dirstate-v2: Add heuristic for when to create a new data file

… instead of appending to the existing one. This is based on keeping track
of how much of the existing data is not used anymore.

Differential Revision: https://phab.mercurial-scm.org/D11097

rust/hg-core/src/dirstate_tree/dirstate_map.rs

@@ -1,1232 +1,1270 b''
 use bytes_cast::BytesCast;
 use micro_timer::timed;
 use std::borrow::Cow;
 use std::convert::TryInto;
 use std::path::PathBuf;
 
 use super::on_disk;
 use super::on_disk::DirstateV2ParseError;
 use super::path_with_basename::WithBasename;
 use crate::dirstate::parsers::pack_entry;
 use crate::dirstate::parsers::packed_entry_size;
 use crate::dirstate::parsers::parse_dirstate_entries;
 use crate::dirstate::parsers::Timestamp;
 use crate::dirstate::MTIME_UNSET;
 use crate::dirstate::SIZE_FROM_OTHER_PARENT;
 use crate::dirstate::SIZE_NON_NORMAL;
 use crate::dirstate::V1_RANGEMASK;
 use crate::matchers::Matcher;
 use crate::utils::hg_path::{HgPath, HgPathBuf};
 use crate::CopyMapIter;
 use crate::DirstateEntry;
 use crate::DirstateError;
 use crate::DirstateParents;
 use crate::DirstateStatus;
 use crate::EntryState;
 use crate::FastHashMap;
 use crate::PatternFileWarning;
 use crate::StateMapIter;
 use crate::StatusError;
 use crate::StatusOptions;
 
+/// Append to an existing data file if the amount of unreachable data (not used
+/// anymore) is less than this fraction of the total amount of existing data.
+const ACCEPTABLE_UNREACHABLE_BYTES_RATIO: f32 = 0.5;
+
 pub struct DirstateMap<'on_disk> {
     /// Contents of the `.hg/dirstate` file
     pub(super) on_disk: &'on_disk [u8],
 
     pub(super) root: ChildNodes<'on_disk>,
 
     /// Number of nodes anywhere in the tree that have `.entry.is_some()`.
     pub(super) nodes_with_entry_count: u32,
 
     /// Number of nodes anywhere in the tree that have
     /// `.copy_source.is_some()`.
     pub(super) nodes_with_copy_source_count: u32,
 
     /// See on_disk::Header
     pub(super) ignore_patterns_hash: on_disk::IgnorePatternsHash,
+
+    /// How many bytes of `on_disk` are not used anymore
+    pub(super) unreachable_bytes: u32,
 }
 
 /// Using a plain `HgPathBuf` of the full path from the repository root as a
 /// map key would also work: all paths in a given map have the same parent
 /// path, so comparing full paths gives the same result as comparing base
 /// names. However `HashMap` would waste time always re-hashing the same
 /// string prefix.
 pub(super) type NodeKey<'on_disk> = WithBasename<Cow<'on_disk, HgPath>>;
 
 /// Similar to `&'tree Cow<'on_disk, HgPath>`, but can also be returned
 /// for on-disk nodes that don’t actually have a `Cow` to borrow.
 pub(super) enum BorrowedPath<'tree, 'on_disk> {
     InMemory(&'tree HgPathBuf),
     OnDisk(&'on_disk HgPath),
 }
 
 pub(super) enum ChildNodes<'on_disk> {
     InMemory(FastHashMap<NodeKey<'on_disk>, Node<'on_disk>>),
     OnDisk(&'on_disk [on_disk::Node]),
 }
 
 pub(super) enum ChildNodesRef<'tree, 'on_disk> {
     InMemory(&'tree FastHashMap<NodeKey<'on_disk>, Node<'on_disk>>),
     OnDisk(&'on_disk [on_disk::Node]),
 }
 
 pub(super) enum NodeRef<'tree, 'on_disk> {
     InMemory(&'tree NodeKey<'on_disk>, &'tree Node<'on_disk>),
     OnDisk(&'on_disk on_disk::Node),
 }
 
 impl<'tree, 'on_disk> BorrowedPath<'tree, 'on_disk> {
     pub fn detach_from_tree(&self) -> Cow<'on_disk, HgPath> {
         match *self {
             BorrowedPath::InMemory(in_memory) => Cow::Owned(in_memory.clone()),
             BorrowedPath::OnDisk(on_disk) => Cow::Borrowed(on_disk),
         }
     }
 }
 
 impl<'tree, 'on_disk> std::ops::Deref for BorrowedPath<'tree, 'on_disk> {
     type Target = HgPath;
 
     fn deref(&self) -> &HgPath {
         match *self {
             BorrowedPath::InMemory(in_memory) => in_memory,
             BorrowedPath::OnDisk(on_disk) => on_disk,
         }
     }
 }
 
 impl Default for ChildNodes<'_> {
     fn default() -> Self {
         ChildNodes::InMemory(Default::default())
     }
 }
 
 impl<'on_disk> ChildNodes<'on_disk> {
     pub(super) fn as_ref<'tree>(
         &'tree self,
     ) -> ChildNodesRef<'tree, 'on_disk> {
         match self {
             ChildNodes::InMemory(nodes) => ChildNodesRef::InMemory(nodes),
             ChildNodes::OnDisk(nodes) => ChildNodesRef::OnDisk(nodes),
         }
     }
 
     pub(super) fn is_empty(&self) -> bool {
         match self {
             ChildNodes::InMemory(nodes) => nodes.is_empty(),
             ChildNodes::OnDisk(nodes) => nodes.is_empty(),
         }
     }
 
-    pub(super) fn make_mut(
+    fn make_mut(
         &mut self,
         on_disk: &'on_disk [u8],
+        unreachable_bytes: &mut u32,
     ) -> Result<
         &mut FastHashMap<NodeKey<'on_disk>, Node<'on_disk>>,
         DirstateV2ParseError,
     > {
         match self {
             ChildNodes::InMemory(nodes) => Ok(nodes),
             ChildNodes::OnDisk(nodes) => {
+                *unreachable_bytes +=
+                    std::mem::size_of_val::<[on_disk::Node]>(nodes) as u32;
                 let nodes = nodes
                     .iter()
                     .map(|node| {
                         Ok((
                             node.path(on_disk)?,
                             node.to_in_memory_node(on_disk)?,
                         ))
                     })
                     .collect::<Result<_, _>>()?;
                 *self = ChildNodes::InMemory(nodes);
                 match self {
                     ChildNodes::InMemory(nodes) => Ok(nodes),
                     ChildNodes::OnDisk(_) => unreachable!(),
                 }
             }
         }
     }
 }
 
 impl<'tree, 'on_disk> ChildNodesRef<'tree, 'on_disk> {
     pub(super) fn get(
         &self,
         base_name: &HgPath,
         on_disk: &'on_disk [u8],
     ) -> Result<Option<NodeRef<'tree, 'on_disk>>, DirstateV2ParseError> {
         match self {
             ChildNodesRef::InMemory(nodes) => Ok(nodes
                 .get_key_value(base_name)
                 .map(|(k, v)| NodeRef::InMemory(k, v))),
             ChildNodesRef::OnDisk(nodes) => {
                 let mut parse_result = Ok(());
                 let search_result = nodes.binary_search_by(|node| {
                     match node.base_name(on_disk) {
                         Ok(node_base_name) => node_base_name.cmp(base_name),
                         Err(e) => {
                             parse_result = Err(e);
                             // Dummy comparison result, `search_result` won’t
                             // be used since `parse_result` is an error
                             std::cmp::Ordering::Equal
                         }
                     }
                 });
                 parse_result.map(|()| {
                     search_result.ok().map(|i| NodeRef::OnDisk(&nodes[i]))
                 })
             }
         }
     }
 
     /// Iterate in undefined order
     pub(super) fn iter(
         &self,
     ) -> impl Iterator<Item = NodeRef<'tree, 'on_disk>> {
         match self {
             ChildNodesRef::InMemory(nodes) => itertools::Either::Left(
                 nodes.iter().map(|(k, v)| NodeRef::InMemory(k, v)),
             ),
             ChildNodesRef::OnDisk(nodes) => {
                 itertools::Either::Right(nodes.iter().map(NodeRef::OnDisk))
             }
         }
     }
 
