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(&self, path: &mut Vec<Frame>, forward: bool) -> Result<Option<Entry>> {
408        self.healthy()?;
409        while let Some(frame) = path.last_mut() {
410            if frame.descend {
411                frame.descend = false;
412                let mut child = frame.node.children.get(frame.index).copied();
413                while let Some(p) = child {
414                    ensure(
415                        path.len() < MAX_DEPTH,
416                        Error::Corrupt("map cursor exceeds height bound"),
417                    )?;
418                    let node = self.read_node(p, path.len())?;
419                    let index = if forward { 0 } else { node.len() };
420                    child = node.children.get(index).copied();
421                    path.push(Frame {
422                        node,
423                        index,
424                        descend: false,
425                    });
426                }
427                continue;
428            }
429            let index = if forward {
430                frame.index
431            } else {
432                if frame.index == 0 {
433                    path.pop();
434                    continue;
435                }
436                frame.index - 1
437            };
438            if index >= frame.node.len() {
439                path.pop();
440                continue;
441            }
442            let entry = frame.node.entry(index);
443            frame.index = if forward { index + 1 } else { index };
444            // A caller may stop at this entry. Read the following subtree only
445            // when another step in this direction is actually requested.
446            frame.descend = !frame.node.children.is_empty();
447            return Ok(Some(entry));
448        }
449        Ok(None)
450    }
451    pub fn range(&self, bounds: impl RangeBounds<K>) -> TreeIter<'_, K, V, M> {
452        let lower = bounds.start_bound().cloned();
453        let upper = bounds.end_bound().cloned();
454        let invalid = match (&lower, &upper) {
455            (Bound::Included(a) | Bound::Excluded(a), Bound::Included(b) | Bound::Excluded(b)) => {
456                a > b
457                    || (a == b
458                        && matches!((&lower, &upper), (Bound::Excluded(_), Bound::Excluded(_))))
459            }
460            _ => false,
461        };
462        TreeIter {
463            tree: self,
464            lower,
465            upper,
466            front: None,
467            back: None,
468            exhausted: false,
469            invalid,
470            keys_only: false,
471        }
472    }
473    pub fn keys_range(
474        &self,
475        bounds: impl RangeBounds<K>,
476    ) -> impl DoubleEndedIterator<Item = Result<K>> + '_ {
477        let mut iter = self.range(bounds);
478        iter.keys_only = true;
479        iter.map(|e| e.map(MapEntry::into_key))
480    }
481}
482struct Frame {
483    node: Rc<ReadNode>,
484    index: usize,
485    descend: bool,
486}
487fn decode<T: Storable>(bytes: Vec<u8>) -> Result<T> {
488    T::check_bounds(&bytes)?;
489    Ok(T::from_bytes(Cow::Owned(bytes))?)
490}
491
492type Split = Option<(Entry, Pointer)>;
493struct Transaction<'a, K, V, M: Memory> {
494    tree: &'a Tree<K, V, M>,
495    root: Root,
496    writes: BTreeMap<u64, Vec<u8>>,
497}
498impl<'a, K: Storable + Ord + Clone, V: Storable, M: Memory> Transaction<'a, K, V, M> {
499    fn new(tree: &'a Tree<K, V, M>) -> Self {
500        Self {
501            tree,
502            root: tree.root.borrow().clone(),
503            writes: BTreeMap::new(),
504        }
505    }
506    fn bytes(&self, p: Pointer, magic: [u8; 8]) -> Result<Vec<u8>> {
507        p.check(self.root.cursor)?;
508        if let Some(bytes) = self.writes.get(&p.offset) {
509            Ok(bytes.clone())
510        } else {
511            record::read_bytes(&self.tree.memory, p.offset, p.capacity(), magic)
512        }
513    }
514    fn node(&self, p: Pointer, depth: usize) -> Result<Node> {
515        Ok(self.read_node(p, depth)?.owned())
516    }
517    fn read_node(&self, p: Pointer, depth: usize) -> Result<Rc<ReadNode>> {
518        if let Some(bytes) = self.writes.get(&p.offset) {
519            Ok(Rc::new(ReadNode::decode::<K, V>(
520                bytes.clone(),
521                self.root.cursor,
522            )?))
