Skip to main content

mnemo_engine/store/
flush.rs

1//! Commits and durability: sealed runs, the WAL and checkpoint, flush, close, and the page-cache and vector-budget settings.
2
3use super::*;
4
5impl Mnemo {
6    /// Seal a serialized buffer into a fresh run of encrypted home pages and
7    /// append a WAL frame for each. Returns the run's `(start_page, pages)`.
8    /// The pages are *not* written to their home locations here — they go to
9    /// the WAL and are folded in later by [`Mnemo::checkpoint`].
10    ///
11    /// `gen` is the run's generation (v9; `0` on v8). On v9 the pages come
12    /// from the free list when an eligible run fits.
13    pub(super) fn seal_run(
14        &mut self,
15        bytes: &[u8],
16        gen: u64,
17        frames: &mut Vec<wal::Frame>,
18    ) -> Result<(u64, u32)> {
19        let pc = bytes.len().div_ceil(PAYLOAD).max(1);
20        let start = self.alloc_pages(pc as u64);
21        self.seal_run_at(start, pc as u64, bytes, gen, frames)?;
22        Ok((start, pc as u32))
23    }
24
25    /// Seal `bytes` into the `pages` already-allocated pages at `start`
26    /// (pages past the data are sealed empty) and append their WAL frames.
27    pub(super) fn seal_run_at(
28        &mut self,
29        start: u64,
30        pages: u64,
31        bytes: &[u8],
32        gen: u64,
33        frames: &mut Vec<wal::Frame>,
34    ) -> Result<()> {
35        // RB-007.6: no v9 page is sealed before this window's lease.
36        debug_assert!(
37            !self.is_v9() || self.gen_leased,
38            "a v9 run sealed before the generation lease"
39        );
40        let direct = self.is_v9() && pages > V9_WAL_RUN_MAX_PAGES;
41        for i in 0..pages as usize {
42            let lo = (i * PAYLOAD).min(bytes.len());
43            let hi = ((i + 1) * PAYLOAD).min(bytes.len());
44            let page_no = start + i as u64;
45            let sealed = self.pager.seal_page(page_no, gen, &bytes[lo..hi])?;
46            if direct {
47                self.pager.write_sealed(page_no, &sealed)?;
48            } else {
49                frames.push((page_no, sealed.to_vec()));
50            }
51        }
52        if direct {
53            self.direct_writes = true;
54        }
55        Ok(())
56    }
57
58    /// Grow the WAL region if `frame_pages` worth of frames would not fit.
59    /// Called at the top of `flush`, where the WAL is always spent (every
60    /// flush ends with a checkpoint), so relocating it cannot strand a
61    /// committed transaction.
62    pub(super) fn ensure_wal_capacity(&mut self, frame_pages: usize) -> Result<()> {
63        let need = wal::txn_byte_len(frame_pages, PAGE_SIZE);
64        if need <= self.header.wal_pages * PAGE_SIZE as u64 {
65            return Ok(());
66        }
67        // Allocate a new, larger region at the file tail (~1.5x headroom).
68        let req = need.div_ceil(PAGE_SIZE as u64);
69        let new_pages = (req + req / 2 + 4).max(DEFAULT_WAL_PAGES);
70        let old_wal_start = self.header.wal_start;
71        let old_wal_pages = self.header.wal_pages;
72        // FMT-9 (a): the new region always comes from the file's tail: the
73        // relocation commits a header-only transaction, so a region taken
74        // from the free list would still be listed as free in the
75        // committed free-list run.
76        let new_start = self.header.next_page;
77        self.header.next_page += new_pages;
78        self.header.wal_start = new_start;
79        self.header.wal_pages = new_pages;
80        self.header.wal_seq += 1;
81        // DUR-1 (R-002.1): WAL-log the header change into the *new* region
82        // first, checkpoint page 0, then invalidate the old region. A crash
83        // anywhere here either keeps the old state (if page 0 wasn't yet
84        // overwritten) or recovers to the new state via the torn-header
85        // scan in `Mnemo::open`, which prefers the highest-txn-id valid
86        // committed header frame.
87        self.relocate_header_through_wal(old_wal_start)?;
88        // FMT-9 (a): once the relocation has committed, the old region is
89        // unreferenced; the next commit frees it.
90        self.release_pages(old_wal_start, old_wal_pages);
91        Ok(())
92    }
93
94    /// Shared helper for every post-`create` page-0 commit: WAL-log the
95    /// new header into the current WAL region, checkpoint page 0, and
96    /// invalidate `old_wal_start` (the previous region, or the current
97    /// region when it hasn't moved). The caller is expected to have
98    /// mutated `self.header` to its target values *including a bumped
99    /// `wal_seq`* (one more than the last on-disk `wal_seq`); this
100    /// routine applies the seal, writes the frame + commit, folds page 0,
101    /// then zeroes the old region's first frame header.
102    ///
103    /// If `old_wal_start == self.header.wal_start`, the invalidation is
104    /// skipped — the region didn't move, so overwriting its first frame
105    /// would destroy the just-committed one.
106    pub(super) fn relocate_header_through_wal(&mut self, old_wal_start: u64) -> Result<()> {
107        self.header.write_counter = self.pager.write_counter;
108        let mut page = self.header.to_page();
109        self.header.apply_seal(&mut page, self.pager.dek())?;
110        let txn_id = self.header.wal_seq;
111        let frames = vec![(0u64, page.to_vec())];
112        wal::commit(
113            self.pager.file_mut(),
114            self.header.wal_start,
115            self.header.wal_pages,
116            txn_id,
117            &frames,
118        )?;
119        self.pager.write_raw(0, &page)?;
120        self.pager.sync()?;
121        if old_wal_start != self.header.wal_start {
122            wal::invalidate_region(self.pager.file_mut(), old_wal_start)?;
123        }
124        Ok(())
125    }
126
127    /// Fold a committed transaction's WAL frames into their home pages.
