armdb 0.1.14

use std::mem::size_of;
use std::ops::Bound;
use std::sync::Arc;

use crate::Key;

use crate::byte_view::ByteView;
use crate::cache::{BlockCache, BlockKey};
use crate::compaction::{CompactionIndex, compact_shard};
use crate::config::Config;
use crate::disk_loc::DiskLoc;
use crate::engine::Engine;
use crate::error::{DbError, DbResult};
use crate::hook::{NoHook, WriteHook};
use crate::io::aligned_buf::AlignedBuf;
use crate::io::direct;
use crate::recovery::recover_var_tree;
use crate::shard::ShardInner;
use crate::skiplist::node::{SkipNode, VarNode, random_height};
use crate::skiplist::{InsertResult, SkipList};
use crate::sync::MutexGuard;

/// A tree with fixed-size keys and variable-length values.
/// Values are stored as `ByteView` (inline ≤20 bytes, heap with ref counting for larger).
/// Disk reads are cached at 4096-byte block granularity via `BlockCache`.
///
/// Each `VarTree` owns its storage engine — one tree = one database directory.
///
/// # Usage
///
/// ```ignore
/// let tree = VarTree::<16>::open("data/messages", Config::default())?;
/// tree.put(&key, b"hello world")?;
/// tree.close()?;
/// ```
///
/// # Iteration
///
/// `iter()`, `range()`, and `prefix_iter()` all return [`VarIter`] which
/// implements `Iterator + DoubleEndedIterator` with `Item = (K, ByteView)`.
/// Each `next()` / `next_back()` may perform disk I/O on a block-cache miss.
///
/// ```ignore
/// let latest = tree.prefix_iter(&group_id).take(50).collect::<Vec<_>>();
/// let oldest = tree.prefix_iter(&group_id).rev().take(10);  // DoubleEndedIterator
/// ```
pub struct VarTree<K: Key, H: WriteHook<K> = NoHook> {
    index: SkipList<VarNode<K>>,
    engine: Engine,
    cache: BlockCache,
    compaction_threshold: f64,
    shard_prefix_bits: usize,
    reversed: bool,
    hook: H,
}

impl<K: Key> VarTree<K> {
    /// Open or create a `VarTree` at the given path.
    /// Recovers the index from existing data files on disk.
    pub fn open(path: impl AsRef<std::path::Path>, config: Config) -> DbResult<Self> {
        Self::open_inner(path, config, NoHook)
    }
}

impl<K: Key, H: WriteHook<K>> VarTree<K, H> {
    /// Open or create a `VarTree` with a write hook for secondary index maintenance.
    pub fn open_hooked(
        path: impl AsRef<std::path::Path>,
        config: Config,
        hook: H,
    ) -> DbResult<Self> {
        Self::open_inner(path, config, hook)
    }

    fn open_inner(path: impl AsRef<std::path::Path>, config: Config, hook: H) -> DbResult<Self> {
        let compaction_threshold = config.compaction_threshold;
        let shard_prefix_bits = config.shard_prefix_bits;
        let reversed = config.reversed;
        let cache = BlockCache::new(&config.cache);
        let engine = Engine::open(path, config)?;

        let tree = Self {
            index: SkipList::new(reversed),
            engine,
            cache,
            compaction_threshold,
            shard_prefix_bits,
            reversed,
            hook,
        };

        // Recover index from disk
        let shard_dirs = tree.engine.shard_dirs();
        let shard_dir_refs = Engine::shard_dir_refs(&shard_dirs);
        let shard_ids = tree.engine.shard_ids();

        let hints = tree.engine.hints();
        let max_gsn = recover_var_tree::<K>(
            &shard_dir_refs,
            &shard_ids,
            tree.index(),
            hints,
            #[cfg(feature = "encryption")]
            tree.engine.cipher(),
        )?;

        tree.engine
            .gsn()
            .fetch_max(max_gsn + 1, std::sync::atomic::Ordering::Relaxed);
        if hints {
            for shard in tree.engine.shards().iter() {
                shard.set_key_len(size_of::<K>());
            }
        }
        tracing::info!(
            key_size = size_of::<K>(),
            entries = tree.len(),
            "var_tree recovered"
        );

        Ok(tree)
    }

    /// Graceful shutdown: write hint files (if enabled), flush write buffers + fsync.
    pub fn close(self) -> DbResult<()> {
        if self.engine.hints() {
            self.sync_hints()?;
        }
        self.engine.flush()
    }

    /// Flush all shard write buffers to disk (without fsync).
    pub fn flush_buffers(&self) -> DbResult<()> {
        self.engine.flush_buffers()
    }

    /// Get the database configuration.
    pub fn config(&self) -> &Config {
        self.engine.config()
    }
}

impl<K: Key, H: WriteHook<K>> CompactionIndex<K> for VarTree<K, H> {
    fn update_if_match(&self, key: &K, old_loc: DiskLoc, new_loc: DiskLoc) -> bool {
        let guard = self.index.collector().enter();
        if let Some(node) = self.index.get(key.as_bytes(), &guard) {
            let current_ptr = node.load_disk_ptr();
            let current_disk = unsafe { *current_ptr };
            if current_disk == old_loc {
                let new_disk_ptr = Box::into_raw(Box::new(new_loc));
                match node.compare_exchange_disk(current_ptr, new_disk_ptr) {
                    Ok(old_ptr) => {
                        unsafe {
                            self.index
                                .collector()
                                .retire(old_ptr, seize::reclaim::boxed::<DiskLoc>);
                        }
                        return true;
                    }
                    Err(_) => {
                        // Concurrent put changed the DiskLoc — entry is no longer
                        // at old_loc, treat as dead in the compacted file.
                        unsafe {
                            drop(Box::from_raw(new_disk_ptr));
                        }
                        return false;
                    }
                }
            }
        }
        false
    }

    fn invalidate_blocks(&self, shard_id: u8, file_id: u32, total_bytes: u64) {
        self.cache.invalidate_file(shard_id, file_id, total_bytes);
    }

    fn contains_key(&self, key: &K) -> bool {
        self.contains(key)
    }
}

impl<K: Key, H: WriteHook<K>> VarTree<K, H> {
    /// Trigger a background compaction pass across all shards.
    pub fn compact(&self) -> DbResult<usize> {
        let mut total_compacted = 0;
        for shard in self.engine.shards().iter() {
            total_compacted += compact_shard(shard, self, self.compaction_threshold)?;
        }
        Ok(total_compacted)
    }