     /// Iterate in parallel in undefined order
     pub(super) fn par_iter(
         &self,
     ) -> impl rayon::iter::ParallelIterator<Item = NodeRef<'tree, 'on_disk>>
     {
         use rayon::prelude::*;
         match self {
             ChildNodesRef::InMemory(nodes) => rayon::iter::Either::Left(
                 nodes.par_iter().map(|(k, v)| NodeRef::InMemory(k, v)),
             ),
             ChildNodesRef::OnDisk(nodes) => rayon::iter::Either::Right(
                 nodes.par_iter().map(NodeRef::OnDisk),
             ),
         }
     }
 
     pub(super) fn sorted(&self) -> Vec<NodeRef<'tree, 'on_disk>> {
         match self {
             ChildNodesRef::InMemory(nodes) => {
                 let mut vec: Vec<_> = nodes
                     .iter()
                     .map(|(k, v)| NodeRef::InMemory(k, v))
                     .collect();
                 fn sort_key<'a>(node: &'a NodeRef) -> &'a HgPath {
                     match node {
                         NodeRef::InMemory(path, _node) => path.base_name(),
                         NodeRef::OnDisk(_) => unreachable!(),
                     }
                 }
                 // `sort_unstable_by_key` doesn’t allow keys borrowing from the
                 // value: https://github.com/rust-lang/rust/issues/34162
                 vec.sort_unstable_by(|a, b| sort_key(a).cmp(sort_key(b)));
                 vec
             }
             ChildNodesRef::OnDisk(nodes) => {
                 // Nodes on disk are already sorted
                 nodes.iter().map(NodeRef::OnDisk).collect()
             }
         }
     }
 }
 
 impl<'tree, 'on_disk> NodeRef<'tree, 'on_disk> {
     pub(super) fn full_path(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<&'tree HgPath, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(path, _node) => Ok(path.full_path()),
             NodeRef::OnDisk(node) => node.full_path(on_disk),
         }
     }
 
     /// Returns a `BorrowedPath`, which can be turned into a `Cow<'on_disk,
     /// HgPath>` detached from `'tree`
     pub(super) fn full_path_borrowed(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<BorrowedPath<'tree, 'on_disk>, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(path, _node) => match path.full_path() {
                 Cow::Borrowed(on_disk) => Ok(BorrowedPath::OnDisk(on_disk)),
                 Cow::Owned(in_memory) => Ok(BorrowedPath::InMemory(in_memory)),
             },
             NodeRef::OnDisk(node) => {
                 Ok(BorrowedPath::OnDisk(node.full_path(on_disk)?))
             }
         }
     }
 
     pub(super) fn base_name(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<&'tree HgPath, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(path, _node) => Ok(path.base_name()),
             NodeRef::OnDisk(node) => node.base_name(on_disk),
         }
     }
 
     pub(super) fn children(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<ChildNodesRef<'tree, 'on_disk>, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(_path, node) => Ok(node.children.as_ref()),
             NodeRef::OnDisk(node) => {
                 Ok(ChildNodesRef::OnDisk(node.children(on_disk)?))
             }
         }
     }
 
     pub(super) fn has_copy_source(&self) -> bool {
         match self {
             NodeRef::InMemory(_path, node) => node.copy_source.is_some(),
             NodeRef::OnDisk(node) => node.has_copy_source(),
         }
     }
 
     pub(super) fn copy_source(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<Option<&'tree HgPath>, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(_path, node) => {
                 Ok(node.copy_source.as_ref().map(|s| &**s))
             }
             NodeRef::OnDisk(node) => node.copy_source(on_disk),
         }
     }
 
     pub(super) fn entry(
         &self,
     ) -> Result<Option<DirstateEntry>, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(_path, node) => {
                 Ok(node.data.as_entry().copied())
             }
             NodeRef::OnDisk(node) => node.entry(),
         }
     }
 
     pub(super) fn state(
         &self,
     ) -> Result<Option<EntryState>, DirstateV2ParseError> {
         match self {
             NodeRef::InMemory(_path, node) => {
                 Ok(node.data.as_entry().map(|entry| entry.state))
             }
             NodeRef::OnDisk(node) => node.state(),
         }
     }
 
     pub(super) fn cached_directory_mtime(
         &self,
     ) -> Option<&'tree on_disk::Timestamp> {
         match self {
             NodeRef::InMemory(_path, node) => match &node.data {
                 NodeData::CachedDirectory { mtime } => Some(mtime),
                 _ => None,
             },
             NodeRef::OnDisk(node) => node.cached_directory_mtime(),
         }
     }
 
     pub(super) fn descendants_with_entry_count(&self) -> u32 {
         match self {
             NodeRef::InMemory(_path, node) => {
                 node.descendants_with_entry_count
             }
             NodeRef::OnDisk(node) => node.descendants_with_entry_count.get(),
         }
     }
 
     pub(super) fn tracked_descendants_count(&self) -> u32 {
         match self {
             NodeRef::InMemory(_path, node) => node.tracked_descendants_count,
             NodeRef::OnDisk(node) => node.tracked_descendants_count.get(),
         }
     }
 }
 
 /// Represents a file or a directory
 #[derive(Default)]
 pub(super) struct Node<'on_disk> {
     pub(super) data: NodeData,
 
     pub(super) copy_source: Option<Cow<'on_disk, HgPath>>,
 
     pub(super) children: ChildNodes<'on_disk>,
 
     /// How many (non-inclusive) descendants of this node have an entry.
     pub(super) descendants_with_entry_count: u32,
 
     /// How many (non-inclusive) descendants of this node have an entry whose
     /// state is "tracked".
     pub(super) tracked_descendants_count: u32,
 }
 
 pub(super) enum NodeData {
     Entry(DirstateEntry),
     CachedDirectory { mtime: on_disk::Timestamp },
     None,
 }
 
 impl Default for NodeData {
     fn default() -> Self {
         NodeData::None
     }
 }
 
 impl NodeData {
     fn has_entry(&self) -> bool {
         match self {
             NodeData::Entry(_) => true,
             _ => false,
         }
     }
 
     fn as_entry(&self) -> Option<&DirstateEntry> {
         match self {
             NodeData::Entry(entry) => Some(entry),
             _ => None,
         }
     }
 }
 
 impl<'on_disk> DirstateMap<'on_disk> {
     pub(super) fn empty(on_disk: &'on_disk [u8]) -> Self {
         Self {
             on_disk,
             root: ChildNodes::default(),
             nodes_with_entry_count: 0,
             nodes_with_copy_source_count: 0,
             ignore_patterns_hash: [0; on_disk::IGNORE_PATTERNS_HASH_LEN],
+            unreachable_bytes: 0,
         }
     }
 
     #[timed]
     pub fn new_v2(
         on_disk: &'on_disk [u8],
         data_size: usize,
     ) -> Result<Self, DirstateError> {
         if let Some(data) = on_disk.get(..data_size) {
             Ok(on_disk::read(data)?)
         } else {
             Err(DirstateV2ParseError.into())
         }
     }
 
     #[timed]
     pub fn new_v1(
         on_disk: &'on_disk [u8],
     ) -> Result<(Self, Option<DirstateParents>), DirstateError> {
         let mut map = Self::empty(on_disk);
         if map.on_disk.is_empty() {
             return Ok((map, None));
         }
 
         let parents = parse_dirstate_entries(
             map.on_disk,
             |path, entry, copy_source| {
                 let tracked = entry.state.is_tracked();
                 let node = Self::get_or_insert_node(
                     map.on_disk,
+                    &mut map.unreachable_bytes,
                     &mut map.root,
                     path,
                     WithBasename::to_cow_borrowed,
                     |ancestor| {
                         if tracked {
                             ancestor.tracked_descendants_count += 1
                         }
                         ancestor.descendants_with_entry_count += 1
                     },
                 )?;
                 assert!(
                     !node.data.has_entry(),
                     "duplicate dirstate entry in read"
                 );
                 assert!(
                     node.copy_source.is_none(),
                     "duplicate dirstate entry in read"
                 );
                 node.data = NodeData::Entry(*entry);
                 node.copy_source = copy_source.map(Cow::Borrowed);
                 map.nodes_with_entry_count += 1;
                 if copy_source.is_some() {
                     map.nodes_with_copy_source_count += 1
                 }
                 Ok(())
             },
         )?;
         let parents = Some(parents.clone());
 
         Ok((map, parents))
     }
 
     /// Assuming dirstate-v2 format, returns whether the next write should
     /// append to the existing data file that contains `self.on_disk` (true),
     /// or create a new data file from scratch (false).
     pub(super) fn write_should_append(&self) -> bool {
-        // Soon this will be a heuristic based on the amount of unreachable
-        // data. For now it’s pseudo-random in order to make tests exercise
-        // both code paths.
-
-        fn bad_rng() -> u32 {
-            std::time::SystemTime::now()
-                .duration_since(std::time::UNIX_EPOCH)
-                .unwrap()
-                .subsec_millis()
-        }
-
-        bad_rng() % 2 == 0
+        let ratio = self.unreachable_bytes as f32 / self.on_disk.len() as f32;
+        ratio < ACCEPTABLE_UNREACHABLE_BYTES_RATIO
     }
 
     fn get_node<'tree>(
         &'tree self,
         path: &HgPath,
     ) -> Result<Option<NodeRef<'tree, 'on_disk>>, DirstateV2ParseError> {
         let mut children = self.root.as_ref();
         let mut components = path.components();
         let mut component =
             components.next().expect("expected at least one components");
         loop {
             if let Some(child) = children.get(component, self.on_disk)? {
                 if let Some(next_component) = components.next() {
                     component = next_component;
                     children = child.children(self.on_disk)?;
                 } else {
                     return Ok(Some(child));
                 }
             } else {
                 return Ok(None);
             }
         }
     }
 
