ic_storage/adapter/
tree.rs

1//! Owned degree-six B-tree. The multi-entry search, split/borrow/merge design and
2//! selective payload access follow the established B-tree techniques used by
3//! ic-stable-structures 0.7.2 (see docs/storage/upstreams.md). This is a new checked
4//! format, not an implementation of its durable layout or compatibility reader.
5//! Routing records contain small inline values; larger values have separate
6//! checksummed records. Mutations stage bytes and reserve before publication.
7use super::{
8    Memory, Storable,
9    error::{CollectionError as Error, Result, ensure},
10    record,
11};
12use serde::{Deserialize, Serialize};
13use std::{
14    borrow::Cow,
15    cell::{Cell, RefCell},
16    collections::BTreeMap,
17    marker::PhantomData,
18    ops::{Bound, RangeBounds},
19    rc::Rc,
20};
21mod read;
22use read::{NodeCache, ReadNode};
23
24pub const MAGIC: [u8; 8] = *b"ICMAP002";
25const NODE: [u8; 8] = *b"ICBTND01";
26const VALUE: [u8; 8] = *b"ICBTVAL1";
27const FREE: [u8; 8] = *b"ICFREE01";
28const BINS: usize = 32;
29const MAX_DEPTH: usize = 32;
30const DEGREE: usize = 6;
31const MAX_ENTRIES: usize = DEGREE * 2 - 1;
32const MIN_ENTRIES: usize = DEGREE - 1;
33const INLINE: usize = 32;
34
35#[derive(Clone, Copy, Debug, Eq, PartialEq, Serialize, Deserialize)]
36struct Pointer {
37    offset: u64,
38    order: u8,
39}
40impl Pointer {
41    fn capacity(self) -> u64 {
42        1_u64 << self.order
43    }
44    fn check(self, cursor: u64) -> Result<()> {
45        ensure(
46            (8..32).contains(&self.order),
47            Error::Corrupt("tree block order"),
48        )?;
49        ensure(
50            self.offset >= record::HEADER_BYTES
51                && self
52                    .offset
53                    .checked_add(self.capacity())
54                    .is_some_and(|end| end <= cursor),
55            Error::Corrupt("tree block bounds"),
56        )
57    }
58    fn put(self, bytes: &mut Vec<u8>) {
59        bytes.extend_from_slice(&self.offset.to_le_bytes());
60        bytes.push(self.order);
61    }
62}
63#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
64struct Root {
65    top: Option<Pointer>,
66    length: u64,
67    cursor: u64,
68    free: Vec<Option<Pointer>>,
69    key_bound: Option<(u32, bool)>,
70    value_bound: Option<(u32, bool)>,
71}
72#[derive(Clone)]
73enum Value {
74    Inline(Vec<u8>),
75    External(Pointer),
76}
77#[derive(Clone)]
78struct Entry {
79    key: Vec<u8>,
80    value: Value,
81}
82#[derive(Clone, Default)]
83struct Node {
84    entries: Vec<Entry>,
85    children: Vec<Pointer>,
86}
87impl Node {
88    fn encode(&self) -> Result<Vec<u8>> {
89        ensure(
90            self.entries.len() <= MAX_ENTRIES,
91            Error::Corrupt("B-tree node capacity"),
92        )?;
93        ensure(
94            self.children.is_empty() || self.children.len() == self.entries.len() + 1,
95            Error::Corrupt("B-tree child count"),
96        )?;
97        let capacity =
98            self.entries
99                .iter()
100                .try_fold(2 + self.children.len() * 9, |size, entry| {
101                    let value_size = match &entry.value {
102                        Value::Inline(v) => v.len(),
103                        Value::External(_) => 9,
104                    };
105                    size.checked_add(5)
106                        .and_then(|n| n.checked_add(entry.key.len()))
107                        .and_then(|n| n.checked_add(value_size))
108                        .ok_or(Error::Budget)
109                })?;
110        record::encode_payload_with(NODE, capacity, |bytes| {
111            bytes.extend_from_slice(&[
112                u8::try_from(self.entries.len()).unwrap(),
113                u8::from(!self.children.is_empty()),
114            ]);
115            for p in &self.children {
116                p.put(bytes);
117            }
118            for entry in &self.entries {
119                bytes.extend_from_slice(
120                    &u32::try_from(entry.key.len())
121                        .map_err(|_| Error::Budget)?