    /// Get a value by key. Checks block cache first, then reads from disk.
    pub fn get(&self, key: &K) -> Option<ByteView> {
        metrics::counter!("armdb.ops", "op" => "get", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.get");
        let guard = self.index.collector().enter();
        let node = match self.index.get(key.as_bytes(), &guard) {
            Some(n) => n,
            None => {
                #[cfg(feature = "hot-path-tracing")]
                tracing::error!(
                    "VarTree get error: index.get returned None for key {:?}",
                    key
                );
                return None;
            }
        };
        let disk = *node.load_disk();
        self.read_value_cached(&disk)
    }

    /// Get a value by key, returning `Err(KeyNotFound)` if absent.
    pub fn get_or_err(&self, key: &K) -> DbResult<ByteView> {
        self.get(key).ok_or(DbError::KeyNotFound)
    }

    /// Insert or update a key-value pair.
    pub fn put(&self, key: &K, value: &[u8]) -> DbResult<()> {
        metrics::counter!("armdb.ops", "op" => "put", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.put");
        let shard_id = self.shard_for(key);
        let mut inner = self.engine.shards()[shard_id].lock();
        let guard = self.index.collector().enter();
        let old_value = if H::NEEDS_OLD_VALUE {
            if let Some(node) = self.index.get(key.as_bytes(), &guard) {
                let disk = *node.load_disk();
                self.read_value_locked(&disk, &inner)
            } else {
                None
            }
        } else {
            None
        };
        self.put_locked(shard_id, &mut inner, &guard, key, value)?;
        self.hook.on_write(key, old_value.as_deref(), Some(value));
        Ok(())
    }

    /// Insert a key-value pair only if the key does not exist.
    /// Returns `Err(KeyExists)` if the key is already present.
    pub fn insert(&self, key: &K, value: &[u8]) -> DbResult<()> {
        metrics::counter!("armdb.ops", "op" => "insert", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.insert");
        let shard_id = self.shard_for(key);
        let mut inner = self.engine.shards()[shard_id].lock();
        let guard = self.index.collector().enter();
        self.insert_locked(shard_id, &mut inner, &guard, key, value)?;
        self.hook.on_write(key, None, Some(value));
        Ok(())
    }

    /// Delete a key. Returns `true` if the key existed.
    pub fn delete(&self, key: &K) -> DbResult<bool> {
        metrics::counter!("armdb.ops", "op" => "delete", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.delete");
        let shard_id = self.shard_for(key);
        let mut inner = self.engine.shards()[shard_id].lock();
        let guard = self.index.collector().enter();
        let old_value = if H::NEEDS_OLD_VALUE {
            if let Some(node) = self.index.get(key.as_bytes(), &guard) {
                let disk = *node.load_disk();
                self.read_value_locked(&disk, &inner)
            } else {
                None
            }
        } else {
            None
        };
        let existed = self.delete_locked(shard_id, &mut inner, &guard, key)?;
        if existed {
            self.hook.on_write(key, old_value.as_deref(), None);
        }
        Ok(existed)
    }

    /// Atomically execute multiple operations on a single shard.
    /// All keys must route to the same shard as `shard_key`.
    /// The closure must be short — shard lock is held for its duration.
    pub fn atomic<R>(
        &self,
        shard_key: &K,
        f: impl FnOnce(&mut VarShard<'_, K, H>) -> DbResult<R>,
    ) -> DbResult<R> {
        let shard_id = self.shard_for(shard_key);
        let inner = self.engine.shards()[shard_id].lock();
        let guard = self.index.collector().enter();
        let mut shard = VarShard {
            tree: self,
            inner,
            shard_id,
            guard,
        };
        f(&mut shard)
    }

    fn put_locked(
        &self,
        shard_id: usize,
        inner: &mut ShardInner,
        guard: &seize::LocalGuard<'_>,
        key: &K,
        value: &[u8],
    ) -> DbResult<()> {
        let (disk_loc, _gsn) = inner.append_entry(shard_id as u8, key.as_bytes(), value, false)?;

        // Fast path: key exists — atomic swap, no write_lock, no node allocation
        if let Some(existing) = self.index.get(key.as_bytes(), guard) {
            let new_disk = Box::into_raw(Box::new(disk_loc));
            let old_disk_ptr = existing.swap_disk(new_disk);
            let old_disk = unsafe { *old_disk_ptr };
            inner.add_dead_bytes(
                old_disk.file_id as u32,
                crate::entry::entry_size(size_of::<K>(), old_disk.len),
            );
            unsafe {
                self.index
                    .collector()
                    .retire(old_disk_ptr, seize::reclaim::boxed::<DiskLoc>);
            }
            return Ok(());
        }

