jeremy/quick_search
jeremy/quick_search
1
0
Fork 0
Code Issues Pull requests Projects Releases 6 Packages Wiki Activity Actions

Parallel per-root content extraction, simpler query parsing, WAL/vacuum maintenance, GUI index size display.

Browse source
This commit is contained in:
= 2026-08-03 03:06:19 -04:00
parent b85c99e277
commit ccbe4f1c8c
43 changed files with 6106 additions and 1712 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
Cargo.lock generated
View file

@ -3179,7 +3179,7 @@ dependencies = [
[[package]]
name = "quicksearch-core"
version = "0.8.5"
version = "0.8.8"
dependencies = [
"argon2",
"ctrlc",
@ -3187,6 +3187,7 @@ dependencies = [
"globset",
"infer",
"kamadak-exif",
"libc",
"lofty",
"lopdf 0.32.0",
"mime_guess",
@ -3207,7 +3208,7 @@ dependencies = [
[[package]]
name = "quicksearch-gui"
version = "0.8.5"
version = "0.8.8"
dependencies = [
"chrono",
"eframe",

2
Cargo.toml
View file

@ -6,7 +6,7 @@ members = [
]
[workspace.package]
version = "0.8.6"
version = "0.8.8"
edition = "2021"
license = "GPL-3.0-or-later"
authors = ["Jeremy <jeremy@karsttech.com>"]

18
README.md
View file

@ -120,7 +120,10 @@ the package is installed.
indexed folder list, full-text extension filters, ignore patterns, and
the indexing options. Stopping switches to manual mode and saves that
(`indexing.auto_index`), so a stopped index stays stopped across
restarts until you return to automatic.
restarts until you return to automatic. The index size beside the
status heading totals the database and its `-wal`/`-shm` sidecars,
refreshed every ten seconds; hovering it lists the ways to make it
smaller.
- **Duplicates**: files sharing a content hash, grouped.
- **Logs**: the lines the app would have printed to a terminal — warnings
from indexing, folder watching and opening files, newest last, with a
@ -264,7 +267,11 @@ Synchronous Rust: `std::thread` + `mpsc` channels, no async runtime.
- **Storage** (`db/`): SQLite via rusqlite (bundled SQLCipher build —
identical to stock SQLite until a key is applied), WAL mode so the
single writer never blocks streaming read-only searches. `files` holds
single writer never blocks streaming read-only searches. A run forces a
`wal_checkpoint(TRUNCATE)` every `processing.maximum_wal_size` bytes of
log, because SQLite's own autocheckpoint can only reset the log at an
instant no reader holds it — and a run keeps a reader per root querying
throughout, so left alone the log grows for the whole run. `files` holds
metadata (name, path, size, mtime, hash, MIME/type bitmask, per-row
index state); `searchabletext` is a *contentless* FTS5 table (postings
only, configurable tokenizer, trigram by default); canonical extracted
@ -284,8 +291,11 @@ Synchronous Rust: `std::thread` + `mpsc` channels, no async runtime.
tags, EXIF; see `extract/`) for FTS. Files no larger than
`processing.hash_length` skip that second pass entirely: the head the walk
reads to hash them is already their whole content, so a plaintext body is
extracted in the same `read` and stored complete. Progress streams through
a polled `IndexingStatus`.
extracted in the same `read` and stored complete. Every run ends — whether
it completed or was stopped — with an optimize pass on its own connection:
checkpoint, VACUUM if the file has at least 10% slack to reclaim, `PRAGMA
optimize`, checkpoint again. Progress streams through a polled
`IndexingStatus`, which reads `Optimizing` for the duration of that pass.
- **Coordinator** (`coordinator.rs`): the object binaries construct.
Owns the `IndexingService`, the debouncing filesystem watcher
(`watcher.rs`), and the mode state machine (Auto / Manual, persisted as

25
config_example.toml
View file

@ -21,16 +21,23 @@ indexing_paths = ["~"]
database_path = "~/.local/share/quicksearch/index.sqlite"
[indexing]
# The indexing mode. true = automatic: filesystem watchers apply changes as
# they happen and a full reindex runs every reindex_interval_minutes (the
# watcher catches changes as they happen, so that interval only needs to be
# The indexing mode. true = automatic: filesystem watchers apply changes a
# couple of seconds after they happen — the queue waits for a burst to settle
# so that deleting a folder costs one operation instead of one per file — and
# a full reindex runs every reindex_interval_minutes (the watcher catches
# changes as they happen, so that interval only needs to be
# often enough to cover whatever the watcher missed). false = manual:
# nothing is indexed until you ask for it. The Stop and Return to Automatic
# buttons on the Manage Index tab write this value, so the mode you left the
# app in is the mode it starts in.
auto_index = true
reindex_interval_minutes = 1440
# Follow symbolic links during directory walks.
# Follow symbolic links during directory walks. Applies to links pointing at
# files as well as at directories: with this off a symlink is not resolved at
# all, so its target is never indexed — which matters because a target can
# live outside every folder listed above. A resolved target is stored under
# its own real path, not the link's. Changing this changes what is in the
# index, so it prompts for a rebuild.
follow_symlinks = false
# Index hidden files and directories. That means dot-files everywhere, and
# additionally anything carrying the Hidden or System attribute on Windows
@ -80,9 +87,17 @@ maximum_text_size = 262144
# Files larger than this skip text extraction entirely (bytes).
maximum_text_file_size = 2097152
# Files per batch during walks / inserts / extraction.
batch_size = 200
batch_size = 500
# Files per transaction for incremental FTS updates.
fts_update_batch_size = 1000
# How large the write-ahead log (index.sqlite-wal) may grow during an
# indexing run before the indexer forces a checkpoint (bytes). SQLite copies
# the log into the index on its own but can only reset it when no reader is
# mid-query, and a run keeps one reader per root busy throughout — so
# unattended the log grows for the whole run and can end up larger than the
# index. Set to 0 to disable forced checkpoints; any other value below
# 16777216 is raised to it.
maximum_wal_size = 536870912
# FTS5 tokenizer: 'trigram' (substring matching, the default; gets
# remove_diacritics 1 appended), 'unicode61', 'porter', or a full FTS5
# option string. See https://www.sqlite.org/fts5.html#tokenizers

13
crates/quicksearch-core/Cargo.toml
View file

@ -43,6 +43,13 @@ zstd = "0.13"
globset = "0.4"
regex = "1"
# `nice()`, for dropping indexing threads to background scheduling priority.
# Linux schedules per task, so it affects only the calling thread. Already in
# the lockfile transitively (rusqlite, getrandom), so naming it directly
# compiles nothing new.
[target.'cfg(unix)'.dependencies]
libc = "0.2"
# `GetDriveTypeW` (a mapped drive letter is the only way to spot an SMB share
# that isn't written as UNC) plus the FILE_ATTRIBUTE_* constants for hidden
# detection. Pinned to 0.52 deliberately: walkdir → winapi-util already
@ -50,8 +57,12 @@ regex = "1"
[target.'cfg(windows)'.dependencies]
# GetDriveTypeW and the FILE_ATTRIBUTE_* constants live in
# Win32_Storage_FileSystem; DRIVE_REMOTE, oddly, is filed under
# Win32_System_WindowsProgramming.
# Win32_System_WindowsProgramming. Win32_System_Threading carries
# SetThreadPriority and THREAD_MODE_BACKGROUND_BEGIN, and both it and
# GetCurrentThread need Win32_Foundation for HANDLE/BOOL.
windows-sys = { version = "0.52", features = [
"Win32_Foundation",
"Win32_Storage_FileSystem",
"Win32_System_Threading",
"Win32_System_WindowsProgramming",
] }

4
crates/quicksearch-core/examples/walkprobe.rs
View file

@ -31,7 +31,7 @@
//! cache (or, on a share, the client's attribute cache), so the second is the
//! one to compare.
use std::sync::atomic::AtomicBool;
use std::sync::{Arc, Mutex};
use std::sync::Arc;
use std::time::{Instant, UNIX_EPOCH};
use quicksearch_core::config::{Config, IgnoreSet};
@ -118,7 +118,7 @@ fn parallel(root: &str, config: &Config, db_path: &str) -> (usize, usize) {
db_path,
config.clone(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
) {

13
crates/quicksearch-core/src/cli.rs
View file

@ -173,10 +173,11 @@ pub fn index_size_breakdown(db_path: &str) -> Result<SizeReport, String> {
})
}
/// Count files with `content_state = 0` (pending). Distinct from
/// `files_row_count − searchabletext_row_count`, which over-counts because
/// files whose content doesn't apply (binary formats, too-large files) sit
/// with `content_state = 3` (NA) forever and never become FTS rows.
/// Count files with `content_state = 0` (pending) — files an extractor claims
/// whose text has not been read yet. Files nothing extracts (binary formats,
/// too-large files) are written `content_state = 3` (NA) when the walk records
/// them and are never counted here, so this is outstanding work rather than
/// `files_row_count − searchabletext_row_count`, which counts those forever.
///
/// Used by the Baloo compat daemon to report the "Files waiting for content
/// indexing" figure both to balooctl and to the LMDB mirror.
@ -263,6 +264,7 @@ mod tests {
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
needs_content: true,
},
)
.unwrap()
@ -281,6 +283,7 @@ mod tests {
mime: None,
ftype: FileType::EMPTY,
hash: None,
needs_content: false,
},
)
.unwrap()
@ -366,6 +369,7 @@ mod tests {
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
needs_content: true,
},
)
.unwrap()
@ -424,6 +428,7 @@ mod tests {
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
needs_content: true,
},
)
.unwrap()

48
crates/quicksearch-core/src/config.rs
View file

@ -61,7 +61,8 @@ pub struct IndexingConfig {
/// supports. Non-empty = only files with these extensions get content
/// extraction/FTS; everything else is still listed for filename search
/// (`content_state = NA`). Entries are case-insensitive, with or
/// without a leading dot.
/// without a leading dot. Applied at walk time, when the row is written —
/// which is why changing this forces a rebuild (see [`diff_actions`]).
pub content_extensions: Vec<String>,
/// Excluded from the index entirely — never even listed. A pattern
/// without `/` matches any single path component (so `.git` prunes
@ -69,9 +70,11 @@ pub struct IndexingConfig {
/// matched against the full path. Glob syntax (`*`, `?`, `[..]`).
pub ignore_patterns: Vec<String>,
/// Per-root walker thread override, keyed by the root string exactly
/// as it appears in `indexing_paths`. Absent or 0 = auto-detect
/// (4 for local storage, 16 for network mounts). Read at run start;
/// a change applies to the next run.
/// as it appears in `indexing_paths` — the indexer resolves both sides
/// to the same canonical path before matching, so a key that spells its
/// root as `~/docs`, `/docs/` or a symlink still applies. Absent or
/// 0 = auto-detect (4 for local storage, 16 for network mounts). Read
/// at run start; a change applies to the next run.
pub root_workers: std::collections::HashMap<String, usize>,
}
@ -95,6 +98,18 @@ pub struct ProcessingConfig {
pub maximum_text_file_size: u64,
pub batch_size: usize,
pub fts_update_batch_size: usize,
/// How large the write-ahead log may grow during a run before the indexer
/// forces a checkpoint, in bytes. `0` disables forced checkpoints;
/// anything else is raised to [`MINIMUM_WAL_SIZE`] at the use site.
///
/// SQLite's own autocheckpoint copies the log into the index continuously
/// but can only *reset* it when no reader is mid-query, and a run keeps a
/// reader per root busy from start to finish — so left alone the log grows
/// for the whole run and can end up larger than the index itself. This is
/// a stall-frequency knob, not a throughput one: by the time a forced
/// checkpoint fires there is almost nothing left to copy, so it costs lock
/// acquisition and little else.
pub maximum_wal_size: u64,
pub tokenize: String,
/// When `true` (default), extracted text is stored zstd-compressed in
/// `documents_text` so search results can render snippet/highlight
@ -107,6 +122,12 @@ pub struct ProcessingConfig {
pub store_text_for_snippets: bool,
}
/// Floor on a non-zero [`ProcessingConfig::maximum_wal_size`]. A checkpoint
/// costs a lock acquisition that can wait on a running search, so a cap set
/// low enough to fire every round would trade a large log for a stalled
/// writer. Zero still means "never force one".
pub const MINIMUM_WAL_SIZE: u64 = 1024 * 1024 * 16;
/// Fuzzy edit distances above this are allowed but warned about: matches
/// become dominated by coincidence and every fuzzy pass slows down.
pub const FUZZY_EDITS_WARN_ABOVE: usize = 3;
@ -177,8 +198,9 @@ impl Default for ProcessingConfig {
hash_length: 1024 * 8,
maximum_text_size: 1024 * 256,
maximum_text_file_size: 1024 * 1024 * 2,
batch_size: 200,
batch_size: 500,
fts_update_batch_size: 1000,
maximum_wal_size: 1024 * 1024 * 512,
tokenize: "trigram".to_string(),
store_text_for_snippets: true,
}
@ -650,6 +672,11 @@ pub fn diff_actions(old: &Config, new: &Config) -> ConfigActions {
let requires_rebuild = old.processing.hash_length != new.processing.hash_length
|| old.processing.tokenize != new.processing.tokenize
|| old.indexing.include_hidden != new.indexing.include_hidden
// Not merely a walk-behaviour knob: with links off, a symlink's target
// is not indexed at all, so turning it off leaves rows for files that
// are no longer in scope — including ones whose real parent lies
// outside every root, which no sweep will ever reach.
|| old.indexing.follow_symlinks != new.indexing.follow_symlinks
|| old.indexing.ignore_patterns != new.indexing.ignore_patterns
|| old.indexing.content_extensions != new.indexing.content_extensions
|| old.security.password_protected != new.security.password_protected
@ -657,9 +684,7 @@ pub fn diff_actions(old: &Config, new: &Config) -> ConfigActions {
|| roots_changed;
ConfigActions {
requires_rebuild,
restart_watcher: requires_rebuild
|| roots_changed
|| old.indexing.follow_symlinks != new.indexing.follow_symlinks,
restart_watcher: requires_rebuild || roots_changed,
search_db_changed: old.paths.database_path != new.paths.database_path,
}
}
@ -727,7 +752,7 @@ mod tests {
fs::write(&path, "[paths]\nindexing_paths=[\"/x\"]\ndatabase_path=\"db.sqlite\"\n").unwrap();
let cfg = Config::load_from(&path).unwrap();
assert_eq!(cfg.paths.indexing_paths, vec!["/x".to_string()]);
assert_eq!(cfg.processing.batch_size, 200, "missing sections default");
assert_eq!(cfg.processing.batch_size, 500, "missing sections default");
assert_eq!(cfg.search.debounce_ms, 150);
assert!((cfg.ui.scale - 1.1).abs() < f32::EPSILON);
fs::remove_dir_all(&dir).ok();
@ -975,10 +1000,12 @@ mod tests {
let a = diff_actions(&base, &c);
assert!(a.requires_rebuild && a.restart_watcher);
// `follow_symlinks` decides what is in the index, not just how the walk
// moves, so it rebuilds as well as restarting the watcher.
let mut c = base.clone();
c.indexing.follow_symlinks = true;
let a = diff_actions(&base, &c);
assert!(!a.requires_rebuild && a.restart_watcher);
assert!(a.requires_rebuild && a.restart_watcher);
let mut c = base.clone();
c.paths.database_path = "/elsewhere.sqlite".into();
@ -988,6 +1015,7 @@ mod tests {
let mut c = base.clone();
c.search.display_limit = 5000;
c.processing.batch_size = 999;
c.processing.maximum_wal_size = 0;
c.indexing.auto_index = false;
c.indexing.reindex_interval_minutes = 5;
let a = diff_actions(&base, &c);