122                        .to_le_bytes(),
123                );
124                bytes.extend_from_slice(&entry.key);
125                match &entry.value {
126                    Value::Inline(value) => {
127                        bytes.push(u8::try_from(value.len()).unwrap());
128                        bytes.extend_from_slice(value);
129                    }
130                    Value::External(p) => {
131                        bytes.push(255);
132                        p.put(bytes);
133                    }
134                }
135            }
136            Ok(())
137        })
138    }
139}
140struct Reader<'a>(&'a [u8]);
141impl<'a> Reader<'a> {
142    fn take(&mut self, n: usize) -> Result<&'a [u8]> {
143        ensure(n <= self.0.len(), Error::Corrupt("truncated B-tree node"))?;
144        let (a, b) = self.0.split_at(n);
145        self.0 = b;
146        Ok(a)
147    }
148    fn byte(&mut self) -> Result<u8> {
149        Ok(self.take(1)?[0])
150    }
151    fn pointer(&mut self, cursor: u64) -> Result<Pointer> {
152        let offset = u64::from_le_bytes(self.take(8)?.try_into().unwrap());
153        let p = Pointer {
154            offset,
155            order: self.byte()?,
156        };
157        p.check(cursor)?;
158        Ok(p)
159    }
160}
161
162pub struct Tree<K, V, M: Memory> {
163    pub memory: M,
164    root: RefCell<Root>,
165    poisoned: Cell<bool>,
166    cache: RefCell<NodeCache>,
167    types: PhantomData<(K, V)>,
168}
169impl<K: Storable + Ord + Clone, V: Storable, M: Memory> Tree<K, V, M> {
170    pub fn reset(memory: M) -> Result<Self> {
171        let root = Root {
172            top: None,
173            length: 0,
174            cursor: record::HEADER_BYTES,
175            free: vec![None; BINS],
176            key_bound: record::bound::<K>(),
177            value_bound: record::bound::<V>(),
178        };
179        let bytes = record::encode(MAGIC, &root)?;
180        ensure(bytes.len() as u64 <= record::HEADER_BYTES, Error::Budget)?;
181        record::reserve(&memory, record::HEADER_BYTES)?;
182        memory.write(0, &bytes);
183        Ok(Self {
184            memory,
185            root: RefCell::new(root),
186            poisoned: Cell::new(false),
187            cache: RefCell::default(),
188            types: PhantomData,
189        })
190    }
191    pub fn load(memory: M) -> Result<Self> {
192        let root: Root = record::read(&memory, 0, record::HEADER_BYTES, MAGIC)?;
193        ensure(
194            root.key_bound == record::bound::<K>() && root.value_bound == record::bound::<V>(),
195            Error::Corrupt("map codec bound changed"),
196        )?;
197        ensure(
198            root.free.len() == BINS
199                && root.cursor >= record::HEADER_BYTES
200                && root.cursor
201                    <= memory
202                        .size()
203                        .checked_mul(crate::WASM_PAGE_SIZE_BYTES)
204                        .ok_or(Error::Budget)?,
205            Error::Corrupt("map allocator bounds"),
206        )?;
207        ensure(
208            root.top.is_none() == (root.length == 0),
209            Error::Corrupt("map root/count disagree"),
210        )?;
211        if let Some(p) = root.top {
212            p.check(root.cursor)?;
213        }
214        for (order, p) in root.free.iter().enumerate() {
215            if let Some(p) = p {
216                p.check(root.cursor)?;
217                ensure(
218                    usize::from(p.order) == order,
219                    Error::Corrupt("free block order"),
220                )?;
221            }
222        }
223        Ok(Self {
224            memory,
225            root: RefCell::new(root),
226            poisoned: Cell::new(false),
227            cache: RefCell::default(),
228            types: PhantomData,
229        })
230    }
231    fn healthy(&self) -> Result<()> {
232        ensure(!self.poisoned.get(), Error::Poisoned)?;
233        // A different allocation can poison the shared governed backing. This
234        // also covers iterator entries already held in a routing frame.
235        let capacity = self
236            .memory
237            .size()
238            .checked_mul(crate::WASM_PAGE_SIZE_BYTES)
239            .ok_or(Error::Budget)?;
240        ensure(
241            self.root.borrow().cursor <= capacity,
242            Error::Corrupt("map allocator bounds"),
243        )
244    }
245    pub fn len(&self) -> u64 {
246        self.healthy().expect("map owner is poisoned");
247        self.root.borrow().length
248    }
249    pub fn set_cache_budget(&mut self, bytes: usize) {
250        self.cache.get_mut().configure(bytes);
251    }
252    fn read_node(&self, p: Pointer, depth: usize) -> Result<Rc<ReadNode>> {
253        let cursor = self.root.borrow().cursor;
254        p.check(cursor)?;
255        if let Some(node) = self.cache.borrow_mut().get(p, depth) {
256            return Ok(node);
257        }
258        let node = Rc::new(ReadNode::decode::<K, V>(
259            record::read_bytes(&self.memory, p.offset, p.capacity(), NODE)?,
260            cursor,
261        )?);
262        self.cache.borrow_mut().put(p, Rc::clone(&node), depth);
263        Ok(node)
264    }
265    fn value(&self, v: Value) -> Result<Vec<u8>> {
266        let bytes = match v {
267            Value::Inline(bytes) => bytes,
268            Value::External(p) => {
269                p.check(self.root.borrow().cursor)?;
270                record::read_payload(&self.memory, p.offset, p.capacity(), VALUE)?