        // Slow path: new key — allocate node + take write_lock via insert
        let height = random_height();
        let node_ptr = VarNode::alloc(*key, disk_loc, height);

        match self.index.insert(node_ptr, guard) {
            InsertResult::Inserted => {}
            InsertResult::Exists(existing) => {
                // Race: another shard inserted same key between get and insert
                let new_disk = Box::into_raw(Box::new(disk_loc));
                let old_disk_ptr = existing.swap_disk(new_disk);
                let old_disk = unsafe { *old_disk_ptr };
                inner.add_dead_bytes(
                    old_disk.file_id as u32,
                    crate::entry::entry_size(size_of::<K>(), old_disk.len),
                );
                unsafe {
                    self.index
                        .collector()
                        .retire(old_disk_ptr, seize::reclaim::boxed::<DiskLoc>);
                }
                unsafe {
                    (*node_ptr)
                        .disk
                        .store(std::ptr::null_mut(), std::sync::atomic::Ordering::Relaxed);
                    VarNode::<K>::dealloc_node(node_ptr);
                }
            }
        }

        Ok(())
    }

    fn insert_locked(
        &self,
        shard_id: usize,
        inner: &mut ShardInner,
        guard: &seize::LocalGuard<'_>,
        key: &K,
        value: &[u8],
    ) -> DbResult<()> {
        if self.index.get(key.as_bytes(), guard).is_some() {
            return Err(DbError::KeyExists);
        }

        let (disk_loc, _gsn) = inner.append_entry(shard_id as u8, key.as_bytes(), value, false)?;
        let height = random_height();
        let node_ptr = VarNode::alloc(*key, disk_loc, height);
        self.index.insert(node_ptr, guard);
        Ok(())
    }

    fn delete_locked(
        &self,
        shard_id: usize,
        inner: &mut ShardInner,
        guard: &seize::LocalGuard<'_>,
        key: &K,
    ) -> DbResult<bool> {
        if self.index.get(key.as_bytes(), guard).is_none() {
            return Ok(false);
        }

        inner.append_entry(shard_id as u8, key.as_bytes(), &[], true)?;

        let removed = self.index.remove(key.as_bytes(), guard);

        if let Some(node_ptr) = removed {
            let disk = *unsafe { &*node_ptr }.load_disk();
            inner.add_dead_bytes(
                disk.file_id as u32,
                crate::entry::entry_size(size_of::<K>(), disk.len),
            );
        }

        Ok(removed.is_some())
    }

    /// Check if a key exists.
    pub fn contains(&self, key: &K) -> bool {
        let guard = self.index.collector().enter();
        self.index.get(key.as_bytes(), &guard).is_some()
    }

    /// Return the first entry in index order, or `None` if empty.
    /// With `reversed=true` (default): the entry with the largest key.
    /// O(1) index lookup. May perform disk I/O on block-cache miss.
    pub fn first(&self) -> Option<(K, ByteView)> {
        let guard = self.index.collector().enter();
        let mut ptr = crate::skiplist::strip_mark(unsafe {
            (*self.index.head_ptr())
                .tower(0)
                .load(std::sync::atomic::Ordering::Acquire)
        });
        while !ptr.is_null() {
            let node = unsafe { &*ptr };
            if !node.is_marked() {
                let disk = *node.load_disk();
                let _ = guard;
                return self.read_value_cached(&disk).map(|v| (node.key, v));
            }
            ptr = crate::skiplist::strip_mark(
                node.tower(0).load(std::sync::atomic::Ordering::Acquire),
            );
        }
        None
    }

    /// Return the last entry in index order, or `None` if empty.
    /// With `reversed=true` (default): the entry with the smallest key.
    /// May perform disk I/O on block-cache miss.
    pub fn last(&self) -> Option<(K, ByteView)> {
        self.iter().next_back()
    }

    // -- Range helpers (front/back pointer positioning) -------------------------

    fn resolve_front_asc(
        &self,
        bound: &Bound<&K>,
        guard: &seize::LocalGuard<'_>,
    ) -> *mut VarNode<K> {
        match bound {
            Bound::Included(k) => self.index.find_first_ge(k.as_bytes(), guard),
            Bound::Excluded(k) => {
                let ge = self.index.find_first_ge(k.as_bytes(), guard);
                if !ge.is_null()
                    && !unsafe { &*ge }.is_marked()
                    && unsafe { &*ge }.key_bytes() == k.as_bytes()
                {
                    crate::skiplist::strip_mark(unsafe {
                        (*ge).tower(0).load(std::sync::atomic::Ordering::Acquire)
                    })
                } else {
                    ge
                }
            }
            Bound::Unbounded => crate::skiplist::strip_mark(unsafe {
                (*self.index.head_ptr())
                    .tower(0)
                    .load(std::sync::atomic::Ordering::Acquire)
            }),
        }
    }

    fn resolve_front_rev(
        &self,
        bound: &Bound<&K>,
        guard: &seize::LocalGuard<'_>,
    ) -> *mut VarNode<K> {
        match bound {
            Bound::Included(k) => self.index.find_first_ge(k.as_bytes(), guard),
            Bound::Excluded(k) => {
                let ge = self.index.find_first_ge(k.as_bytes(), guard);
                if !ge.is_null()
                    && !unsafe { &*ge }.is_marked()
                    && unsafe { &*ge }.key_bytes() == k.as_bytes()
                {
                    crate::skiplist::strip_mark(unsafe {
                        (*ge).tower(0).load(std::sync::atomic::Ordering::Acquire)
                    })
                } else {
                    ge
                }
            }
            Bound::Unbounded => crate::skiplist::strip_mark(unsafe {
                (*self.index.head_ptr())
                    .tower(0)
                    .load(std::sync::atomic::Ordering::Acquire)
            }),
        }
    }

    fn prefix_bounds(&self, prefix: &[u8]) -> (K, Bound<K>) {
        if self.reversed {
            let mut search = K::zeroed();
            search.as_bytes_mut().fill(0xFF);
            search.as_bytes_mut()[..prefix.len()].copy_from_slice(prefix);
            let mut end_key = K::zeroed();
            end_key.as_bytes_mut()[..prefix.len()].copy_from_slice(prefix);
            (search, Bound::Included(end_key))
        } else {
            let mut search = K::zeroed();
            search.as_bytes_mut()[..prefix.len()].copy_from_slice(prefix);
            let end = prefix_to_end_bound::<K>(prefix);
            (search, end)
        }
    }