634
crates/quicksearch-core/src/content.rs Normal file
View file

@ -0,0 +1,634 @@
//! Parallel content extraction for one indexing root.
//!
//! The second half of a root's pipeline, and the sibling of [`crate::walk`]:
//! a pool of worker threads produces finished work over a bounded channel, and
//! the single writer drains it round-robin against every other root.
//!
//! That shape is the whole point. Extraction used to run *on* the writer
//! thread, a batch of files at a time, with the database connection held for
//! the duration — so one root reading a network share or a stack of large PDFs
//! stopped every other root dead for as long as its batch took, and blocked
//! the watcher's incremental writes with it. Roots are meant to be
//! independent; walking already was, and this makes the rest of the pipeline
//! match.
//!
//! The split within the pass mirrors the walk's too: **one feeder thread owns
//! the only database connection**, paging through the root's pending rows,
//! while N workers do nothing but filesystem work. A connection per worker
//! would multiply SQLite's page cache by the pool size, which is exactly what
//! [`crate::db::schema::PRAGMAS_WALK_READER`] exists to avoid.
//!
//! Deliberately std-only (`std::thread` + `std::sync::mpsc`), matching the
//! house style of [`crate::walk`] and [`crate::watcher`].
use std::sync::atomic::AtomicBool;
use std::sync::{mpsc, Arc, Condvar, Mutex};
use std::thread::{self, JoinHandle};
use crate::config::Config;
use crate::extract::Registry;
use crate::file_handling::{decide_content, ContentOutcome, ExtractCursor};
use crate::indexing::should_abort;
use crate::walk::{try_recv_next, TryNext, WorkerStats};
/// Finished rows waiting for the writer.
///
/// Far shallower than the walk's 4096. An [`ExtractedRow`] carries up to
/// `maximum_text_size` of extracted text — 256 KiB by default, some 32× a
/// `WalkedFile` — so the walk's depth would put gigabytes in flight. At 32 the
/// ceiling is ~8 MiB per root, plus whatever the workers hold mid-document.
const READY_CAP: usize = 32;
/// Rows fetched but not yet claimed by a worker.
///
/// Small for the same reason [`crate::walk`] caps its prefetch: the feeder
/// must not run arbitrarily far ahead of a pool that is bound by file I/O.
/// These are only ids and paths, so the cost is the strings, not the text.
const QUEUE_AHEAD: usize = 256;
/// How many rows the feeder fetches per query. Large enough that a slow root
/// is not paying a round trip per file, small enough to stay inside
/// [`QUEUE_AHEAD`].
const FEED_PAGE: usize = 128;
/// One file's extracted content, ready to be written.
#[derive(Debug)]
pub struct ExtractedRow {
pub file_id: i64,
/// The `files.name` the FTS row is indexed under.
pub name: String,
pub outcome: ContentOutcome,
}
/// A row the feeder handed to the pool: everything a worker needs, and nothing
/// that would make it touch the database.
#[derive(Debug)]
struct Pending {
file_id: i64,
name: String,
path: String,
mime: Option<String>,
}
#[derive(Default)]
struct Queue {
rows: Vec<Pending>,
/// Set while the feeder is mid-query, holding rows that are in neither the
/// queue nor a worker. Without it a worker could see an empty queue
/// between two pages and declare the pass finished early.
feeding: bool,
/// True once the feeder has read the last page.
drained: bool,
done: bool,
}
struct Shared {
queue: Mutex<Queue>,
idle: Condvar,
}
impl Shared {
/// Claim a row, blocking while the feeder might still produce more.
///
/// `None` only when the queue is empty *and* the feeder is finished — at
/// that instant nobody is left who could add another row.
fn take(&self) -> Option<Pending> {
let mut q = self.queue.lock().unwrap();
loop {
if q.done {
return None;
}
if let Some(row) = q.rows.pop() {
// The feeder may be parked behind QUEUE_AHEAD.
self.idle.notify_all();
return Some(row);
}
if q.drained && !q.feeding {
q.done = true;
self.idle.notify_all();
return None;
}
q = self.idle.wait(q).unwrap();
}
}
/// Claim the right to fetch one page, or `None` once the pass is over.
/// Parks while the queue is already [`QUEUE_AHEAD`] deep.
fn take_feed_slot(&self) -> Option<()> {
let mut q = self.queue.lock().unwrap();
loop {
if q.done || q.drained {
return None;
}
if q.rows.len() < QUEUE_AHEAD {
q.feeding = true;
return Some(());
}
q = self.idle.wait(q).unwrap();
}
}
/// Publish a page and clear the in-flight flag together, under one lock —
/// the indivisibility [`Shared::take`]'s end-of-pass test relies on.
fn finish_feed(&self, rows: Vec<Pending>, last_page: bool) {
let mut q = self.queue.lock().unwrap();
// Reversed: `take` pops from the back, and rows should reach workers
// in id order so a partial run leaves a contiguous prefix done.
q.rows.extend(rows.into_iter().rev());
q.feeding = false;
if last_page {
q.drained = true;
}
self.idle.notify_all();
}
fn shutdown(&self) {
let mut q = self.queue.lock().unwrap();
q.done = true;
q.drained = true;
self.idle.notify_all();
}
}
/// A running content pass. Draining it yields finished rows; dropping it stops
/// the workers and joins them.
pub struct ContentPass {
rx: Option<mpsc::Receiver<ExtractedRow>>,
handles: Vec<JoinHandle<()>>,
feeder: Option<JoinHandle<()>>,
shared: Arc<Shared>,
stats: WorkerStats,
}
impl ContentPass {
/// Non-blocking pull, for the writer loop multiplexing several roots.
pub fn try_next(&mut self) -> TryNext<ExtractedRow> {
try_recv_next(self.rx.as_ref())
}
/// A cheap, cloneable handle for reading pool activity while the pass is
/// mutably borrowed by the writer loop. The sibling of
/// [`crate::walk::ParallelWalk::worker_stats`], and for the same reason:
/// once a root reaches extraction its walk pool is gone, so the progress
/// display has to read this one instead.
pub fn worker_stats(&self) -> WorkerStats {
self.stats.clone()
}
/// Join the workers and report whether every one finished cleanly.
///
/// The caller needs this for the same reason the walk does: a dead worker
/// and a finished worker both close the channel, so from the receiving end
/// they are indistinguishable.
pub fn finish(&mut self) -> bool {
// Dropping the receiver first releases any worker parked in `send`.
self.rx = None;
self.shared.shutdown();
let mut clean = true;
for handle in self.handles.drain(..) {
if handle.join().is_err() {
clean = false;
}
}
if let Some(handle) = self.feeder.take() {
if handle.join().is_err() {
clean = false;
}
}
clean
}
}
impl Drop for ContentPass {
fn drop(&mut self) {
self.shared.shutdown();
// No-op if the caller already called `finish`.
self.finish();
}
}
/// Page the root's pending rows into the queue from one read-only connection.
///
/// A failed query ends the pass rather than retrying: the rows stay
/// `content_state = 0` and the next run picks them up, which is the same
/// outcome as being interrupted.
fn feeder(shared: &Shared, db_path: &str, mut cursor: ExtractCursor, max_size: i64) {
let conn = match crate::db::open::open_walk_reader(db_path) {
Ok(conn) => conn,
Err(e) => {
crate::log_warn!("content reader: {}", e);
shared.shutdown();
return;
}
};
while shared.take_feed_slot().is_some() {
let page = match crate::db::repo::pending_content_page(
&conn,
&cursor,
max_size,
FEED_PAGE as i64,
) {
Ok(page) => page,
Err(e) => {
crate::log_warn!("{}", e);
shared.shutdown();
return;
}
};
let last_page = page.len() < FEED_PAGE;
if let Some((id, _, _, _)) = page.last() {
cursor.last_id = *id;
}
let rows = page
.into_iter()
.map(|(file_id, name, path, mime)| Pending {
file_id,
name,
path,
mime,
})
.collect();
shared.finish_feed(rows, last_page);
if last_page {
return;
}
}
}
fn worker(
shared: &Shared,
tx: &mpsc::SyncSender<ExtractedRow>,
registry: &Registry,
config: &Config,
stop_flag: &Arc<AtomicBool>,
suspend_flag: &Arc<AtomicBool>,
stats: &WorkerStats,
) {
while let Some(row) = shared.take() {
// Held for the whole of `decide_content` — reading and parsing a
// document is exactly the work the progress line is reporting.
let _busy = stats.enter();
if should_abort(stop_flag, suspend_flag) {
shared.shutdown();
return;
}
let outcome = decide_content(&row.path, row.mime.as_deref(), registry, config);
let sent = tx.send(ExtractedRow {
file_id: row.file_id,
name: row.name,
outcome,
});
if sent.is_err() {
// Receiver gone: the run was stopped or failed. Not an error.
shared.shutdown();
return;
}
}
}
/// Extract every pending row under `cursor`'s range, in parallel.
///
/// `workers` is the root's own count — the same value its walk uses, because
/// extraction over a share is round-trip bound for the same reason walking is.
/// Clamped to 1..=64.
#[allow(clippy::too_many_arguments)]
pub fn extract_content(
db_path: &str,
cursor: &ExtractCursor,
registry: Arc<Registry>,
config: Config,
stop_flag: Arc<AtomicBool>,
suspend_flag: Arc<AtomicBool>,
workers: usize,
) -> ContentPass {
let shared = Arc::new(Shared {
queue: Mutex::new(Queue::default()),
idle: Condvar::new(),
});
let max_size = i64::try_from(config.processing.maximum_text_file_size).unwrap_or(i64::MAX);
let (tx, rx) = mpsc::sync_channel(READY_CAP);
let stats = WorkerStats::new(workers.clamp(1, 64));
let handles = (0..stats.total())
.map(|_| {
let (shared, tx) = (shared.clone(), tx.clone());
let (registry, config) = (registry.clone(), config.clone());
let (stop_flag, suspend_flag) = (stop_flag.clone(), suspend_flag.clone());
let stats = stats.clone();
thread::spawn(move || {
crate::platform::set_background_priority();
worker(&shared, &tx, &registry, &config, &stop_flag, &suspend_flag, &stats)
})
})
.collect();
// The workers must hold the only senders, or `try_recv` never reports the
// end of the pass. The feeder deliberately holds none: it produces rows to
// extract, not extracted rows.
drop(tx);
let feeder_handle = {
let (shared, db_path, cursor) = (shared.clone(), db_path.to_string(), cursor.clone());
thread::spawn(move || {
crate::platform::set_background_priority();
feeder(&shared, &db_path, cursor, max_size)
})
};
ContentPass {
rx: Some(rx),
handles,
feeder: Some(feeder_handle),
shared,
stats,
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::db::open_or_recreate;
use crate::db::repo::{self, insert_file, NewFile};
use crate::file_handling::{extract_scope_prepare, store_extracted};
use crate::mime::FileType;
use std::path::{Path, PathBuf};
use std::time::UNIX_EPOCH;
fn tmp(tag: &str) -> PathBuf {
let mut p = std::env::temp_dir();
p.push(format!(
"qs-content-{}-{}-{}",
tag,
std::process::id(),
std::time::SystemTime::now()
.duration_since(UNIX_EPOCH)
.unwrap()
.as_nanos()
));
p
}
/// A tree of `n` text files under `root/sub`, plus an index holding a
/// pending row for each.
fn seed(tag: &str, dirs: &[(&str, usize)]) -> (PathBuf, PathBuf) {
let tree = tmp(&format!("{}-tree", tag));
let db = tmp(&format!("{}-db", tag));
let mut conn = open_or_recreate(db.to_str().unwrap(), "trigram").unwrap();
let tx = conn.transaction().unwrap();
for (dir, n) in dirs {
let d = tree.join(dir);
std::fs::create_dir_all(&d).unwrap();
for i in 0..*n {
let f = d.join(format!("f{:04}.txt", i));
std::fs::write(&f, format!("sphinx of black quartz {} {}", dir, i)).unwrap();
insert_file(
&tx,
&NewFile {
name: f.file_name().unwrap().to_str().unwrap(),
path: f.to_str().unwrap(),
parent: d.to_str().unwrap(),
size: std::fs::metadata(&f).unwrap().len(),
mtime: 1,
inode: None,
device_id: None,
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
needs_content: true,
},
)
.unwrap()
.expect("unique path");
}
}
tx.commit().unwrap();
drop(conn);
(tree, db)
}
fn pass_for(tree: &Path, db: &Path, sub: &str, workers: usize) -> ContentPass {
extract_content(
db.to_str().unwrap(),
&ExtractCursor::for_root(tree.join(sub).to_str().unwrap()),
Arc::new(Registry::default_set()),
Config::default(),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
workers,
)
}
/// Drain a pass to exhaustion, blocking between polls the way the writer
/// loop's outer sleep does.
fn drain(pass: &mut ContentPass) -> Vec<ExtractedRow> {
let mut out = Vec::new();
loop {
match pass.try_next() {
TryNext::Item(row) => out.push(row),
TryNext::Empty => thread::sleep(std::time::Duration::from_millis(1)),
TryNext::Finished => return out,
}
}
}
#[test]
fn every_pending_row_is_yielded_exactly_once() {
let (tree, db) = seed("once", &[("r1", 250)]);
let mut pass = pass_for(&tree, &db, "r1", 4);
let rows = drain(&mut pass);
assert!(pass.finish(), "no worker panicked");
assert_eq!(rows.len(), 250);
let ids: std::collections::HashSet<i64> = rows.iter().map(|r| r.file_id).collect();
assert_eq!(ids.len(), 250, "no id may be yielded twice");
assert!(
rows.iter()
.all(|r| matches!(r.outcome, ContentOutcome::Done { .. })),
"every plaintext file extracts"
);
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
/// The pass is scoped by the cursor's path range, so a sibling root's rows
/// are never touched. (Moved here with the extraction itself: the scoping
/// is now the feeder's query, not the writer's.)
#[test]
fn the_pass_is_scoped_to_its_root_range() {
let (tree, db) = seed("scope", &[("r1", 3), ("r2", 3)]);
let conn_mutex = Arc::new(Mutex::new(
open_or_recreate(db.to_str().unwrap(), "trigram").unwrap(),
));
let config = Config::default();
let cursor = ExtractCursor::for_root(tree.join("r1").to_str().unwrap());
let scope = extract_scope_prepare(&conn_mutex, &cursor, &config).unwrap();
assert_eq!(scope.pending, 3, "only r1's files are in range");
assert_eq!(scope.already_done, 0, "nothing extracted yet");
let mut pass = pass_for(&tree, &db, "r1", 2);
let rows = drain(&mut pass);
assert!(pass.finish());
assert_eq!(rows.len(), 3);
let stop = Arc::new(AtomicBool::new(false));
assert_eq!(
store_extracted(&conn_mutex, &rows, &stop, &config).unwrap(),
3
);
let state = |p: &Path| -> i64 {
conn_mutex
.lock()
.unwrap()
.query_row(
"SELECT content_state FROM files WHERE path = ?1",
rusqlite::params![p.to_str().unwrap()],
|r| r.get(0),
)
.unwrap()
};
assert_eq!(state(&tree.join("r1/f0000.txt")), repo::STATE_DONE);
assert_eq!(
state(&tree.join("r2/f0000.txt")),
repo::STATE_PENDING,
"out-of-range row untouched"
);
// A second run over the unchanged root reports it already extracted,
// so progress reads "3 of 3" rather than "0 of 0".
let scope2 = extract_scope_prepare(&conn_mutex, &cursor, &config).unwrap();
assert_eq!((scope2.pending, scope2.already_done), (0, 3));
let hits: i64 = conn_mutex
.lock()
.unwrap()
.query_row(
"SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH '\"sphinx\"'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(hits, 3);
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
#[test]
fn an_empty_range_terminates_immediately() {
// The "nothing to do at t=0" corner: every worker must observe the pass
// as finished rather than waiting for rows that will never arrive.
let (tree, db) = seed("empty", &[("r1", 2)]);
let mut pass = pass_for(&tree, &db, "nonexistent", 4);
assert!(drain(&mut pass).is_empty());
assert!(pass.finish());
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
#[test]
fn an_already_stopped_pass_does_not_run_to_completion() {
let (tree, db) = seed("stop", &[("r1", 400)]);
let mut pass = extract_content(
db.to_str().unwrap(),
&ExtractCursor::for_root(tree.join("r1").to_str().unwrap()),
Arc::new(Registry::default_set()),
Config::default(),
Arc::new(AtomicBool::new(true)),
Arc::new(AtomicBool::new(false)),
4,
);
assert!(drain(&mut pass).len() < 400);
assert!(pass.finish());
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
#[test]
fn dropping_the_pass_early_does_not_hang() {
// Workers blocked in `send` must be released by the receiver going
// away, or `Drop` would join threads that never wake.
let (tree, db) = seed("early-drop", &[("r1", 500)]);
let mut pass = pass_for(&tree, &db, "r1", 4);
// Pull one, leave the rest queued and the channel full.
loop {
match pass.try_next() {
TryNext::Item(_) => break,
TryNext::Empty => thread::sleep(std::time::Duration::from_millis(1)),
TryNext::Finished => break,
}
}
drop(pass); // must return, not deadlock
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
#[test]
fn repeated_passes_agree_on_the_result_set() {
// The termination protocol is racy by nature; run it enough times under
// real contention that a premature exit would show up.
let (tree, db) = seed("repeat", &[("r1", 120)]);
for run in 0..20 {
let mut pass = pass_for(&tree, &db, "r1", 4);
let rows = drain(&mut pass);
assert!(pass.finish(), "run {}", run);
assert_eq!(rows.len(), 120, "run {}", run);
}
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
/// The pass counts its own busy threads, which is what the progress line
/// shows once a root leaves the walk behind.
#[test]
fn the_pool_reports_its_own_activity() {
let (tree, db) = seed("stats", &[("r1", 300)]);
let mut pass = pass_for(&tree, &db, "r1", 4);
let stats = pass.worker_stats();
assert_eq!(stats.total(), 4);
// Nothing is drained here, so the ready channel fills and every worker
// ends up parked mid-row inside `send` — busy by the definition the
// display uses, and there are more rows than the channel holds, so all
// four get there.
let mut peak = 0;
for _ in 0..500 {
peak = peak.max(stats.active());
if peak == 4 {
break;
}
thread::sleep(std::time::Duration::from_millis(2));
}
assert_eq!(peak, 4, "every worker busy while the channel is full");
assert_eq!(drain(&mut pass).len(), 300);
assert!(pass.finish());
assert_eq!(stats.active(), 0, "a finished pool is idle");
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
/// A file that vanished between the walk and extraction is a failure with
/// a reason, not a silent skip: the row records why so it is not retried
/// forever.
#[test]
fn a_missing_file_is_reported_as_failed() {
let (tree, db) = seed("missing", &[("r1", 2)]);
std::fs::remove_file(tree.join("r1/f0000.txt")).unwrap();
let mut pass = pass_for(&tree, &db, "r1", 2);
let rows = drain(&mut pass);
assert!(pass.finish());
assert_eq!(rows.len(), 2, "both rows are still reported");
assert_eq!(
rows.iter()
.filter(|r| matches!(r.outcome, ContentOutcome::Failed(_)))
.count(),
1
);
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_file(&db).ok();
}
}