271            }
272        };
273        V::check_bounds(&bytes)?;
274        Ok(bytes)
275    }
276    fn find(&self, key: &K) -> Result<Option<(Rc<ReadNode>, usize)>> {
277        self.healthy()?;
278        let mut p = self.root.borrow().top;
279        for depth in 0..MAX_DEPTH {
280            let Some(current) = p else { return Ok(None) };
281            let node = self.read_node(current, depth)?;
282            match node.search(key)? {
283                Ok(i) => return Ok(Some((node, i))),
284                Err(i) => p = node.children.get(i).copied(),
285            }
286        }
287        Err(Error::Corrupt("map path exceeds height bound"))
288    }
289    pub fn get(&self, key: &K) -> Result<Option<V>> {
290        let Some((node, index)) = self.find(key)? else {
291            return Ok(None);
292        };
293        if let Some(bytes) = node.inline_value(index) {
294            V::check_bounds(bytes)?;
295            Ok(Some(V::from_bytes(Cow::Borrowed(bytes))?))
296        } else {
297            let value = node.value(index);
298            // External decoding does not borrow the routing buffer. Release it
299            // before allocating the payload, as in the original lookup path.
300            drop(node);
301            decode(self.value(value)?).map(Some)
302        }
303    }
304    pub fn contains_key(&self, key: &K) -> Result<bool> {
305        Ok(self.find(key)?.is_some())
306    }
307    pub fn insert(&mut self, key: K, value: V) -> Result<Option<V>> {
308        self.healthy()?;
309        let key_bytes = key.to_bytes_checked()?.into_owned();
310        let bytes = value.into_bytes_checked()?;
311        let mut tx = Transaction::new(self);
312        let (top, split, old) = tx.insert(tx.root.top, &key, key_bytes, bytes, 0)?;
313        tx.root.top = Some(if let Some((entry, right)) = split {
314            tx.save(
315                None,
316                &Node {
317                    entries: vec![entry],
318                    children: vec![top, right],
319                },
320            )?
321        } else {
322            top
323        });
324        drop(key);
325        let old = old.map(decode).transpose()?;
326        tx.commit()?;
327        Ok(old)
328    }
329    pub fn remove(&mut self, key: &K) -> Result<Option<V>> {
330        self.healthy()?;
331        let mut tx = Transaction::new(self);
332        let Some(top) = tx.root.top else {
333            return Ok(None);
334        };
335        let (changed, old) = tx.remove(top, key, 0)?;
336        let Some(old) = old else { return Ok(None) };
337        let value = decode(old)?;
338        if let Some(node) = changed {
339            tx.root.top = if node.entries.is_empty() {
340                tx.free(top)?;
341                node.children.first().copied()
342            } else {
343                Some(tx.save(Some(top), &node)?)
344            };
345        }
346        tx.root.length -= 1;
347        tx.commit()?;
348        Ok(Some(value))
349    }
350    pub fn clear(&mut self) -> Result<()> {
351        self.healthy()?;
352        let mut tx = Transaction::new(self);
353        tx.root.top = None;
354        tx.root.length = 0;
355        tx.root.cursor = record::HEADER_BYTES;
356        tx.root.free.fill(None);
357        tx.commit()
358    }
359    fn seek_path(&self, bound: &Bound<K>, forward: bool) -> Result<Vec<Frame>> {
360        self.healthy()?;
361        let mut path = Vec::new();
362        let mut pointer = self.root.borrow().top;
363        while let Some(p) = pointer {
364            ensure(
365                path.len() < MAX_DEPTH,
366                Error::Corrupt("map cursor exceeds height bound"),
367            )?;
368            let node = self.read_node(p, path.len())?;
369            let index = match bound {
370                Bound::Unbounded => {
371                    if forward {
372                        0
373                    } else {
374                        node.len()
375                    }
376                }
377                Bound::Included(k) => match node.search(k)? {
378                    Ok(i) => {
379                        if forward {
380                            i
381                        } else {
382                            i + 1
383                        }
384                    }
385                    Err(i) => i,
386                },
387                Bound::Excluded(k) => match node.search(k)? {
388                    Ok(i) => {
389                        if forward {
390                            i + 1
391                        } else {
392                            i
393                        }
394                    }
395                    Err(i) => i,
396                },
397            };
398            pointer = node.children.get(index).copied();
399            path.push(Frame {
400                node,
401                index,
402                descend: false,
403            });
404        }
405        Ok(path)
406    }
407    fn advance_path(
408        &self,
409        path: &mut Vec<Frame>,
410        forward: bool,
411        keys_only: bool,
412    ) -> Result<Option<(K, Option<Value>)>> {
413        self.healthy()?;
414        while let Some(frame) = path.last_mut() {
415            if frame.descend {
416                frame.descend = false;
417                let mut child = frame.node.children.get(frame.index).copied();
418                while let Some(p) = child {
419                    ensure(
420                        path.len() < MAX_DEPTH,
421                        Error::Corrupt("map cursor exceeds height bound"),
422                    )?;
423                    let node = self.read_node(p, path.len())?;
424                    let index = if forward { 0 } else { node.len() };
425                    child = node.children.get(index).copied();
426                    path.push(Frame {
427                        node,
428                        index,
429                        descend: false,
430                    });
431                }
432                continue;
433            }
434            let index = if forward {
435                frame.index
436            } else {
437                if frame.index == 0 {
438                    path.pop();
439                    continue;
440                }
441                frame.index - 1
442            };
443            if index >= frame.node.len() {
444                path.pop();
445                continue;
446            }
447            let bytes = frame.node.key_bytes(index);
448            K::check_bounds(bytes)?;
449            let key = K::from_bytes(Cow::Borrowed(bytes))?;
450            let value = (!keys_only).then(|| frame.node.value(index));
451            frame.index = if forward { index + 1 } else { index };
452            // A caller may stop at this entry. Read the following subtree only
453            // when another step in this direction is actually requested.