    /// Iterate entries whose keys start with `prefix`.
    ///
    /// `reversed=true` (default): yields matching keys in DESC order.
    /// `next()` is O(1), `next_back()` is O(log n). Both may perform disk I/O on cache miss.
    pub fn prefix_iter(&self, prefix: &[u8]) -> VarIter<'_, K, H> {
        let guard = self.index.collector().enter();
        let (search_key, end) = self.prefix_bounds(prefix);
        let front = self.index.find_first_ge(search_key.as_bytes(), &guard);
        VarIter {
            tree: self,
            front,
            back: None,
            end,
            start: Bound::Included(search_key),
            reversed: self.reversed,
            done: false,
            _guard: guard,
        }
    }

    /// Iterate all entries in index order.
    ///
    /// `reversed=true` (default): DESC. `reversed=false`: ASC.
    /// `next()` is O(1), `next_back()` is O(log n). Both may perform disk I/O on cache miss.
    pub fn iter(&self) -> VarIter<'_, K, H> {
        let guard = self.index.collector().enter();
        let front = crate::skiplist::strip_mark(unsafe {
            (*self.index.head_ptr())
                .tower(0)
                .load(std::sync::atomic::Ordering::Acquire)
        });
        VarIter {
            tree: self,
            front,
            back: None,
            end: Bound::Unbounded,
            start: Bound::Unbounded,
            reversed: self.reversed,
            done: false,
            _guard: guard,
        }
    }

    /// Iterate entries in `[start, end)` — start inclusive, end exclusive.
    ///
    /// `reversed=true` (default): DESC within range. `reversed=false`: ASC.
    /// `next()` is O(1), `next_back()` is O(log n). Both may perform disk I/O on cache miss.
    pub fn range(&self, start: &K, end: &K) -> VarIter<'_, K, H> {
        self.range_bounds(Bound::Included(start), Bound::Excluded(end))
    }

    /// Iterate entries in range defined by `start` and `end` bounds.
    ///
    /// Unlike [`range()`](Self::range), allows `Included`, `Excluded`, or `Unbounded`
    /// for each bound independently.
    ///
    /// `reversed=true` (default): DESC within range. `reversed=false`: ASC.
    /// `next()` is O(1), `next_back()` is O(log n). Both may perform disk I/O on cache miss.
    pub fn range_bounds(&self, start: Bound<&K>, end: Bound<&K>) -> VarIter<'_, K, H> {
        let guard = self.index.collector().enter();
        if self.reversed {
            let front = self.resolve_front_rev(&end, &guard);
            VarIter {
                tree: self,
                front,
                back: None,
                end: bound_owned(&start),
                start: bound_owned(&end),
                reversed: true,
                done: false,
                _guard: guard,
            }
        } else {
            let front = self.resolve_front_asc(&start, &guard);
            VarIter {
                tree: self,
                front,
                back: None,
                end: bound_owned(&end),
                start: bound_owned(&start),
                reversed: false,
                done: false,
                _guard: guard,
            }
        }
    }

    pub fn len(&self) -> usize {
        self.index.len()
    }

    pub fn is_empty(&self) -> bool {
        self.index.is_empty()
    }

    /// Write hint files for all active shard files. Call during graceful shutdown.
    pub fn sync_hints(&self) -> DbResult<()> {
        for shard in self.engine.shards().iter() {
            shard.write_active_hint(size_of::<K>())?;
        }
        Ok(())
    }

    /// Pre-populate the block cache with blocks containing live values.
    ///
    /// Walks the index to collect unique block offsets, sorts them for sequential I/O,
    /// then reads each block into the cache. Only blocks with live data are loaded —
    /// dead entries and garbage are skipped.
    pub fn warmup(&self) -> DbResult<()> {
        use std::collections::BTreeSet;

        let guard = self.index.collector().enter();

        // Collect unique (shard_id, file_id, block_offset) from live index entries
        let mut blocks: BTreeSet<(u8, u32, u64)> = BTreeSet::new();
        let mut current = crate::skiplist::strip_mark(unsafe {
            (*self.index.head_ptr())
                .tower(0)
                .load(std::sync::atomic::Ordering::Acquire)
        });
        while !current.is_null() {
            let node = unsafe { &*current };
            current = crate::skiplist::strip_mark(
                node.tower(0).load(std::sync::atomic::Ordering::Acquire),
            );
            if node.is_marked() {
                continue;
            }
            let disk = *node.load_disk();
            let block_offset = disk.offset as u64 & !4095;
            blocks.insert((disk.shard_id, disk.file_id as u32, block_offset));
        }
        drop(guard);

        // Read blocks in sorted order (sequential I/O per shard/file)
        for (shard_id, file_id, block_offset) in &blocks {
            let key = BlockKey {
                shard_id: *shard_id,
                file_id: *file_id,
                block_offset: *block_offset,
            };
            if self.cache.get(&key).is_some() {
                continue;
            }
            let shard = &self.engine.shards()[*shard_id as usize];
            let (buf, _) = shard.read_block(*file_id, *block_offset)?;
            self.cache.insert(key, Arc::new(buf));
        }