128    pub(super) fn checkpoint(&mut self, frames: &[wal::Frame]) -> Result<()> {
129        for (page_no, bytes) in frames {
130            let mut img = [0u8; PAGE_SIZE];
131            img.copy_from_slice(bytes);
132            self.pager.write_sealed(*page_no, &img)?;
133        }
134        self.pager.sync()?;
135        Ok(())
136    }
137
138    /// Persist all pending changes as one **write-ahead-logged transaction**.
139    ///
140    /// Sequence:
141    ///
142    /// 1. **Prepare.** Serialize the control plane (catalog, ANN index),
143    ///    grow the WAL if needed, and **lease** counter and page slots —
144    ///    bump `header.write_counter` and `header.next_page` by upper
145    ///    bounds on this transaction's consumption and persist that
146    ///    *leased* header to disk before any encrypted page is written.
147    /// 2. **Data pages.** [`Pager::flush`] writes dirty record pages
148    ///    copy-on-write under fresh nonces and fsyncs them. The orphan-page
149    ///    nonce-reuse window (see [Phase 1.1 of the improvement plan])
150    ///    is closed because the leased header on disk already records a
151    ///    `write_counter` past anything this step can produce.
152    /// 3. **Control plane.** Seal the new catalog / ANN index / snapshot
153    ///    manifest into fresh home page runs.
154    /// 4. **Commit.** Log all of step 3's frames plus the final header
155    ///    frame into the WAL and fsync — the durability point.
156    /// 5. **Checkpoint.** Fold the WAL into the home pages.
157    ///
158    /// A crash before step 4 leaves the previous state intact. A crash
159    /// after step 4 is repaired by [`Mnemo::open`] replaying the WAL.
160    /// Safe to call repeatedly.
161    pub fn flush(&mut self) -> Result<()> {
162        if self.read_only {
163            return Err(MnemoError::ReadOnly);
164        }
165        // Snapshot the thread-local fsync counter. Every `sync_all` the
166        // pager / wal / checkpoint path makes between here and the end
167        // of this function bumps the same counter; `delta` lands on
168        // `self.fsyncs_last_flush` so `Stats::fsyncs_last_flush` can
169        // report it. The planner wants this to decide whether cutting
170        // fsync count is a viable write-latency win before we touch
171        // the write path's durability model.
172        let fsync_before = crate::fsync_count();
173        self.commit_point = None;
174        // PERF-1 step 3: while the index is unloaded, records written since
175        // its run pile up for the next load to re-assign. Once there are
176        // INDEX_PERSIST_EVERY_CHANGES of them, load it now so this flush
177        // persists a current index and keeps that work bounded.
178        if self.ann_unloaded && self.dirty_catalog {
179            let (index_start, index_gen) = (self.header.index_start, self.index_gen_now());
180            let v9 = self.is_v9();
181            let pending = self
182                .catalog
183                .iter()
184                .filter(|e| !e.deleted && written_after_index(e, v9, index_start, index_gen))
185                .count();
186            if pending >= INDEX_PERSIST_EVERY_CHANGES {
187                self.ensure_ann()?;
188            }
189        }
190        // PERF-1: persist the (derived) ANN index only when forced or due.
191        // Otherwise the header keeps pointing at the previous index run, and
192        // records written since then are re-assigned when it's next loaded
193        // (`reconcile_ann_with_catalog`). `ann == None` with `dirty_index` set is a
194        // dropped index, which is persisted at once.
195        let persist_index = if self.is_v9() {
196            self.plan_index_v9()
197        } else {
198            self.dirty_index
199                && (self.index_persist.forced
200                    || self.ann.is_none()
201                    || self.index_persist.commits + 1 >= INDEX_PERSIST_EVERY_COMMITS
202                    || self.index_persist.changes >= INDEX_PERSIST_EVERY_CHANGES)
203        };
204        let free_work =
205            self.is_v9() && (self.free_dirty || !self.pending_free.is_empty() || self.dirty_keys);
206        if !self.dirty_catalog && !persist_index && !self.dirty_cache && !free_work {
207            // No control-plane changes pending. Because every `put` that
208            // dirties a data page also dirties the catalog, and every
209            // `cache_put` that dirties a data page also dirties
210            // `dirty_cache`, there can't be dirty data pages either —
211            // nothing to do.
212            self.fsyncs_last_flush = 0;
213            return Ok(());
214        }
215        // v0.14.0 step 1: nothing is written to a file that was replaced at
216        // this handle's path.
217        self.ensure_not_replaced()?;
218
219        // FIX-1 R-001.8 part 3: snapshot the memory/index-dirty flags
220        // before step 3b decides whether to push a restore point.
221        // Only flushes that change the memory catalog or ANN index
222        // need a snapshot — a cache-only flush (`dirty_cache` set,
223        // `dirty_catalog` + `dirty_index` both false) must not push,
224        // because snapshots don't capture cache state and every push
225        // counts against `max_snapshots` (default 256), so a cache-
226        // heavy workload would otherwise evict every real memory
227        // restore point within ~256 cache writes.