304
crates/quicksearch-core/src/coordinator.rs
View file

@ -27,7 +27,7 @@ use std::path::PathBuf;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{mpsc, Arc, Mutex};
use std::thread::JoinHandle;
use std::time::Duration;
use std::time::{Duration, Instant};
use rusqlite::Connection;
@ -142,6 +142,8 @@ impl IndexCoordinator {
watcher_rx: None,
watcher_gen: 0,
pending: HashMap::new(),
last_event_at: None,
pending_since: None,
needs_full_run: false,
saw_running: false,
write_conn: None,
@ -259,6 +261,50 @@ fn enqueue(pending: &mut HashMap<PathBuf, FsEvent>, event: FsEvent) {
}
}
/// Whether a queued event is a removal. The queue only ever holds
/// Create/Modify/Remove — [`enqueue`] splits renames into their halves.
fn is_removal(event: &FsEvent) -> bool {
matches!(event, FsEvent::Remove(_))
}
/// Drop queued removals that a queued removal of one of their ancestors
/// already covers, in place.
///
/// `rm -rf dir/` reports `dir` and every path beneath it; applying `dir` sweeps
/// the whole range, so the descendants are duplicate work. Collapsing here —
/// rather than at application time — is what keeps a mass deletion from
/// tripping [`PENDING_OVERFLOW`] and forcing a redundant full run.
///
/// Only removals collapse against removals. A `Create` under a removed
/// directory is a re-creation and must survive: removals are applied first, so
/// it lands afterwards and the row is correct either way.
fn collapse_pending_removals(pending: &mut HashMap<PathBuf, FsEvent>) {
if pending.values().filter(|e| is_removal(e)).take(2).count() < 2 {
return;
}
let removed: std::collections::HashSet<PathBuf> = pending
.iter()
.filter(|(_, ev)| is_removal(ev))
.map(|(p, _)| p.clone())
.collect();
// `Path::ancestors` walks whole components, so `/a/bc` is never treated as
// living under `/a/b` — the rule `remove_tree` and `UnreadableDirs::covers`
// also use.
pending.retain(|path, ev| {
!is_removal(ev) || !path.ancestors().skip(1).any(|a| removed.contains(a))
});
}
/// How long [`Inner::apply_pending`] may hold the command loop before handing
/// the rest of the queue to the next tick.
///
/// A long run over a busy tree can accumulate tens of thousands of events, and
/// draining them all inline would block Stop and Apply behind the backlog. A
/// deadline bounds that latency without throttling throughput the way a fixed
/// event budget would; `queued_events` already reports the remainder, so a
/// multi-tick drain is visible rather than silent.
const APPLY_BUDGET: Duration = Duration::from_millis(250);
struct Inner {
config: Config,
indexing: Arc<IndexingService>,
@ -271,6 +317,12 @@ struct Inner {
watcher_rx: Option<mpsc::Receiver<(u64, Result<Watcher, WatchError>)>>,
watcher_gen: u64,
pending: HashMap<PathBuf, FsEvent>,
/// When the most recent event arrived; the burst is over once this is
/// `pending_settle` old.
last_event_at: Option<Instant>,
/// When the oldest un-applied event arrived, so a steady trickle cannot
/// defer application past `pending_max_defer`.
pending_since: Option<Instant>,
needs_full_run: bool,
/// A start was requested; set false once the service reports running,
/// so idle-after-running transitions are detectable.
@ -363,9 +415,13 @@ impl Inner {
let status = self.indexing.get_status();
match status {
IndexingStatus::Running { .. } | IndexingStatus::Stopping => {
IndexingStatus::Running { .. }
| IndexingStatus::Stopping
| IndexingStatus::Optimizing => {
// Single-writer rule: never touch the DB while a full run
// is active; the queue drains on a later tick.
// is active; the queue drains on a later tick. Optimizing
// counts — it holds a write transaction over the whole file
// for as long as the rewrite takes.
self.saw_running = true;
return;
}
@ -384,12 +440,12 @@ impl Inner {
if self.mode != IndexMode::Auto {
if self.mode == IndexMode::ManualStopped {
self.pending.clear();
self.clear_pending();
}
return;
}
if !self.pending.is_empty() && !self.needs_full_run {
if !self.pending.is_empty() && !self.needs_full_run && self.pending_settled() {
self.apply_pending();
}
@ -399,19 +455,58 @@ impl Inner {
}
fn drain_events(&mut self) {
let mut received = false;
while let Ok(ev) = self.event_rx.try_recv() {
enqueue(&mut self.pending, ev);
received = true;
}
if received {
let now = Instant::now();
self.last_event_at = Some(now);
self.pending_since.get_or_insert(now);
// Before the overflow test, not after: an `rm -rf` of half a
// million files collapses to a handful of directory roots, and
// measuring the queue by its raw event count would throw all of
// them away and schedule a full run instead.
collapse_pending_removals(&mut self.pending);
}
if self.pending.len() > PENDING_OVERFLOW {
// Replaying a storm one file at a time is slower than one
// incremental full run (unchanged files skip on mtime).
self.pending.clear();
self.clear_pending();
self.needs_full_run = true;
}
}
/// Drop the queue and the timers that describe it, so a stale
/// `pending_since` cannot force an immediate apply of the next event.
fn clear_pending(&mut self) {
self.pending.clear();
self.last_event_at = None;
self.pending_since = None;
}
/// Whether the queue has gone quiet long enough to be worth applying, or
/// has waited long enough that it must be applied regardless.
fn pending_settled(&self) -> bool {
let quiet = self
.last_event_at
.is_none_or(|t| t.elapsed() >= self.watcher_config.pending_settle);
let overdue = self
.pending_since
.is_some_and(|t| t.elapsed() >= self.watcher_config.pending_max_defer);
quiet || overdue
}
/// Apply as much of the queue as fits in [`APPLY_BUDGET`], removals first.
///
/// Removals lead deliberately. The queue is an unordered map, so the old
/// arbitrary application order could delete a row a `Create` in the same
/// batch had just written. With removals first both interleavings converge
/// on the truth: a delete-then-recreate ends with the file present, and a
/// create-then-delete ends with it absent, because the upsert half consults
/// the filesystem and finds nothing there.
fn apply_pending(&mut self) {
let events: Vec<FsEvent> = self.pending.drain().map(|(_, ev)| ev).collect();
let conn = match self.ensure_write_conn() {
Ok(conn) => conn,
Err(e) => {
@ -423,13 +518,57 @@ impl Inner {
};
// Borrow dance: pull the connection out while applying.
let mut conn = conn;
for ev in &events {
if let Err(e) = apply_fs_event(&mut conn, ev, &self.config, &self.ignore, &self.registry)
{
crate::log_warn!("coordinator: apply {:?}: {}", ev, e);
let deadline = Instant::now() + APPLY_BUDGET;
let chunk = self.config.processing.batch_size.max(1);
let removals: Vec<PathBuf> = self
.pending
.iter()
.filter(|(_, ev)| is_removal(ev))
.map(|(p, _)| p.clone())
.collect();
for batch in removals.chunks(chunk) {
if let Err(e) = crate::incremental::remove_paths(&mut conn, batch, chunk) {
crate::log_warn!("coordinator: remove: {}", e);
}
for path in batch {
self.pending.remove(path);
}
if Instant::now() >= deadline {
break;
}
}
if Instant::now() < deadline {
let upserts: Vec<PathBuf> = self
.pending
.iter()
.filter(|(_, ev)| !is_removal(ev))
.map(|(p, _)| p.clone())
.collect();
for path in upserts {
let Some(ev) = self.pending.remove(&path) else {
continue;
};
if let Err(e) =
apply_fs_event(&mut conn, &ev, &self.config, &self.ignore, &self.registry)
{
crate::log_warn!("coordinator: apply {:?}: {}", ev, e);
}
if Instant::now() >= deadline {
break;
}
}
}
self.write_conn = Some(conn);
// Whatever is left goes to the next tick, and goes immediately: the
// pause was ours, not the filesystem's, so it must not re-arm the
// settle window.
self.last_event_at = None;
if self.pending.is_empty() {
self.pending_since = None;
}
}
fn ensure_write_conn(&mut self) -> Result<Connection, String> {
@ -478,7 +617,18 @@ impl Inner {
// The full run owns the DB (and may wipe/rebuild the file).
self.write_conn = None;
self.needs_full_run = false;
self.pending.clear();
// Queued creates and modifies are dropped — the walk about to start
// rediscovers every one of them. Removals are kept, because it cannot:
// a deletion under a directory the walk fails to read is deliberately
// left alone by `unreadable.covers`, and one of a symlink target whose
// real parent lies outside every root is exempted by `aliased_paths`.
// Those rows would otherwise leak until a rebuild. Re-applying a
// removal the walk did happen to catch is a harmless no-op.
self.pending.retain(|_, ev| is_removal(ev));
if self.pending.is_empty() {
self.last_event_at = None;
self.pending_since = None;
}
if let Err(e) = self
.indexing
.start_indexing(roots, self.db_path(), self.config.clone())
@ -486,18 +636,10 @@ impl Inner {
crate::log_warn!("coordinator: start indexing: {}", e);
return;
}
// Give the service's command thread a moment to flip the status;
// small trees can finish between two coordinator ticks, and the
// finished-run bookkeeping keys off `saw_running`.
for _ in 0..200 {
if !matches!(
self.indexing.get_status(),
IndexingStatus::Idle | IndexingStatus::Error(_)
) {
break;
}
std::thread::sleep(Duration::from_millis(10));
}
// `start_indexing` claims the Running status before it returns, so
// `get_status()` is already authoritative — no poll, and no window in
// which this thread could believe the service idle and start writing
// to a database the run is about to reopen.
self.saw_running = true;
}
@ -516,7 +658,7 @@ impl Inner {
self.mode = IndexMode::ManualStopped;
self.config.indexing.auto_index = false;
self.stop_watcher();
self.pending.clear();
self.clear_pending();
let status = self.indexing.get_status();
if !matches!(status, IndexingStatus::Idle | IndexingStatus::Error(_)) {
// Signal only — waiting up to 5 s here would stall every
@ -775,11 +917,13 @@ mod tests {
}
/// Short debounce windows so trailing-edge events flush within test
/// timeouts (production default is a 30 s window).
/// timeouts (production defaults are 30 s / 2 s).
fn fast_watcher() -> WatcherConfig {
WatcherConfig {
throttle_window: Duration::from_millis(300),
tick_interval: Duration::from_millis(100),
pending_settle: Duration::from_millis(200),
pending_max_defer: Duration::from_secs(3),
..WatcherConfig::default()
}
}
@ -983,6 +1127,114 @@ mod tests {
std::fs::remove_dir_all(&extra_root).ok();
}
/// A `rm -rf` reports the directory *and* everything under it. Only the
/// directory needs applying — its range sweep covers the rest — and
/// collapsing before the overflow test is what stops a large deletion from
/// discarding the queue and forcing a full run.
#[test]
fn collapsing_reduces_a_tree_deletion_to_its_root() {
let mut pending = HashMap::new();
let dir = PathBuf::from("/x/tree");
enqueue(&mut pending, FsEvent::Remove(dir.clone()));
for i in 0..500 {
enqueue(
&mut pending,
FsEvent::Remove(dir.join(format!("sub{}/f{}.txt", i % 5, i))),
);
enqueue(&mut pending, FsEvent::Remove(dir.join(format!("sub{}", i % 5))));
}
// Not under the removed tree, and not a removal: both must survive.
enqueue(&mut pending, FsEvent::Remove(PathBuf::from("/x/treehouse")));
enqueue(&mut pending, FsEvent::Create(dir.join("reborn.txt")));
collapse_pending_removals(&mut pending);
let mut left: Vec<String> = pending
.keys()
.map(|p| p.to_string_lossy().into_owned())
.collect();
left.sort();
assert_eq!(
left,
vec![
"/x/tree".to_string(),
"/x/tree/reborn.txt".to_string(),
"/x/treehouse".to_string(),
],
"only the removal root, the re-creation, and the prefix sibling remain"
);
}
#[test]
fn collapsing_a_queue_without_removals_is_a_no_op() {
let mut pending = HashMap::new();
enqueue(&mut pending, FsEvent::Create(PathBuf::from("/x/a")));
enqueue(&mut pending, FsEvent::Modify(PathBuf::from("/x/a/b")));
collapse_pending_removals(&mut pending);
assert_eq!(pending.len(), 2);
}
/// Deleting a populated directory must land as one queued removal, not one
/// per file, and the rows must actually go.
#[test]
fn a_deleted_directory_is_applied_as_a_single_collapsed_removal() {
let f = Fixture::new(true);
let tree = f.dir.join("tree");
for i in 0..40 {
let p = tree.join(format!("sub{}/f{:03}.txt", i % 4, i));
std::fs::create_dir_all(p.parent().unwrap()).unwrap();
std::fs::write(&p, format!("body {}", i)).unwrap();
}
std::fs::write(f.dir.join("keep.txt"), "survivor").unwrap();
let coord =
IndexCoordinator::start_with_watcher_config(f.config.clone(), fast_watcher()).unwrap();
wait_for("initial index", Duration::from_secs(30), || {
f.file_count() == 41
});
std::fs::remove_dir_all(&tree).unwrap();
wait_for("subtree removed from the index", Duration::from_secs(30), || {
f.file_count() == 1
});
coord.shutdown();
}
/// The queue must not be applied while a full run owns the database, and
/// the events must survive to be applied once it finishes.
#[test]
fn a_deletion_during_a_full_run_is_queued_then_applied() {
let f = Fixture::new(false); // manual: runs happen only when asked
for i in 0..300 {
let p = f.dir.join(format!("d{}/f{:03}.txt", i % 6, i));
std::fs::create_dir_all(p.parent().unwrap()).unwrap();
std::fs::write(&p, format!("body {}", i)).unwrap();
}
let coord =
IndexCoordinator::start_with_watcher_config(f.config.clone(), fast_watcher()).unwrap();
coord.reindex_now();
wait_for("first run", Duration::from_secs(30), || {
coord.state().last_full_index.is_some() && f.file_count() == 300
});
// Auto mode so the watcher is live, then delete while a run is going.
coord.set_mode(IndexMode::Auto);
wait_for("watcher active", Duration::from_secs(30), || {
matches!(coord.state().watcher, WatcherStatus::Active { .. })
});
coord.reindex_now();
std::fs::remove_dir_all(f.dir.join("d0")).unwrap();
// 300 files, 50 of them under d0.
wait_for("deletion applied after the run", Duration::from_secs(60), || {
f.file_count() == 250
});
coord.shutdown();
}
#[test]
fn enqueue_last_wins_and_rename_splits() {
let mut pending = HashMap::new();

4
crates/quicksearch-core/src/db/mod.rs
View file

@ -15,6 +15,6 @@ pub mod schema;
pub use key::{process_key_hex, set_process_key};
pub use open::{
open_existing, open_or_recreate, verify_process_key, CURRENT_SCHEMA_VERSION,
KEY_MISMATCH_PREFIX,
index_needs_rebuild, open_existing, open_or_recreate, verify_process_key,
CURRENT_SCHEMA_VERSION, KEY_MISMATCH_PREFIX,
};

152
crates/quicksearch-core/src/db/open.rs
View file

@ -18,8 +18,8 @@ use std::path::Path;
use rusqlite::{params, Connection, OpenFlags, OptionalExtension};
use super::schema::{
effective_tokenizer, fts_create_sql, PRAGMAS_FAST, PRAGMAS_READONLY, PRAGMAS_WALK_READER,
SCHEMA_CURRENT,
effective_tokenizer, fts_create_sql, PRAGMAS_FAST, PRAGMAS_MAINTENANCE, PRAGMAS_READONLY,
PRAGMAS_WALK_READER, SCHEMA_CURRENT,
};
use crate::security::IndexKey;
@ -33,7 +33,7 @@ pub const KEY_MISMATCH_PREFIX: &str = "KEY_MISMATCH: ";
/// a way that makes an old DB unreadable by new code. Any such bump
/// causes existing indexes to be wiped on next open — there's no
/// migration path by design.
pub const CURRENT_SCHEMA_VERSION: u32 = 3;
pub const CURRENT_SCHEMA_VERSION: u32 = 4;
/// Open `db_path`, applying fast-path pragmas, and ensure the on-disk
/// schema matches what this build expects. If it doesn't, delete the
@ -114,6 +114,17 @@ pub fn open_walk_reader(db_path: &str) -> Result<Connection, String> {
)
}
/// A writable connection for post-run compaction, and the only one that may
/// VACUUM. See [`PRAGMAS_MAINTENANCE`] for why it cannot be the indexer's.
pub fn open_maintenance(db_path: &str) -> Result<Connection, String> {
open_keyed_with_pragmas(
db_path,
true,
super::key::process_key().as_ref(),
PRAGMAS_MAINTENANCE,
)
}
pub(crate) fn open_existing_keyed(
db_path: &str,
write: bool,
@ -155,8 +166,63 @@ fn open_keyed_with_pragmas(
/// index. Used by the GUI unlock screen and the CLI prompt loop before any
/// service starts; the error carries [`KEY_MISMATCH_PREFIX`] on a wrong
/// password.
///
/// Answers **only** the key question. It deliberately does not go through
/// [`open_existing`], which additionally demands a current schema — a
/// different question, with a different owner. Whether the stored schema is
/// current is the *indexer's* business, and its answer is to wipe and rebuild
/// ([`open_or_recreate`]); an unlock screen has nothing useful to do with it.
///
/// Conflating the two made every schema bump present itself to anyone using
/// password protection as an unlock failure, with no way past the gate even
/// with the correct password:
///
/// ```text
/// index at …/index.sqlite is not a compatible QuickSearch index
/// (schema v4 expected); refusing to modify it. Re-index to rebuild.
/// ```
///
/// An unprotected install in the same state starts fine and rebuilds on its
/// first run; this keeps the protected one behaving the same way.
pub fn verify_process_key(db_path: &str) -> Result<(), String> {
open_existing(db_path, false).map(|_| ())
verify_key(db_path, super::key::process_key().as_ref())
}
/// Whether an existing index will be discarded and rebuilt by the next
/// indexing run because it was written under a different schema version.
///
/// The GUI asks this at startup so it can *say so*. The wipe is otherwise
/// silent: `open_or_recreate` replaces the file, every search fails in the
/// meantime, and the only visible explanation is a schema-version error
/// string — which reads like data loss rather than a version upgrade.
///
/// `false` for anything this cannot positively establish: no file yet, a file
/// the process key does not open, or a database that cannot be queried at all.
/// Announcing a reset that is not happening would be worse than saying nothing.
pub fn index_needs_rebuild(db_path: &str) -> bool {
let Ok(conn) = Connection::open_with_flags(
db_path,
OpenFlags::SQLITE_OPEN_NO_MUTEX | OpenFlags::SQLITE_OPEN_READ_ONLY,
) else {
return false;
};
if key_and_probe(&conn, db_path, super::key::process_key().as_ref()).is_err() {
return false;
}
// Only `Ok(false)` — a version we read and it differed, or a layout with no
// `schema_info` at all. An `Err` means we could not tell.
matches!(schema_version_current(&conn), Ok(false))
}
pub(crate) fn verify_key(db_path: &str, key: Option<&IndexKey>) -> Result<(), String> {
// Read-only and no CREATE: verifying a key must never bring a database
// into existence, and must never modify one.
let conn = Connection::open_with_flags(
db_path,
OpenFlags::SQLITE_OPEN_NO_MUTEX | OpenFlags::SQLITE_OPEN_READ_ONLY,
)
.map_err(|e| format!("Failed to open database at {}: {}", db_path, e))?;
key_and_probe(&conn, db_path, key)
}
/// Apply the SQLCipher key (if any) and force the first page off disk.
@ -627,6 +693,84 @@ mod tests {
std::fs::remove_file(&p).ok();
}
/// The one difference that makes the maintenance profile exist.
///
/// SQLCipher is compiled `-DSQLITE_TEMP_STORE=2`, under which SQLite puts
/// temporary databases in memory for any `temp_store` but an explicit
/// `FILE` (1). VACUUM builds the whole replacement index in that temporary
/// database, so on the indexer's connection — which sets `MEMORY` — it
/// would try to hold a rebuilt multi-gigabyte index in RAM.
#[test]
fn maintenance_opens_keep_temporaries_on_disk() {
let p = tmp_db_path();
{
let conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
let indexer: i64 = conn.query_row("PRAGMA temp_store", [], |r| r.get(0)).unwrap();
assert_eq!(indexer, 2, "the indexer's own profile is MEMORY");
}
let conn = open_maintenance(p.to_str().unwrap()).unwrap();
let store: i64 = conn.query_row("PRAGMA temp_store", [], |r| r.get(0)).unwrap();
assert_eq!(store, 1, "maintenance must build its temporaries on disk");
// And the directory those temporaries land in is steerable, which is
// what keeps them off a RAM-backed /tmp. Deprecated but present.
let dir = p.parent().unwrap().to_string_lossy().into_owned();
conn.execute_batch(&format!("PRAGMA temp_store_directory = '{}';", dir))
.unwrap();
let set: String = conn
.query_row("PRAGMA temp_store_directory", [], |r| r.get(0))
.unwrap();
assert_eq!(set, dir);
conn.execute_batch("PRAGMA temp_store_directory = '';").unwrap();
drop(conn);
std::fs::remove_file(&p).ok();
}
/// Drives the GUI's "your index is being reset" modal, so a false positive
/// announces a wipe that is not happening and a false negative lets one
/// happen in silence.
#[test]
fn index_needs_rebuild_only_when_the_schema_really_differs() {
let p = tmp_db_path();
assert!(
!index_needs_rebuild(p.to_str().unwrap()),
"no file yet is a fresh install, not a reset"
);
{
let _ = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
}
assert!(
!index_needs_rebuild(p.to_str().unwrap()),
"a current index is not going to be rebuilt"
);
// Age it, exactly as a version bump does.
{
let conn = open_existing(p.to_str().unwrap(), true).unwrap();
conn.execute("UPDATE schema_info SET value = '1' WHERE key = 'version'", [])
.unwrap();
}
assert!(index_needs_rebuild(p.to_str().unwrap()));
// A pre-`schema_info` layout counts too.
std::fs::remove_file(&p).ok();
{
let conn = Connection::open(&p).unwrap();
conn.execute("CREATE TABLE files (id INTEGER PRIMARY KEY, name TEXT)", [])
.unwrap();
}
assert!(index_needs_rebuild(p.to_str().unwrap()));
// Not a database at all: we cannot tell, so we say nothing.
std::fs::write(&p, [0x5a; 4096]).unwrap();
assert!(!index_needs_rebuild(p.to_str().unwrap()));
std::fs::remove_file(&p).ok();
}
#[test]
fn open_existing_errors_on_missing_file() {
let p = tmp_db_path();