454            frame.descend = !frame.node.children.is_empty();
455            return Ok(Some((key, value)));
456        }
457        Ok(None)
458    }
459    pub fn range(&self, bounds: impl RangeBounds<K>) -> TreeIter<'_, K, V, M> {
460        let lower = bounds.start_bound().cloned();
461        let upper = bounds.end_bound().cloned();
462        let invalid = match (&lower, &upper) {
463            (Bound::Included(a) | Bound::Excluded(a), Bound::Included(b) | Bound::Excluded(b)) => {
464                a > b
465                    || (a == b
466                        && matches!((&lower, &upper), (Bound::Excluded(_), Bound::Excluded(_))))
467            }
468            _ => false,
469        };
470        TreeIter {
471            tree: self,
472            lower,
473            upper,
474            front: None,
475            back: None,
476            exhausted: false,
477            invalid,
478            keys_only: false,
479        }
480    }
481    pub fn keys_range(
482        &self,
483        bounds: impl RangeBounds<K>,
484    ) -> impl DoubleEndedIterator<Item = Result<K>> + '_ {
485        let mut iter = self.range(bounds);
486        iter.keys_only = true;
487        iter.map(|e| e.map(MapEntry::into_key))
488    }
489}
490struct Frame {
491    node: Rc<ReadNode>,
492    index: usize,
493    descend: bool,
494}
495fn decode<T: Storable>(bytes: Vec<u8>) -> Result<T> {
496    T::check_bounds(&bytes)?;
497    Ok(T::from_bytes(Cow::Owned(bytes))?)
498}
499
500type Split = Option<(Entry, Pointer)>;
501struct Transaction<'a, K, V, M: Memory> {
502    tree: &'a Tree<K, V, M>,
503    root: Root,
504    writes: BTreeMap<u64, Vec<u8>>,
505}
506impl<'a, K: Storable + Ord + Clone, V: Storable, M: Memory> Transaction<'a, K, V, M> {
507    fn new(tree: &'a Tree<K, V, M>) -> Self {
508        Self {
509            tree,
510            root: tree.root.borrow().clone(),
511            writes: BTreeMap::new(),
512        }
513    }
514    fn bytes(&self, p: Pointer, magic: [u8; 8]) -> Result<Vec<u8>> {
515        p.check(self.root.cursor)?;
516        if let Some(bytes) = self.writes.get(&p.offset) {
517            Ok(bytes.clone())
518        } else {
519            record::read_bytes(&self.tree.memory, p.offset, p.capacity(), magic)
520        }
521    }
522    fn node(&self, p: Pointer, depth: usize) -> Result<Node> {
523        Ok(self.read_node(p, depth)?.owned())
524    }
525    fn read_node(&self, p: Pointer, depth: usize) -> Result<Rc<ReadNode>> {
526        if let Some(bytes) = self.writes.get(&p.offset) {
527            Ok(Rc::new(ReadNode::decode::<K, V>(
528                bytes.clone(),
529                self.root.cursor,
530            )?))
531        } else {
532            self.tree.read_node(p, depth)
533        }
534    }
535    fn value(&self, v: &Value) -> Result<Vec<u8>> {
536        let bytes = match v {
537            Value::Inline(b) => b.clone(),
538            Value::External(p) => {
539                p.check(self.root.cursor)?;
540                if let Some(bytes) = self.writes.get(&p.offset) {
541                    record::decode_payload(VALUE, bytes)?.to_vec()
542                } else {
543                    record::read_payload(&self.tree.memory, p.offset, p.capacity(), VALUE)?