228        let is_memory_or_index_flush = self.dirty_catalog || persist_index;
229
230        // FMT-9: a v9 session's first sealed page needs the generation
231        // lease (normally taken by its first record write already).
232        self.ensure_gen_lease()?;
233        // FMT-9e: turns that aged past the cutoff are rewritten compressed
234        // (bounded per commit), then packed with the other waiting records.
235        self.settle_cold_turns()?;
236        // FMT-9b: column entries first; they move rows' vector locations,
237        // so they must land before the catalog is serialized.
238        self.pack_records()?;
239        self.pack_vectors()?;
240
241        // 1. Serialize control plane, ensure WAL capacity, lease counter +
242        //    page slots, and persist the leased header *before* any data
243        //    page hits the disk.
244        let ctx = self.prepare_for_flush(persist_index)?;
245
246        // 2. Record (vector) data pages: copy-on-write to fresh pages, fsynced.
247        //    Safe: the lease guarantees that even if we crash here, reopen
248        //    will see a `write_counter` past every nonce this step uses.
249        self.pager.flush()?;
250
251        // 3. Seal the catalog / index control pages into fresh home runs.
252        //    FMT-9: `txn_id` is this transaction's id and the generation
253        //    of every run it seals. A v9 run replaced here is freed by this
254        //    same commit (`superseded`); v8 never reuses pages.
255        let txn_id = self.header.wal_seq + 1;
256        let gen = self.write_gen();
257        debug_assert!(!self.is_v9() || gen == txn_id);
258        let v9 = self.is_v9();
259        let mut superseded: Vec<(u64, u64)> = Vec::new();
260        let mut frames: Vec<wal::Frame> = Vec::new();
261        // FMT-9c.2: a delta run goes on the chain; a new base replaces the
262        // old base and the whole chain (all freed by this commit). The
263        // chain is adopted once the commit succeeds.
264        let mut new_chain: Option<(Vec<RunPtr>, usize)> = None;
265        if let Some(bytes) = &ctx.cat_bytes {
266            if ctx.cat_delta {
267                let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
268                let ptr = RunPtr {
269                    start,
270                    pages: pages as u64,
271                    len: bytes.len() as u64,
272                    gen,
273                };
274                self.header.catalog_delta = ptr;
275                let mut chain = self.delta_chain.clone();
276                chain.push(ptr);
277                new_chain = Some((chain, self.delta_rows + self.dirty_rows.len()));
278            } else {
279                if v9 && self.header.catalog_pages > 0 {
280                    superseded.push((self.header.catalog_start, self.header.catalog_pages));
281                }
282                if v9 {
283                    for r in &self.delta_chain {
284                        superseded.push((r.start, r.pages));
285                    }
286                    self.header.catalog_delta = RunPtr::default();
287                    new_chain = Some((Vec::new(), 0));
288                }
289                let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
290                self.header.catalog_start = start;
291                self.header.catalog_pages = pages as u64;
292                self.header.catalog_len = bytes.len() as u64;
293                self.header.catalog_gen = gen;
294            }
295        }
296        // FMT-9d: the BM25 segment, then the manifest that lists it.
297        let mut new_bm25: Option<Bm25Segments> = None;
298        if let Some((bytes, replace_from)) = &ctx.bm25_seg {
299            let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
300            let mut segs = self.bm25.manifest.segments.clone();
301            let mut seg_docs = self.bm25.manifest.seg_docs.clone();
302            seg_docs.resize(segs.len(), 0);
303            let docs = if let Some(from) = *replace_from {
304                for t in segs.drain(from..) {
305                    superseded.push((t.0, t.1));
306                }
307                seg_docs.truncate(from);
308                self.bm25.merged_docs
309            } else {
310                self.bm25.pending.len() as u32
311            };
312            segs.push((start, pages as u64, bytes.len() as u64, gen));
313            seg_docs.push(docs);
314            let man = Bm25Manifest {
315                tokenizer: text::TOKENIZER_V1,
316                next_doc: self.bm25.manifest.next_doc,
317                segments: segs.clone(),
318                seg_docs: seg_docs.clone(),
319            };
320            let mbytes =
321                rmp_serde::to_vec(&man).map_err(|e| MnemoError::Serialize(e.to_string()))?;
322            if !self.header.bm25.is_empty() {
323                superseded.push((self.header.bm25.start, self.header.bm25.pages));
324            }
325            let (ms, mp) = self.seal_run(&mbytes, gen, &mut frames)?;
326            self.header.bm25 = RunPtr {
327                start: ms,
328                pages: mp as u64,
329                len: mbytes.len() as u64,
330                gen,
331            };
332            new_bm25 = Some((segs, seg_docs));
333        }
334        let mut new_ichain: Option<(Vec<RunPtr>, usize)> = None;
335        if persist_index && ctx.idx_delta {
336            if let Some(bytes) = &ctx.idx_bytes {
337                let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
338                let ptr = RunPtr {
339                    start,
340                    pages: pages as u64,
341                    len: bytes.len() as u64,
342                    gen,
343                };
344                self.header.index_delta = ptr;
345                let mut chain = self.index_chain.clone();
346                chain.push(ptr);
347                let n = self.index_delta_ops.as_ref().map_or(0, |o| o.len());
348                new_ichain = Some((chain, self.index_chain_ops + n));
349            }
350        } else if persist_index {
351            if v9 && self.header.index_pages > 0 {
352                superseded.push((self.header.index_start, self.header.index_pages));
353            }
354            if v9 {
355                for r in &self.index_chain {
356                    superseded.push((r.start, r.pages));
357                }
358                self.header.index_delta = RunPtr::default();
359                new_ichain = Some((Vec::new(), 0));
360            }
361            match &ctx.idx_bytes {
362                Some(bytes) => {
363                    let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
364                    self.header.index_start = start;
365                    self.header.index_pages = pages as u64;
366                    self.header.index_len = bytes.len() as u64;
367                    self.header.index_gen = gen;
368                }
369                None => {
370                    self.header.index_start = 0;
371                    self.header.index_pages = 0;
372                    self.header.index_len = 0;
373                    self.header.index_gen = 0;
374                }
375            }
376        }
377        // v8+ result-cache directory (Phase 10.1). Seal the serialized
378        // directory into a fresh page run and update the pointer in the
379        // header. The seal AAD already covers cache_start/pages/len for
380        // v8+ (see `header_seal_aad`), so tampering with the pointer
381        // trips the header seal on the next open.