        Ok(())
    }

    pub(crate) fn index(&self) -> &SkipList<VarNode<K>> {
        &self.index
    }

    /// Iterate all entries and optionally mutate them. Call once at startup.
    ///
    /// The callback receives each (key, value_bytes) and returns `MigrateAction`.
    /// Block cache is populated naturally — no need to call `warmup()` separately.
    /// Hook is called for every entry, including `Keep`.
    ///
    /// Returns the number of mutated entries.
    pub fn migrate(
        &self,
        f: impl Fn(&K, &[u8]) -> crate::MigrateAction<ByteView>,
    ) -> DbResult<usize> {
        use crate::MigrateAction;

        let _guard = self.index.collector().enter();
        let mut current = crate::skiplist::strip_mark(unsafe {
            (*self.index.head_ptr())
                .tower(0)
                .load(std::sync::atomic::Ordering::Acquire)
        });
        let mut count = 0;
        while !current.is_null() {
            let node = unsafe { &*current };
            current = crate::skiplist::strip_mark(
                node.tower(0).load(std::sync::atomic::Ordering::Acquire),
            );
            if node.is_marked() {
                continue;
            }
            let disk = *node.load_disk();
            let value = match self.read_value_cached(&disk) {
                Some(v) => v,
                None => continue,
            };
            match f(&node.key, &value) {
                MigrateAction::Keep => {
                    self.hook.on_write(&node.key, None, Some(&value));
                }
                MigrateAction::Update(new_value) => {
                    self.put(&node.key, &new_value)?;
                    count += 1;
                }
                MigrateAction::Delete => {
                    self.delete(&node.key)?;
                    count += 1;
                }
            }
        }

        tracing::info!(mutations = count, "var_tree migration complete");
        Ok(count)
    }

    /// Read a value via block cache. Checks write buffer for unflushed data.
    /// Returns None on I/O error.
    fn read_value_cached(&self, disk: &DiskLoc) -> Option<ByteView> {
        let block_offset = disk.offset as u64 & !4095;
        let start = (disk.offset & 4095) as usize;
        let len = disk.len as usize;
        let cache_key = BlockKey {
            shard_id: disk.shard_id,
            file_id: disk.file_id as u32,
            block_offset,
        };

        // 1. Block Cache (lock-free)
        if let Some(block) = self.cache.get(&cache_key) {
            metrics::counter!("armdb.cache.hit").increment(1);
            return Self::extract_from_block(&block, start, len, || {
                self.get_or_read_block(disk.shard_id, disk.file_id as u32, block_offset + 4096)
            });
        }

        // 2. Write buffer — for unflushed data in active file (brief shard lock)
        {
            let shard = &self.engine.shards()[disk.shard_id as usize];
            let inner = shard.lock();
            if inner.active.file_id == disk.file_id as u32
                && let Some(bytes) = inner.write_buf.read(disk.offset as u64, len)
            {
                return Some(ByteView::new(bytes));
            }
        }

        // 3. Disk read + cache
        metrics::counter!("armdb.cache.miss").increment(1);
        let block = self.get_or_read_block(disk.shard_id, disk.file_id as u32, block_offset)?;
        Self::extract_from_block(&block, start, len, || {
            self.get_or_read_block(disk.shard_id, disk.file_id as u32, block_offset + 4096)
        })
    }

    fn extract_from_block(
        block: &AlignedBuf,
        start: usize,
        len: usize,
        next_block: impl FnOnce() -> Option<Arc<AlignedBuf>>,
    ) -> Option<ByteView> {
        if start + len <= 4096 {
            Some(ByteView::new(&block[start..start + len]))
        } else {
            let next = next_block()?;
            let first_part = &block[start..];
            let second_len = len - first_part.len();
            let mut combined = Vec::with_capacity(len);
            combined.extend_from_slice(first_part);
            combined.extend_from_slice(&next[..second_len]);
            Some(ByteView::from_vec(combined))
        }
    }

    /// Get a block from cache, or read it from disk and cache it.
    fn get_or_read_block(
        &self,
        shard_id: u8,
        file_id: u32,
        block_offset: u64,
    ) -> Option<Arc<AlignedBuf>> {
        let key = BlockKey {
            shard_id,
            file_id,
            block_offset,
        };
        if let Some(cached) = self.cache.get(&key) {
            return Some(cached);
        }
        let shard = &self.engine.shards()[shard_id as usize];
        let (buf, is_full_block) = match shard.read_block(file_id, block_offset) {
            Ok(b) => b,
            Err(e) => {
                #[cfg(feature = "hot-path-tracing")]
                tracing::error!("VarTree block read error: {:?}", e);
                let _ = e;
                return None;
            }
        };
        let arc = Arc::new(buf);
        // Only cache full blocks (entirely within the file's data region).
        // Partial blocks at the end of a file have zero-padded tails that
        // would become stale after file rotation or further writes.
        if is_full_block {
            self.cache.insert(key, arc.clone());
        }
        Some(arc)
    }