523        } else {
524            self.tree.read_node(p, depth)
525        }
526    }
527    fn value(&self, v: &Value) -> Result<Vec<u8>> {
528        let bytes = match v {
529            Value::Inline(b) => b.clone(),
530            Value::External(p) => {
531                p.check(self.root.cursor)?;
532                if let Some(bytes) = self.writes.get(&p.offset) {
533                    record::decode_payload(VALUE, bytes)?.to_vec()
534                } else {
535                    record::read_payload(&self.tree.memory, p.offset, p.capacity(), VALUE)?
536                }
537            }
538        };
539        V::check_bounds(&bytes)?;
540        Ok(bytes)
541    }
542    fn free(&mut self, p: Pointer) -> Result<()> {
543        let index = usize::from(p.order);
544        self.writes
545            .insert(p.offset, record::encode(FREE, &self.root.free[index])?);
546        self.root.free[index] = Some(p);
547        Ok(())
548    }
549    fn allocate(&mut self, len: usize) -> Result<Pointer> {
550        let size = len
551            .max(256)
552            .checked_next_power_of_two()
553            .ok_or(Error::Budget)?;
554        let order = size.trailing_zeros() as usize;
555        ensure(order < BINS, Error::Budget)?;
556        for bin in order..BINS {
557            if let Some(mut p) = self.root.free[bin] {
558                let next: Option<Pointer> = record::decode(FREE, &self.bytes(p, FREE)?)?;
559                if let Some(next) = next {
560                    next.check(self.root.cursor)?;
561                    ensure(
562                        next.order == p.order && next != p,
563                        Error::Corrupt("map free-list link"),
564                    )?;
565                }
566                self.root.free[bin] = next;
567                // Return only the requested capacity. Each unused half remains
568                // an ordinary free-list block, staged with the same mutation.
569                while usize::from(p.order) > order {
570                    p.order -= 1;
571                    let sibling = Pointer {
572                        offset: p.offset + p.capacity(),
573                        order: p.order,
574                    };
575                    self.free(sibling)?;
576                }
577                return Ok(p);
578            }
579        }
580        let p = Pointer {
581            offset: self.root.cursor,
582            order: u8::try_from(order).unwrap(),
583        };
584        self.root.cursor = self
585            .root
586            .cursor
587            .checked_add(p.capacity())
588            .ok_or(Error::Budget)?;
589        Ok(p)
590    }
591    fn save_bytes(&mut self, previous: Option<Pointer>, bytes: Vec<u8>) -> Result<Pointer> {
592        let p = if let Some(p) = previous.filter(|p| p.capacity() >= bytes.len() as u64) {
593            let needed = bytes.len().max(256).next_power_of_two() as u64;
594            // Reclaim only into already reusable smaller capacity. Splitting the
595            // live block itself loses its large size class and causes unbounded
596            // growth under repeated large/small replacements without coalescing.
597            let smaller_available = (needed.trailing_zeros() as usize..usize::from(p.order))
598                .any(|order| self.root.free[order].is_some());
599            if p.capacity() / needed >= 4 && smaller_available {
600                let smaller = self.allocate(bytes.len())?;
601                self.free(p)?;
602                smaller
603            } else {
604                p
605            }
606        } else {
607            if let Some(p) = previous {
608                self.free(p)?;
609            }
610            self.allocate(bytes.len())?
611        };
612        self.writes.insert(p.offset, bytes);
613        Ok(p)
614    }
615    fn save(&mut self, p: Option<Pointer>, node: &Node) -> Result<Pointer> {
616        self.save_bytes(p, node.encode()?)