382        if self.dirty_cache {
383            if v9 && self.header.cache_pages > 0 {
384                superseded.push((self.header.cache_start, self.header.cache_pages));
385            }
386            match &ctx.cache_bytes {
387                Some(bytes) if !self.cache_dir.is_empty() => {
388                    let (start, pages) = self.seal_run(bytes, gen, &mut frames)?;
389                    self.header.cache_start = start;
390                    self.header.cache_pages = pages as u64;
391                    self.header.cache_len = bytes.len() as u64;
392                    self.header.cache_gen = gen;
393                }
394                _ => {
395                    // Either the directory serialized to nothing (all
396                    // entries deleted after tombstone-only migration) or
397                    // it's empty — collapse the pointer to zero.
398                    self.header.cache_start = 0;
399                    self.header.cache_pages = 0;
400                    self.header.cache_len = 0;
401                    self.header.cache_gen = 0;
402                }
403            }
404        }
405
406        // 3b. Record this transaction as a restorable snapshot. The manifest
407        //     update is staged in a local and adopted only once the commit
408        //     below succeeds, so a failed flush leaves no phantom entry.
409        //
410        //     The cap (`self.max_snapshots`) is applied to the staged local
411        //     before the new entry lands, so after appending the manifest
412        //     holds at most `max_snapshots` entries — the most-recent N.
413        //     `max_snapshots == 0` disables the cap. On v8, pages
414        //     referenced only by pruned snapshots stay on disk until
415        //     `compact_file`; on v9 they become reusable once no retained
416        //     snapshot is older than the transaction that freed them.
417        //
418        //     FIX-1 R-001.8 part 3: skip the snapshot push on a
419        //     cache-only flush. Snapshots don't capture the cache
420        //     anyway, and every push counts against `max_snapshots`
421        //     (default 256) + rewrites the manifest run — so a
422        //     cache-heavy workload would otherwise evict every real
423        //     memory restore point within ~256 cache writes. On a
424        //     cache-only flush we reuse the current manifest
425        //     verbatim; the manifest run already lives at
426        //     `self.header.manifest_{start,pages,len}` (set by the
427        //     last memory-touching flush, or at `create` time).
428        let live = self.len() as u64;
429        let mut manifest = self.manifest.clone();
430        if is_memory_or_index_flush {
431            if self.max_snapshots > 0 && !v9 {
432                // Drop the oldest entries so there's room for the new one.
433                while manifest.len() >= self.max_snapshots {
434                    manifest.remove(0);
435                }
436            }
437            manifest.push(Snapshot {
438                txn_id,
439                created_at: memory::now_secs(),
440                catalog_start: self.header.catalog_start,
441                catalog_pages: self.header.catalog_pages,
442                catalog_len: self.header.catalog_len,
443                index_start: self.header.index_start,
444                index_pages: self.header.index_pages,
445                index_len: self.header.index_len,
446                memory_count: live,
447                gens: v9.then_some(SnapshotGens {
448                    catalog: self.header.catalog_gen,
449                    index: self.header.index_gen,
450                    delta: ptr_tuple(self.header.catalog_delta),
451                    index_delta: ptr_tuple(self.header.index_delta),
452                    bm25: ptr_tuple(self.header.bm25),
453                }),
454            });
455            if v9 {
456                let extra = self.pending_free.iter().map(|r| r.1).sum::<u64>()
457                    + superseded.iter().map(|r| r.1).sum::<u64>()
458                    + self.header.manifest_pages;
459                self.apply_retention(&mut manifest, extra);
460            }
461            let man_bytes =
462                rmp_serde::to_vec(&manifest).map_err(|e| MnemoError::Serialize(e.to_string()))?;
463            if v9 && self.header.manifest_pages > 0 {
464                superseded.push((self.header.manifest_start, self.header.manifest_pages));
465            }
466            let (m_start, m_pages) = self.seal_run(&man_bytes, gen, &mut frames)?;
467            self.header.manifest_start = m_start;
468            self.header.manifest_pages = m_pages as u64;
469            self.header.manifest_len = man_bytes.len() as u64;
470            self.header.manifest_gen = gen;
471        }
472
473        // 3c. FMT-9 (b): the key-slot run, rewritten copy-on-write when a
474        //     slot changed. The header seal covers its pointer.