     /// Returns a mutable reference to the node at `path` if it exists
     ///
     /// This takes `root` instead of `&mut self` so that callers can mutate
     /// other fields while the returned borrow is still valid
     fn get_node_mut<'tree>(
         on_disk: &'on_disk [u8],
+        unreachable_bytes: &mut u32,
         root: &'tree mut ChildNodes<'on_disk>,
         path: &HgPath,
     ) -> Result<Option<&'tree mut Node<'on_disk>>, DirstateV2ParseError> {
         let mut children = root;
         let mut components = path.components();
         let mut component =
             components.next().expect("expected at least one components");
         loop {
-            if let Some(child) = children.make_mut(on_disk)?.get_mut(component)
+            if let Some(child) = children
+                .make_mut(on_disk, unreachable_bytes)?
+                .get_mut(component)
             {
                 if let Some(next_component) = components.next() {
                     component = next_component;
                     children = &mut child.children;
                 } else {
                     return Ok(Some(child));
                 }
             } else {
                 return Ok(None);
             }
         }
     }
 
     pub(super) fn get_or_insert<'tree, 'path>(
         &'tree mut self,
         path: &HgPath,
     ) -> Result<&'tree mut Node<'on_disk>, DirstateV2ParseError> {
         Self::get_or_insert_node(
             self.on_disk,
+            &mut self.unreachable_bytes,
             &mut self.root,
             path,
             WithBasename::to_cow_owned,
             |_| {},
         )
     }
 
-    pub(super) fn get_or_insert_node<'tree, 'path>(
+    fn get_or_insert_node<'tree, 'path>(
         on_disk: &'on_disk [u8],
+        unreachable_bytes: &mut u32,
         root: &'tree mut ChildNodes<'on_disk>,
         path: &'path HgPath,
         to_cow: impl Fn(
             WithBasename<&'path HgPath>,
         ) -> WithBasename<Cow<'on_disk, HgPath>>,
         mut each_ancestor: impl FnMut(&mut Node),
     ) -> Result<&'tree mut Node<'on_disk>, DirstateV2ParseError> {
         let mut child_nodes = root;
         let mut inclusive_ancestor_paths =
             WithBasename::inclusive_ancestors_of(path);
         let mut ancestor_path = inclusive_ancestor_paths
             .next()
             .expect("expected at least one inclusive ancestor");
         loop {
             // TODO: can we avoid allocating an owned key in cases where the
             // map already contains that key, without introducing double
             // lookup?
             let child_node = child_nodes
-                .make_mut(on_disk)?
+                .make_mut(on_disk, unreachable_bytes)?
                 .entry(to_cow(ancestor_path))
                 .or_default();
             if let Some(next) = inclusive_ancestor_paths.next() {
                 each_ancestor(child_node);
                 ancestor_path = next;
                 child_nodes = &mut child_node.children;
             } else {
                 return Ok(child_node);
             }
         }
     }
 
     fn add_or_remove_file(
         &mut self,
         path: &HgPath,
         old_state: EntryState,
         new_entry: DirstateEntry,
     ) -> Result<(), DirstateV2ParseError> {
         let had_entry = old_state != EntryState::Unknown;
         let tracked_count_increment =
             match (old_state.is_tracked(), new_entry.state.is_tracked()) {
                 (false, true) => 1,
                 (true, false) => -1,
                 _ => 0,
             };
 
         let node = Self::get_or_insert_node(
             self.on_disk,
+            &mut self.unreachable_bytes,
             &mut self.root,
             path,
             WithBasename::to_cow_owned,
             |ancestor| {
                 if !had_entry {
                     ancestor.descendants_with_entry_count += 1;
                 }
 
                 // We can’t use `+= increment` because the counter is unsigned,
                 // and we want debug builds to detect accidental underflow
                 // through zero
                 match tracked_count_increment {
                     1 => ancestor.tracked_descendants_count += 1,
                     -1 => ancestor.tracked_descendants_count -= 1,
                     _ => {}
                 }
             },
         )?;
         if !had_entry {
             self.nodes_with_entry_count += 1
         }
         node.data = NodeData::Entry(new_entry);
         Ok(())
     }
 
     fn iter_nodes<'tree>(
         &'tree self,
     ) -> impl Iterator<
         Item = Result<NodeRef<'tree, 'on_disk>, DirstateV2ParseError>,
     > + 'tree {
         // Depth first tree traversal.
         //
         // If we could afford internal iteration and recursion,
         // this would look like:
         //
         // ```
         // fn traverse_children(
         //     children: &ChildNodes,
         //     each: &mut impl FnMut(&Node),
         // ) {
         //     for child in children.values() {
         //         traverse_children(&child.children, each);
         //         each(child);
         //     }
         // }
         // ```
         //
         // However we want an external iterator and therefore can’t use the
         // call stack. Use an explicit stack instead:
         let mut stack = Vec::new();
         let mut iter = self.root.as_ref().iter();
         std::iter::from_fn(move || {
             while let Some(child_node) = iter.next() {
                 let children = match child_node.children(self.on_disk) {
                     Ok(children) => children,
                     Err(error) => return Some(Err(error)),
                 };
                 // Pseudo-recursion
                 let new_iter = children.iter();
                 let old_iter = std::mem::replace(&mut iter, new_iter);
                 stack.push((child_node, old_iter));
             }
             // Found the end of a `children.iter()` iterator.
             if let Some((child_node, next_iter)) = stack.pop() {
                 // "Return" from pseudo-recursion by restoring state from the
                 // explicit stack
                 iter = next_iter;
 
                 Some(Ok(child_node))
             } else {
                 // Reached the bottom of the stack, we’re done
                 None
             }
         })
     }
 
     fn clear_known_ambiguous_mtimes(
         &mut self,
         paths: &[impl AsRef<HgPath>],
     ) -> Result<(), DirstateV2ParseError> {
         for path in paths {
             if let Some(node) = Self::get_node_mut(
                 self.on_disk,
+                &mut self.unreachable_bytes,
                 &mut self.root,
                 path.as_ref(),
             )? {
                 if let NodeData::Entry(entry) = &mut node.data {
                     entry.clear_mtime();
                 }
             }
         }
         Ok(())
     }
 
     /// Return a faillilble iterator of full paths of nodes that have an
     /// `entry` for which the given `predicate` returns true.
     ///
     /// Fallibility means that each iterator item is a `Result`, which may
     /// indicate a parse error of the on-disk dirstate-v2 format. Such errors
     /// should only happen if Mercurial is buggy or a repository is corrupted.
     fn filter_full_paths<'tree>(
         &'tree self,
         predicate: impl Fn(&DirstateEntry) -> bool + 'tree,
     ) -> impl Iterator<Item = Result<&HgPath, DirstateV2ParseError>> + 'tree
     {
         filter_map_results(self.iter_nodes(), move |node| {
             if let Some(entry) = node.entry()? {
                 if predicate(&entry) {
                     return Ok(Some(node.full_path(self.on_disk)?));
                 }
             }
             Ok(None)
         })
     }
+
+    fn count_dropped_path(unreachable_bytes: &mut u32, path: &Cow<HgPath>) {
+        if let Cow::Borrowed(path) = path {
+            *unreachable_bytes += path.len() as u32
+        }
+    }
 }
 