879
crates/quicksearch-core/src/db/repo.rs
View file

File diff suppressed because it is too large Load diff

35
crates/quicksearch-core/src/db/schema.rs
View file

@ -11,7 +11,9 @@
/// writers exist (full index runs and the coordinator's incremental
/// updates) and they're serialized by design; `busy_timeout` is a backstop,
/// not a coordination mechanism. A clean shutdown truncates the log via
/// [`super::repo::checkpoint_and_close`].
/// [`super::repo::checkpoint_and_close`], and a long run truncates it
/// periodically as it goes (see `maximum_wal_size`) — SQLite's own
/// autocheckpoint backfills but cannot reset a log that readers are touching.
pub const PRAGMAS_FAST: &str = "
PRAGMA journal_mode = WAL;
PRAGMA synchronous = NORMAL;
@ -21,6 +23,27 @@ pub const PRAGMAS_FAST: &str = "
PRAGMA foreign_keys = ON;
";
/// Pragmas for the connection that compacts the index after a run.
///
/// [`PRAGMAS_FAST`] but for `temp_store`, and that one difference is the whole
/// reason this profile exists. VACUUM builds the replacement database in the
/// temp store; SQLCipher is compiled `-DSQLITE_TEMP_STORE=2`, under which
/// anything but an explicit `FILE` puts that database in memory. Vacuuming a
/// multi-gigabyte index on the indexer's connection would try to hold the
/// entire rebuilt index in RAM. See [`super::repo::maintain`], which also
/// points the temp directory at the index's own volume.
///
/// The smaller page cache is because this connection does one bulk copy and
/// then closes; the 40 MiB the indexer keeps hot buys it nothing.
pub const PRAGMAS_MAINTENANCE: &str = "
PRAGMA journal_mode = WAL;
PRAGMA synchronous = NORMAL;
PRAGMA busy_timeout = 5000;
PRAGMA cache_size = 2000;
PRAGMA temp_store = FILE;
PRAGMA foreign_keys = ON;
";
/// Pragmas safe to apply on a read-only connection, where `journal_mode`
/// and `synchronous` can't be changed on the file. Used by
/// [`super::open::open_existing`] for read-only opens; write paths get the
@ -84,7 +107,15 @@ CREATE TABLE files (
hash BLOB
);
CREATE INDEX idx_files_parent ON files(parent);
-- Covering, not just `(parent)`. The walk's row prefetcher issues
-- `SELECT name, mtime FROM files WHERE parent = ?` once per directory — the
-- hottest read in a full run — and with the bare index that is an index probe
-- plus a table-row fetch per entry. Those fetches are cold by design: the walk
-- reader deliberately runs on a 1 MiB page cache (see `PRAGMAS_WALK_READER`).
-- Carrying `name` and `mtime` in the index makes it an index-only scan.
-- `parent` stays leading, so `SELECT DISTINCT parent` range scans and
-- `paths_in_dir` are unaffected.
CREATE INDEX idx_files_parent ON files(parent, name, mtime);
CREATE INDEX idx_files_mtime ON files(mtime);
CREATE INDEX idx_files_type ON files(type);
CREATE INDEX idx_files_mime ON files(mime);

74
crates/quicksearch-core/src/extract/mod.rs
View file

@ -121,6 +121,27 @@ impl Registry {
self
}
/// The extractor that claims `mime`, if any. The one place dispatch
/// happens, so every question about a MIME — "who handles it", "does
/// anyone handle it", "handle it from these bytes" — gets the same answer.
fn find(&self, mime: &str) -> Option<&dyn Extractor> {
let lower = mime.to_ascii_lowercase();
self.extractors
.iter()
.find(|e| e.supports(&lower))
.map(|e| &**e)
}
/// Whether any extractor claims `mime` — the same question
/// [`Registry::extract`] answers by returning `Ok(None)`, asked without
/// touching the file. This is what lets the walk decide a row's
/// `content_state` up front instead of queueing it for a worker that will
/// only mark it not-applicable (see
/// [`crate::file_handling::content_extractable`]).
pub fn supports(&self, mime: &str) -> bool {
self.find(mime).is_some()
}
/// Look up a handler for `mime` and run it against `path`. Returns
/// `Ok(None)` if no extractor claims the MIME — the caller should then
/// decide whether the file is "not applicable" (text state NA).
@ -129,13 +150,7 @@ impl Registry {
path: &Path,
mime: &str,
) -> Result<Option<ExtractedContent>, ExtractError> {
let lower = mime.to_ascii_lowercase();
for e in &self.extractors {
if e.supports(&lower) {
return e.extract(path).map(Some);
}
}
Ok(None)
self.find(mime).map(|e| e.extract(path)).transpose()
}
/// [`Registry::extract`] for a file whose complete contents the caller
@ -151,13 +166,7 @@ impl Registry {
mime: &str,
head: &[u8],
) -> Option<Result<ExtractedContent, ExtractError>> {
let lower = mime.to_ascii_lowercase();
for e in &self.extractors {
if e.supports(&lower) {
return e.extract_from_head(path, head);
}
}
None
self.find(mime).and_then(|e| e.extract_from_head(path, head))
}
/// The default set wired up for Set A: plaintext, office docs, PDF,
@ -223,6 +232,43 @@ mod tests {
}
}
#[test]
fn supports_agrees_with_extract_dispatch() {
// `supports` is the cheap form of the question `extract` answers with
// `Ok(None)`. They must agree for every MIME, or the walk would write
// a content state the content pass then contradicts. The path does not
// exist, so a claimed MIME surfaces as `Err`, not `Ok(None)` — which is
// exactly the distinction under test.
let r = Registry::default_set();
let missing = Path::new("/nonexistent/quicksearch-supports-probe");
for mime in [
"text/plain",
"TEXT/PLAIN",
"text/x-rust",
"application/json",
"APPLICATION/PDF",
"application/pdf",
"audio/mpeg",
"Image/JPEG",
"application/msword",
"application/vnd.oasis.opendocument.text",
// Real MIMEs with no extractor: the population the fix is about.
"video/mp4",
"application/zip",
"application/x-executable",
"application/octet-stream",
"",
] {
let claimed = !matches!(r.extract(missing, mime), Ok(None));
assert_eq!(
r.supports(mime),
claimed,
"supports and extract disagree about {:?}",
mime
);
}
}
#[test]
fn properties_sorted_is_deterministic() {
let c = ExtractedContent::with_text("hi")