475        if v9 && self.dirty_keys {
476            if !self.header.keys.is_empty() {
477                superseded.push((self.header.keys.start, self.header.keys.pages));
478            }
479            if self.key_slots.slots.is_empty() {
480                self.header.keys = RunPtr::default();
481            } else {
482                let bytes = rmp_serde::to_vec(&self.key_slots)
483                    .map_err(|e| MnemoError::Serialize(e.to_string()))?;
484                let (start, pages) = self.seal_run(&bytes, gen, &mut frames)?;
485                self.header.keys = RunPtr {
486                    start,
487                    pages: pages as u64,
488                    len: bytes.len() as u64,
489                    gen,
490                };
491            }
492        }
493
494        // 3d. FMT-9 (a): the free list. Everything this transaction
495        //     stopped referencing (`pending_free`, the runs replaced above,
496        //     and the previous free-list run itself) is freed *at this
497        //     transaction's id*, in this transaction's own commit, so it
498        //     can't be reused until this commit is durable and no retained
499        //     snapshot predates it. The list is staged in a local and
500        //     adopted only once the commit succeeds (a failed flush that
501        //     is retried must not free the same runs twice).
502        let mut shrink = false;
503        let staged_free =
504            if v9 && (self.free_dirty || !self.pending_free.is_empty() || !superseded.is_empty()) {
505                let mut staged = self.free.clone();
506                for &(p, c) in self.pending_free.iter().chain(superseded.iter()) {
507                    staged.free(txn_id, p, c)?;
508                }
509                if !self.header.free.is_empty() {
510                    staged.free(txn_id, self.header.free.start, self.header.free.pages)?;
511                }
512                // FMT-9c: give eligible free runs at the file's tail back
513                // to the filesystem: the header's `next_page` moves down in
514                // this commit, and the file is cut after it.
515                let min_snap = self.min_snapshot_txn();
516                while let Some(p) =
517                    staged.take_tail(self.header.next_page, self.header.wal_seq, min_snap)
518                {
519                    self.header.next_page = p;
520                    shrink = true;
521                }
522                // Size the run with room for the one extra entry that
523                // allocating it out of the list can create (a split), then
524                // allocate it from the list (eligible runs only) or the tail.
525                let est = rmp_serde::to_vec(&staged.to_runs())
526                    .map_err(|e| MnemoError::Serialize(e.to_string()))?
527                    .len()
528                    + 64;
529                let pc = est.div_ceil(PAYLOAD).max(1) as u64;
530                let start = match staged.alloc(pc, self.header.wal_seq, self.min_snapshot_txn()) {
531                    Some(p) => p,
532                    None => {
533                        let p = self.header.next_page;
534                        self.header.next_page += pc;
535                        p
536                    }
537                };
538                let bytes = rmp_serde::to_vec(&staged.to_runs())
539                    .map_err(|e| MnemoError::Serialize(e.to_string()))?;
540                if bytes.len() as u64 > pc * PAYLOAD as u64 {
541                    return Err(MnemoError::Invalid(
542                        "free list outgrew its reserved run".into(),
543                    ));
544                }
545                self.seal_run_at(start, pc, &bytes, gen, &mut frames)?;
546                self.header.free = RunPtr {
547                    start,
548                    pages: pc,
549                    len: bytes.len() as u64,
550                    gen,
551                };
552                Some(staged)
553            } else {
554                None
555            };
556
557        // 3e. The header is the transaction's final frame; stamp the new id
558        //     and the *actual* (post-flush) write counter, which is <= the
559        //     leased value persisted in `prepare_for_flush`. This is the
560        //     value that ultimately replaces the leased header on checkpoint.
561        self.header.vector_count = live;
562        self.header.write_counter = self.pager.write_counter;
563        self.header.wal_seq = txn_id;
564        let mut hpage = self.header.to_page();
565        self.header.apply_seal(&mut hpage, self.pager.dek())?;
566        frames.push((0, hpage.to_vec()));
567
568        // 4. COMMIT — log the transaction and fsync. Nothing at a home page
569        //    other than the leased header has changed yet; this single
570        //    fsync is what makes the rest durable.
571        // FMT-9c: big runs went straight to their home pages; make them
572        // durable before the header that references them can be.
573        if self.direct_writes {
574            self.pager.sync()?;
575            self.direct_writes = false;
576        }
577        let (wal_start, wal_pages) = (self.header.wal_start, self.header.wal_pages);
578        // 12.4: from here a failure may leave the transaction committed.
579        self.commit_point = Some(txn_id);
580        wal::commit(self.pager.file_mut(), wal_start, wal_pages, txn_id, &frames)?;
581
582        // 5. Checkpoint — fold the WAL into the home pages.
583        self.checkpoint(&frames)?;
584        // FMT-9c: the committed header no longer covers the tail; cut it.
585        // A crash before this leaves a longer file, which is harmless.
586        if shrink {
587            let len = self.header.next_page * PAGE_SIZE as u64;
588            self.pager.file_mut().set_len(len)?;
589        }
590
591        self.manifest = manifest;
592        if let Some(staged) = staged_free {
593            self.free = staged;
594        }
595        self.pending_free.clear();
596        self.free_dirty = false;
597        self.dirty_keys = false;
598        if let Some((chain, rows)) = new_chain {
599            self.delta_chain = chain;
600            self.delta_rows = rows;
601        }
602        if let Some((chain, n)) = new_ichain {
603            self.index_chain = chain;
604            self.index_chain_ops = n;
605        }
606        if let Some((segs, seg_docs)) = new_bm25 {
607            // FMT-9f: keep the loaded dicts and read only the new segment's
608            // (a merge replaced the segments from `kept` on), instead of
609            // dropping them all and reloading every dict on next use.
