                            mtime,
+                            size: meta.len(),
+                        };
+                        Some((path, pd))
+                    })
+            })
+            .collect();
+        let mut paths = Vec::new();
+        for (new_path, new_data) in &new_path_data {
+            match self.path_data.get(new_path) {
+                Some(old_data) => {
+                    if new_data != old_data {
+                        paths.push(new_path.to_path_buf());
+                    }
+                }
+                None => {
+                    paths.push(new_path.clone());
+                }
+            }
+        }
+        for old_path in self.path_data.keys() {
+            if !new_path_data.contains_key(old_path) {
+                paths.push(old_path.to_path_buf());
+            }
+        }
+        self.path_data = new_path_data;
+        paths
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Helper for testing the watcher.
+    fn check_watch_behavior(
+        gitignore_path: &str,
+        gitignore: &str,
+        book_root_path: &str,
+        ignored: &[&str],
+        not_ignored: &[&str],
+        extra_setup: &dyn Fn(&Path),
+    ) {
+        // Create the book and initialize things.
+        let temp = tempfile::Builder::new()
+            .prefix("mdbook-")
+            .tempdir()
+            .unwrap();
+        let root = temp.path();
+        let book_root = root.join(book_root_path);
+        // eprintln!("book_root={book_root:?}",);
+        MDBook::init(&book_root).build().unwrap();
+        std::fs::write(root.join(gitignore_path), gitignore).unwrap();
+        let create = |paths: &[&str]| {
+            let mut paths = paths
+                .iter()
+                .map(|path| root.join(path))
+                .inspect(|path| {
+                    std::fs::create_dir_all(path.parent().unwrap()).unwrap();
+                    std::fs::write(path, "initial content").unwrap();
+                })
+                .map(|path| path.canonicalize().unwrap())
+                .collect::<Vec<_>>();
+            paths.sort();
+            paths
+        };
+        let ignored = create(ignored);
+        let not_ignored = create(not_ignored);
+        extra_setup(&book_root);
+        // Create a watcher and check its behavior.
+        let book = MDBook::load(&book_root).unwrap();
+        let mut watcher = Watcher::new(&book_root);
+        watcher.set_roots(&book);
+        // Do an initial scan to initialize its state.
+        watcher.scan();
+        // Verify the steady state is empty.
+        let changed = watcher.scan();
+        assert_eq!(changed, Vec::<PathBuf>::new());
+        // Modify all files, and verify that only not_ignored are detected.
+        for path in ignored.iter().chain(not_ignored.iter()) {
+            std::fs::write(path, "modified").unwrap();
+        }
+        let changed = watcher.scan();
+        let mut changed = changed
+            .into_iter()
+            .map(|p| p.canonicalize().unwrap())
+            .collect::<Vec<_>>();
+        changed.sort();
+        assert_eq!(changed, not_ignored);
+        // Verify again that steady state is empty.
+        let changed = watcher.scan();
+        assert_eq!(changed, Vec::<PathBuf>::new());
+    }
+
+    #[test]
+    fn test_ignore() {
+        // Basic gitignore test.
+        check_watch_behavior(
+            "foo/.gitignore",
+            "*.tmp",
+            "foo",
+            &["foo/src/somefile.tmp"],
+            &["foo/src/chapter.md"],
+            &|_book_root| {},
+        );
+    }
+
+    #[test]
+    fn test_ignore_in_parent() {
+        // gitignore is in the parent of the book
+        check_watch_behavior(
+            ".gitignore",
+            "*.tmp\nsomedir/\n/inroot\n/foo/src/inbook\n",
+            "foo",
+            &[
+                "foo/src/somefile.tmp",
+                "foo/src/somedir/somefile",
+                "inroot/somefile",
+                "foo/src/inbook/somefile",
+            ],
+            &["foo/src/inroot/somefile"],
+            &|_book_root| {},
+        );
+    }
+
+    #[test]
+    fn test_ignore_canonical() {
+        // test with path with ..
+        check_watch_behavior(
+            ".gitignore",
+            "*.tmp\nsomedir/\n/foo/src/inbook\n",
+            "bar/../foo",
+            &[
+                "foo/src/somefile.tmp",
+                "foo/src/somedir/somefile",
+                "foo/src/inbook/somefile",
+            ],
+            &["foo/src/chapter.md"],
+            &|_book_root| {},
+        );
+    }
+
+    #[test]
+    fn test_scan_extra_watch() {
+        // Check behavior with extra-watch-dirs
+        check_watch_behavior(
+            ".gitignore",
+            "*.tmp\n/outside-root/ignoreme\n/foo/examples/ignoreme\n",
+            "foo",
+            &[
+                "foo/src/somefile.tmp",
+                "foo/examples/example.tmp",
+                "outside-root/somefile.tmp",
+                "outside-root/ignoreme",
+                "foo/examples/ignoreme",
+            ],
+            &[
+                "foo/src/chapter.md",
+                "foo/examples/example.rs",
+                "foo/examples/example2.rs",
+