655
crates/quicksearch-core/src/file_handling.rs
View file

@ -1,4 +1,4 @@
use std::sync::atomic::AtomicBool;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Mutex, Arc};
use std::fs::File;
use std::io::Read;
@ -532,16 +532,16 @@ fn get_file_hash(
Ok((hasher.finalize().to_vec(), head))
}
/// Nudge FTS5 to merge its index segments. Best-effort optimization; any
/// error is logged but not fatal.
pub fn fts_finalize_after_text_indexing(conn: &Connection) -> Result<(), String> {
/// Nudge FTS5 to merge its index segments. Best-effort: the only failure
/// possible is logged and swallowed, so there is nothing for a caller to
/// handle — hence no `Result`.
pub fn fts_finalize_after_text_indexing(conn: &Connection) {
if let Err(e) = conn.execute(
"INSERT INTO searchabletext(searchabletext, rank) VALUES('automerge', 8)",
[],
) {
crate::log_warn!("FTS automerge failed (non-fatal): {}", e);
}
Ok(())
}
/// An owned, fully-derived file record: everything needed to insert or
@ -566,6 +566,11 @@ pub struct OwnedNewFile {
/// one, and the walk's channel is bounded at `CHANNEL_CAP`, so this adds
/// at most `CHANNEL_CAP * hash_length` of in-flight memory.
pub inline_text: Option<String>,
/// Whether the content pass has anything to do for this file — see
/// [`content_extractable`], plus the `maximum_text_file_size` gate. `false`
/// means the row is born `STATE_NA`, so "pending" counts only real work and
/// the pass never queues a file just to write it off.
pub needs_content: bool,
}
impl OwnedNewFile {
@ -581,6 +586,7 @@ impl OwnedNewFile {
mime: self.mime.as_deref(),
ftype: self.ftype,
hash: Some(&self.hash),
needs_content: self.needs_content,
}
}
}
@ -636,27 +642,27 @@ pub fn prepare_file_record(
let mime = guess_mime_from_head(Path::new(path), &head);
let ftype = mime.as_deref().map(mime_to_type).unwrap_or(FileType::EMPTY);
// Decided once, here, where the sniffed MIME, the size, the config and the
// registry are all in hand — and stored as the row's `content_state`, so
// "pending" downstream means a file that really needs reading rather than
// one the content pass would only mark not-applicable.
let needs_content = size <= config.processing.maximum_text_file_size
&& content_extractable(Path::new(path), mime.as_deref(), config, registry);
// When the head is the whole file, an extractor that works from bytes can
// finish the job now and spare the content pass an open/read/close. Any
// condition that does not hold simply leaves this `None`, and the file
// stays pending exactly as before — including invalid UTF-8, which the
// content pass records as a failure with a reason.
let inline_text = mime.as_deref().and_then(|m| {
// The head is the whole file only up to `hash_length`. The
// `maximum_text_file_size` gate is the content pass's own (see
// `extract_scope_prepare`), repeated so both paths agree even when a
// config sets it below `hash_length`.
let inline_text = mime.as_deref().filter(|_| needs_content).and_then(|m| {
// The head is the whole file only up to `hash_length`.
//
// Size 0 is excluded rather than treated as "trivially complete":
// procfs, sysfs and some FUSE mounts report it for files that do have
// content, and inlining would store empty text for them. The content
// pass reads those correctly, and an actually-empty file costs the
// same there as it ever did.
if size == 0
|| size > config.processing.hash_length as u64
|| size > config.processing.maximum_text_file_size
|| !crate::config::content_allowed(Path::new(path), config)
{
if size == 0 || size > config.processing.hash_length as u64 {
return None;
}
match registry.extract_complete_head(Path::new(path), m, &head) {
@ -686,6 +692,7 @@ pub fn prepare_file_record(
ftype,
hash,
inline_text,
needs_content,
})
}
@ -729,10 +736,71 @@ pub fn extract_and_store(
registry: &Registry,
config: &Config,
) -> Result<(), String> {
let outcome = decide_content(path, mime, registry, config);
store_content_outcome(tx, file_id, name, &outcome, config)
}
/// What should be written for one file's content, decided without touching
/// the database.
///
/// Split from the write so the deciding — which opens the file, reads it and
/// runs an extractor over it — can happen on a worker thread while the single
/// writer holds nothing. See [`crate::content`].
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum ContentOutcome {
/// Text (already truncated to `maximum_text_size`) and sorted properties.
Done {
text: String,
properties: Vec<(String, String)>,
},
/// No extractor claims the MIME, or `content_extensions` excludes it.
NotApplicable,
/// The extractor ran and failed; the reason goes on the row.
Failed(String),
}
/// Whether extraction will produce anything for this file: the
/// `content_extensions` filter allows it, and some extractor claims its MIME.
///
/// The single predicate behind both the `content_state` a row is born with
/// (see [`prepare_file_record`]) and [`decide_content`]'s not-applicable
/// early-out, so the two can never drift. That matters more than it looks: if
/// they disagreed, a file the walk wrote off as NA would silently never be
/// full-text indexed.
///
/// Deliberately size-free. The content pass's workers never see a size — the
/// `maximum_text_file_size` gate lives in the feeder's query
/// ([`crate::db::repo::pending_content_page`]) and in [`extract_scope_prepare`]
/// — so folding it in here would be a gate only one of the two callers could
/// honour.
pub fn content_extractable(
path: &Path,
mime: Option<&str>,
config: &Config,
registry: &Registry,
) -> bool {
crate::config::content_allowed(path, config)
&& mime.is_some_and(|m| registry.supports(m))
}
/// Read `path` and decide what its content row should say. No database access,
/// no locks held — this is the expensive half.
///
/// `mime` is authoritative, including when it is `None`: see
/// [`extract_and_store`].
pub fn decide_content(
path: &str,
mime: Option<&str>,
registry: &Registry,
config: &Config,
) -> ContentOutcome {
let p = Path::new(path);
if !crate::config::content_allowed(p, config) {
return repo::set_content_na(tx, file_id);
if !content_extractable(p, mime, config, registry) {
return ContentOutcome::NotApplicable;
}
// `content_extractable` just established that some extractor claims this
// MIME, so the `Ok(None)` arm below is unreachable; it stays as the honest
// answer should that ever stop holding.
let result = match mime {
Some(m) => registry.extract(p, m),
None => Ok(None),
@ -740,20 +808,39 @@ pub fn extract_and_store(
match result {
Ok(Some(mut content)) => {
if content.text.len() > config.processing.maximum_text_size {
content.text = safe_truncate_string(&content.text, config.processing.maximum_text_size);
content.text =
safe_truncate_string(&content.text, config.processing.maximum_text_size);
}
ContentOutcome::Done {
properties: content.properties_sorted(),
text: content.text,
}
let props = content.properties_sorted();
repo::set_content_done(
tx,
file_id,
name,
&content.text,
&props,
config.processing.store_text_for_snippets,
)
}
Ok(None) => repo::set_content_na(tx, file_id),
Err(reason) => repo::set_content_failed(tx, file_id, &reason),
Ok(None) => ContentOutcome::NotApplicable,
Err(reason) => ContentOutcome::Failed(reason),
}
}
/// Apply a decision from [`decide_content`]. The cheap half: pure database
/// writes, so this is all that runs with the connection held.
pub fn store_content_outcome(
tx: &rusqlite::Transaction<'_>,
file_id: i64,
name: &str,
outcome: &ContentOutcome,
config: &Config,
) -> Result<(), String> {
match outcome {
ContentOutcome::Done { text, properties } => repo::set_content_done(
tx,
file_id,
name,
text,
properties,
config.processing.store_text_for_snippets,
),
ContentOutcome::NotApplicable => repo::set_content_na(tx, file_id),
ContentOutcome::Failed(reason) => repo::set_content_failed(tx, file_id, reason),
}
}
@ -769,7 +856,7 @@ pub fn extract_and_store(
pub fn process_batch_updates(
conn_mutex: &Arc<Mutex<Connection>>,
files_to_update: &[OwnedNewFile],
stop_flag: &Arc<Mutex<bool>>,
stop_flag: &Arc<AtomicBool>,
config: &Config,
) -> Result<(), String> {
if files_to_update.is_empty() {
@ -779,7 +866,7 @@ pub fn process_batch_updates(
let fts_batch = config.processing.fts_update_batch_size.max(1);
for batch in files_to_update.chunks(fts_batch) {
if *stop_flag.lock().unwrap() {
if stop_flag.load(Ordering::Relaxed) {
return Ok(());
}
@ -789,22 +876,13 @@ pub fn process_batch_updates(
.map_err(|e| format!("Failed to begin transaction: {}", e))?;
for rec in batch.iter() {
if *stop_flag.lock().unwrap() {
if stop_flag.load(Ordering::Relaxed) {
drop(tx);
drop(conn);
return Ok(());
}
let updated = repo::update_file_basic(
&tx,
&rec.path,
rec.size,
rec.mtime,
Some(rec.hash.as_slice()),
rec.mime.as_deref(),
rec.ftype,
)
.map_err(|e| {
let updated = repo::update_file_basic(&tx, &rec.as_new_file()).map_err(|e| {
format!(
"Failed to update file record + clear stale content for {}: {}",
rec.path, e
@ -867,7 +945,7 @@ pub(crate) fn store_inline_text(
pub fn process_batch_inserts(
conn_mutex: &Arc<Mutex<Connection>>,
files_to_insert: &[OwnedNewFile],
stop_flag: &Arc<Mutex<bool>>,
stop_flag: &Arc<AtomicBool>,
config: &Config,
) -> Result<(), String> {
if files_to_insert.is_empty() {
@ -875,7 +953,7 @@ pub fn process_batch_inserts(
}
for batch in files_to_insert.chunks(config.processing.batch_size) {
if *stop_flag.lock().unwrap() {
if stop_flag.load(Ordering::Relaxed) {
return Ok(());
}
@ -885,7 +963,7 @@ pub fn process_batch_inserts(
.map_err(|e| format!("Failed to begin transaction: {}", e))?;
for rec in batch.iter() {
if *stop_flag.lock().unwrap() {
if stop_flag.load(Ordering::Relaxed) {
drop(tx);
drop(conn);
return Ok(());
@ -904,52 +982,76 @@ pub fn process_batch_inserts(
Ok(())
}
/// Delete the rows a completed run found no file behind, in chunked
/// transactions. Returns how many went.
///
/// The chunking is not just about transaction size: the connection is shared
/// with every search and status query, and `should_abort` *blocks* while the
/// indexer is suspended. Checking it inside the transaction — as this used to —
/// pinned the connection for the whole suspension and froze the GUI behind it.
/// So the stop flag, which never blocks, is what guards the inner loop, and
/// suspension is only ever observed between chunks with nothing held.
pub fn cleanup_stale_index_entries(
conn_mutex: &Arc<Mutex<Connection>>,
stale_paths: &[String],
stop_flag: &Arc<Mutex<bool>>,
stop_flag: &Arc<AtomicBool>,
suspend_flag: &Arc<AtomicBool>,
config: &Config,
) -> Result<usize, String> {
if stale_paths.is_empty() {
return Ok(0);
}
let conn = conn_mutex.lock().unwrap();
let tx = conn
.unchecked_transaction()
.map_err(|e| format!("Failed to begin stale cleanup transaction: {}", e))?;
let chunk = config.processing.batch_size.max(1);
let mut deleted_count = 0usize;
for path in stale_paths {
for batch in stale_paths.chunks(chunk) {
// Outside the lock, so a suspend parks here rather than mid-transaction.
if should_abort(stop_flag, suspend_flag) {
let _ = tx.commit();
drop(conn);
return Ok(deleted_count);
}
if repo::delete_file_by_path(&tx, path).map_err(|e| {
format!(
"Failed to remove stale index entry for {}: {}",
path, e
)
})? {
deleted_count += 1;
let conn = conn_mutex.lock().unwrap();
let tx = conn
.unchecked_transaction()
.map_err(|e| format!("Failed to begin stale cleanup transaction: {}", e))?;
for path in batch {
if stop_flag.load(Ordering::Relaxed) {
break;
}
if repo::delete_file_by_path(&tx, path)
.map_err(|e| format!("Failed to remove stale index entry for {}: {}", path, e))?
{
deleted_count += 1;
}
}
tx.commit()
.map_err(|e| format!("Failed to commit stale cleanup transaction: {}", e))?;
if stop_flag.load(Ordering::Relaxed) {
return Ok(deleted_count);
}
}
tx.commit()
.map_err(|e| format!("Failed to commit stale cleanup transaction: {}", e))?;
if deleted_count > 0 && !should_abort(stop_flag, suspend_flag) {
fts_finalize_after_text_indexing(&conn)?;
let conn = conn_mutex.lock().unwrap();
fts_finalize_after_text_indexing(&conn);
}
Ok(deleted_count)
}
/// Keyset cursor for per-root content extraction. `lo`/`hi` bound the
/// root's path range: `[root + "/", root + "0")` — `'0'` is `'/' + 1`, so
/// the pair is a pure index range on `UNIQUE(files.path)`.
/// Keyset cursor bounding everything stored beneath one directory.
///
/// `lo`/`hi` are the half-open path range `[dir + SEP, dir + (SEP + 1))`, so
/// the pair is a pure index range on `UNIQUE(files.path)` — `SEARCH … (path>?
/// AND path<?)` rather than a scan. Used by content extraction, by the
/// vanished-directory sweep, and by the watcher's subtree deletion, all of
/// which need the same answer to "which rows live under this directory".
///
/// The separator is load-bearing and must be the platform's own. `files.path`
/// stores native separators (see [`path_to_db_string`]), and the successor of
/// `/` (`0x2F`) is `'0'` while the successor of `\` (`0x5C`) is `']'` — using
/// the Unix pair on Windows yields `hi = "C:\Users\me0"`, which every stored
/// path sorts *above*. The range would be empty, silently disabling content
/// extraction and the vanished-directory sweep on that platform.
#[derive(Debug, Clone)]
pub struct ExtractCursor {
pub last_id: i64,
@ -960,11 +1062,16 @@ pub struct ExtractCursor {
impl ExtractCursor {
/// Cursor covering everything under `root`.
pub fn for_root(root: &str) -> ExtractCursor {
let base = root.trim_end_matches('/');
const SEP: char = std::path::MAIN_SEPARATOR;
// Both separators are trimmed, not just the platform's: a config or a
// watcher event may spell a directory either way, and a trailing one
// would otherwise be doubled into the bounds.
let base = root.trim_end_matches(['/', '\\']);
let next = char::from_u32(SEP as u32 + 1).expect("separator successor is a valid char");
ExtractCursor {
last_id: 0,
lo: format!("{}/", base),
hi: format!("{}0", base),
lo: format!("{}{}", base, SEP),
hi: format!("{}{}", base, next),
}
}
}
@ -973,6 +1080,11 @@ impl ExtractCursor {
/// and rows whose text is already searchable from earlier runs. Progress
/// displays show their sum so an unchanged root reads as fully extracted
/// rather than "extracted 0".
///
/// Both halves count only files an extractor claims: rows nothing will extract
/// are written `NA` at walk time (see [`content_extractable`]) and so appear in
/// neither. That is what makes the sum a denominator for the *work*, rather
/// than for every file under the root.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct ExtractScope {
pub pending: usize,
@ -980,8 +1092,15 @@ pub struct ExtractScope {
}
/// Prepare a root's extraction scope: flip oversize pending rows to NA
/// (idempotent; also handles a `maximum_text_file_size` lowered between
/// runs) and count what is pending vs. already extracted in the range.
/// (idempotent) and count what is pending vs. already extracted in the range.
///
/// [`prepare_file_record`] already applies `maximum_text_file_size` when it
/// writes a row, so the sweep finds nothing on a database this build filled.
/// It stays for the two cases that decision cannot cover: a
/// `maximum_text_file_size` *lowered* between runs (which, unlike
/// `content_extensions`, does not force a rebuild — see
/// [`crate::config::diff_actions`]), and rows left pending by an older build
/// that marked every file pending regardless.
pub fn extract_scope_prepare(
conn_mutex: &Arc<Mutex<Connection>>,
cursor: &ExtractCursor,
@ -1017,87 +1136,47 @@ pub fn extract_scope_prepare(
})
}
/// Run ONE bounded batch of content extraction within the cursor's range.
/// Returns rows processed; 0 means the range is drained (or the run is
/// stopping). `on_file` receives each file's name for progress display.
/// Designed to be pumped by the per-root writer loop, so one root's
/// extraction interleaves with other roots' walks and extractions.
pub fn extract_one_batch(
/// Write a batch of already-extracted rows.
///
/// The cheap half of the content pass: [`crate::content`] does the reading and
/// the extracting on its own worker threads, and this is all that runs with the
/// connection held. Chunked so each transaction — and so each hold of the
/// connection lock — stays short, exactly as [`process_batch_inserts`] does.
///
/// Returns how many rows were written. A row whose write fails is logged and
/// skipped rather than failing the run: its `content_state` stays pending, so
/// the next run retries it.
pub fn store_extracted(
conn_mutex: &Arc<Mutex<Connection>>,
cursor: &mut ExtractCursor,
registry: &Registry,
rows: &[crate::content::ExtractedRow],
stop_flag: &Arc<AtomicBool>,
config: &Config,
stop_flag: &Arc<Mutex<bool>>,
suspend_flag: &Arc<AtomicBool>,
on_file: &mut dyn FnMut(&str),
) -> Result<usize, String> {
if should_abort(stop_flag, suspend_flag) {
if rows.is_empty() {
return Ok(0);
}
let max_size = i64::try_from(config.processing.maximum_text_file_size).unwrap_or(i64::MAX);
let batch_limit = config.processing.batch_size.max(1) as i64;
let batch: Vec<(i64, String, String, Option<String>)> = {
let mut written = 0usize;
for batch in rows.chunks(config.processing.batch_size.max(1)) {
if stop_flag.load(Ordering::Relaxed) {
return Ok(written);
}
let conn = conn_mutex.lock().unwrap();
let mut stmt = conn
.prepare(
"SELECT id, name, path, mime FROM files
WHERE content_state = 0 AND size <= ?1 AND id > ?2
AND path >= ?3 AND path < ?4
ORDER BY id
LIMIT ?5",
)
.map_err(|e| format!("Failed to prepare text indexing query: {}", e))?;
let rows = stmt
.query_map(
rusqlite::params![max_size, cursor.last_id, cursor.lo, cursor.hi, batch_limit],
|row| {
Ok((
row.get::<_, i64>(0)?,
row.get::<_, String>(1)?,
row.get::<_, String>(2)?,
row.get::<_, Option<String>>(3)?,
))
},
)
.map_err(|e| format!("Failed to query files for text indexing: {}", e))?;
rows.collect::<Result<Vec<_>, _>>()
.map_err(|e| format!("Failed to read file row: {}", e))?
};
if batch.is_empty() {
return Ok(0);
}
let mut processed = 0usize;
let conn = conn_mutex.lock().unwrap();
let tx = conn
.unchecked_transaction()
.map_err(|e| format!("Failed to begin transaction: {}", e))?;
for (file_id, fname, fpath, fmime) in batch.iter() {
if *stop_flag.lock().unwrap() {
break;
let tx = conn
.unchecked_transaction()
.map_err(|e| format!("Failed to begin transaction: {}", e))?;
for row in batch {
if let Err(e) =
store_content_outcome(&tx, row.file_id, &row.name, &row.outcome, config)
{
crate::log_warn!("content indexing for {}: {}", row.name, e);
continue;
}
written += 1;
}
on_file(fname);
if let Err(e) = extract_and_store(
&tx,
*file_id,
fname,
fpath,
fmime.as_deref(),
registry,
config,
) {
crate::log_warn!("content indexing for {}: {}", fpath, e);
}
// Advance only past what was actually processed, so a stop
// mid-batch never skips rows (they stay pending for the next run).
cursor.last_id = *file_id;
processed += 1;
tx.commit()
.map_err(|e| format!("Failed to commit transaction: {}", e))?;
}
tx.commit()
.map_err(|e| format!("Failed to commit transaction: {}", e))?;
Ok(processed)
Ok(written)
}
#[cfg(test)]
@ -1310,6 +1389,54 @@ mod tests {
);
}
/// The range must bracket paths spelled with the *platform's* separator.
///
/// The bug this pins: `hi` was hard-coded as `'/' + 1` = `'0'`, so on
/// Windows every stored `C:\Users\me\…` path sorted above it and the range
/// was empty — content extraction and the vanished-directory sweep both
/// silently did nothing.
#[test]
fn extract_cursor_brackets_paths_under_its_root() {
use std::path::MAIN_SEPARATOR as SEP;
let root = format!("{}Users{}me", root_prefix(), SEP);
let c = ExtractCursor::for_root(&root);
let inside = format!("{}{}docs{}a.txt", root, SEP, SEP);
assert!(
inside.as_str() >= c.lo.as_str() && inside.as_str() < c.hi.as_str(),
"{:?} must fall inside [{:?}, {:?})",
inside,
c.lo,
c.hi
);
// The directory itself is *not* in the range (the range is what lives
// beneath it), and a prefix sibling is outside it.
assert!(root.as_str() < c.lo.as_str());
let sibling = format!("{}Users{}mexico{}a.txt", root_prefix(), SEP, SEP);
assert!(
!(sibling.as_str() >= c.lo.as_str() && sibling.as_str() < c.hi.as_str()),
"{:?} is a sibling of {:?}, not a child",
sibling,
root
);
// A trailing separator of either flavour must not double up.
for spelled in [format!("{}/", root), format!("{}\\", root)] {
let c2 = ExtractCursor::for_root(&spelled);
assert_eq!((&c2.lo, &c2.hi), (&c.lo, &c.hi), "spelled {:?}", spelled);
}
}
/// An absolute-path prefix for the running platform.
fn root_prefix() -> String {
if cfg!(windows) {
r"C:\".to_string()
} else {
"/".to_string()
}
}
/// A Remove event names a path that is already gone, so the key for it has
/// to be built from the deepest ancestor that still resolves.
#[test]
@ -1437,9 +1564,6 @@ mod tests {
#[cfg(test)]
mod count_and_extract_tests {
use super::*;
use crate::db::open_or_recreate;
use crate::db::repo::{insert_file, NewFile};
use crate::mime::FileType;
use std::sync::atomic::{AtomicBool, Ordering};
fn tmp(tag: &str) -> std::path::PathBuf {
@ -1492,102 +1616,161 @@ mod count_and_extract_tests {
assert!(cancel.load(Ordering::Relaxed));
}
/// The invariant the whole "decide at walk time" design rests on. If these
/// two ever disagree, a file the walk wrote off as NA would silently never
/// be full-text indexed — a far worse bug than the inflated denominator
/// this replaced. Fails the moment someone edits one predicate and not the
/// other.
#[test]
fn extract_one_batch_is_scoped_to_its_root_range() {
let tree = tmp("extract-tree");
std::fs::create_dir_all(tree.join("r1")).unwrap();
std::fs::create_dir_all(tree.join("r2")).unwrap();
let f1 = tree.join("r1/inside.txt");
let f2 = tree.join("r2/outside.txt");
std::fs::write(&f1, "sphinx of black quartz").unwrap();
std::fs::write(&f2, "judge my vow").unwrap();
let db = tmp("extract-db");
let mut conn = open_or_recreate(db.to_str().unwrap(), "trigram").unwrap();
{
let tx = conn.transaction().unwrap();
for f in [&f1, &f2] {
insert_file(
&tx,
&NewFile {
name: f.file_name().unwrap().to_str().unwrap(),
path: f.to_str().unwrap(),
parent: f.parent().unwrap().to_str().unwrap(),
size: std::fs::metadata(f).unwrap().len(),
mtime: 1,
inode: None,
device_id: None,
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
},
)
.unwrap()
.expect("unique");
}
tx.commit().unwrap();
}
let conn_mutex = Arc::new(Mutex::new(conn));
fn content_extractable_is_decide_contents_not_applicable() {
let root = tmp("extractable");
std::fs::create_dir_all(&root).unwrap();
let mut cfg = Config::default();
let registry = Registry::default_set();
let config = Config::default();
let stop = Arc::new(Mutex::new(false));
let suspend = Arc::new(AtomicBool::new(false));
let mut cursor = ExtractCursor::for_root(tree.join("r1").to_str().unwrap());
let scope = extract_scope_prepare(&conn_mutex, &cursor, &config).unwrap();
assert_eq!(scope.pending, 1, "only r1's file is in range");
assert_eq!(scope.already_done, 0, "nothing extracted yet");
// Real files, because `decide_content` runs the extractor for anything
// it claims and its answer must be a genuine one.
let cases = [
"notes.txt",
"data.json",
"schema.sql",
"song.mp3",
"photo.jpg",
"movie.mp4",
"archive.zip",
"blob.bin",
"noextension",
];
for name in cases {
let p = root.join(name);
std::fs::write(&p, b"plain bytes with no magic").unwrap();
}
let mut seen_names = Vec::new();
loop {
let n = extract_one_batch(
&conn_mutex,
&mut cursor,
&registry,
&config,
&stop,
&suspend,
&mut |name| seen_names.push(name.to_string()),
)
.unwrap();
if n == 0 {
break;
// Once with the filter off (the default: everything the registry
// claims), once with it narrowed to `.txt`.
for filter in [Vec::new(), vec!["txt".to_string()]] {
cfg.indexing.content_extensions = filter.clone();
for name in cases {
let p = root.join(name);
let path = p.to_str().unwrap();
let mime = guess_mime_from_head(&p, b"plain bytes with no magic");
let claimed = content_extractable(&p, mime.as_deref(), &cfg, &registry);
let outcome = decide_content(path, mime.as_deref(), &registry, &cfg);
assert_eq!(
claimed,
outcome != ContentOutcome::NotApplicable,
"{} (mime {:?}, filter {:?}): predicate says {}, decide_content says {:?}",
name,
mime,
filter,
claimed,
outcome
);
}
}
assert_eq!(seen_names, vec!["inside.txt".to_string()]);
let conn = conn_mutex.lock().unwrap();
let state = |path: &std::path::Path| -> i64 {
conn.query_row(
"SELECT content_state FROM files WHERE path = ?1",
rusqlite::params![path.to_str().unwrap()],
|r| r.get(0),
)
.unwrap()
std::fs::remove_dir_all(&root).ok();
}
#[test]
fn prepare_file_record_marks_only_claimable_files() {
let root = tmp("claimable");
std::fs::create_dir_all(&root).unwrap();
let registry = Registry::default_set();
let needs = |cfg: &Config, name: &str| -> bool {
let p = root.join(name);
let meta = std::fs::metadata(&p).unwrap();
prepare_file_record(p.to_str().unwrap(), &meta, cfg, &registry)
.expect("regular file")
.needs_content
};
assert_eq!(state(&f1), repo::STATE_DONE, "in-range row extracted");
assert_eq!(state(&f2), repo::STATE_PENDING, "out-of-range row untouched");
drop(conn);
// A second run over the unchanged root must report the file as
// already extracted, so progress reads "1 of 1", never "0 of 0".
let cursor2 = ExtractCursor::for_root(tree.join("r1").to_str().unwrap());
let scope2 = extract_scope_prepare(&conn_mutex, &cursor2, &config).unwrap();
assert_eq!(scope2.pending, 0);
assert_eq!(scope2.already_done, 1);
let conn = conn_mutex.lock().unwrap();
let hits: i64 = conn
.query_row(
"SELECT COUNT(*) FROM searchabletext WHERE searchabletext MATCH '\"sphinx\"'",
[],
|r| r.get(0),
)
for name in ["notes.txt", "song.mp3", "movie.mp4", "blob.bin"] {
std::fs::write(root.join(name), b"body").unwrap();
}
let big = root.join("huge.txt");
std::fs::write(&big, vec![b'x'; 4096]).unwrap();
let cfg = Config::default();
assert!(needs(&cfg, "notes.txt"), "plaintext is claimed");
assert!(needs(&cfg, "song.mp3"), "audio tags are content too");
assert!(!needs(&cfg, "movie.mp4"), "no extractor claims video");
assert!(!needs(&cfg, "blob.bin"), "unsniffable: no MIME, no extractor");
// Over `maximum_text_file_size`, so the content pass would never read
// it even though plaintext claims the MIME.
let mut small_cap = Config::default();
small_cap.processing.maximum_text_file_size = 1024;
assert!(!needs(&small_cap, "huge.txt"));
assert!(needs(&cfg, "huge.txt"), "and claimed under the default cap");
// The `content_extensions` allowlist excludes it.
let mut only_md = Config::default();
only_md.indexing.content_extensions = vec!["md".to_string()];
assert!(!needs(&only_md, "notes.txt"));
std::fs::remove_dir_all(&root).ok();
}
/// The headline regression test: the number the manage-index tab shows as
/// the extraction denominator, measured where `indexing.rs` measures it —
/// after the walk's inserts, before any content pass runs.
#[test]
fn extract_scope_counts_only_files_an_extractor_claims() {
let root = tmp("denominator");
std::fs::create_dir_all(&root).unwrap();
let mut db = root.clone();
db.set_extension("sqlite");
// 3 files an extractor claims, 5 it never will. Big enough that the
// old behaviour (every row pending) can't coincide with the new one.
let claimed = ["a.txt", "b.json", "c.mp3"];
let unclaimed = ["d.mp4", "e.zip", "f.bin", "g.exe", "h"];
for name in claimed.iter().chain(unclaimed.iter()) {
std::fs::write(root.join(name), b"body bytes, no magic").unwrap();
}
let config = Config::default();
let registry = Registry::default_set();
let records: Vec<OwnedNewFile> = claimed
.iter()
.chain(unclaimed.iter())
.map(|name| {
let p = root.join(name);
let meta = std::fs::metadata(&p).unwrap();
prepare_file_record(p.to_str().unwrap(), &meta, &config, &registry)
.expect("regular file")
})
.collect();
let conn_mutex = Arc::new(Mutex::new(
crate::db::open_or_recreate(db.to_str().unwrap(), "trigram").unwrap(),
));
let stop = Arc::new(AtomicBool::new(false));
process_batch_inserts(&conn_mutex, &records, &stop, &config).unwrap();
let cursor = ExtractCursor::for_root(root.to_str().unwrap());
let scope = extract_scope_prepare(&conn_mutex, &cursor, &config).unwrap();
// `extract_total` in the GUI. The two small text-ish files were
// finished inline by the walk so they land in `already_done`; the mp3
// needs the disk pass. Either way the denominator is the claimed set —
// before this was decided at walk time it read 8, every indexed file.
assert_eq!(
(scope.pending, scope.already_done),
(1, 2),
"denominator must be the files needing text, not every indexed file"
);
assert_eq!(scope.pending + scope.already_done, claimed.len());
let total: i64 = conn_mutex
.lock()
.unwrap()
.query_row("SELECT COUNT(*) FROM files", [], |r| r.get(0))
.unwrap();
assert_eq!(hits, 1);
assert_eq!(total as usize, claimed.len() + unclaimed.len());
drop(conn);
std::fs::remove_dir_all(&tree).ok();
std::fs::remove_dir_all(&root).ok();
std::fs::remove_file(&db).ok();
}
}