610            let kept = self
611                .bm25
612                .manifest
613                .segments
614                .iter()
615                .zip(&segs)
616                .take_while(|(a, b)| a == b)
617                .count();
618            if let Some(mut dicts) = self.bm25.dicts.take() {
619                dicts.truncate(kept);
620                let mut ok = dicts.len() == kept;
621                for &t in &segs[kept..] {
622                    match load_segment_dict(&mut self.pager, tuple_ptr(t)) {
623                        Ok(d) if ok => dicts.push(d),
624                        _ => ok = false,
625                    }
626                }
627                if ok {
628                    self.bm25.dicts = Some(dicts);
629                }
630            }
631            self.bm25.manifest.segments = segs;
632            self.bm25.manifest.seg_docs = seg_docs;
633            self.bm25.pending.clear();
634            self.bm25.pending_tokens = None;
635        }
636        if persist_index {
637            self.index_rebase = false;
638            self.index_delta_ops = None;
639        }
640        if ctx.cat_bytes.is_some() {
641            self.dirty_rows.clear();
642            self.catalog_rewrite = false;
643        }
644        self.dirty_catalog = false;
645        if persist_index {
646            self.dirty_index = false;
647            self.index_persist = IndexPersist::default();
648        } else if self.dirty_index {
649            self.index_persist.commits += 1;
650        }
651        self.dirty_cache = false;
652        // Batched policy: successful flush consumes the pending batch.
653        self.cache_batch.reset();
654        // Record how many fsyncs this flush issued, for Stats reporting.
655        // Thread-local counter guarantees we see only this handle's own
656        // sync_all calls between the snapshot above and this read.
657        self.fsyncs_last_flush =
658            (crate::fsync_count().wrapping_sub(fsync_before)).min(u32::MAX as u64) as u32;
659        // 12.4 shared mode (RB-007.5): this commit is in the file.
660        self.window_committed = true;
661        Ok(())
662    }
663
664    /// Pre-flush prelude shared by [`Mnemo::flush`] and the
665    /// `__crash_partial_flush_for_testing` hook.
666    ///
667    /// Serializes the control plane (catalog, ANN index, cache directory)
668    /// and grows the WAL if needed. That's it: this used to persist a
669    /// leased header too (SEC-1 obsoleted it; see DUR-1 R-002.1 note in
670    /// the body for the full rationale), but after DUR-1 the only page-0
671    /// writes outside `create` go through the WAL via
672    /// [`Mnemo::relocate_header_through_wal`].
673    ///
674    /// The historical pre-stamping used to close the nonce-reuse window
675    /// flagged in Phase 1.1 of the improvement plan: a crash anywhere
676    /// between here and the WAL commit left a file whose on-disk header
677    /// recorded a `write_counter` past every nonce the rest of this
678    /// transaction could produce, so the next reopen started from a
679    /// counter that guaranteed uniqueness against any orphan data pages
680    /// still on disk. SEC-1's per-session random nonce prefix gives the
681    /// same guarantee without needing a persisted counter, which is why
682    /// this routine no longer writes page 0.
683    pub(super) fn prepare_for_flush(&mut self, persist_index: bool) -> Result<FlushPrelude> {
684        // Serialize.
685        // FMT-9c.2: on v9 usually just the changed rows, as a delta run.
686        let cat_delta = self.dirty_catalog && self.catalog_as_delta();
687        let cat_bytes = if !self.dirty_catalog {
688            None
689        } else if cat_delta {
690            let rows: Vec<(u32, CatalogEntry)> = self
691                .dirty_rows
692                .iter()
693                .filter(|&&i| i < self.catalog.len())
694                .map(|&i| (i as u32, self.catalog[i].clone()))
695                .collect();
696            let d = CatalogDelta {
697                prev: ptr_tuple(self.header.catalog_delta),
698                rows,
699            };
700            Some(rmp_serde::to_vec(&d).map_err(|e| MnemoError::Serialize(e.to_string()))?)
701        } else {
702            if self.is_v9() {
703                self.compact_tombstones();
704            }
705            Some(
706                rmp_serde::to_vec(&self.catalog)
707                    .map_err(|e| MnemoError::Serialize(e.to_string()))?,
708            )
709        };
710        // FMT-9c.3: a v9 commit usually writes just its index changes.
711        let idx_delta = persist_index && self.index_delta_ops.is_some();
712        let idx_bytes: Option<Vec<u8>> = if !persist_index {
713            None
714        } else if let Some(ops) = &self.index_delta_ops {
715            let d = IndexDeltaRef {
716                prev: ptr_tuple(self.header.index_delta),
717                ops,
718            };
719            Some(rmp_serde::to_vec(&d).map_err(|e| MnemoError::Serialize(e.to_string()))?)
720        } else {
721            match &self.ann {
722                Some(ann) => {
723                    let mut b =
724                        rmp_serde::to_vec(ann).map_err(|e| MnemoError::Serialize(e.to_string()))?;
725                    // FMT-9: v9 index runs carry a layout byte first.
726                    if self.is_v9() {
727                        b.insert(0, self.index_layout_override.unwrap_or(V9_INDEX_LAYOUT));
728                    }
729                    Some(b)
730                }
731                None => None,
732            }
733        };
734        // Serialize the result-cache directory when it's dirty. An
735        // empty directory serializes to a small header + 0-element
736        // array; still cheap enough that we don't special-case it.