    /// Compare-and-swap: if current value == expected, replace with new_value.
    /// Returns `Ok(())` on success, `Err(CasMismatch)` if current != expected,
    /// `Err(KeyNotFound)` if key doesn't exist.
    ///
    /// **Caveat:** Holds the shard lock while reading the current value. On a
    /// block-cache miss this performs disk I/O under the lock, blocking all
    /// writes to the same shard. Pre-warm the cache to avoid latency spikes.
    pub fn cas(&self, key: &K, expected: &[u8], new_value: &[u8]) -> DbResult<()> {
        metrics::counter!("armdb.ops", "op" => "cas", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.cas");
        let shard_id = self.shard_for(key);
        let shard = &self.engine.shards()[shard_id];
        let mut inner = shard.lock();

        let guard = self.index.collector().enter();
        let existing = self
            .index
            .get(key.as_bytes(), &guard)
            .ok_or(DbError::KeyNotFound)?;

        let disk = *existing.load_disk();
        let current = self
            .read_value_locked(&disk, &inner)
            .ok_or(DbError::KeyNotFound)?;
        if current.as_ref() != expected {
            return Err(DbError::CasMismatch);
        }

        let (new_disk_loc, _gsn) =
            inner.append_entry(shard_id as u8, key.as_bytes(), new_value, false)?;

        let new_disk = Box::into_raw(Box::new(new_disk_loc));
        let old_disk_ptr = existing.swap_disk(new_disk);
        let old_disk = unsafe { *old_disk_ptr };
        inner.add_dead_bytes(
            old_disk.file_id as u32,
            crate::entry::entry_size(size_of::<K>(), old_disk.len),
        );
        unsafe {
            self.index
                .collector()
                .retire(old_disk_ptr, seize::reclaim::boxed::<DiskLoc>);
        }

        self.hook.on_write(key, Some(&*current), Some(new_value));
        Ok(())
    }

    /// Atomically read-modify-write. Returns `Some(new_value)` if key existed, `None` otherwise.
    /// The closure receives the current value and returns a new ByteView.
    /// The closure must not be heavy (shard lock is held).
    ///
    /// **Caveat:** Holds the shard lock while reading the current value. On a
    /// block-cache miss this performs disk I/O under the lock, blocking all
    /// writes to the same shard. Pre-warm the cache to avoid latency spikes.
    pub fn update(&self, key: &K, f: impl FnOnce(&[u8]) -> ByteView) -> DbResult<Option<ByteView>> {
        self.update_inner(key, f, false)
    }

    /// Like [`update()`](Self::update), but returns `Some(old_value)` instead of the new one.
    pub fn fetch_update(
        &self,
        key: &K,
        f: impl FnOnce(&[u8]) -> ByteView,
    ) -> DbResult<Option<ByteView>> {
        self.update_inner(key, f, true)
    }

    fn update_inner(
        &self,
        key: &K,
        f: impl FnOnce(&[u8]) -> ByteView,
        return_old: bool,
    ) -> DbResult<Option<ByteView>> {
        metrics::counter!("armdb.ops", "op" => "update", "tree" => "var_tree").increment(1);
        #[cfg(feature = "hot-path-tracing")]
        tracing::trace!("var_tree.update");
        let shard_id = self.shard_for(key);
        let shard = &self.engine.shards()[shard_id];
        let mut inner = shard.lock();

        let guard = self.index.collector().enter();
        let existing = match self.index.get(key.as_bytes(), &guard) {
            Some(n) => n,
            None => return Ok(None),
        };

        let disk = *existing.load_disk();
        let current = match self.read_value_locked(&disk, &inner) {
            Some(v) => v,
            None => return Ok(None),
        };

        let new_value = f(&current);

        let (new_disk_loc, _gsn) =
            inner.append_entry(shard_id as u8, key.as_bytes(), &new_value, false)?;

        let new_disk = Box::into_raw(Box::new(new_disk_loc));
        let old_disk_ptr = existing.swap_disk(new_disk);
        let old_disk = unsafe { *old_disk_ptr };
        inner.add_dead_bytes(
            old_disk.file_id as u32,
            crate::entry::entry_size(size_of::<K>(), old_disk.len),
        );
        unsafe {
            self.index
                .collector()
                .retire(old_disk_ptr, seize::reclaim::boxed::<DiskLoc>);
        }

        self.hook.on_write(key, Some(&*current), Some(&*new_value));
        Ok(Some(if return_old { current } else { new_value }))
    }

    /// Read value when shard lock is already held. Used by CAS/update.
    fn read_value_locked(&self, disk: &DiskLoc, inner: &ShardInner) -> Option<ByteView> {
        let len = disk.len as usize;

        // 1. Write buffer (for unflushed data)
        if inner.active.file_id == disk.file_id as u32
            && let Some(bytes) = inner.write_buf.read(disk.offset as u64, len)
        {
            return Some(ByteView::new(bytes));
        }

        // 2. Block cache (lock-free)
        let block_offset = disk.offset as u64 & !4095;
        let start = (disk.offset & 4095) as usize;
        if start + len <= 4096 {
            let cache_key = BlockKey {
                shard_id: disk.shard_id,
                file_id: disk.file_id as u32,
                block_offset,
            };
            if let Some(block) = self.cache.get(&cache_key) {
                return Some(ByteView::new(&block[start..start + len]));
            }
        }

        // 3. Disk read (file handle from inner, no re-lock)
        let file: &std::fs::File = if inner.active.file_id == disk.file_id as u32 {
            &inner.active.read_file
        } else {
            let imm = inner
                .immutable
                .iter()
                .find(|f| f.file_id == disk.file_id as u32)?;
            &imm.file
        };
        let bytes = direct::pread_value(file, disk.offset as u64, len).ok()?;
        Some(ByteView::new(&bytes))
    }

    pub fn shard_for(&self, key: &K) -> usize {
        if self.shard_prefix_bits == 0 || self.shard_prefix_bits >= size_of::<K>() * 8 {
            let hash = xxhash_rust::xxh3::xxh3_64(key.as_bytes());
            return (hash as usize) % self.engine.shards().len();
        }

        let full_bytes = self.shard_prefix_bits / 8;
        let extra_bits = self.shard_prefix_bits % 8;

        let hash = if extra_bits == 0 {
            xxhash_rust::xxh3::xxh3_64(&key.as_bytes()[..full_bytes])
        } else {
            let mut buf = K::zeroed();
            buf.as_bytes_mut()[..full_bytes + 1].copy_from_slice(&key.as_bytes()[..full_bytes + 1]);
            let mask = !((1u8 << (8 - extra_bits)) - 1);
            buf.as_bytes_mut()[full_bytes] = key.as_bytes()[full_bytes] & mask;
            xxhash_rust::xxh3::xxh3_64(&buf.as_bytes()[..full_bytes + 1])
        };