617    }
618    fn save_value(&mut self, previous: Option<&Value>, bytes: Vec<u8>) -> Result<Value> {
619        let old = match previous {
620            Some(Value::External(p)) => Some(*p),
621            _ => None,
622        };
623        if bytes.len() <= INLINE {
624            if let Some(p) = old {
625                self.free(p)?;
626            }
627            Ok(Value::Inline(bytes))
628        } else {
629            Ok(Value::External(
630                self.save_bytes(old, record::encode_payload(VALUE, &bytes)?)?,
631            ))
632        }
633    }
634    fn insert(
635        &mut self,
636        p: Option<Pointer>,
637        key: &K,
638        key_bytes: Vec<u8>,
639        bytes: Vec<u8>,
640        depth: usize,
641    ) -> Result<(Pointer, Split, Option<Vec<u8>>)> {
642        ensure(
643            depth < MAX_DEPTH,
644            Error::Corrupt("map path exceeds height bound"),
645        )?;
646        let read = p.map(|p| self.read_node(p, depth)).transpose()?;
647        let search = read.as_ref().map_or(Ok(Err(0)), |node| node.search(key))?;
648        let (mut node, old) = match search {
649            Ok(i) => {
650                let read = read.as_ref().unwrap();
651                let previous = read.value(i);
652                let old = self.value(&previous)?;
653                let value = self.save_value(Some(&previous), bytes)?;
654                let same_pointer =
655                    matches!((&previous,&value),(Value::External(a),Value::External(b)) if a==b);
656                if same_pointer {
657                    return Ok((p.unwrap(), None, Some(old)));
658                }
659                let mut node = read.owned();
660                node.entries[i].value = value;
661                (node, Some(old))
662            }
663            Err(i) => {
664                if let Some(child) = read.as_ref().and_then(|node| node.children.get(i)).copied() {
665                    let (next, split, old) =
666                        self.insert(Some(child), key, key_bytes, bytes, depth + 1)?;
667                    if split.is_none() && next == child {
668                        return Ok((p.unwrap(), None, old));
669                    }
670                    let mut node = read.as_ref().unwrap().owned();
671                    node.children[i] = next;
672                    if let Some((entry, right)) = split {
673                        node.entries.insert(i, entry);
674                        node.children.insert(i + 1, right);
675                    }
676                    (node, old)
677                } else {
678                    let value = self.save_value(None, bytes)?;
679                    let mut node = read
680                        .as_ref()
681                        .map_or_else(Node::default, |node| node.owned());
682                    node.entries.insert(
683                        i,
684                        Entry {
685                            key: key_bytes,
686                            value,
687                        },
688                    );
689                    self.root.length = self.root.length.checked_add(1).ok_or(Error::Budget)?;
690                    (node, None)
691                }
692            }
693        };
694        let split = if node.entries.len() > MAX_ENTRIES {
695            let right_entries = node.entries.split_off(DEGREE + 1);
696            let median = node.entries.pop().unwrap();
697            let children = if node.children.is_empty() {
698                vec![]
699            } else {
700                node.children.split_off(DEGREE + 1)
701            };
702            let right = self.save(
703                None,
704                &Node {
705                    entries: right_entries,
706                    children,
707                },
708            )?;
709            Some((median, right))
710        } else {
711            None
712        };
713        Ok((self.save(p, &node)?, split, old))
714    }
715    // Return a modified node without encoding it: its parent may still repair
716    // an underflow. None means its persisted routing record remains unchanged.
717    fn remove(
718        &mut self,
719        p: Pointer,
720        key: &K,
721        depth: usize,
722    ) -> Result<(Option<Node>, Option<Vec<u8>>)> {
723        ensure(
724            depth < MAX_DEPTH,
725            Error::Corrupt("map path exceeds height bound"),
726        )?;
727        let read = self.read_node(p, depth)?;
728        match read.search(key)? {
729            Ok(i) => {
730                let mut node = read.owned();
731                let old = self.value(&node.entries[i].value)?;
732                if let Value::External(p) = node.entries[i].value {
733                    self.free(p)?;
734                }
735                if node.children.is_empty() {
736                    node.entries.remove(i);
737                } else {
738                    let child_pointer = node.children[i + 1];
739                    let (child, entry) = self.extract_min(child_pointer, depth + 1)?;
740                    node.entries[i] = entry;
741                    if let Some(child) = child {
742                        self.finish_child(&mut node, i + 1, child, depth)?;
743                    }
744                }
745                Ok((Some(node), Some(old)))
746            }
747            Err(i) => {
748                let Some(child_pointer) = read.children.get(i).copied() else {
749                    return Ok((None, None));
750                };
751                let (child, old) = self.remove(child_pointer, key, depth + 1)?;
752                let node = child
753                    .map(|child| self.changed_parent(&read, i, child, depth))
754                    .transpose()?