 /// Like `Iterator::filter_map`, but over a fallible iterator of `Result`s.
 ///
 /// The callback is only called for incoming `Ok` values. Errors are passed
 /// through as-is. In order to let it use the `?` operator the callback is
 /// expected to return a `Result` of `Option`, instead of an `Option` of
 /// `Result`.
 fn filter_map_results<'a, I, F, A, B, E>(
     iter: I,
     f: F,
 ) -> impl Iterator<Item = Result<B, E>> + 'a
 where
     I: Iterator<Item = Result<A, E>> + 'a,
     F: Fn(A) -> Result<Option<B>, E> + 'a,
 {
     iter.filter_map(move |result| match result {
         Ok(node) => f(node).transpose(),
         Err(e) => Some(Err(e)),
     })
 }
 
 impl<'on_disk> super::dispatch::DirstateMapMethods for DirstateMap<'on_disk> {
     fn clear(&mut self) {
         self.root = Default::default();
         self.nodes_with_entry_count = 0;
         self.nodes_with_copy_source_count = 0;
     }
 
     fn add_file(
         &mut self,
         filename: &HgPath,
         entry: DirstateEntry,
         added: bool,
         merged: bool,
         from_p2: bool,
         possibly_dirty: bool,
     ) -> Result<(), DirstateError> {
         let mut entry = entry;
         if added {
             assert!(!possibly_dirty);
             assert!(!from_p2);
             entry.state = EntryState::Added;
             entry.size = SIZE_NON_NORMAL;
             entry.mtime = MTIME_UNSET;
         } else if merged {
             assert!(!possibly_dirty);
             assert!(!from_p2);
             entry.state = EntryState::Merged;
             entry.size = SIZE_FROM_OTHER_PARENT;
             entry.mtime = MTIME_UNSET;
         } else if from_p2 {
             assert!(!possibly_dirty);
             entry.state = EntryState::Normal;
             entry.size = SIZE_FROM_OTHER_PARENT;
             entry.mtime = MTIME_UNSET;
         } else if possibly_dirty {
             entry.state = EntryState::Normal;
             entry.size = SIZE_NON_NORMAL;
             entry.mtime = MTIME_UNSET;
         } else {
             entry.state = EntryState::Normal;
             entry.size = entry.size & V1_RANGEMASK;
             entry.mtime = entry.mtime & V1_RANGEMASK;
         }
 
         let old_state = match self.get(filename)? {
             Some(e) => e.state,
             None => EntryState::Unknown,
         };
 
         Ok(self.add_or_remove_file(filename, old_state, entry)?)
     }
 
     fn remove_file(
         &mut self,
         filename: &HgPath,
         in_merge: bool,
     ) -> Result<(), DirstateError> {
         let old_entry_opt = self.get(filename)?;
         let old_state = match old_entry_opt {
             Some(e) => e.state,
             None => EntryState::Unknown,
         };
         let mut size = 0;
         if in_merge {
             // XXX we should not be able to have 'm' state and 'FROM_P2' if not
             // during a merge. So I (marmoute) am not sure we need the
             // conditionnal at all. Adding double checking this with assert
             // would be nice.
             if let Some(old_entry) = old_entry_opt {
                 // backup the previous state
                 if old_entry.state == EntryState::Merged {
                     size = SIZE_NON_NORMAL;
                 } else if old_entry.state == EntryState::Normal
                     && old_entry.size == SIZE_FROM_OTHER_PARENT
                 {
                     // other parent
                     size = SIZE_FROM_OTHER_PARENT;
                 }
             }
         }
         if size == 0 {
             self.copy_map_remove(filename)?;
         }
         let entry = DirstateEntry {
             state: EntryState::Removed,
             mode: 0,
             size,
             mtime: 0,
         };
         Ok(self.add_or_remove_file(filename, old_state, entry)?)
     }
 
     fn drop_file(&mut self, filename: &HgPath) -> Result<bool, DirstateError> {
         let old_state = match self.get(filename)? {
             Some(e) => e.state,
             None => EntryState::Unknown,
         };
         struct Dropped {
             was_tracked: bool,
             had_entry: bool,
             had_copy_source: bool,
         }
 
         /// If this returns `Ok(Some((dropped, removed)))`, then
         ///
         /// * `dropped` is about the leaf node that was at `filename`
         /// * `removed` is whether this particular level of recursion just
         ///   removed a node in `nodes`.
         fn recur<'on_disk>(
             on_disk: &'on_disk [u8],
+            unreachable_bytes: &mut u32,
             nodes: &mut ChildNodes<'on_disk>,
             path: &HgPath,
         ) -> Result<Option<(Dropped, bool)>, DirstateV2ParseError> {
             let (first_path_component, rest_of_path) =
                 path.split_first_component();
-            let node = if let Some(node) =
-                nodes.make_mut(on_disk)?.get_mut(first_path_component)
+            let nodes = nodes.make_mut(on_disk, unreachable_bytes)?;
+            let node = if let Some(node) = nodes.get_mut(first_path_component)
             {
                 node
             } else {
                 return Ok(None);
             };
             let dropped;
             if let Some(rest) = rest_of_path {
-                if let Some((d, removed)) =
-                    recur(on_disk, &mut node.children, rest)?
-                {
+                if let Some((d, removed)) = recur(
+                    on_disk,
+                    unreachable_bytes,
+                    &mut node.children,
+                    rest,
+                )? {
                     dropped = d;
                     if dropped.had_entry {
                         node.descendants_with_entry_count -= 1;
                     }
                     if dropped.was_tracked {
                         node.tracked_descendants_count -= 1;
                     }
 
                     // Directory caches must be invalidated when removing a
                     // child node
                     if removed {
                         if let NodeData::CachedDirectory { .. } = &node.data {
                             node.data = NodeData::None
                         }
                     }
                 } else {
                     return Ok(None);
                 }
             } else {
                 let had_entry = node.data.has_entry();
                 if had_entry {
                     node.data = NodeData::None
                 }
+                if let Some(source) = &node.copy_source {
+                    DirstateMap::count_dropped_path(unreachable_bytes, source)
+                }
                 dropped = Dropped {
                     was_tracked: node
                         .data
                         .as_entry()
                         .map_or(false, |entry| entry.state.is_tracked()),
                     had_entry,
                     had_copy_source: node.copy_source.take().is_some(),
                 };
             }
             // After recursion, for both leaf (rest_of_path is None) nodes and
             // parent nodes, remove a node if it just became empty.
             let remove = !node.data.has_entry()
                 && node.copy_source.is_none()
                 && node.children.is_empty();
             if remove {
-                nodes.make_mut(on_disk)?.remove(first_path_component);
+                let (key, _) =
+                    nodes.remove_entry(first_path_component).unwrap();
+                DirstateMap::count_dropped_path(
+                    unreachable_bytes,
+                    key.full_path(),
+                )
             }
             Ok(Some((dropped, remove)))
         }
 
-        if let Some((dropped, _removed)) =
-            recur(self.on_disk, &mut self.root, filename)?
-        {
+        if let Some((dropped, _removed)) = recur(
+            self.on_disk,
+            &mut self.unreachable_bytes,
+            &mut self.root,
+            filename,
+        )? {
             if dropped.had_entry {
                 self.nodes_with_entry_count -= 1
             }
             if dropped.had_copy_source {
                 self.nodes_with_copy_source_count -= 1
             }
             Ok(dropped.had_entry)
         } else {
             debug_assert!(!old_state.is_tracked());
             Ok(false)
         }
     }
 
     fn clear_ambiguous_times(
         &mut self,
         filenames: Vec<HgPathBuf>,
         now: i32,
     ) -> Result<(), DirstateV2ParseError> {
         for filename in filenames {
-            if let Some(node) =
-                Self::get_node_mut(self.on_disk, &mut self.root, &filename)?
-            {
+            if let Some(node) = Self::get_node_mut(
+                self.on_disk,
+                &mut self.unreachable_bytes,
+                &mut self.root,
+                &filename,
+            )? {
                 if let NodeData::Entry(entry) = &mut node.data {
                     entry.clear_ambiguous_mtime(now);
                 }
             }
         }
         Ok(())
     }
 
     fn non_normal_entries_contains(
         &mut self,
         key: &HgPath,
     ) -> Result<bool, DirstateV2ParseError> {
         Ok(if let Some(node) = self.get_node(key)? {
             node.entry()?.map_or(false, |entry| entry.is_non_normal())
         } else {
             false
         })
     }
 
     fn non_normal_entries_remove(&mut self, _key: &HgPath) {
         // Do nothing, this `DirstateMap` does not have a separate "non normal
         // entries" set that need to be kept up to date
     }
 
     fn non_normal_or_other_parent_paths(
         &mut self,
     ) -> Box<dyn Iterator<Item = Result<&HgPath, DirstateV2ParseError>> + '_>
     {
         Box::new(self.filter_full_paths(|entry| {
             entry.is_non_normal() || entry.is_from_other_parent()
         }))
     }
 
     fn set_non_normal_other_parent_entries(&mut self, _force: bool) {
         // Do nothing, this `DirstateMap` does not have a separate "non normal
         // entries" and "from other parent" sets that need to be recomputed
     }
 
     fn iter_non_normal_paths(
         &mut self,
     ) -> Box<
         dyn Iterator<Item = Result<&HgPath, DirstateV2ParseError>> + Send + '_,
     > {
         self.iter_non_normal_paths_panic()
     }
 