        (hash as usize) % self.engine.shards().len()
    }
}

#[cfg(feature = "replication")]
impl<K: Key, H: WriteHook<K>> crate::replication::ReplicationTarget for VarTree<K, H> {
    fn apply_entry(
        &self,
        shard_id: u8,
        file_id: u32,
        entry_offset: u64,
        header: &crate::entry::EntryHeader,
        key: &[u8],
        _value: &[u8],
    ) -> DbResult<()> {
        let key: K = K::from_bytes(key);

        let value_offset =
            entry_offset + size_of::<crate::entry::EntryHeader>() as u64 + size_of::<K>() as u64;
        let disk = DiskLoc::new(
            shard_id,
            file_id as u16,
            value_offset as u32,
            header.value_len,
        );

        if header.is_tombstone() {
            let guard = self.index.collector().enter();
            self.index.remove(key.as_bytes(), &guard);
        } else {
            let guard = self.index.collector().enter();
            let height = random_height();
            let node_ptr = VarNode::alloc(key, disk, height);
            match self.index.insert(node_ptr, &guard) {
                InsertResult::Inserted => {}
                InsertResult::Exists(existing) => {
                    let new_disk_ptr = Box::into_raw(Box::new(disk));
                    let old_disk_ptr = existing.swap_disk(new_disk_ptr);
                    unsafe {
                        self.index
                            .collector()
                            .retire(old_disk_ptr, seize::reclaim::boxed::<DiskLoc>);
                    }
                    // Deallocate the unused node
                    unsafe {
                        (*node_ptr)
                            .disk
                            .store(std::ptr::null_mut(), std::sync::atomic::Ordering::Relaxed);
                        VarNode::<K>::dealloc_node(node_ptr);
                    }
                }
            }
        }

        Ok(())
    }

    fn try_apply_entry(
        &self,
        shard_id: u8,
        file_id: u32,
        entry_offset: u64,
        header: &crate::entry::EntryHeader,
        raw_after_header: &[u8],
    ) -> DbResult<bool> {
        if raw_after_header.len() < size_of::<K>() + header.value_len as usize {
            return Ok(false);
        }
        let key = &raw_after_header[..size_of::<K>()];
        let value = &raw_after_header[size_of::<K>()..size_of::<K>() + header.value_len as usize];
        let crc = crate::entry::compute_crc32(header.gsn, header.value_len, key, value);
        if crc != header.crc32 {
            return Ok(false);
        }
        self.apply_entry(shard_id, file_id, entry_offset, header, key, value)?;
        Ok(true)
    }

    fn key_len(&self) -> usize {
        size_of::<K>()
    }
}

/// Handle for atomic multi-key operations within a single shard.
/// Obtained via [`VarTree::atomic`]. The shard lock is held for the
/// lifetime of this struct — keep the closure short.
pub struct VarShard<'a, K: Key, H: WriteHook<K> = NoHook> {
    tree: &'a VarTree<K, H>,
    inner: MutexGuard<'a, ShardInner>,
    shard_id: usize,
    guard: seize::LocalGuard<'a>,
}

impl<K: Key, H: WriteHook<K>> VarShard<'_, K, H> {
    pub fn put(&mut self, key: &K, value: &[u8]) -> DbResult<()> {
        self.check_shard(key)?;
        self.tree
            .put_locked(self.shard_id, &mut self.inner, &self.guard, key, value)
    }

    pub fn insert(&mut self, key: &K, value: &[u8]) -> DbResult<()> {
        self.check_shard(key)?;
        self.tree
            .insert_locked(self.shard_id, &mut self.inner, &self.guard, key, value)
    }

    pub fn delete(&mut self, key: &K) -> DbResult<bool> {
        self.check_shard(key)?;
        self.tree
            .delete_locked(self.shard_id, &mut self.inner, &self.guard, key)
    }

    pub fn get(&self, key: &K) -> Option<ByteView> {
        let node = self.tree.index.get(key.as_bytes(), &self.guard)?;
        let disk = *node.load_disk();
        self.tree.read_value_locked(&disk, &self.inner)
    }

    pub fn get_or_err(&self, key: &K) -> DbResult<ByteView> {
        self.get(key).ok_or(DbError::KeyNotFound)
    }

    pub fn contains(&self, key: &K) -> bool {
        self.tree.index.get(key.as_bytes(), &self.guard).is_some()
    }

    fn check_shard(&self, key: &K) -> DbResult<()> {
        if self.tree.shard_for(key) != self.shard_id {
            return Err(DbError::ShardMismatch);
        }
        Ok(())
    }
}

fn bound_owned<K: Copy>(b: &Bound<&K>) -> Bound<K> {
    match b {
        Bound::Included(k) => Bound::Included(**k),
        Bound::Excluded(k) => Bound::Excluded(**k),
        Bound::Unbounded => Bound::Unbounded,
    }
}

fn prefix_to_end_bound<K: Key>(prefix: &[u8]) -> Bound<K> {
    let mut incremented = prefix.to_vec();
    let mut carry = true;
    for byte in incremented.iter_mut().rev() {
        if carry {
            if *byte == 0xFF {
                *byte = 0x00;
            } else {
                *byte += 1;
                carry = false;
                break;
            }
        }
    }
    if carry {
        Bound::Unbounded
    } else {
        let mut end = K::zeroed();
        end.as_bytes_mut()[..incremented.len()].copy_from_slice(&incremented);
        Bound::Excluded(end)
    }
}