544                }
545            }
546        };
547        V::check_bounds(&bytes)?;
548        Ok(bytes)
549    }
550    fn free(&mut self, p: Pointer) -> Result<()> {
551        let index = usize::from(p.order);
552        self.writes
553            .insert(p.offset, record::encode(FREE, &self.root.free[index])?);
554        self.root.free[index] = Some(p);
555        Ok(())
556    }
557    fn allocate(&mut self, len: usize) -> Result<Pointer> {
558        let size = len
559            .max(256)
560            .checked_next_power_of_two()
561            .ok_or(Error::Budget)?;
562        let order = size.trailing_zeros() as usize;
563        ensure(order < BINS, Error::Budget)?;
564        for bin in order..BINS {
565            if let Some(mut p) = self.root.free[bin] {
566                let next: Option<Pointer> = record::decode(FREE, &self.bytes(p, FREE)?)?;
567                if let Some(next) = next {
568                    next.check(self.root.cursor)?;
569                    ensure(
570                        next.order == p.order && next != p,
571                        Error::Corrupt("map free-list link"),
572                    )?;
573                }
574                self.root.free[bin] = next;
575                // Return only the requested capacity. Each unused half remains
576                // an ordinary free-list block, staged with the same mutation.
577                while usize::from(p.order) > order {
578                    p.order -= 1;
579                    let sibling = Pointer {
580                        offset: p.offset + p.capacity(),
581                        order: p.order,
582                    };
583                    self.free(sibling)?;
584                }
585                return Ok(p);
586            }
587        }
588        let p = Pointer {
589            offset: self.root.cursor,
590            order: u8::try_from(order).unwrap(),
591        };
592        self.root.cursor = self
593            .root
594            .cursor
595            .checked_add(p.capacity())
596            .ok_or(Error::Budget)?;
597        Ok(p)
598    }
599    fn save_bytes(&mut self, previous: Option<Pointer>, bytes: Vec<u8>) -> Result<Pointer> {
600        let p = if let Some(p) = previous.filter(|p| p.capacity() >= bytes.len() as u64) {
601            let needed = bytes.len().max(256).next_power_of_two() as u64;
602            // Reclaim only into already reusable smaller capacity. Splitting the
603            // live block itself loses its large size class and causes unbounded
604            // growth under repeated large/small replacements without coalescing.
605            let smaller_available = (needed.trailing_zeros() as usize..usize::from(p.order))
606                .any(|order| self.root.free[order].is_some());
607            if p.capacity() / needed >= 4 && smaller_available {
608                let smaller = self.allocate(bytes.len())?;
609                self.free(p)?;
610                smaller
611            } else {
612                p
613            }
614        } else {
615            if let Some(p) = previous {
616                self.free(p)?;
617            }
618            self.allocate(bytes.len())?
619        };
620        self.writes.insert(p.offset, bytes);
621        Ok(p)
622    }
623    fn save(&mut self, p: Option<Pointer>, node: &Node) -> Result<Pointer> {
624        self.save_bytes(p, node.encode()?)
625    }
626    fn save_value(&mut self, previous: Option<&Value>, bytes: Vec<u8>) -> Result<Value> {
627        let old = match previous {
628            Some(Value::External(p)) => Some(*p),
629            _ => None,
630        };
631        if bytes.len() <= INLINE {
632            if let Some(p) = old {
633                self.free(p)?;
634            }
635            Ok(Value::Inline(bytes))
636        } else {
637            Ok(Value::External(
638                self.save_bytes(old, record::encode_payload(VALUE, &bytes)?)?,
639            ))
640        }
641    }
642    fn insert(
643        &mut self,
644        p: Option<Pointer>,
645        key: &K,
646        key_bytes: Vec<u8>,
647        bytes: Vec<u8>,
648        depth: usize,
649    ) -> Result<(Pointer, Split, Option<Vec<u8>>)> {
650        ensure(
651            depth < MAX_DEPTH,
652            Error::Corrupt("map path exceeds height bound"),
653        )?;
654        let read = p.map(|p| self.read_node(p, depth)).transpose()?;
655        let search = read.as_ref().map_or(Ok(Err(0)), |node| node.search(key))?;
656        let (mut node, old) = match search {
657            Ok(i) => {
658                let read = read.as_ref().unwrap();
659                let previous = read.value(i);
660                let old = self.value(&previous)?;
661                let value = self.save_value(Some(&previous), bytes)?;
662                let same_pointer =
663                    matches!((&previous,&value),(Value::External(a),Value::External(b)) if a==b);
664                if same_pointer {
665                    return Ok((p.unwrap(), None, Some(old)));
666                }
667                if let (Value::Inline(_), Value::Inline(bytes)) = (&previous, &value) {
668                    let next = self.save_bytes(p, read.replace_inline(i, bytes)?)?;
669                    return Ok((next, None, Some(old)));