     fn iter_non_normal_paths_panic(
         &self,
     ) -> Box<
         dyn Iterator<Item = Result<&HgPath, DirstateV2ParseError>> + Send + '_,
     > {
         Box::new(self.filter_full_paths(|entry| entry.is_non_normal()))
     }
 
     fn iter_other_parent_paths(
         &mut self,
     ) -> Box<
         dyn Iterator<Item = Result<&HgPath, DirstateV2ParseError>> + Send + '_,
     > {
         Box::new(self.filter_full_paths(|entry| entry.is_from_other_parent()))
     }
 
     fn has_tracked_dir(
         &mut self,
         directory: &HgPath,
     ) -> Result<bool, DirstateError> {
         if let Some(node) = self.get_node(directory)? {
             // A node without a `DirstateEntry` was created to hold child
             // nodes, and is therefore a directory.
             let state = node.state()?;
             Ok(state.is_none() && node.tracked_descendants_count() > 0)
         } else {
             Ok(false)
         }
     }
 
     fn has_dir(&mut self, directory: &HgPath) -> Result<bool, DirstateError> {
         if let Some(node) = self.get_node(directory)? {
             // A node without a `DirstateEntry` was created to hold child
             // nodes, and is therefore a directory.
             let state = node.state()?;
             Ok(state.is_none() && node.descendants_with_entry_count() > 0)
         } else {
             Ok(false)
         }
     }
 
     #[timed]
     fn pack_v1(
         &mut self,
         parents: DirstateParents,
         now: Timestamp,
     ) -> Result<Vec<u8>, DirstateError> {
         let now: i32 = now.0.try_into().expect("time overflow");
         let mut ambiguous_mtimes = Vec::new();
         // Optizimation (to be measured?): pre-compute size to avoid `Vec`
         // reallocations
         let mut size = parents.as_bytes().len();
         for node in self.iter_nodes() {
             let node = node?;
             if let Some(entry) = node.entry()? {
                 size += packed_entry_size(
                     node.full_path(self.on_disk)?,
                     node.copy_source(self.on_disk)?,
                 );
                 if entry.mtime_is_ambiguous(now) {
                     ambiguous_mtimes.push(
                         node.full_path_borrowed(self.on_disk)?
                             .detach_from_tree(),
                     )
                 }
             }
         }
         self.clear_known_ambiguous_mtimes(&ambiguous_mtimes)?;
 
         let mut packed = Vec::with_capacity(size);
         packed.extend(parents.as_bytes());
 
         for node in self.iter_nodes() {
             let node = node?;
             if let Some(entry) = node.entry()? {
                 pack_entry(
                     node.full_path(self.on_disk)?,
                     &entry,
                     node.copy_source(self.on_disk)?,
                     &mut packed,
                 );
             }
         }
         Ok(packed)
     }
 
     /// Returns new data together with whether that data should be appended to
     /// the existing data file whose content is at `self.on_disk` (true),
     /// instead of written to a new data file (false).
     #[timed]
     fn pack_v2(
         &mut self,
         now: Timestamp,
         can_append: bool,
     ) -> Result<(Vec<u8>, bool), DirstateError> {
         // TODO: how do we want to handle this in 2038?
         let now: i32 = now.0.try_into().expect("time overflow");
         let mut paths = Vec::new();
         for node in self.iter_nodes() {
             let node = node?;
             if let Some(entry) = node.entry()? {
                 if entry.mtime_is_ambiguous(now) {
                     paths.push(
                         node.full_path_borrowed(self.on_disk)?
                             .detach_from_tree(),
                     )
                 }
             }
         }
         // Borrow of `self` ends here since we collect cloned paths
 
         self.clear_known_ambiguous_mtimes(&paths)?;
 
         on_disk::write(self, can_append)
     }
 
     fn status<'a>(
         &'a mut self,
         matcher: &'a (dyn Matcher + Sync),
         root_dir: PathBuf,
         ignore_files: Vec<PathBuf>,
         options: StatusOptions,
     ) -> Result<(DirstateStatus<'a>, Vec<PatternFileWarning>), StatusError>
     {
         super::status::status(self, matcher, root_dir, ignore_files, options)
     }
 
     fn copy_map_len(&self) -> usize {
         self.nodes_with_copy_source_count as usize
     }
 
     fn copy_map_iter(&self) -> CopyMapIter<'_> {
         Box::new(filter_map_results(self.iter_nodes(), move |node| {
             Ok(if let Some(source) = node.copy_source(self.on_disk)? {
                 Some((node.full_path(self.on_disk)?, source))
             } else {
                 None
             })
         }))
     }
 
     fn copy_map_contains_key(
         &self,
         key: &HgPath,
     ) -> Result<bool, DirstateV2ParseError> {
         Ok(if let Some(node) = self.get_node(key)? {
             node.has_copy_source()
         } else {
             false
         })
     }
 
     fn copy_map_get(
         &self,
         key: &HgPath,
     ) -> Result<Option<&HgPath>, DirstateV2ParseError> {
         if let Some(node) = self.get_node(key)? {
             if let Some(source) = node.copy_source(self.on_disk)? {
                 return Ok(Some(source));
             }
         }
         Ok(None)
     }
 
     fn copy_map_remove(
         &mut self,
         key: &HgPath,
     ) -> Result<Option<HgPathBuf>, DirstateV2ParseError> {
         let count = &mut self.nodes_with_copy_source_count;
-        Ok(
-            Self::get_node_mut(self.on_disk, &mut self.root, key)?.and_then(
-                |node| {
-                    if node.copy_source.is_some() {
-                        *count -= 1
-                    }
-                    node.copy_source.take().map(Cow::into_owned)
-                },
-            ),
-        )
+        let unreachable_bytes = &mut self.unreachable_bytes;
+        Ok(Self::get_node_mut(
+            self.on_disk,
+            unreachable_bytes,
+            &mut self.root,
+            key,
+        )?
+        .and_then(|node| {
+            if let Some(source) = &node.copy_source {
+                *count -= 1;
+                Self::count_dropped_path(unreachable_bytes, source);
+            }
+            node.copy_source.take().map(Cow::into_owned)
+        }))
     }
 
     fn copy_map_insert(
         &mut self,
         key: HgPathBuf,
         value: HgPathBuf,
     ) -> Result<Option<HgPathBuf>, DirstateV2ParseError> {
         let node = Self::get_or_insert_node(
             self.on_disk,
+            &mut self.unreachable_bytes,
             &mut self.root,
             &key,
             WithBasename::to_cow_owned,
             |_ancestor| {},
         )?;
         if node.copy_source.is_none() {
             self.nodes_with_copy_source_count += 1
         }
         Ok(node.copy_source.replace(value.into()).map(Cow::into_owned))
     }
 
     fn len(&self) -> usize {
         self.nodes_with_entry_count as usize
     }
 
     fn contains_key(
         &self,
         key: &HgPath,
     ) -> Result<bool, DirstateV2ParseError> {
         Ok(self.get(key)?.is_some())
     }
 
     fn get(
         &self,
         key: &HgPath,
     ) -> Result<Option<DirstateEntry>, DirstateV2ParseError> {
         Ok(if let Some(node) = self.get_node(key)? {
             node.entry()?
         } else {
             None
         })
     }
 
     fn iter(&self) -> StateMapIter<'_> {
         Box::new(filter_map_results(self.iter_nodes(), move |node| {
             Ok(if let Some(entry) = node.entry()? {
                 Some((node.full_path(self.on_disk)?, entry))
             } else {
                 None
             })
         }))
     }
 
     fn iter_directories(
         &self,
     ) -> Box<
         dyn Iterator<
                 Item = Result<
                     (&HgPath, Option<Timestamp>),
                     DirstateV2ParseError,
                 >,
             > + Send
             + '_,
     > {
         Box::new(filter_map_results(self.iter_nodes(), move |node| {
             Ok(if node.state()?.is_none() {
                 Some((
                     node.full_path(self.on_disk)?,
                     node.cached_directory_mtime()
                         .map(|mtime| Timestamp(mtime.seconds())),
                 ))
             } else {
                 None
             })
         }))
     }
 }

rust/hg-core/src/dirstate_tree/on_disk.rs

@@ -1,748 +1,756 b''
 //! The "version 2" disk representation of the dirstate
 //!
 //! # File format
 //!
 //! In dirstate-v2 format, the `.hg/dirstate` file is a "docket that starts
 //! with a fixed-sized header whose layout is defined by the `DocketHeader`
 //! struct, followed by the data file identifier.
 //!
 //! A separate `.hg/dirstate.{uuid}.d` file contains most of the data. That
 //! file may be longer than the size given in the docket, but not shorter. Only
 //! the start of the data file up to the given size is considered. The
 //! fixed-size "root" of the dirstate tree whose layout is defined by the
 //! `Root` struct is found at the end of that slice of data.
 //!
 //! Its `root_nodes` field contains the slice (offset and length) to
 //! the nodes representing the files and directories at the root of the
 //! repository. Each node is also fixed-size, defined by the `Node` struct.
 //! Nodes in turn contain slices to variable-size paths, and to their own child
 //! nodes (if any) for nested files and directories.
 