737        let cache_bytes: Option<Vec<u8>> = if self.dirty_cache {
738            Some(
739                rmp_serde::to_vec(&self.cache_dir)
740                    .map_err(|e| MnemoError::Serialize(e.to_string()))?,
741            )
742        } else {
743            None
744        };
745
746        let cat_pc = cat_bytes
747            .as_ref()
748            .map_or(0, |b| b.len().div_ceil(PAYLOAD).max(1));
749        let idx_pc = idx_bytes
750            .as_ref()
751            .map_or(0, |b| b.len().div_ceil(PAYLOAD).max(1));
752        let cache_pc = cache_bytes
753            .as_ref()
754            .map_or(0, |b| b.len().div_ceil(PAYLOAD).max(1));
755        // Upper bound on the manifest run: one extra entry, <=82 bytes each
756        // (rmp_serde encodes a Snapshot as a 9-element fixarray of i64/u64,
757        // worst case 81 bytes plus a 1-byte array header).
758        // v9 snapshots carry two more u64s (≤ 20 bytes with the array header).
759        let per_snapshot = if self.is_v9() { 102 } else { 82 };
760        let man_upper = 9 + (self.manifest.len() + 1) * per_snapshot;
761        let man_pc_est = man_upper.div_ceil(PAYLOAD).max(1);
762        // FMT-9: the free-list run (≤ 28 bytes per run: every pending and
763        // superseded run plus a split and a relocated WAL region) and the
764        // key-slot run.
765        let (free_pc_est, keys_pc) = if self.is_v9() {
766            let runs = self.free.len() + self.pending_free.len() + 12;
767            let free_pc = (runs * 28 + 64).div_ceil(PAYLOAD).max(1);
768            let keys_pc = if self.dirty_keys {
769                rmp_serde::to_vec(&self.key_slots)
770                    .map_err(|e| MnemoError::Serialize(e.to_string()))?
771                    .len()
772                    .div_ceil(PAYLOAD)
773                    .max(1)
774            } else {
775                0
776            };
777            (free_pc, keys_pc)
778        } else {
779            (0, 0)
780        };
781
782        // Grow WAL if needed (DUR-1: relocation is WAL-logged and does
783        // its own fsyncs). The cache directory rides in the same
784        // transaction so its page count contributes to the reservation.
785        // FMT-9d: the BM25 segment (and its manifest, one page).
786        let bm25_seg = self.bm25_next_segment()?;
787        let bm25_pc = bm25_seg
788            .as_ref()
789            .map_or(0, |(b, _)| b.len().div_ceil(PAYLOAD).max(1));
790        // FMT-9c: on v9 only runs of up to V9_WAL_RUN_MAX_PAGES go through
791        // the WAL; longer ones are written in place.
792        let v9 = self.is_v9();
793        let in_wal = |pc: usize| {
794            if v9 && pc as u64 > V9_WAL_RUN_MAX_PAGES {
795                0
796            } else {
797                pc
798            }
799        };
800        self.ensure_wal_capacity(
801            in_wal(cat_pc)
802                + in_wal(idx_pc)
803                + in_wal(cache_pc)
804                + in_wal(man_pc_est)
805                + in_wal(free_pc_est)
806                + in_wal(keys_pc)
807                + in_wal(bm25_pc)
808                + usize::from(bm25_pc > 0)
809                + 1,
810        )?;
811
812        // DUR-1 (R-002.1): the pre-flush "leased header" write that used
813        // to live here is gone. It existed to close the Phase 1.1 nonce
814        // reuse window (reopen must see a `write_counter` past any nonce
815        // this flush could produce before data pages hit disk). SEC-1
816        // superseded that: nonces now carry a per-session random prefix
817        // and `write_counter` is no longer a nonce input, so the leased
818        // write bought us nothing but an extra fsync per flush. Orphan
819        // data pages past `header.next_page` on a crashed flush are
820        // still safe — the next session's random prefix makes any later
821        // write at the same page_no uniquely nonced regardless of
822        // `write_counter`.
823
824        Ok(FlushPrelude {
825            bm25_seg,
826            cat_bytes,
827            cat_delta,
828            idx_delta,
829            idx_bytes,
830            cache_bytes,
831        })
832    }
833
834    /// Flush and close: consumes the handle and releases the file and its
835    /// OS advisory lock before returning, so the same file can be reopened
836    /// right away (by this process or another). On a read-write handle it
837    /// flushes first, persisting a dirty ANN index too; on a read-only
838    /// handle there's nothing to persist.
839    ///
840    /// Using the handle after `close` is a compile error. If the flush
841    /// fails, the error is returned and the handle is dropped all the same:
842    /// changes since the last successful [`Mnemo::flush`] are lost, as
843    /// they would be on a plain `drop` (which never flushes).
844    ///
845    /// A struct that owns a handle and closes it from `&mut self` keeps it
846    /// in an `Option` and `take()`s it:
847    ///
848    /// ```no_run
849    /// use mnemo_engine::{Mnemo, Result};
850    ///
851    /// struct App {
852    ///     db: Option<Mnemo>,
853    /// }
854    ///
855    /// impl App {
856    ///     fn shutdown(&mut self) -> Result<()> {
857    ///         match self.db.take() {
858    ///             Some(db) => db.close(),
859    ///             None => Ok(()), // already closed
860    ///         }
861    ///     }
862    /// }
863    /// ```
864    ///
865    /// Changed in v0.12.0: `close` took `&mut self` and left the file and
866    /// lock held until the handle was dropped.
867    ///
868    /// **If the final flush fails, the handle is dropped with it**, and the
869    /// writes that weren't flushed are lost (a full disk on a phone, say).
870    /// [`Mnemo::try_close`] gives the handle back on that path, so the caller
871    /// can free space and retry.
872    pub fn close(mut self) -> Result<()> {
873        if !self.read_only {
874            // PERF-1: closing persists a dirty ANN index too, so the next
875            // open has nothing to re-assign.