/// Iterator over entries in a `VarTree`. Returned by `iter()`, `range()`, and `prefix_iter()`.
///
/// Weakly-consistent: concurrent inserts/updates may be visible during iteration.
/// Deleted entries are skipped. The `seize` guard prevents use-after-free.
/// Each `next()` may perform disk I/O on a block-cache miss.
pub struct VarIter<'a, K: Key, H: WriteHook<K> = NoHook> {
    tree: &'a VarTree<K, H>,
    front: *mut VarNode<K>,
    /// `None` = not yet resolved (lazy). Computed on first `next_back()` call.
    back: Option<*mut VarNode<K>>,
    end: Bound<K>,
    start: Bound<K>,
    reversed: bool,
    done: bool,
    _guard: seize::LocalGuard<'a>,
}

// SAFETY: VarIter holds a seize guard that prevents node reclamation.
// The raw pointer is only dereferenced while the guard is alive.
unsafe impl<K: Key, H: WriteHook<K>> Send for VarIter<'_, K, H> {}

impl<K: Key, H: WriteHook<K>> Iterator for VarIter<'_, K, H> {
    type Item = (K, ByteView);

    fn next(&mut self) -> Option<Self::Item> {
        loop {
            if self.done || self.front.is_null() {
                return None;
            }
            let node = unsafe { &*self.front };
            let converged = self.back.is_some_and(|back| std::ptr::eq(self.front, back));
            self.front = crate::skiplist::strip_mark(
                node.tower(0).load(std::sync::atomic::Ordering::Acquire),
            );
            if converged {
                self.done = true;
            }
            if node.is_marked() {
                if converged {
                    return None;
                }
                continue;
            }
            if !self.check_end(&node.key) {
                self.done = true;
                return None;
            }
            let disk = *node.load_disk();
            match self.tree.read_value_cached(&disk) {
                Some(value) => return Some((node.key, value)),
                None => {
                    if converged {
                        return None;
                    }
                    continue;
                }
            }
        }
    }
}

impl<K: Key, H: WriteHook<K>> DoubleEndedIterator for VarIter<'_, K, H> {
    fn next_back(&mut self) -> Option<Self::Item> {
        if self.back.is_none() {
            self.back = Some(self.resolve_back());
            if self.front.is_null() {
                self.done = true;
            }
        }
        loop {
            let back = self.back.unwrap_or(std::ptr::null_mut());
            if self.done || back.is_null() {
                return None;
            }
            let node = unsafe { &*back };
            let key = node.key;
            let converged = std::ptr::eq(self.front, back);
            self.back = Some(self.tree.index().find_last_lt(key.as_bytes()));
            if converged {
                self.done = true;
            }
            if node.is_marked() {
                if converged {
                    return None;
                }
                continue;
            }
            if !self.check_start(&key) {
                self.done = true;
                return None;
            }
            let disk = *node.load_disk();
            match self.tree.read_value_cached(&disk) {
                Some(value) => return Some((key, value)),
                None => {
                    if converged {
                        return None;
                    }
                    continue;
                }
            }
        }
    }
}

impl<K: Key, H: WriteHook<K>> VarIter<'_, K, H> {
    /// Lazily resolve the back pointer for DoubleEndedIterator.
    fn resolve_back(&self) -> *mut VarNode<K> {
        let index = self.tree.index();
        match &self.end {
            Bound::Unbounded => index.find_last(),
            Bound::Excluded(k) => index.find_last_lt(k.as_bytes()),
            Bound::Included(k) => {
                let ge = index.find_first_ge(k.as_bytes(), &self._guard);
                if !ge.is_null()
                    && !unsafe { &*ge }.is_marked()
                    && unsafe { &*ge }.key_bytes() == k.as_bytes()
                {
                    ge
                } else {
                    index.find_last_lt(k.as_bytes())
                }
            }
        }
    }

    #[inline(always)]
    fn check_end(&self, key: &K) -> bool {
        match &self.end {
            Bound::Unbounded => true,
            Bound::Excluded(end) => {
                if self.reversed {
                    key.as_bytes() > end.as_bytes()
                } else {
                    key.as_bytes() < end.as_bytes()
                }
            }
            Bound::Included(end) => {
                if self.reversed {
                    key.as_bytes() >= end.as_bytes()
                } else {
                    key.as_bytes() <= end.as_bytes()
                }
            }
        }
    }

    #[inline(always)]
    fn check_start(&self, key: &K) -> bool {
        match &self.start {
            Bound::Unbounded => true,
            Bound::Excluded(s) => {
                if self.reversed {
                    key.as_bytes() < s.as_bytes()
                } else {
                    key.as_bytes() > s.as_bytes()
                }
            }
            Bound::Included(s) => {
                if self.reversed {
                    key.as_bytes() <= s.as_bytes()
                } else {
                    key.as_bytes() >= s.as_bytes()
                }
            }
        }
    }
    /// Collect only the keys. Convenience for backward compatibility.
    pub fn collect_keys(&mut self) -> Vec<K> {
        self.map(|(k, _)| k).collect()
    }

    /// Collect all remaining entries. Convenience for backward compatibility.
    pub fn collect_entries(&mut self) -> Vec<(K, ByteView)> {
        self.collect()
    }
}