755                    .flatten();
756                Ok((node, old))
757            }
758        }
759    }
760    fn extract_min(&mut self, p: Pointer, depth: usize) -> Result<(Option<Node>, Entry)> {
761        ensure(
762            depth < MAX_DEPTH,
763            Error::Corrupt("map path exceeds height bound"),
764        )?;
765        let read = self.read_node(p, depth)?;
766        ensure(read.len() != 0, Error::Corrupt("empty successor node"))?;
767        if read.children.is_empty() {
768            let mut node = read.owned();
769            let entry = node.entries.remove(0);
770            Ok((Some(node), entry))
771        } else {
772            let (child, entry) = self.extract_min(read.children[0], depth + 1)?;
773            let node = child
774                .map(|child| self.changed_parent(&read, 0, child, depth))
775                .transpose()?
776                .flatten();
777            Ok((node, entry))
778        }
779    }
780    // Keep unchanged ancestors borrowed, including when the child was rewritten
781    // in place. Only a moved child or an underflow requires owning the parent.
782    fn changed_parent(
783        &mut self,
784        read: &ReadNode,
785        i: usize,
786        child: Node,
787        depth: usize,
788    ) -> Result<Option<Node>> {
789        if child.entries.len() >= MIN_ENTRIES {
790            let previous = read.children[i];
791            let next = self.save(Some(previous), &child)?;
792            if next == previous {
793                return Ok(None);
794            }
795            let mut parent = read.owned();
796            parent.children[i] = next;
797            Ok(Some(parent))
798        } else {
799            let mut parent = read.owned();
800            self.rebalance(&mut parent, i, child, depth)?;
801            Ok(Some(parent))
802        }
803    }
804    fn finish_child(
805        &mut self,
806        parent: &mut Node,
807        i: usize,
808        child: Node,
809        depth: usize,
810    ) -> Result<()> {
811        if child.entries.len() >= MIN_ENTRIES {
812            parent.children[i] = self.save(Some(parent.children[i]), &child)?;
813            Ok(())
814        } else {
815            self.rebalance(parent, i, child, depth)
816        }
817    }
818    fn rebalance(
819        &mut self,
820        parent: &mut Node,
821        i: usize,
822        mut child: Node,
823        parent_depth: usize,
824    ) -> Result<()> {
825        let depth = parent_depth + 1;
826        let p = parent.children[i];
827        let mut left = if i > 0 {
828            Some(self.node(parent.children[i - 1], depth)?)
829        } else {
830            None
831        };
832        if let Some(left) = &mut left {
833            ensure(
834                left.children.is_empty() == child.children.is_empty(),
835                Error::Corrupt("B-tree sibling levels"),
836            )?;
837            if left.entries.len() > MIN_ENTRIES {
838                child.entries.insert(
839                    0,
840                    std::mem::replace(&mut parent.entries[i - 1], left.entries.pop().unwrap()),
841                );
842                if let Some(p) = left.children.pop() {
843                    child.children.insert(0, p);
844                }
845                parent.children[i - 1] = self.save(Some(parent.children[i - 1]), left)?;
846                parent.children[i] = self.save(Some(p), &child)?;
847                return Ok(());