 use crate::dirstate_tree::dirstate_map::{self, DirstateMap, NodeRef};
 use crate::dirstate_tree::path_with_basename::WithBasename;
 use crate::errors::HgError;
 use crate::utils::hg_path::HgPath;
 use crate::DirstateEntry;
 use crate::DirstateError;
 use crate::DirstateParents;
 use crate::EntryState;
 use bytes_cast::unaligned::{I32Be, I64Be, U16Be, U32Be};
 use bytes_cast::BytesCast;
 use format_bytes::format_bytes;
 use std::borrow::Cow;
 use std::convert::{TryFrom, TryInto};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
 /// Added at the start of `.hg/dirstate` when the "v2" format is used.
 /// This a redundant sanity check more than an actual "magic number" since
 /// `.hg/requires` already governs which format should be used.
 pub const V2_FORMAT_MARKER: &[u8; 12] = b"dirstate-v2\n";
 
 /// Keep space for 256-bit hashes
 const STORED_NODE_ID_BYTES: usize = 32;
 
 /// … even though only 160 bits are used for now, with SHA-1
 const USED_NODE_ID_BYTES: usize = 20;
 
 pub(super) const IGNORE_PATTERNS_HASH_LEN: usize = 20;
 pub(super) type IgnorePatternsHash = [u8; IGNORE_PATTERNS_HASH_LEN];
 
 // Must match `HEADER` in `mercurial/dirstateutils/docket.py`
 #[derive(BytesCast)]
 #[repr(C)]
 struct DocketHeader {
     marker: [u8; V2_FORMAT_MARKER.len()],
     parent_1: [u8; STORED_NODE_ID_BYTES],
     parent_2: [u8; STORED_NODE_ID_BYTES],
 
     /// Counted in bytes
     data_size: Size,
 
     uuid_size: u8,
 }
 
 pub struct Docket<'on_disk> {
     header: &'on_disk DocketHeader,
     uuid: &'on_disk [u8],
 }
 
 #[derive(BytesCast)]
 #[repr(C)]
 struct Root {
     root_nodes: ChildNodes,
     nodes_with_entry_count: Size,
     nodes_with_copy_source_count: Size,
 
+    /// How many bytes of this data file are not used anymore
+    unreachable_bytes: Size,
+
     /// If non-zero, a hash of ignore files that were used for some previous
     /// run of the `status` algorithm.
     ///
     /// We define:
     ///
     /// * "Root" ignore files are `.hgignore` at the root of the repository if
     ///   it exists, and files from `ui.ignore.*` config. This set of files is
     ///   then sorted by the string representation of their path.
     /// * The "expanded contents" of an ignore files is the byte string made
     ///   by concatenating its contents with the "expanded contents" of other
     ///   files included with `include:` or `subinclude:` files, in inclusion
     ///   order. This definition is recursive, as included files can
     ///   themselves include more files.
     ///
     /// This hash is defined as the SHA-1 of the concatenation (in sorted
     /// order) of the "expanded contents" of each "root" ignore file.
     /// (Note that computing this does not require actually concatenating byte
     /// strings into contiguous memory, instead SHA-1 hashing can be done
     /// incrementally.)
     ignore_patterns_hash: IgnorePatternsHash,
 }
 
 #[derive(BytesCast)]
 #[repr(C)]
 pub(super) struct Node {
     full_path: PathSlice,
 
     /// In bytes from `self.full_path.start`
     base_name_start: PathSize,
 
     copy_source: OptPathSlice,
     children: ChildNodes,
     pub(super) descendants_with_entry_count: Size,
     pub(super) tracked_descendants_count: Size,
 
     /// Depending on the value of `state`:
     ///
     /// * A null byte: `data` is not used.
     ///
     /// * A `n`, `a`, `r`, or `m` ASCII byte: `state` and `data` together
     ///   represent a dirstate entry like in the v1 format.
     ///
     /// * A `d` ASCII byte: the bytes of `data` should instead be interpreted
     ///   as the `Timestamp` for the mtime of a cached directory.
     ///
     ///   The presence of this state means that at some point, this path in
     ///   the working directory was observed:
     ///
     ///   - To be a directory
     ///   - With the modification time as given by `Timestamp`
     ///   - That timestamp was already strictly in the past when observed,
     ///     meaning that later changes cannot happen in the same clock tick
     ///     and must cause a different modification time (unless the system
     ///     clock jumps back and we get unlucky, which is not impossible but
     ///     but deemed unlikely enough).
     ///   - All direct children of this directory (as returned by
     ///     `std::fs::read_dir`) either have a corresponding dirstate node, or
     ///     are ignored by ignore patterns whose hash is in
     ///     `Root::ignore_patterns_hash`.
     ///
     ///   This means that if `std::fs::symlink_metadata` later reports the
     ///   same modification time and ignored patterns haven’t changed, a run
     ///   of status that is not listing ignored   files can skip calling
     ///   `std::fs::read_dir` again for this directory,   iterate child
     ///   dirstate nodes instead.
     state: u8,
     data: Entry,
 }
 
 #[derive(BytesCast, Copy, Clone)]
 #[repr(C)]
 struct Entry {
     mode: I32Be,
     mtime: I32Be,
     size: I32Be,
 }
 
 /// Duration since the Unix epoch
 #[derive(BytesCast, Copy, Clone, PartialEq)]
 #[repr(C)]
 pub(super) struct Timestamp {
     seconds: I64Be,
 
     /// In `0 .. 1_000_000_000`.
     ///
     /// This timestamp is later or earlier than `(seconds, 0)` by this many
     /// nanoseconds, if `seconds` is non-negative or negative, respectively.
     nanoseconds: U32Be,
 }
 
 /// Counted in bytes from the start of the file
 ///
 /// NOTE: not supporting `.hg/dirstate` files larger than 4 GiB.
 type Offset = U32Be;
 
 /// Counted in number of items
 ///
 /// NOTE: we choose not to support counting more than 4 billion nodes anywhere.
 type Size = U32Be;
 
 /// Counted in bytes
 ///
 /// NOTE: we choose not to support file names/paths longer than 64 KiB.
 type PathSize = U16Be;
 
 /// A contiguous sequence of `len` times `Node`, representing the child nodes
 /// of either some other node or of the repository root.
 ///
 /// Always sorted by ascending `full_path`, to allow binary search.
 /// Since nodes with the same parent nodes also have the same parent path,
 /// only the `base_name`s need to be compared during binary search.
 #[derive(BytesCast, Copy, Clone)]
 #[repr(C)]
 struct ChildNodes {
     start: Offset,
     len: Size,
 }
 
 /// A `HgPath` of `len` bytes
 #[derive(BytesCast, Copy, Clone)]
 #[repr(C)]
 struct PathSlice {
     start: Offset,
     len: PathSize,
 }
 
 /// Either nothing if `start == 0`, or a `HgPath` of `len` bytes
 type OptPathSlice = PathSlice;
 
 /// Make sure that size-affecting changes are made knowingly
 fn _static_assert_size_of() {
     let _ = std::mem::transmute::<DocketHeader, [u8; 81]>;
-    let _ = std::mem::transmute::<Root, [u8; 36]>;
+    let _ = std::mem::transmute::<Root, [u8; 40]>;
     let _ = std::mem::transmute::<Node, [u8; 43]>;
 }
 