205
crates/quicksearch-core/src/incremental.rs
View file

@ -17,15 +17,14 @@ use std::path::Path;
use rusqlite::{Connection, OptionalExtension};
use crate::config::{content_allowed, Config, IgnoreSet};
use crate::config::{Config, IgnoreSet};
use crate::db::repo;
use crate::extract::Registry;
use crate::file_handling::{
db_key_for_missing_path, extract_and_store, filtered_walk, UnreadableDirs,
db_key_for_missing_path, extract_and_store, filtered_walk, ExtractCursor, UnreadableDirs,
prepare_file_record_from_path, store_inline_text,
};
use crate::platform::path_has_hidden_component_under;
use crate::query::translator::like_subtree_pattern;
use crate::watcher::FsEvent;
/// Apply one filesystem event to the index. Missing files are treated as
@ -120,15 +119,7 @@ fn upsert_file(
let file_id = match existing {
Some((_, mtime)) if mtime.max(0) as u64 == rec.mtime => return Ok(()),
Some((id, _)) => {
repo::update_file_basic(
&tx,
&rec.path,
rec.size,
rec.mtime,
Some(&rec.hash),
rec.mime.as_deref(),
rec.ftype,
)?;
repo::update_file_basic(&tx, &rec.as_new_file())?;
id
}
None => match repo::insert_file(&tx, &rec.as_new_file())? {
@ -139,9 +130,13 @@ fn upsert_file(
},
};
if rec.size > config.processing.maximum_text_file_size
|| !content_allowed(Path::new(&rec.path), config)
{
// The insert/update above already wrote NA; re-asserting it costs one
// indexed UPDATE on a path that handles one file per event, and keeps this
// correct on its own terms rather than on `insert_file`'s. It is also the
// only size gate on this path — `decide_content` has none, so falling
// through would hand a multi-gigabyte `.txt` to the plaintext extractor,
// which reads the whole file into memory.
if !rec.needs_content {
repo::set_content_na(&tx, file_id)?;
} else if let Some(text) = rec.inline_text.as_deref() {
// Small enough that `prepare_file_record_from_path` already read the
@ -163,34 +158,68 @@ fn upsert_file(
}
fn remove_path(conn: &mut Connection, path: &Path) -> Result<(), String> {
// The insert side stores a canonicalized path, so the raw event spelling
// is not a usable key — but the file is already gone, so `canonicalize`
// cannot be called on it directly either.
let path_str = db_key_for_missing_path(path);
let tx = conn
.transaction()
.map_err(|e| format!("begin incremental tx: {}", e))?;
remove_paths(conn, std::slice::from_ref(&path.to_path_buf()), 1)
}
repo::delete_file_by_path(&tx, &path_str)?;
// Directory removals surface as one event for the directory itself —
// sweep everything indexed beneath it.
let subtree: Vec<String> = {
let mut stmt = tx
.prepare("SELECT path FROM files WHERE path LIKE ?1 ESCAPE '\\'")
.map_err(|e| e.to_string())?;
let rows = stmt
.query_map(rusqlite::params![like_subtree_pattern(&path_str)], |r| {
r.get::<_, String>(0)
})
.map_err(|e| e.to_string())?;
rows.collect::<Result<Vec<_>, _>>().map_err(|e| e.to_string())?
};
for p in &subtree {
repo::delete_file_by_path(&tx, p)?;
/// Drop removals that a removal of one of their ancestors already covers.
///
/// `rm -rf dir/` reports `dir` *and* every file beneath it. Since removing
/// `dir` sweeps its whole path range, each descendant event is duplicate work —
/// 10 000 of them for a 10 000-file tree.
///
/// Ancestor membership is tested against a set rather than by sorting, which
/// keeps this obviously correct: no ordering argument to get wrong, and
/// `Path::ancestors` walks whole components, so `/a/bc` is never treated as
/// living under `/a/b` — the same rule `remove_tree` and
/// `UnreadableDirs::covers` use. Paths are shallow, so the cost is linear.
pub fn collapse_removal_roots(paths: Vec<std::path::PathBuf>) -> Vec<std::path::PathBuf> {
if paths.len() < 2 {
return paths;
}
let all: std::collections::HashSet<&Path> = paths.iter().map(|p| p.as_path()).collect();
paths
.iter()
.filter(|p| !p.ancestors().skip(1).any(|a| all.contains(a)))
.cloned()
.collect()
}
tx.commit().map_err(|e| format!("commit incremental tx: {}", e))
/// Delete `paths` and everything indexed beneath them, in transactions of at
/// most `chunk` paths.
///
/// A mass deletion is the case this exists for. Callers hand over the *roots*
/// of the removal set (see [`collapse_removal_roots`]), so `rm -rf dir/` is one
/// path here rather than one per file, and each one costs a fixed handful of
/// range-driven statements ([`repo::delete_subtree`]) instead of a full table
/// scan plus five statements per file.
///
/// Chunking bounds how long any single transaction holds the connection, the
/// same way `process_batch_inserts` bounds the indexer's writes.
pub fn remove_paths(
conn: &mut Connection,
paths: &[std::path::PathBuf],
chunk: usize,
) -> Result<(), String> {
for batch in paths.chunks(chunk.max(1)) {
let tx = conn
.transaction()
.map_err(|e| format!("begin incremental tx: {}", e))?;
for path in batch {
// The insert side stores a canonicalized path, so the raw event
// spelling is not a usable key — but the file is already gone, so
// `canonicalize` cannot be called on it directly either.
let path_str = db_key_for_missing_path(path);
// The path itself, whether it was a file or a directory...
repo::delete_file_by_path(&tx, &path_str)?;
// ...then everything beneath it, for a directory removal, which
// surfaces as a single event for the directory.
let range = ExtractCursor::for_root(&path_str);
repo::delete_subtree(&tx, &range.lo, &range.hi)?;
}
tx.commit()
.map_err(|e| format!("commit incremental tx: {}", e))?;
}
Ok(())
}
#[cfg(test)]
@ -297,6 +326,81 @@ mod tests {
}
}
fn collapse(paths: &[&str]) -> Vec<String> {
let mut out: Vec<String> =
collapse_removal_roots(paths.iter().map(std::path::PathBuf::from).collect())
.iter()
.map(|p| p.to_string_lossy().into_owned())
.collect();
out.sort();
out
}
#[test]
fn removal_roots_collapse_to_the_shallowest_ancestor() {
// `rm -rf dir/` reports the directory and every file beneath it;
// removing the directory already sweeps its whole range.
assert_eq!(
collapse(&["/dir", "/dir/a.txt", "/dir/b/c.txt", "/dir/b"]),
vec!["/dir"]
);
// Component-wise, so a name-prefix sibling is not swallowed.
assert_eq!(
collapse(&["/a/b", "/a/bc"]),
vec!["/a/b", "/a/bc"],
"/a/bc does not live under /a/b"
);
assert_eq!(
collapse(&["/a/b", "/a/b.txt"]),
vec!["/a/b", "/a/b.txt"],
"a sibling file sorting between a dir and its children survives"
);
// Unrelated removals all survive; order of input does not matter.
assert_eq!(
collapse(&["/x/deep/f", "/y", "/x"]),
vec!["/x", "/y"],
"/x/deep/f is covered by /x, /y is independent"
);
// Degenerate inputs.
assert!(collapse(&[]).is_empty());
assert_eq!(collapse(&["/only"]), vec!["/only"]);
}
/// The collapse must not change what ends up deleted — only how much work
/// it takes to get there.
#[test]
fn collapsed_removal_deletes_the_same_rows_as_the_full_set() {
let mut f = Fixture::new();
f.write("tree/a.txt", "alpha");
f.write("tree/deep/b.txt", "beta");
f.write("tree2/keep.txt", "survivor");
let tree = f.dir.join("tree");
f.apply(&FsEvent::Create(tree.clone()));
f.apply(&FsEvent::Create(f.dir.join("tree2")));
assert_eq!(f.counts().0, 3);
let canonical_tree = f.canonical(&tree);
std::fs::remove_dir_all(&tree).unwrap();
// What a real `rm -rf` produces: the directory plus every path under it.
let reported: Vec<std::path::PathBuf> = vec![
canonical_tree.clone().into(),
format!("{}/deep", canonical_tree).into(),
format!("{}/a.txt", canonical_tree).into(),
format!("{}/deep/b.txt", canonical_tree).into(),
];
let roots = collapse_removal_roots(reported);
assert_eq!(roots.len(), 1, "one range covers the whole tree");
remove_paths(&mut f.conn, &roots, 200).unwrap();
assert_eq!(f.counts(), (1, 1, 1), "only tree2 survives");
let survivor = f.canonical(&f.dir.join("tree2").join("keep.txt"));
assert!(f.row(&survivor).is_some());
}
#[test]
fn create_indexes_file_and_content() {
let mut f = Fixture::new();
@ -414,6 +518,27 @@ mod tests {
assert_eq!(f.counts(), (1, 0, 0));
}
/// The third way a file can end up NA — no extractor claims its MIME —
/// which the filter and oversize tests either side of this one don't
/// cover. A row left pending here would be re-fed to the content pass on
/// every subsequent run and counted in the extraction denominator forever.
#[test]
fn an_unclaimed_mime_is_na_in_the_watcher_path() {
let mut f = Fixture::new();
// `.mp4` sniffs as video/mp4 by extension; nothing extracts video.
let vid = f.write("clip.mp4", "not really an mp4, and it needn't be");
f.apply(&FsEvent::Create(vid.clone()));
let canonical = f.canonical(&vid);
let (_, _, content_state) = f.row(&canonical).expect("row listed");
assert_eq!(
content_state,
repo::STATE_NA,
"filename indexed, nothing to extract"
);
assert_eq!(f.counts(), (1, 0, 0));
}
/// A Remove event whose path is spelled differently from the stored key
/// must still delete the row. `dir/./f.txt` and `dir/f.txt` are the same
/// file; only the canonicalized spelling is in the index.

916
crates/quicksearch-core/src/indexing.rs
View file

File diff suppressed because it is too large Load diff

1
crates/quicksearch-core/src/lib.rs
View file

@ -1,5 +1,6 @@
pub mod cli;
pub mod config;
pub mod content;
pub mod coordinator;
pub mod db;
pub mod document_extraction;

45
crates/quicksearch-core/src/platform.rs
View file

@ -227,6 +227,41 @@ pub const WATCH_ROOTS_RECURSIVELY: bool = cfg!(windows);
/// on both sides, consistently.
pub const PATH_COLLATION: &str = if cfg!(windows) { "NOCASE" } else { "BINARY" };
/// Drop the **calling thread** to background scheduling priority.
///
/// Per-thread, not per-process. The GUI shares this process, so lowering the
/// process would slow the very window the user is watching progress in — the
/// point is to yield to the foreground, not to throttle ourselves. Called by
/// the threads that do indexing work and by nobody else; the search worker,
/// the coordinator and the watcher exist to answer promptly and keep normal
/// priority.
///
/// Best-effort and idempotent: a refusal is not worth reporting, since the
/// only consequence is that indexing competes on equal terms.
pub fn set_background_priority() {
#[cfg(target_os = "linux")]
{
// Linux schedules per task, so `nice` moves this thread alone.
// Deliberately not `setpriority(PRIO_PROCESS, 0, …)`, which is
// process-wide on the BSDs and would take the GUI with it.
unsafe { libc::nice(10) };
}
#[cfg(windows)]
{
use windows_sys::Win32::System::Threading::{
GetCurrentThread, SetThreadPriority, THREAD_MODE_BACKGROUND_BEGIN,
};
// Background *mode*, not merely a lower priority number: it drops I/O
// priority as well, which is what actually keeps a walk from starving
// the foreground on a spinning disk.
unsafe { SetThreadPriority(GetCurrentThread(), THREAD_MODE_BACKGROUND_BEGIN) };
}
// Elsewhere (macOS, BSD): deliberately nothing. `nice` there applies to the
// whole process, so it would hit the GUI. The right call is
// `pthread_set_qos_class_self_np(QOS_CLASS_UTILITY, 0)`, which is worth
// adding on its own terms rather than approximating here.
}
/// How long to keep retrying a delete that fails because something else holds
/// the file open.
#[cfg(windows)]
@ -351,6 +386,16 @@ mod tests {
assert!(!is_unc_string(""));
}
/// Not observable cross-platform beyond "it did not blow up", which is
/// still worth pinning: the call is `unsafe` on both real targets, and it
/// runs at the top of every walker thread, so it must also be safe to
/// repeat.
#[test]
fn background_priority_is_best_effort_and_repeatable() {
set_background_priority();
set_background_priority();
}
#[test]
fn collation_matches_like_case_folding() {
// LIKE folds ASCII case on every platform; the `=` half of a path