670                }
671                let mut node = read.owned();
672                node.entries[i].value = value;
673                (node, Some(old))
674            }
675            Err(i) => {
676                if let Some(child) = read.as_ref().and_then(|node| node.children.get(i)).copied() {
677                    let (next, split, old) =
678                        self.insert(Some(child), key, key_bytes, bytes, depth + 1)?;
679                    if split.is_none() && next == child {
680                        return Ok((p.unwrap(), None, old));
681                    }
682                    let mut node = read.as_ref().unwrap().owned();
683                    node.children[i] = next;
684                    if let Some((entry, right)) = split {
685                        node.entries.insert(i, entry);
686                        node.children.insert(i + 1, right);
687                    }
688                    (node, old)
689                } else {
690                    let value = self.save_value(None, bytes)?;
691                    let mut node = read
692                        .as_ref()
693                        .map_or_else(Node::default, |node| node.owned());
694                    node.entries.insert(
695                        i,
696                        Entry {
697                            key: key_bytes,
698                            value,
699                        },
700                    );
701                    self.root.length = self.root.length.checked_add(1).ok_or(Error::Budget)?;
702                    (node, None)
703                }
704            }
705        };
706        let split = if node.entries.len() > MAX_ENTRIES {
707            let right_entries = node.entries.split_off(DEGREE + 1);
708            let median = node.entries.pop().unwrap();
709            let children = if node.children.is_empty() {
710                vec![]
711            } else {
712                node.children.split_off(DEGREE + 1)
713            };
714            let right = self.save(
715                None,
716                &Node {
717                    entries: right_entries,
718                    children,
719                },
720            )?;
721            Some((median, right))
722        } else {
723            None
724        };
725        Ok((self.save(p, &node)?, split, old))
726    }
727    // Return a modified node without encoding it: its parent may still repair
728    // an underflow. None means its persisted routing record remains unchanged.
729    fn remove(
730        &mut self,
731        p: Pointer,
732        key: &K,
733        depth: usize,
734    ) -> Result<(Option<Node>, Option<Vec<u8>>)> {
735        ensure(
736            depth < MAX_DEPTH,
737            Error::Corrupt("map path exceeds height bound"),
738        )?;
739        let read = self.read_node(p, depth)?;
740        match read.search(key)? {
741            Ok(i) => {
742                let mut node = read.owned();
743                let old = self.value(&node.entries[i].value)?;
744                if let Value::External(p) = node.entries[i].value {
745                    self.free(p)?;
746                }
747                if node.children.is_empty() {
748                    node.entries.remove(i);
749                } else {
750                    let child_pointer = node.children[i + 1];
751                    let (child, entry) = self.extract_min(child_pointer, depth + 1)?;
752                    node.entries[i] = entry;
753                    if let Some(child) = child {
754                        self.finish_child(&mut node, i + 1, child, depth)?;
755                    }
756                }
757                Ok((Some(node), Some(old)))
758            }
759            Err(i) => {
760                let Some(child_pointer) = read.children.get(i).copied() else {
761                    return Ok((None, None));
762                };
763                let (child, old) = self.remove(child_pointer, key, depth + 1)?;
764                let node = child
765                    .map(|child| self.changed_parent(&read, i, child, depth))
766                    .transpose()?
767                    .flatten();
768                Ok((node, old))
769            }
770        }
771    }
772    fn extract_min(&mut self, p: Pointer, depth: usize) -> Result<(Option<Node>, Entry)> {
773        ensure(
774            depth < MAX_DEPTH,
775            Error::Corrupt("map path exceeds height bound"),
776        )?;
777        let read = self.read_node(p, depth)?;
778        ensure(read.len() != 0, Error::Corrupt("empty successor node"))?;
779        if read.children.is_empty() {
780            let mut node = read.owned();
781            let entry = node.entries.remove(0);
782            Ok((Some(node), entry))
783        } else {
784            let (child, entry) = self.extract_min(read.children[0], depth + 1)?;
785            let node = child
786                .map(|child| self.changed_parent(&read, 0, child, depth))
787                .transpose()?
788                .flatten();
789            Ok((node, entry))
790        }
791    }
792    // Keep unchanged ancestors borrowed, including when the child was rewritten
793    // in place. Only a moved child or an underflow requires owning the parent.