 /// Unexpected file format found in `.hg/dirstate` with the "v2" format.
 ///
 /// This should only happen if Mercurial is buggy or a repository is corrupted.
 #[derive(Debug)]
 pub struct DirstateV2ParseError;
 
 impl From<DirstateV2ParseError> for HgError {
     fn from(_: DirstateV2ParseError) -> Self {
         HgError::corrupted("dirstate-v2 parse error")
     }
 }
 
 impl From<DirstateV2ParseError> for crate::DirstateError {
     fn from(error: DirstateV2ParseError) -> Self {
         HgError::from(error).into()
     }
 }
 
 impl<'on_disk> Docket<'on_disk> {
     pub fn parents(&self) -> DirstateParents {
         use crate::Node;
         let p1 = Node::try_from(&self.header.parent_1[..USED_NODE_ID_BYTES])
             .unwrap()
             .clone();
         let p2 = Node::try_from(&self.header.parent_2[..USED_NODE_ID_BYTES])
             .unwrap()
             .clone();
         DirstateParents { p1, p2 }
     }
 
     pub fn data_size(&self) -> usize {
         // This `unwrap` could only panic on a 16-bit CPU
         self.header.data_size.get().try_into().unwrap()
     }
 
     pub fn data_filename(&self) -> String {
         String::from_utf8(format_bytes!(b"dirstate.{}.d", self.uuid)).unwrap()
     }
 }
 
 pub fn read_docket(
     on_disk: &[u8],
 ) -> Result<Docket<'_>, DirstateV2ParseError> {
     let (header, uuid) =
         DocketHeader::from_bytes(on_disk).map_err(|_| DirstateV2ParseError)?;
     let uuid_size = header.uuid_size as usize;
     if header.marker == *V2_FORMAT_MARKER && uuid.len() == uuid_size {
         Ok(Docket { header, uuid })
     } else {
         Err(DirstateV2ParseError)
     }
 }
 
 fn read_root<'on_disk>(
     on_disk: &'on_disk [u8],
 ) -> Result<&'on_disk Root, DirstateV2ParseError> {
     // Find the `Root` at the end of the given slice
     let root_offset = on_disk
         .len()
         .checked_sub(std::mem::size_of::<Root>())
         // A non-empty slice too short is an error
         .ok_or(DirstateV2ParseError)?;
     let (root, _) = Root::from_bytes(&on_disk[root_offset..])
         .map_err(|_| DirstateV2ParseError)?;
     Ok(root)
 }
 
 pub(super) fn read<'on_disk>(
     on_disk: &'on_disk [u8],
 ) -> Result<DirstateMap<'on_disk>, DirstateV2ParseError> {
     if on_disk.is_empty() {
         return Ok(DirstateMap::empty(on_disk));
     }
     let root = read_root(on_disk)?;
+    let mut unreachable_bytes = root.unreachable_bytes.get();
+    // Each append writes a new `Root`, so it’s never reused
+    unreachable_bytes += std::mem::size_of::<Root>() as u32;
     let dirstate_map = DirstateMap {
         on_disk,
         root: dirstate_map::ChildNodes::OnDisk(read_nodes(
             on_disk,
             root.root_nodes,
         )?),
         nodes_with_entry_count: root.nodes_with_entry_count.get(),
         nodes_with_copy_source_count: root.nodes_with_copy_source_count.get(),
         ignore_patterns_hash: root.ignore_patterns_hash,
+        unreachable_bytes,
     };
     Ok(dirstate_map)
 }
 
 impl Node {
     pub(super) fn full_path<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<&'on_disk HgPath, DirstateV2ParseError> {
         read_hg_path(on_disk, self.full_path)
     }
 
     pub(super) fn base_name_start<'on_disk>(
         &self,
     ) -> Result<usize, DirstateV2ParseError> {
         let start = self.base_name_start.get();
         if start < self.full_path.len.get() {
             let start = usize::try_from(start)
                 // u32 -> usize, could only panic on a 16-bit CPU
                 .expect("dirstate-v2 base_name_start out of bounds");
             Ok(start)
         } else {
             Err(DirstateV2ParseError)
         }
     }
 
     pub(super) fn base_name<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<&'on_disk HgPath, DirstateV2ParseError> {
         let full_path = self.full_path(on_disk)?;
         let base_name_start = self.base_name_start()?;
         Ok(HgPath::new(&full_path.as_bytes()[base_name_start..]))
     }
 
     pub(super) fn path<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<dirstate_map::NodeKey<'on_disk>, DirstateV2ParseError> {
         Ok(WithBasename::from_raw_parts(
             Cow::Borrowed(self.full_path(on_disk)?),
             self.base_name_start()?,
         ))
     }
 
     pub(super) fn has_copy_source<'on_disk>(&self) -> bool {
         self.copy_source.start.get() != 0
     }
 
     pub(super) fn copy_source<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<Option<&'on_disk HgPath>, DirstateV2ParseError> {
         Ok(if self.has_copy_source() {
             Some(read_hg_path(on_disk, self.copy_source)?)
         } else {
             None
         })
     }
 
     pub(super) fn node_data(
         &self,
     ) -> Result<dirstate_map::NodeData, DirstateV2ParseError> {
         let entry = |state| {
             dirstate_map::NodeData::Entry(self.entry_with_given_state(state))
         };
 
         match self.state {
             b'\0' => Ok(dirstate_map::NodeData::None),
             b'd' => Ok(dirstate_map::NodeData::CachedDirectory {
                 mtime: *self.data.as_timestamp(),
             }),
             b'n' => Ok(entry(EntryState::Normal)),
             b'a' => Ok(entry(EntryState::Added)),
             b'r' => Ok(entry(EntryState::Removed)),
             b'm' => Ok(entry(EntryState::Merged)),
             _ => Err(DirstateV2ParseError),
         }
     }
 
     pub(super) fn cached_directory_mtime(&self) -> Option<&Timestamp> {
         if self.state == b'd' {
             Some(self.data.as_timestamp())
         } else {
             None
         }
     }
 
     pub(super) fn state(
         &self,
     ) -> Result<Option<EntryState>, DirstateV2ParseError> {
         match self.state {
             b'\0' | b'd' => Ok(None),
             b'n' => Ok(Some(EntryState::Normal)),
             b'a' => Ok(Some(EntryState::Added)),
             b'r' => Ok(Some(EntryState::Removed)),
             b'm' => Ok(Some(EntryState::Merged)),
             _ => Err(DirstateV2ParseError),
         }
     }
 
     fn entry_with_given_state(&self, state: EntryState) -> DirstateEntry {
         DirstateEntry {
             state,
             mode: self.data.mode.get(),
             mtime: self.data.mtime.get(),
             size: self.data.size.get(),
         }
     }
 
     pub(super) fn entry(
         &self,
     ) -> Result<Option<DirstateEntry>, DirstateV2ParseError> {
         Ok(self
             .state()?
             .map(|state| self.entry_with_given_state(state)))
     }
 
     pub(super) fn children<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<&'on_disk [Node], DirstateV2ParseError> {
         read_nodes(on_disk, self.children)
     }
 
     pub(super) fn to_in_memory_node<'on_disk>(
         &self,
         on_disk: &'on_disk [u8],
     ) -> Result<dirstate_map::Node<'on_disk>, DirstateV2ParseError> {
         Ok(dirstate_map::Node {
             children: dirstate_map::ChildNodes::OnDisk(
                 self.children(on_disk)?,
             ),
             copy_source: self.copy_source(on_disk)?.map(Cow::Borrowed),
             data: self.node_data()?,
             descendants_with_entry_count: self
                 .descendants_with_entry_count
                 .get(),
             tracked_descendants_count: self.tracked_descendants_count.get(),
         })
     }
 }
 
 impl Entry {
     fn from_timestamp(timestamp: Timestamp) -> Self {
         // Safety: both types implement the `ByteCast` trait, so we could
         // safely use `as_bytes` and `from_bytes` to do this conversion. Using
         // `transmute` instead makes the compiler check that the two types
         // have the same size, which eliminates the error case of
         // `from_bytes`.
         unsafe { std::mem::transmute::<Timestamp, Entry>(timestamp) }
     }
 
     fn as_timestamp(&self) -> &Timestamp {
         // Safety: same as above in `from_timestamp`
         unsafe { &*(self as *const Entry as *const Timestamp) }
     }
 }
 
 impl Timestamp {
     pub fn seconds(&self) -> i64 {
         self.seconds.get()
     }
 }
 
 impl From<SystemTime> for Timestamp {
     fn from(system_time: SystemTime) -> Self {
         let (secs, nanos) = match system_time.duration_since(UNIX_EPOCH) {
             Ok(duration) => {
                 (duration.as_secs() as i64, duration.subsec_nanos())
             }
             Err(error) => {
                 let negative = error.duration();
                 (-(negative.as_secs() as i64), negative.subsec_nanos())
             }
         };
         Timestamp {
             seconds: secs.into(),
             nanoseconds: nanos.into(),
         }
     }
 }
 
 impl From<&'_ Timestamp> for SystemTime {
     fn from(timestamp: &'_ Timestamp) -> Self {
         let secs = timestamp.seconds.get();
         let nanos = timestamp.nanoseconds.get();
         if secs >= 0 {
             UNIX_EPOCH + Duration::new(secs as u64, nanos)
         } else {
             UNIX_EPOCH - Duration::new((-secs) as u64, nanos)
         }
     }
 }
 