848            }
849        }
850        let mut right = if i + 1 < parent.children.len() {
851            Some(self.node(parent.children[i + 1], depth)?)
852        } else {
853            None
854        };
855        if let Some(right) = &mut right {
856            ensure(
857                right.children.is_empty() == child.children.is_empty(),
858                Error::Corrupt("B-tree sibling levels"),
859            )?;
860            if right.entries.len() > MIN_ENTRIES {
861                child.entries.push(std::mem::replace(
862                    &mut parent.entries[i],
863                    right.entries.remove(0),
864                ));
865                if !right.children.is_empty() {
866                    child.children.push(right.children.remove(0));
867                }
868                parent.children[i + 1] = self.save(Some(parent.children[i + 1]), right)?;
869                parent.children[i] = self.save(Some(p), &child)?;
870                return Ok(());
871            }
872        }
873        let (left_index, mut left, right) = if let Some(left) = left {
874            (i - 1, left, child)
875        } else {
876            (
877                i,
878                child,
879                right.ok_or(Error::Corrupt("B-tree missing separator"))?,
880            )
881        };
882        ensure(
883            left_index < parent.entries.len(),
884            Error::Corrupt("B-tree missing separator"),
885        )?;
886        let left_p = parent.children[left_index];
887        let right_p = parent.children[left_index + 1];
888        left.entries.push(parent.entries.remove(left_index));
889        left.entries.extend(right.entries);
890        left.children.extend(right.children);
891        self.free(right_p)?;
892        parent.children[left_index] = self.save(Some(left_p), &left)?;
893        parent.children.remove(left_index + 1);
894        Ok(())
895    }
896    fn commit(self) -> Result<()> {
897        let changed = self.root != *self.tree.root.borrow();
898        let header = if changed {
899            Some(record::encode(MAGIC, &self.root)?)
900        } else {
901            None
902        };
903        if let Some(h) = &header {
904            ensure(h.len() as u64 <= record::HEADER_BYTES, Error::Budget)?;
905        }
906        record::reserve(&self.tree.memory, self.root.cursor)?;
907        self.tree.poisoned.set(true);
908        if self.root.cursor < self.tree.root.borrow().cursor
909            || self.root.top != self.tree.root.borrow().top
910        {
911            self.tree.cache.borrow_mut().clear();
912        }
913        for (offset, bytes) in self.writes {
914            self.tree.cache.borrow_mut().invalidate(offset);
915            self.tree.memory.write(offset, &bytes);
916        }
917        if let Some(h) = header {
918            self.tree.memory.write(0, &h);
919        }
920        *self.tree.root.borrow_mut() = self.root;
921        self.tree.poisoned.set(false);
922        Ok(())
923    }
924}
925///
926/// `MapEntry`
927///
928/// An owned key and lazily decoded value from an ordered scan.
929///
930pub struct MapEntry<K, V> {
931    key: K,
932    bytes: Vec<u8>,
933    value: PhantomData<V>,
934}
935impl<K, V: Storable> MapEntry<K, V> {
936    /// Borrow the decoded key.
937    pub const fn key(&self) -> &K {
938        &self.key
939    }
940    /// Decode the value using its declared codec.
941    ///
942    /// # Errors
943    /// Returns the application codec or declared-bound error for this value.
944    pub fn value(&self) -> Result<V> {
945        V::check_bounds(&self.bytes)?;
946        Ok(V::from_bytes(Cow::Borrowed(&self.bytes))?)
947    }
948    /// Consume the entry and decode its value.
949    ///
950    /// # Errors
951    /// Returns the application codec or declared-bound error for this value.
952    pub fn into_pair(self) -> Result<(K, V)> {
953        Ok((self.key, decode(self.bytes)?))
954    }
955    /// Consume the entry without decoding its value.
956    pub fn into_key(self) -> K {
957        self.key
958    }
959    /// Consume the entry and decode only its value.
960    ///
961    /// # Errors
962    /// Returns the application codec or declared-bound error for this value.
963    pub fn into_value(self) -> Result<V> {
964        decode(self.bytes)
965    }
966}
967
968///
969/// `TreeIter`
970///
971/// Lazy bidirectional traversal with bounded routing-node stacks. External values
972/// are read only for in-range entries and decoded when requested by the caller.