794    fn changed_parent(
795        &mut self,
796        read: &ReadNode,
797        i: usize,
798        child: Node,
799        depth: usize,
800    ) -> Result<Option<Node>> {
801        if child.entries.len() >= MIN_ENTRIES {
802            let previous = read.children[i];
803            let next = self.save(Some(previous), &child)?;
804            if next == previous {
805                return Ok(None);
806            }
807            let mut parent = read.owned();
808            parent.children[i] = next;
809            Ok(Some(parent))
810        } else {
811            let mut parent = read.owned();
812            self.rebalance(&mut parent, i, child, depth)?;
813            Ok(Some(parent))
814        }
815    }
816    fn finish_child(
817        &mut self,
818        parent: &mut Node,
819        i: usize,
820        child: Node,
821        depth: usize,
822    ) -> Result<()> {
823        if child.entries.len() >= MIN_ENTRIES {
824            parent.children[i] = self.save(Some(parent.children[i]), &child)?;
825            Ok(())
826        } else {
827            self.rebalance(parent, i, child, depth)
828        }
829    }
830    fn rebalance(
831        &mut self,
832        parent: &mut Node,
833        i: usize,
834        mut child: Node,
835        parent_depth: usize,
836    ) -> Result<()> {
837        let depth = parent_depth + 1;
838        let p = parent.children[i];
839        let mut left = if i > 0 {
840            Some(self.node(parent.children[i - 1], depth)?)
841        } else {
842            None
843        };
844        if let Some(left) = &mut left {
845            ensure(
846                left.children.is_empty() == child.children.is_empty(),
847                Error::Corrupt("B-tree sibling levels"),
848            )?;
849            if left.entries.len() > MIN_ENTRIES {
850                child.entries.insert(
851                    0,
852                    std::mem::replace(&mut parent.entries[i - 1], left.entries.pop().unwrap()),
853                );
854                if let Some(p) = left.children.pop() {
855                    child.children.insert(0, p);
856                }
857                parent.children[i - 1] = self.save(Some(parent.children[i - 1]), left)?;
858                parent.children[i] = self.save(Some(p), &child)?;
859                return Ok(());
860            }
861        }
862        let mut right = if i + 1 < parent.children.len() {
863            Some(self.node(parent.children[i + 1], depth)?)
864        } else {
865            None
866        };
867        if let Some(right) = &mut right {
868            ensure(
869                right.children.is_empty() == child.children.is_empty(),
870                Error::Corrupt("B-tree sibling levels"),
871            )?;
872            if right.entries.len() > MIN_ENTRIES {
873                child.entries.push(std::mem::replace(
874                    &mut parent.entries[i],
875                    right.entries.remove(0),
876                ));
877                if !right.children.is_empty() {
878                    child.children.push(right.children.remove(0));
879                }
880                parent.children[i + 1] = self.save(Some(parent.children[i + 1]), right)?;
881                parent.children[i] = self.save(Some(p), &child)?;
882                return Ok(());
883            }
884        }
885        let (left_index, mut left, right) = if let Some(left) = left {
886            (i - 1, left, child)
887        } else {
888            (
889                i,
890                child,
891                right.ok_or(Error::Corrupt("B-tree missing separator"))?,
892            )
893        };
894        ensure(
895            left_index < parent.entries.len(),
896            Error::Corrupt("B-tree missing separator"),
897        )?;
898        let left_p = parent.children[left_index];
899        let right_p = parent.children[left_index + 1];
900        left.entries.push(parent.entries.remove(left_index));
901        left.entries.extend(right.entries);
902        left.children.extend(right.children);
903        self.free(right_p)?;
904        parent.children[left_index] = self.save(Some(left_p), &left)?;
905        parent.children.remove(left_index + 1);
906        Ok(())
907    }
908    fn commit(self) -> Result<()> {
909        let changed = self.root != *self.tree.root.borrow();
910        let header = if changed {
911            Some(record::encode(MAGIC, &self.root)?)
912        } else {
913            None
914        };
915        if let Some(h) = &header {
916            ensure(h.len() as u64 <= record::HEADER_BYTES, Error::Budget)?;
917        }
918        record::reserve(&self.tree.memory, self.root.cursor)?;
919        self.tree.poisoned.set(true);
920        if self.root.cursor < self.tree.root.borrow().cursor
921            || self.root.top != self.tree.root.borrow().top
922        {
923            self.tree.cache.borrow_mut().clear();
924        }
925        for (offset, bytes) in self.writes {
926            self.tree.cache.borrow_mut().invalidate(offset);
927            self.tree.memory.write(offset, &bytes);
928        }
929        if let Some(h) = header {
930            self.tree.memory.write(0, &h);
931        }
932        *self.tree.root.borrow_mut() = self.root;
933        self.tree.poisoned.set(false);
934        Ok(())
935    }
936}
937///
938/// `MapEntry`
939///
940/// An owned key and lazily decoded value from an ordered scan.
941///
942pub struct MapEntry<K, V> {
943    key: K,
944    bytes: Vec<u8>,
945    value: PhantomData<V>,
946}
947impl<K, V: Storable> MapEntry<K, V> {
948    /// Borrow the decoded key.
949    pub const fn key(&self) -> &K {
950        &self.key
951    }
952    /// Decode the value using its declared codec.
953    ///
954    /// # Errors
955    /// Returns the application codec or declared-bound error for this value.