876            self.index_persist.forced = true;
877            self.flush()?;
878        }
879        // Dropping the handle closes the file, which releases the lock.
880        drop(self);
881        Ok(())
882    }
883
884    /// [`Mnemo::close`] that keeps the data when the final flush fails: the
885    /// error carries the handle back, with every unflushed write still in
886    /// it, as [`std::io::BufWriter::into_inner`] does. Retry with
887    /// [`Mnemo::flush`] or another `try_close` once the cause is fixed.
888    /// A read-only handle never flushes, so it always closes. Added in
889    /// v0.12.2 (RB-001.22).
890    ///
891    /// ```no_run
892    /// # use mnemo_engine::Mnemo;
893    /// # fn run(db: Mnemo) -> mnemo_engine::Result<()> {
894    /// match db.try_close() {
895    ///     Ok(()) => {}
896    ///     Err(e) => {
897    ///         eprintln!("close failed: {}", e.error());
898    ///         let mut db = e.into_inner(); // still open, still holding the writes
899    ///         db.flush()?;                  // e.g. after freeing space
900    ///         db.close()?;
901    ///     }
902    /// }
903    /// # Ok(()) }
904    /// ```
905    pub fn try_close(mut self) -> std::result::Result<(), CloseError> {
906        if !self.read_only {
907            self.index_persist.forced = true;
908            if let Err(error) = self.flush() {
909                return Err(CloseError {
910                    handle: Box::new(self),
911                    error,
912                });
913            }
914        }
915        drop(self);
916        Ok(())
917    }
918
919    /// PERF-1: flush, persisting the ANN index even if the
920    /// [`INDEX_PERSIST_EVERY_COMMITS`] / [`INDEX_PERSIST_EVERY_CHANGES`]
921    /// policy wouldn't yet. Plain [`Mnemo::flush`] writes only the catalog
922    /// between persists; the index is derived, so a crash costs nothing but
923    /// re-assigning the newer records on the next load.
924    pub fn flush_index(&mut self) -> Result<()> {
925        if self.read_only {
926            return Err(MnemoError::ReadOnly);
927        }
928        self.index_persist.forced = true;
929        self.flush()
930    }
931
932    /// **TEST ONLY — do not use.** Reproduces a crash inside
933    /// [`Mnemo::flush`] *after* the prelude has leased counter+page slots
934    /// and persisted the leased header, but *before* the WAL is committed.
935    /// Runs `prepare_for_flush` + `pager.flush` and returns — leaving
936    /// data pages and a leased header on disk, with no commit frame.
937    ///
938    /// Used by `tests/integration.rs::nonce_unique_after_crashed_data_flush`
939    /// to verify that this window (Phase 1.1 of the improvement plan) does
940    /// not enable AES-GCM nonce reuse. Calling this in production code
941    /// strands data pages; never expose it from a binding.
942    #[doc(hidden)]
943    #[cfg(feature = "dev-hooks")]
944    pub fn __crash_partial_flush_for_testing(&mut self) -> Result<()> {
945        if self.read_only {
946            return Err(MnemoError::ReadOnly);
947        }
948        if !self.dirty_catalog && !self.dirty_index {
949            return Ok(());
950        }
951        // FMT-9b: v9 records and vectors reach their pages at the top of a
952        // flush; do that part too.
953        self.ensure_gen_lease()?;
954        self.pack_records()?;
955        self.pack_vectors()?;
956        let _prelude = self.prepare_for_flush(self.dirty_index)?;
957        self.pager.flush()
958    }
959
960    /// Bound the in-memory page cache to `pages` decrypted pages.
961    ///
962    /// The cache holds decrypted page payloads to speed repeated reads. By
963    /// default it is capped at 1024 pages (~8 MiB; 8192 before v0.10.0).
964    /// Lower the cap to trade hit rate for a smaller footprint, or raise
965    /// it for a hotter cache. The
966    /// cap governs *clean* pages — pages with un-flushed writes are always
967    /// retained until [`Mnemo::flush`], regardless of the cap.
968    pub fn set_cache_capacity(&mut self, pages: usize) {
969        self.pager.set_cache_capacity(pages);
970    }
971
972    /// FMT-9b (v9 files): keep the fp16 vector column in memory while it
973    /// takes at most `bytes` (default [`DEFAULT_VECTOR_CACHE_BUDGET`],
974    /// 96 MiB, about 125K vectors at 384 dims). Above it, recall reads
975    /// vector entries through the page cache instead: less memory, slower
976    /// recall. Lowering it below the resident size drops the resident copy.
977    pub fn set_vector_cache_budget(&mut self, bytes: usize) {
978        self.vector_cache_budget = bytes;
979        let need = self.col.scored_rows * veccol::entry_len(self.dimensions);
980        if self.col.resident && need > bytes {
981            self.col.data = Vec::new();
982            self.col.norms = Vec::new();
983            self.col.have = Vec::new();
984            self.col.resident = false;
985        }
986    }
987
988    /// Page-cache occupancy: `(pages_cached, capacity)`. Renamed from
989    /// `cache_stats` in v0.4.0: the result-cache API (Phase 10.1) took
990    /// that name for its per-namespace [`CacheStats`] summary, and the
991    /// page-cache pair is the older/less-user-facing of the two, so it
992    /// gets the more-specific name.
993    pub fn page_cache_stats(&self) -> (usize, usize) {
994        self.pager.cache_stats()
995    }
996
997    // === Result cache (Phase 10.1 + 10.3) ============================
998}