973///
974pub struct TreeIter<'a, K, V, M: Memory> {
975    tree: &'a Tree<K, V, M>,
976    lower: Bound<K>,
977    upper: Bound<K>,
978    front: Option<Vec<Frame>>,
979    back: Option<Vec<Frame>>,
980    exhausted: bool,
981    invalid: bool,
982    keys_only: bool,
983}
984impl<K: Storable + Ord + Clone, V: Storable, M: Memory> TreeIter<'_, K, V, M> {
985    fn step(&mut self, forward: bool) -> Result<Option<MapEntry<K, V>>> {
986        ensure(!self.invalid, Error::Bounds)?;
987        self.tree.healthy()?;
988        let empty = match (&self.lower, &self.upper) {
989            (Bound::Included(a), Bound::Included(b)) => a > b,
990            (Bound::Included(a) | Bound::Excluded(a), Bound::Included(b) | Bound::Excluded(b)) => {
991                a >= b
992            }
993            _ => false,
994        };
995        if empty {
996            return Ok(None);
997        }
998        let (path, bound) = if forward {
999            (&mut self.front, &self.lower)
1000        } else {
1001            (&mut self.back, &self.upper)
1002        };
1003        if path.is_none() {
1004            *path = Some(self.tree.seek_path(bound, forward)?);
1005        }
1006        let entry = self.tree.advance_path(path.as_mut().unwrap(), forward)?;
1007        if let Some(entry) = entry {
1008            let key: K = decode(entry.key)?;
1009            let after = match &self.lower {
1010                Bound::Unbounded => true,
1011                Bound::Included(k) => key >= *k,
1012                Bound::Excluded(k) => key > *k,
1013            };
1014            let before = match &self.upper {
1015                Bound::Unbounded => true,
1016                Bound::Included(k) => key <= *k,
1017                Bound::Excluded(k) => key < *k,
1018            };
1019            ensure(
1020                if forward { after } else { before },
1021                Error::Corrupt("map cursor key order"),
1022            )?;
1023            if !(after && before) {
1024                return Ok(None);
1025            }
1026            if forward {
1027                self.lower = Bound::Excluded(key.clone());
1028            } else {
1029                self.upper = Bound::Excluded(key.clone());
1030            }
1031            return Ok(Some(MapEntry {
1032                key,
1033                bytes: if self.keys_only {
1034                    vec![]
1035                } else {
1036                    self.tree.value(entry.value)?
1037                },
1038                value: PhantomData,
1039            }));
1040        }
1041        Ok(None)
1042    }
1043    fn advance(&mut self, forward: bool) -> Option<Result<MapEntry<K, V>>> {
1044        if self.exhausted {
1045            return None;
1046        }
1047        match self.step(forward) {
1048            Ok(Some(entry)) => Some(Ok(entry)),
1049            Ok(None) => {
1050                self.exhausted = true;
1051                None
1052            }
1053            Err(error) => {
1054                self.exhausted = true;
1055                Some(Err(error))
1056            }
1057        }
1058    }
1059}
1060impl<K: Storable + Ord + Clone, V: Storable, M: Memory> Iterator for TreeIter<'_, K, V, M> {
1061    type Item = Result<MapEntry<K, V>>;
1062    fn next(&mut self) -> Option<Self::Item> {
1063        self.advance(true)
1064    }
1065}
1066impl<K: Storable + Ord + Clone, V: Storable, M: Memory> DoubleEndedIterator
1067    for TreeIter<'_, K, V, M>
1068{
1069    fn next_back(&mut self) -> Option<Self::Item> {
1070        self.advance(false)
1071    }
1072}
1073impl<K: Storable + Ord + Clone, V: Storable, M: Memory> std::iter::FusedIterator
1074    for TreeIter<'_, K, V, M>
1075{
1076}
1077
1078#[cfg(test)]
1079mod tests;
ic_storage/adapter/tree.rs

ic_storage/adapter/
tree.rs