 fn read_hg_path(
     on_disk: &[u8],
     slice: PathSlice,
 ) -> Result<&HgPath, DirstateV2ParseError> {
     read_slice(on_disk, slice.start, slice.len.get()).map(HgPath::new)
 }
 
 fn read_nodes(
     on_disk: &[u8],
     slice: ChildNodes,
 ) -> Result<&[Node], DirstateV2ParseError> {
     read_slice(on_disk, slice.start, slice.len.get())
 }
 
 fn read_slice<T, Len>(
     on_disk: &[u8],
     start: Offset,
     len: Len,
 ) -> Result<&[T], DirstateV2ParseError>
 where
     T: BytesCast,
     Len: TryInto<usize>,
 {
     // Either `usize::MAX` would result in "out of bounds" error since a single
     // `&[u8]` cannot occupy the entire addess space.
     let start = start.get().try_into().unwrap_or(std::usize::MAX);
     let len = len.try_into().unwrap_or(std::usize::MAX);
     on_disk
         .get(start..)
         .and_then(|bytes| T::slice_from_bytes(bytes, len).ok())
         .map(|(slice, _rest)| slice)
         .ok_or_else(|| DirstateV2ParseError)
 }
 
 pub(crate) fn for_each_tracked_path<'on_disk>(
     on_disk: &'on_disk [u8],
     mut f: impl FnMut(&'on_disk HgPath),
 ) -> Result<(), DirstateV2ParseError> {
     let root = read_root(on_disk)?;
     fn recur<'on_disk>(
         on_disk: &'on_disk [u8],
         nodes: ChildNodes,
         f: &mut impl FnMut(&'on_disk HgPath),
     ) -> Result<(), DirstateV2ParseError> {
         for node in read_nodes(on_disk, nodes)? {
             if let Some(state) = node.state()? {
                 if state.is_tracked() {
                     f(node.full_path(on_disk)?)
                 }
             }
             recur(on_disk, node.children, f)?
         }
         Ok(())
     }
     recur(on_disk, root.root_nodes, &mut f)
 }
 
 /// Returns new data together with whether that data should be appended to the
 /// existing data file whose content is at `dirstate_map.on_disk` (true),
 /// instead of written to a new data file (false).
 pub(super) fn write(
     dirstate_map: &mut DirstateMap,
     can_append: bool,
 ) -> Result<(Vec<u8>, bool), DirstateError> {
     let append = can_append && dirstate_map.write_should_append();
 
     // This ignores the space for paths, and for nodes without an entry.
     // TODO: better estimate? Skip the `Vec` and write to a file directly?
     let size_guess = std::mem::size_of::<Root>()
         + std::mem::size_of::<Node>()
             * dirstate_map.nodes_with_entry_count as usize;
 
     let mut writer = Writer {
         dirstate_map,
         append,
         out: Vec::with_capacity(size_guess),
     };
 
     let root_nodes = writer.write_nodes(dirstate_map.root.as_ref())?;
 
     let root = Root {
         root_nodes,
         nodes_with_entry_count: dirstate_map.nodes_with_entry_count.into(),
         nodes_with_copy_source_count: dirstate_map
             .nodes_with_copy_source_count
             .into(),
+        unreachable_bytes: dirstate_map.unreachable_bytes.into(),
         ignore_patterns_hash: dirstate_map.ignore_patterns_hash,
     };
     writer.out.extend(root.as_bytes());
     Ok((writer.out, append))
 }
 
 struct Writer<'dmap, 'on_disk> {
     dirstate_map: &'dmap DirstateMap<'on_disk>,
     append: bool,
     out: Vec<u8>,
 }
 
 impl Writer<'_, '_> {
     fn write_nodes(
         &mut self,
         nodes: dirstate_map::ChildNodesRef,
     ) -> Result<ChildNodes, DirstateError> {
         // Reuse already-written nodes if possible
         if self.append {
             if let dirstate_map::ChildNodesRef::OnDisk(nodes_slice) = nodes {
                 let start = self.on_disk_offset_of(nodes_slice).expect(
                     "dirstate-v2 OnDisk nodes not found within on_disk",
                 );
                 let len = child_nodes_len_from_usize(nodes_slice.len());
                 return Ok(ChildNodes { start, len });
             }
         }
 
         // `dirstate_map::ChildNodes::InMemory` contains a `HashMap` which has
         // undefined iteration order. Sort to enable binary search in the
         // written file.
         let nodes = nodes.sorted();
         let nodes_len = nodes.len();
 
         // First accumulate serialized nodes in a `Vec`
         let mut on_disk_nodes = Vec::with_capacity(nodes_len);
         for node in nodes {
             let children =
                 self.write_nodes(node.children(self.dirstate_map.on_disk)?)?;
             let full_path = node.full_path(self.dirstate_map.on_disk)?;
             let full_path = self.write_path(full_path.as_bytes());
             let copy_source = if let Some(source) =
                 node.copy_source(self.dirstate_map.on_disk)?
             {
                 self.write_path(source.as_bytes())
             } else {
                 PathSlice {
                     start: 0.into(),
                     len: 0.into(),
                 }
             };
             on_disk_nodes.push(match node {
                 NodeRef::InMemory(path, node) => {
                     let (state, data) = match &node.data {
                         dirstate_map::NodeData::Entry(entry) => (
                             entry.state.into(),
                             Entry {
                                 mode: entry.mode.into(),
                                 mtime: entry.mtime.into(),
                                 size: entry.size.into(),
                             },
                         ),
                         dirstate_map::NodeData::CachedDirectory { mtime } => {
                             (b'd', Entry::from_timestamp(*mtime))
                         }
                         dirstate_map::NodeData::None => (
                             b'\0',
                             Entry {
                                 mode: 0.into(),
                                 mtime: 0.into(),
                                 size: 0.into(),
                             },
                         ),
                     };
                     Node {
                         children,
                         copy_source,
                         full_path,
                         base_name_start: u16::try_from(path.base_name_start())
                             // Could only panic for paths over 64 KiB
                             .expect("dirstate-v2 path length overflow")
                             .into(),
                         descendants_with_entry_count: node
                             .descendants_with_entry_count
                             .into(),
                         tracked_descendants_count: node
                             .tracked_descendants_count
                             .into(),
                         state,
                         data,
                     }
                 }
                 NodeRef::OnDisk(node) => Node {
                     children,
                     copy_source,
                     full_path,
                     ..*node
                 },
             })
         }
         // … so we can write them contiguously, after writing everything else
         // they refer to.
         let start = self.current_offset();
         let len = child_nodes_len_from_usize(nodes_len);
         self.out.extend(on_disk_nodes.as_bytes());
         Ok(ChildNodes { start, len })
     }
 
     /// If the given slice of items is within `on_disk`, returns its offset
     /// from the start of `on_disk`.
     fn on_disk_offset_of<T>(&self, slice: &[T]) -> Option<Offset>
     where
         T: BytesCast,
     {
         fn address_range(slice: &[u8]) -> std::ops::RangeInclusive<usize> {
             let start = slice.as_ptr() as usize;
             let end = start + slice.len();
             start..=end
         }
         let slice_addresses = address_range(slice.as_bytes());
         let on_disk_addresses = address_range(self.dirstate_map.on_disk);
         if on_disk_addresses.contains(slice_addresses.start())
             && on_disk_addresses.contains(slice_addresses.end())
         {
             let offset = slice_addresses.start() - on_disk_addresses.start();
             Some(offset_from_usize(offset))
         } else {
             None
         }
     }
 
     fn current_offset(&mut self) -> Offset {
         let mut offset = self.out.len();
         if self.append {
             offset += self.dirstate_map.on_disk.len()
         }
         offset_from_usize(offset)
     }
 
     fn write_path(&mut self, slice: &[u8]) -> PathSlice {
         let len = path_len_from_usize(slice.len());
         // Reuse an already-written path if possible
         if self.append {
             if let Some(start) = self.on_disk_offset_of(slice) {
                 return PathSlice { start, len };
             }
         }
         let start = self.current_offset();
         self.out.extend(slice.as_bytes());
         PathSlice { start, len }
     }
 }
 
 fn offset_from_usize(x: usize) -> Offset {
     u32::try_from(x)
         // Could only panic for a dirstate file larger than 4 GiB
         .expect("dirstate-v2 offset overflow")
         .into()
 }
 
 fn child_nodes_len_from_usize(x: usize) -> Size {
     u32::try_from(x)
         // Could only panic with over 4 billion nodes
         .expect("dirstate-v2 slice length overflow")
         .into()
 }
 
 fn path_len_from_usize(x: usize) -> PathSize {
     u16::try_from(x)
         // Could only panic for paths over 64 KiB
         .expect("dirstate-v2 path length overflow")
         .into()
 }