44
crates/quicksearch-core/src/query/ast.rs
View file

@ -1,5 +1,6 @@
//! Query AST shared between parser and translator.
//! Shared query vocabulary.
/// The comparison in a `key op value` filter, as the lexer emits it.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Op {
/// `key:value` — substring / FTS MATCH semantics.
@ -15,44 +16,3 @@ pub enum Op {
/// `key>=value`
Ge,
}
#[derive(Debug, Clone, PartialEq)]
pub enum Term {
/// An unquoted word or quoted phrase. Feeds the FTS MATCH expression.
Literal(String),
/// A structured filter such as `type:Audio` or `modified:>=2024-01-01`.
Property { key: String, op: Op, value: String },
And(Vec<Term>),
Or(Vec<Term>),
}
impl Term {
/// Combine two terms with AND, flattening to avoid a deeply nested tree.
pub fn and(a: Term, b: Term) -> Term {
match (a, b) {
(Term::And(mut xs), Term::And(ys)) => {
xs.extend(ys);
Term::And(xs)
}
(Term::And(mut xs), other) | (other, Term::And(mut xs)) => {
xs.push(other);
Term::And(xs)
}
(a, b) => Term::And(vec![a, b]),
}
}
pub fn or(a: Term, b: Term) -> Term {
match (a, b) {
(Term::Or(mut xs), Term::Or(ys)) => {
xs.extend(ys);
Term::Or(xs)
}
(Term::Or(mut xs), other) | (other, Term::Or(mut xs)) => {
xs.push(other);
Term::Or(xs)
}
(a, b) => Term::Or(vec![a, b]),
}
}
}

40
crates/quicksearch-core/src/query/mod.rs
View file

@ -1,33 +1,33 @@
//! Structured query parser and SQL translator.
//! Search-box input → the ranked cascade's term plus its structured filters.
//!
//! Input syntax (a deliberate subset of KDE Baloo's query language — just
//! enough to be useful standalone; full Baloo grammar lives in the Set B
//! compat layer):
//! The entry point is [`split_for_cascade`], which splits raw input into a
//! single search phrase and zero or more `key:value` filters. There is no
//! boolean grammar by design: the cascade ranks one phrase, so `AND`, `OR` and
//! parentheses are ordinary words rather than operators (see [`split`]).
//!
//! - plain words: `foo bar` (implicit AND)
//! - quoted phrases: `"hello world"`
//! - boolean operators: `AND`, `OR` (case-sensitive)
//! - grouping: `(a OR b)`
//! - structured filters:
//! - `type:Audio` / `type:Image` / `type:Document` / `type:Text` / `type:Video`
//! / `type:Archive` / `type:Spreadsheet` / `type:Presentation` / `type:Folder`
//! - `modified:>=2024-01-01`, `modified:<2023-12-01`, `modified:=2024-05-20`
//! (also accepts `modified>=2024-01-01` without the colon)
//! - `path:/some/dir` — matches files with that directory as their parent
//! or any ancestor.
//! Recognized filters:
//!
//! The entry point is [`parse_and_build`], which accepts a query string and
//! returns a ready-to-execute [`SqlQuery`].
//! - `type:Audio` / `Image` / `Video` / `Document` / `Text` / `Archive` /
//! `Spreadsheet` / `Presentation` / `Folder`
//! - `modified:>=2024-01-01`, `modified:<2023-12-01`, `modified:=2024-05-20`
//! (`mtime:` is a synonym; `=` matches the whole day)
//! - `path:/some/dir` — that directory or any beneath it (`folder:`,
//! `includefolder:` are synonyms)
//! - `name:report` / `filename:report` — filename substring, `*` globs
//! - `mime:application/pdf` — exact MIME type
//! - `regex:…` — compiled in [`pattern`] and matched in Rust, never in SQL
//!
//! Each recognized filter becomes a [`translator::FilterFragment`] that the
//! cascade ANDs onto every stage; everything else joins the term.
pub mod ast;
pub mod lexer;
pub mod parser;
pub mod pattern;
pub mod split;
pub mod translator;
pub use ast::{Op, Term};
pub use ast::Op;
pub use lexer::tokenize_spanned;
pub use pattern::{RegexQuery, TermPattern};
pub use split::{split_for_cascade, CascadeQuery};
pub use translator::{parse_and_build, SqlQuery, TranslateError};
pub use translator::TranslateError;

284
crates/quicksearch-core/src/query/parser.rs
View file

@ -1,284 +0,0 @@
//! Recursive-descent parser: token stream → [`Term`] tree.
//!
//! Grammar (implicit AND between adjacent terms):
//! ```text
//! expr := or_expr
//! or_expr := and_expr ("OR" and_expr)*
//! and_expr := atom ("AND"? atom)*
//! atom := "(" expr ")" | property | literal
//! property := ident ":"|"="|"<"|"<="|">"|">=" value
//! literal := WORD | QUOTED
//! ```
use super::ast::{Op, Term};
use super::lexer::{tokenize, LexError, Token};
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum ParseError {
Lex(LexError),
Unexpected { at: usize, reason: String },
Empty,
}
impl std::fmt::Display for ParseError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
ParseError::Lex(e) => write!(f, "{}", e),
ParseError::Unexpected { at, reason } => {
write!(f, "parse error at token {}: {}", at, reason)
}
ParseError::Empty => write!(f, "empty query"),
}
}
}
impl std::error::Error for ParseError {}
pub fn parse(input: &str) -> Result<Term, ParseError> {
let tokens = tokenize(input).map_err(ParseError::Lex)?;
let trimmed: Vec<_> = tokens.into_iter().collect();
if trimmed.is_empty() {
return Err(ParseError::Empty);
}
let mut p = Parser { tokens: trimmed, pos: 0 };
let term = p.parse_or()?;
if p.pos < p.tokens.len() {
return Err(ParseError::Unexpected {
at: p.pos,
reason: format!("trailing token {:?}", p.tokens[p.pos]),
});
}
Ok(term)
}
struct Parser {
tokens: Vec<Token>,
pos: usize,
}
impl Parser {
fn peek(&self) -> Option<&Token> {
self.tokens.get(self.pos)
}
fn bump(&mut self) -> Option<Token> {
let t = self.tokens.get(self.pos).cloned();
if t.is_some() {
self.pos += 1;
}
t
}
fn parse_or(&mut self) -> Result<Term, ParseError> {
let mut left = self.parse_and()?;
while matches!(self.peek(), Some(Token::Or)) {
self.bump();
let right = self.parse_and()?;
left = Term::or(left, right);
}
Ok(left)
}
fn parse_and(&mut self) -> Result<Term, ParseError> {
let mut left = self.parse_atom()?;
loop {
match self.peek() {
Some(Token::And) => {
self.bump();
let right = self.parse_atom()?;
left = Term::and(left, right);
}
// Implicit AND: adjacent atoms without an operator.
Some(Token::Word(_))
| Some(Token::Quoted(_))
| Some(Token::LParen) => {
let right = self.parse_atom()?;
left = Term::and(left, right);
}
_ => break,
}
}
Ok(left)
}
fn parse_atom(&mut self) -> Result<Term, ParseError> {
match self.bump() {
Some(Token::LParen) => {
let inner = self.parse_or()?;
match self.bump() {
Some(Token::RParen) => Ok(inner),
other => Err(ParseError::Unexpected {
at: self.pos,
reason: format!("expected ')' got {:?}", other),
}),
}
}
Some(Token::Quoted(s)) => Ok(Term::Literal(s)),
Some(Token::Word(w)) => {
// If followed by an operator, this word is a property key.
if let Some(Token::Op(op)) = self.peek().cloned() {
self.bump();
// `key:>=value` — an Op followed by another Op becomes the
// effective comparator; the original colon is "separator".
let effective_op = if op == Op::Contains {
if let Some(Token::Op(inner_op)) = self.peek().cloned() {
self.bump();
inner_op
} else {
Op::Contains
}
} else {
op
};
let value = match self.bump() {
Some(Token::Word(v)) => v,
Some(Token::Quoted(v)) => v,
other => {
return Err(ParseError::Unexpected {
at: self.pos,
reason: format!("expected property value got {:?}", other),
})
}
};
Ok(Term::Property {
key: w,
op: effective_op,
value,
})
} else {
Ok(Term::Literal(w))
}
}
other => Err(ParseError::Unexpected {
at: self.pos,
reason: format!("expected term, got {:?}", other),
}),
}
}
}
#[cfg(test)]
mod tests {
use super::*;
fn lit(s: &str) -> Term {
Term::Literal(s.into())
}
fn prop(k: &str, op: Op, v: &str) -> Term {
Term::Property {
key: k.into(),
op,
value: v.into(),
}
}
#[test]
fn single_word() {
assert_eq!(parse("foo").unwrap(), lit("foo"));
}
#[test]
fn quoted_phrase() {
assert_eq!(parse(r#""hello world""#).unwrap(), lit("hello world"));
}
#[test]
fn implicit_and() {
assert_eq!(
parse("foo bar").unwrap(),
Term::And(vec![lit("foo"), lit("bar")])
);
}
#[test]
fn explicit_and() {
assert_eq!(
parse("foo AND bar").unwrap(),
Term::And(vec![lit("foo"), lit("bar")])
);
}
#[test]
fn or_has_lower_precedence_than_and() {
assert_eq!(
parse("a b OR c").unwrap(),
Term::Or(vec![
Term::And(vec![lit("a"), lit("b")]),
lit("c")
])
);
}
#[test]
fn parens_override_precedence() {
assert_eq!(
parse("a (b OR c)").unwrap(),
Term::And(vec![lit("a"), Term::Or(vec![lit("b"), lit("c")])])
);
}
#[test]
fn property_contains() {
assert_eq!(
parse("type:Audio").unwrap(),
prop("type", Op::Contains, "Audio")
);
}
#[test]
fn property_comparator_via_colon() {
assert_eq!(
parse("modified:>=2024-01-01").unwrap(),
prop("modified", Op::Ge, "2024-01-01")
);
}
#[test]
fn property_comparator_bare() {
assert_eq!(
parse("modified>=2024-01-01").unwrap(),
prop("modified", Op::Ge, "2024-01-01")
);
}
#[test]
fn mixed_filter_and_literal() {
assert_eq!(
parse("type:Audio beatles").unwrap(),
Term::And(vec![
prop("type", Op::Contains, "Audio"),
lit("beatles")
])
);
}
#[test]
fn nested_or() {
assert_eq!(
parse("(a OR b) AND (c OR d)").unwrap(),
Term::And(vec![
Term::Or(vec![lit("a"), lit("b")]),
Term::Or(vec![lit("c"), lit("d")])
])
);
}
#[test]
fn empty_query_is_error() {
assert!(matches!(parse(" "), Err(ParseError::Empty)));
}
#[test]
fn trailing_token_is_error() {
assert!(parse("a )").is_err());
}
#[test]
fn property_value_may_be_quoted() {
assert_eq!(
parse(r#"path:"/tmp with space""#).unwrap(),
prop("path", Op::Contains, "/tmp with space")
);
}
}

32
crates/quicksearch-core/src/query/pattern.rs
View file

@ -236,6 +236,38 @@ impl TermPattern {
}
}
/// Case-insensitive [`find_first`] against an already-folded haystack.
///
/// The literal path would otherwise fold the haystack itself, and the
/// cascade's full-text passes need the same fold for counting, searching
/// and snippet extraction — three copies of a document that can run to
/// `maximum_text_size`. Folding is byte-length preserving, so the returned
/// range is valid in the unfolded original too.
pub fn find_first_folded(&self, folded: &str) -> Option<Range<usize>> {
match self {
TermPattern::Empty => None,
TermPattern::Literal(l) => {
let pos = folded.find(&l.folded)?;
Some(pos..pos + l.text.len())
}
// The regex engine folds as it matches, so it needs no help and
// allocates nothing either way.
TermPattern::Wildcard(w) => w.search_ci.find(folded).map(|m| m.range()),
}
}
/// Case-insensitive [`count`] against an already-folded haystack. See
/// [`TermPattern::find_first_folded`].
pub fn count_folded(&self, folded: &str) -> usize {
match self {
TermPattern::Empty => 0,
// Both sides are already folded, so an exact scan *is* the
// case-insensitive one — and it allocates nothing.
TermPattern::Literal(l) => snippet::count_occurrences(folded, &l.folded, true),
TermPattern::Wildcard(w) => w.search_ci.find_iter(folded).take(COUNT_CAP).count(),
}
}
/// Non-overlapping occurrence count, capped at 1000 (the cascade's
/// `count_frac` saturates there anyway).
pub fn count(&self, text: &str, case_insensitive: bool) -> usize {