956    pub fn value(&self) -> Result<V> {
957        V::check_bounds(&self.bytes)?;
958        Ok(V::from_bytes(Cow::Borrowed(&self.bytes))?)
959    }
960    /// Consume the entry and decode its value.
961    ///
962    /// # Errors
963    /// Returns the application codec or declared-bound error for this value.
964    pub fn into_pair(self) -> Result<(K, V)> {
965        Ok((self.key, decode(self.bytes)?))
966    }
967    /// Consume the entry without decoding its value.
968    pub fn into_key(self) -> K {
969        self.key
970    }
971    /// Consume the entry and decode only its value.
972    ///
973    /// # Errors
974    /// Returns the application codec or declared-bound error for this value.
975    pub fn into_value(self) -> Result<V> {
976        decode(self.bytes)
977    }
978}
979
980///
981/// `TreeIter`
982///
983/// Lazy bidirectional traversal with bounded routing-node stacks. External values
984/// are read only for in-range entries and decoded when requested by the caller.
985///
986pub struct TreeIter<'a, K, V, M: Memory> {
987    tree: &'a Tree<K, V, M>,
988    lower: Bound<K>,
989    upper: Bound<K>,
990    front: Option<Vec<Frame>>,
991    back: Option<Vec<Frame>>,
992    exhausted: bool,
993    invalid: bool,
994    keys_only: bool,
995}
996impl<K: Storable + Ord + Clone, V: Storable, M: Memory> TreeIter<'_, K, V, M> {
997    fn step(&mut self, forward: bool) -> Result<Option<MapEntry<K, V>>> {
998        ensure(!self.invalid, Error::Bounds)?;
999        self.tree.healthy()?;
1000        let empty = match (&self.lower, &self.upper) {
1001            (Bound::Included(a), Bound::Included(b)) => a > b,
1002            (Bound::Included(a) | Bound::Excluded(a), Bound::Included(b) | Bound::Excluded(b)) => {
1003                a >= b
1004            }
1005            _ => false,
1006        };
1007        if empty {
1008            return Ok(None);
1009        }
1010        let (path, bound) = if forward {
1011            (&mut self.front, &self.lower)
1012        } else {
1013            (&mut self.back, &self.upper)
1014        };
1015        if path.is_none() {
1016            *path = Some(self.tree.seek_path(bound, forward)?);
1017        }
1018        let entry = self
1019            .tree
1020            .advance_path(path.as_mut().unwrap(), forward, self.keys_only)?;
1021        if let Some((key, value)) = entry {
1022            let after = match &self.lower {
1023                Bound::Unbounded => true,
1024                Bound::Included(k) => key >= *k,
1025                Bound::Excluded(k) => key > *k,
1026            };
1027            let before = match &self.upper {
1028                Bound::Unbounded => true,
1029                Bound::Included(k) => key <= *k,
1030                Bound::Excluded(k) => key < *k,
1031            };
1032            ensure(
1033                if forward { after } else { before },
1034                Error::Corrupt("map cursor key order"),
1035            )?;
1036            if !(after && before) {
1037                return Ok(None);
1038            }
1039            if forward {
1040                self.lower = Bound::Excluded(key.clone());
1041            } else {
1042                self.upper = Bound::Excluded(key.clone());
1043            }
1044            return Ok(Some(MapEntry {
1045                key,
1046                bytes: value
1047                    .map(|value| self.tree.value(value))
1048                    .transpose()?
1049                    .unwrap_or_default(),
1050                value: PhantomData,
1051            }));
1052        }
1053        Ok(None)
1054    }
1055    fn advance(&mut self, forward: bool) -> Option<Result<MapEntry<K, V>>> {
1056        if self.exhausted {
1057            return None;
1058        }
1059        match self.step(forward) {
1060            Ok(Some(entry)) => Some(Ok(entry)),
1061            Ok(None) => {
1062                self.exhausted = true;
1063                None
1064            }
1065            Err(error) => {
1066                self.exhausted = true;
1067                Some(Err(error))
1068            }
1069        }
1070    }
1071}
1072impl<K: Storable + Ord + Clone, V: Storable, M: Memory> Iterator for TreeIter<'_, K, V, M> {
1073    type Item = Result<MapEntry<K, V>>;
1074    fn next(&mut self) -> Option<Self::Item> {
1075        self.advance(true)
1076    }
1077}
1078impl<K: Storable + Ord + Clone, V: Storable, M: Memory> DoubleEndedIterator
1079    for TreeIter<'_, K, V, M>
1080{
1081    fn next_back(&mut self) -> Option<Self::Item> {
1082        self.advance(false)
1083    }
1084}
1085impl<K: Storable + Ord + Clone, V: Storable, M: Memory> std::iter::FusedIterator
1086    for TreeIter<'_, K, V, M>
1087{
1088}
1089
1090#[cfg(test)]
1091mod tests;
ic_storage/adapter/tree.rs

ic_storage/adapter/
tree.rs