586
crates/quicksearch-core/src/query/translator.rs
View file

@ -1,38 +1,16 @@
//! AST → executable SQL.
//! `key op value` filters → composable SQL fragments.
//!
//! One filter becomes one [`FilterFragment`]: a WHERE fragment over table alias
//! `f` with anonymous `?` placeholders, plus the values to bind. The search
//! cascade ANDs those fragments onto every stage's query, so they have to
//! compose by plain appending — which anonymous placeholders do and numbered
//! ones would not.
use super::ast::{Op, Term};
use super::parser::{parse, ParseError};
use super::ast::Op;
use crate::mime::FileType;
/// A prepared SQL statement plus its positional parameters. Parameters are
/// rusqlite `Value` for convenience at the call site. Every `?N` in `sql`
/// corresponds to `params[N-1]`.
#[derive(Debug, Clone)]
pub struct SqlQuery {
pub sql: String,
pub params: Vec<rusqlite::types::Value>,
}
/// Sort strategy for the result set.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Sort {
/// Newest first by file modification time.
ByMtimeDesc,
/// FTS5 relevance rank. Only sensible when the query has FTS terms.
ByRank,
/// No ORDER BY clause.
None,
}
impl Default for Sort {
fn default() -> Self {
Sort::ByMtimeDesc
}
}
#[derive(Debug, Clone)]
pub enum TranslateError {
Parse(ParseError),
UnknownProperty(String),
BadDate(String),
BadRegex(String),
@ -45,7 +23,6 @@ pub enum TranslateError {
impl std::fmt::Display for TranslateError {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
TranslateError::Parse(e) => write!(f, "{}", e),
TranslateError::UnknownProperty(k) => write!(f, "unknown property '{}'", k),
TranslateError::BadDate(s) => write!(f, "bad date '{}'", s),
TranslateError::BadRegex(s) => write!(f, "regex error: {}", s),
@ -58,121 +35,6 @@ impl std::fmt::Display for TranslateError {
impl std::error::Error for TranslateError {}
/// One-shot: parse the input string and build the SQL. Limit/offset are
/// applied at the SQL level; `None` means "no limit".
pub fn parse_and_build(
input: &str,
limit: Option<u32>,
offset: u32,
sort: Sort,
) -> Result<SqlQuery, TranslateError> {
let ast = parse(input).map_err(TranslateError::Parse)?;
build(&ast, limit, offset, sort)
}
/// Translate an already-parsed AST into SQL.
pub fn build(
ast: &Term,
limit: Option<u32>,
offset: u32,
sort: Sort,
) -> Result<SqlQuery, TranslateError> {
let mut b = Builder::default();
let where_sql = b.translate(ast)?;
// The FTS branch joins a CTE of matching rowids; the structured branch
// stands alone. A query may have both — in that case we intersect at the
// file.id level.
let needs_fts_join = !b.fts_parts.is_empty();
// Structured placeholders were numbered assuming no FTS param. When an
// FTS MATCH is prepended as ?1, bump every `?N` in the generated WHERE
// clause by 1 so the positional bindings line up.
let where_sql = if needs_fts_join {
shift_placeholders(&where_sql)
} else {
where_sql
};
let mut sql = String::new();
if needs_fts_join {
sql.push_str(
"WITH fts_hits AS (SELECT rowid FROM searchabletext WHERE searchabletext MATCH ?1) ",
);
b.all_params
.insert(0, rusqlite::types::Value::Text(fts_expr(&b.fts_parts)));
}
sql.push_str("SELECT f.id, f.name, f.path FROM files f ");
if needs_fts_join {
sql.push_str("JOIN fts_hits ON fts_hits.rowid = f.id ");
}
if !where_sql.is_empty() {
sql.push_str("WHERE ");
sql.push_str(&where_sql);
sql.push(' ');
}
match sort {
Sort::ByMtimeDesc => sql.push_str("ORDER BY f.mtime DESC "),
Sort::ByRank if needs_fts_join => sql.push_str("ORDER BY rank "),
Sort::ByRank | Sort::None => {}
}
if let Some(n) = limit {
sql.push_str(&format!("LIMIT {} ", n));
}
if offset > 0 {
sql.push_str(&format!("OFFSET {}", offset));
}
Ok(SqlQuery {
sql: sql.trim_end().to_string(),
params: b.all_params,
})
}
#[derive(Default)]
struct Builder {
/// Tokens that will be joined into the single FTS MATCH expression.
fts_parts: Vec<FtsFragment>,
/// Everything else (structured filters) as a WHERE clause.
all_params: Vec<rusqlite::types::Value>,
}
#[derive(Debug, Clone)]
enum FtsFragment {
/// A plain word or phrase to be included as an AND'd token.
And(String),
/// Grouped alternation, already rendered with internal ORs.
OrGroup(Vec<String>),
}
fn fts_expr(parts: &[FtsFragment]) -> String {
// Render `(a AND b AND (c OR d))` where a/b/c/d are quoted phrases.
let mut out = String::new();
let mut first = true;
for p in parts {
if !first {
out.push_str(" AND ");
}
first = false;
match p {
FtsFragment::And(s) => out.push_str(&quote_phrase(s)),
FtsFragment::OrGroup(items) => {
out.push('(');
let mut first_item = true;
for item in items {
if !first_item {
out.push_str(" OR ");
}
first_item = false;
out.push_str(&quote_phrase(item));
}
out.push(')');
}
}
}
out
}
/// Escape a phrase for FTS5 MATCH. FTS5 itself uses doubled quotes for
/// literal quotes inside a quoted phrase; wrapping in quotes renders all
/// other MATCH metacharacters (`( ) * :` etc.) inert. Injection-safe by
@ -190,92 +52,6 @@ pub fn quote_phrase(s: &str) -> String {
buf
}
impl Builder {
/// Translate a sub-term, returning its contribution to the SQL WHERE
/// clause. FTS parts are accumulated in `self.fts_parts` (not returned
/// here) since they collapse into a single MATCH expression at the top.
fn translate(&mut self, t: &Term) -> Result<String, TranslateError> {
match t {
Term::Literal(s) => {
self.fts_parts.push(FtsFragment::And(s.clone()));
Ok(String::new())
}
Term::Property { key, op, value } => self.translate_property(key, *op, value),
Term::And(children) => {
let mut pieces = Vec::new();
for c in children {
let p = self.translate(c)?;
if !p.is_empty() {
pieces.push(p);
}
}
Ok(join_with("AND", &pieces))
}
Term::Or(children) => {
// An OR of pure-literal children collapses into one FTS OR-group.
if children.iter().all(|c| matches!(c, Term::Literal(_))) {
let group: Vec<String> = children
.iter()
.map(|c| match c {
Term::Literal(s) => s.clone(),
_ => unreachable!(),
})
.collect();
self.fts_parts.push(FtsFragment::OrGroup(group));
return Ok(String::new());
}
// Mixed OR — each branch becomes a separate sub-Builder whose
// WHERE fragments we OR together. FTS branches cannot mix
// with structured branches cleanly at the SQL level here;
// keep it simple by requiring that mixed-OR branches produce
// structured-only WHERE fragments.
let mut pieces = Vec::new();
for c in children {
let before = self.fts_parts.len();
let p = self.translate(c)?;
if self.fts_parts.len() > before {
return Err(TranslateError::UnknownProperty(
"OR mixing FTS and structured terms is not supported in Set A".into(),
));
}
if !p.is_empty() {
pieces.push(p);
}
}
Ok(format!("({})", join_with("OR", &pieces)))
}
}
}
/// Delegate to the shared [`build_filter`] fragment builder, then
/// convert its anonymous `?` placeholders to this builder's numbered
/// scheme (params[0] is reserved for the FTS MATCH when one exists;
/// `build` shifts numbers afterwards).
fn translate_property(
&mut self,
key: &str,
op: Op,
value: &str,
) -> Result<String, TranslateError> {
let frag = build_filter(key, op, value, false)?;
let mut params = frag.params.into_iter();
let mut out = String::with_capacity(frag.sql.len() + 8);
for c in frag.sql.chars() {
if c == '?' {
let v = params
.next()
.expect("FilterFragment placeholder/param counts match");
self.all_params.push(v);
out.push('?');
out.push_str(&self.all_params.len().to_string());
} else {
out.push(c);
}
}
Ok(out)
}
}
/// A structured-filter fragment over table alias `f`: SQL with anonymous
/// `?` placeholders plus the values they bind. Anonymous placeholders
/// compose by simple appending — the search cascade tacks fragments onto
@ -295,9 +71,8 @@ pub fn is_filter_key(key: &str) -> bool {
)
}
/// Translate one `key op value` filter into a [`FilterFragment`]. The
/// single source of filter semantics, shared by the legacy numbered
/// [`build`] path and the cascade's [`super::split`].
/// Translate one `key op value` filter into a [`FilterFragment`]. The single
/// source of filter semantics, reached through [`super::split`].
///
/// `glob` marks a value whose unquoted `*` should act as a wildcard — only
/// `name:`/`filename:` honor it; every other key treats the star literally.
@ -457,14 +232,6 @@ pub fn like_subtree_pattern(dir: &str) -> String {
)
}
fn join_with(sep: &str, pieces: &[String]) -> String {
pieces
.iter()
.map(|p| p.clone())
.collect::<Vec<_>>()
.join(&format!(" {} ", sep))
}
/// Parse a date string. Accepts `YYYY-MM-DD`. Returns unix seconds at 00:00 UTC.
fn parse_date_to_unix(s: &str) -> Option<i64> {
// Minimal parser: split on '-' into y/m/d integers.
@ -475,130 +242,144 @@ fn parse_date_to_unix(s: &str) -> Option<i64> {
let y: i64 = parts[0].parse().ok()?;
let m: u32 = parts[1].parse().ok()?;
let d: u32 = parts[2].parse().ok()?;
if !(1..=12).contains(&m) || !(1..=31).contains(&d) || !(1970..=9999).contains(&y) {
if !(1..=12).contains(&m) || !(1970..=9999).contains(&y) {
return None;
}
// Against the real month length, not a flat 1..=31: `civil_to_unix` happily
// rolls 2024-02-31 over into March, so a typo would silently filter on a
// date the user never typed rather than being reported.
if d < 1 || d > days_in_month(y, m) {
return None;
}
// Compute unix seconds using the days-since-epoch formula.
Some(civil_to_unix(y, m as i64, d as i64))
}
/// Days in a Gregorian month, February by the leap rule.
fn days_in_month(year: i64, month: u32) -> u32 {
match month {
1 | 3 | 5 | 7 | 8 | 10 | 12 => 31,
4 | 6 | 9 | 11 => 30,
2 if year % 4 == 0 && (year % 100 != 0 || year % 400 == 0) => 29,
2 => 28,
_ => 0,
}
}
/// Howard Hinnant's civil-from-days algorithm, converting (year, month, day)
/// in the Gregorian calendar to days since 1970-01-01.
fn civil_to_unix(y: i64, m: i64, d: i64) -> i64 {
let y = if m <= 2 { y - 1 } else { y };
let era = if y >= 0 { y } else { y - 399 } / 400;
let yoe = (y - era * 400) as i64;
let yoe = y - era * 400;
let doy = (153 * (if m > 2 { m - 3 } else { m + 9 }) + 2) / 5 + d - 1;
let doe = yoe * 365 + yoe / 4 - yoe / 100 + doy;
let days = era * 146_097 + doe - 719_468;
days * 86_400
}
/// Rewrite placeholder indices after an FTS MATCH parameter was inserted at
/// position 0. Called automatically in [`build`] when needed — but since we
/// append placeholders numerically during translation and prepend the FTS
/// param afterwards, every explicit `?N` in the generated SQL is off by one
/// when FTS is present.
fn shift_placeholders(sql: &str) -> String {
// Find every `?N` and bump N by 1. Only ASCII digits after `?`; skip
// anonymous `?` (which rusqlite won't mix with numbered anyway).
let bytes = sql.as_bytes();
let mut out = String::with_capacity(sql.len());
let mut i = 0;
while i < bytes.len() {
if bytes[i] == b'?' && i + 1 < bytes.len() && bytes[i + 1].is_ascii_digit() {
out.push('?');
let mut j = i + 1;
while j < bytes.len() && bytes[j].is_ascii_digit() {
j += 1;
}
let n: u64 = std::str::from_utf8(&bytes[i + 1..j])
.unwrap()
.parse()
.unwrap();
out.push_str(&(n + 1).to_string());
i = j;
} else {
out.push(bytes[i] as char);
i += 1;
}
}
out
}
#[cfg(test)]
mod tests {
use super::*;
fn build_q(input: &str) -> SqlQuery {
parse_and_build(input, None, 0, Sort::None).expect("build")
/// The filter builder is reached through `split_for_cascade`, which is
/// where these values come from in production; calling it directly here
/// keeps the tests about filter semantics rather than about tokenizing.
fn frag(key: &str, op: Op, value: &str) -> FilterFragment {
build_filter(key, op, value, false).expect("builds")
}
#[test]
fn single_literal() {
let q = build_q("foo");
assert!(q.sql.contains("searchabletext MATCH ?"));
fn type_filter_masks_the_bits() {
let f = frag("type", Op::Contains, "Audio");
assert_eq!(f.sql, "(f.type & ?) != 0");
assert_eq!(
q.params,
vec![rusqlite::types::Value::Text("\"foo\"".into())]
);
}
#[test]
fn implicit_and_joins_as_fts_and() {
let q = build_q("foo bar");
assert!(q.sql.contains("searchabletext MATCH ?"));
let m: &rusqlite::types::Value = &q.params[0];
if let rusqlite::types::Value::Text(s) = m {
assert_eq!(s, "\"foo\" AND \"bar\"");
} else {
panic!();
}
}
#[test]
fn or_of_literals_becomes_fts_group() {
let q = build_q("a OR b");
if let rusqlite::types::Value::Text(s) = &q.params[0] {
assert_eq!(s, "(\"a\" OR \"b\")");
} else {
panic!();
}
}
#[test]
fn type_filter_only() {
let q = build_q("type:Audio");
assert!(q.sql.contains("(f.type & ?"));
assert_eq!(
q.params,
f.params,
vec![rusqlite::types::Value::Integer(FileType::AUDIO.bits() as i64)]
);
}
#[test]
fn modified_ge() {
let q = build_q("modified:>=2024-01-01");
assert!(q.sql.contains("f.mtime >= ?"));
assert_eq!(q.params.len(), 1);
fn modified_comparisons_and_day_ranges() {
let f = frag("modified", Op::Ge, "2024-01-01");
assert!(f.sql.contains("f.mtime >= ?"));
assert_eq!(f.params.len(), 1);
// `=` means the whole day, not the second.
let f = frag("modified", Op::Eq, "2024-05-20");
assert!(f.sql.contains("f.mtime >= ?") && f.sql.contains("f.mtime < ?"));
assert_eq!(f.params.len(), 2);
}
#[test]
fn modified_eq_is_day_range() {
let q = build_q("modified:=2024-05-20");
assert!(q.sql.contains("f.mtime >= ?") && q.sql.contains("f.mtime < ?"));
assert_eq!(q.params.len(), 2);
fn path_filter_covers_the_folder_and_its_subtree() {
let f = frag("path", Op::Contains, "/home/me/docs");
assert!(f.sql.contains("f.parent = ?"));
assert!(f.sql.contains("f.parent LIKE ?"));
// The LIKE half must declare its escape character; without the clause
// a Windows separator would be eaten as an escape.
assert!(f.sql.contains("ESCAPE '\\'"), "{}", f.sql);
}
#[test]
fn path_filter() {
let q = build_q("path:/home/me/docs");
assert!(q.sql.contains("f.parent = ?"));
assert!(q.sql.contains("f.parent LIKE ?"));
// The LIKE half must be escaped and declare its escape character;
// without the clause a Windows separator would be eaten as an escape.
assert!(q.sql.contains("ESCAPE '\\'"), "{}", q.sql);
fn unsupported_shapes_error() {
assert!(matches!(
build_filter("artist", Op::Contains, "beatles", false),
Err(TranslateError::UnknownProperty(_))
));
assert!(matches!(
build_filter("modified", Op::Ge, "not-a-date", false),
Err(TranslateError::BadDate(_))
));
assert!(matches!(
build_filter("type", Op::Contains, "NotAThing", false),
Err(TranslateError::UnknownProperty(_))
));
// Only `name:` accepts a bare `contains`; an ordering operator on it
// is meaningless.
assert!(matches!(
build_filter("name", Op::Ge, "report", false),
Err(TranslateError::UnsupportedOp { .. })
));
}
/// `civil_to_unix` rolls an impossible day over into the next month, so a
/// typo would otherwise filter on a date the user never typed — silently,
/// since the query still runs.
#[test]
fn impossible_dates_are_rejected_rather_than_rolled_over() {
for bad in [
"2024-02-30",
"2023-02-29", // not a leap year
"2024-04-31",
"2024-01-00",
"2024-13-01",
"1969-12-31", // before the epoch
] {
assert!(
parse_date_to_unix(bad).is_none(),
"{} should be rejected",
bad
);
}
for good in ["2024-02-29", "2000-02-29", "2024-01-31", "2024-04-30"] {
assert!(parse_date_to_unix(good).is_some(), "{} should parse", good);
}
// Centurial leap rule: 1900 is not a leap year, 2000 is.
assert!(parse_date_to_unix("1900-02-29").is_none());
}
#[test]
fn is_filter_key_covers_exactly_the_supported_keys() {
for key in [
"type", "modified", "mtime", "path", "folder", "includefolder", "name", "filename",
"mime", "TYPE", "Path",
] {
assert!(is_filter_key(key), "{} should be a filter key", key);
}
for key in ["artist", "regex", "size", ""] {
assert!(!is_filter_key(key), "{} should not be a filter key", key);
}
}
/// The subtree pattern is the one place the separator and the LIKE escape
@ -693,31 +474,6 @@ mod tests {
}
}
#[test]
fn combined_fts_and_type() {
let q = build_q("type:Audio beatles");
assert!(q.sql.contains("searchabletext MATCH ?"));
assert!(q.sql.contains("(f.type & ?"));
// FTS param first, then type bits.
assert_eq!(q.params.len(), 2);
assert_eq!(
q.params[0],
rusqlite::types::Value::Text("\"beatles\"".into())
);
}
#[test]
fn unknown_property_errors() {
let err = parse_and_build("artist:beatles", None, 0, Sort::None).unwrap_err();
assert!(matches!(err, TranslateError::UnknownProperty(_)));
}
#[test]
fn bad_date_errors() {
let err = parse_and_build("modified:>=not-a-date", None, 0, Sort::None).unwrap_err();
assert!(matches!(err, TranslateError::BadDate(_)));
}
#[test]
fn civil_conversion() {
// 2024-01-01 → 1704067200 unix
@ -728,122 +484,4 @@ mod tests {
assert_eq!(civil_to_unix(2000, 2, 29), 951_782_400);
}
#[test]
fn limit_offset_sort_render() {
let q = parse_and_build("foo", Some(10), 20, Sort::ByMtimeDesc).unwrap();
assert!(q.sql.contains("ORDER BY f.mtime DESC"));
assert!(q.sql.contains("LIMIT 10"));
assert!(q.sql.contains("OFFSET 20"));
}
// End-to-end: run a generated query against a real SQLite DB to verify
// placeholder shifting after FTS prepending is correct.
#[test]
fn end_to_end_combined_filter_executes() {
use crate::db::{
open_or_recreate,
repo::{insert_file, set_content_done, NewFile},
};
use crate::mime::FileType;
use rusqlite::types::ToSql;
let mut p = std::env::temp_dir();
p.push(format!(
"qs-query-e2e-{}-{}.sqlite",
std::process::id(),
std::time::SystemTime::now()
.duration_since(std::time::UNIX_EPOCH)
.unwrap()
.as_nanos()
));
let mut conn = open_or_recreate(p.to_str().unwrap(), "trigram").unwrap();
// Two audio files, one document. Only the "audio with beatles" file
// should be returned for `type:Audio beatles`.
let ids: Vec<i64> = {
let tx = conn.transaction().unwrap();
let a = insert_file(
&tx,
&NewFile {
name: "beatles-track.mp3",
path: "/music/beatles-track.mp3",
parent: "/music",
size: 100,
mtime: 1_700_000_000,
inode: None,
device_id: None,
mime: Some("audio/mpeg"),
ftype: FileType::AUDIO,
hash: None,
},
)
.unwrap()
.expect("unique path");
set_content_done(
&tx,
a,
"beatles-track.mp3",
"beatles hey jude",
&[("artist".into(), "The Beatles".into())],
true,
)
.unwrap();
let b = insert_file(
&tx,
&NewFile {
name: "bach.flac",
path: "/music/bach.flac",
parent: "/music",
size: 100,
mtime: 1_700_000_000,
inode: None,
device_id: None,
mime: Some("audio/flac"),
ftype: FileType::AUDIO,
hash: None,
},
)
.unwrap()
.expect("unique path");
set_content_done(&tx, b, "bach.flac", "bach prelude", &[], true).unwrap();
let c = insert_file(
&tx,
&NewFile {
name: "notes.txt",
path: "/docs/notes.txt",
parent: "/docs",
size: 50,
mtime: 1_700_000_000,
inode: None,
device_id: None,
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
},
)
.unwrap()
.expect("unique path");
set_content_done(&tx, c, "notes.txt", "beatles biography", &[], true).unwrap();
tx.commit().unwrap();
vec![a, b, c]
};
let q = parse_and_build("type:Audio beatles", None, 0, Sort::ByMtimeDesc).unwrap();
let rows: Vec<(i64, String, String)> = {
let mut stmt = conn.prepare(&q.sql).expect("prepare");
let bind: Vec<&dyn ToSql> = q.params.iter().map(|v| v as &dyn ToSql).collect();
stmt.query_map(bind.as_slice(), |r| {
Ok((r.get::<_, i64>(0)?, r.get::<_, String>(1)?, r.get::<_, String>(2)?))
})
.unwrap()
.map(|r| r.unwrap())
.collect()
};
assert_eq!(rows.len(), 1, "expected only beatles-track.mp3; got {:?}", rows);
assert_eq!(rows[0].0, ids[0]);
drop(conn);
std::fs::remove_file(&p).ok();
}
}

149
crates/quicksearch-core/src/search/cascade.rs
View file

@ -1,7 +1,6 @@
//! The ranked search cascade.
//!
//! One term, four table scans, eleven ranks. Rank base = stage number, so
//! later stages only ever append to a rank-sorted result list:
//! One term, four table scans, eleven ranks. Rank base = stage number:
//!
//! | rank | meaning | scan |
//! |-----:|----------------------------------|------|
@ -45,9 +44,23 @@
//! Every scan appends the caller's structured-filter SQL (anonymous
//! placeholders over alias `f`) and checks the generation counter as it
//! streams; a bumped generation aborts mid-statement.
//!
//! # Batches are ordered within themselves, not against each other
//!
//! A pass hands hits over *while it scans* (see [`FLUSH_INTERVAL`]) rather than
//! at the end. On a large index a single pass runs for seconds, and holding its
//! results back means the UI shows nothing for that whole time even though a
//! rank-1 filename match may have turned up in the first few milliseconds.
//!
//! The cost is that a scan finds hits in table order, not rank order: batch two
//! can contain something better than anything in batch one. Each batch is
//! sorted before it goes, but the consumer owns the ordering *across* batches —
//! the GUI keeps its table sorted by whichever column is keyed and places each
//! arrival where that sort requires. Do not assume arrival order is rank order.
use std::collections::HashSet;
use std::sync::atomic::{AtomicU64, Ordering};
use std::time::{Duration, Instant};
use rusqlite::Connection;
@ -206,6 +219,19 @@ struct Deferred {
overflowed: bool,
}
/// Longest a pass may sit on hits before handing them over.
///
/// A pass is a whole-table scan, and on a large index that runs for seconds.
/// Waiting for it to finish means the GUI shows nothing for that whole time,
/// even when a rank-1 filename match turned up in the first few milliseconds.
/// Draining on a clock rather than only on a full buffer is what makes a
/// *sparse* query responsive too — six matches out of seven million still
/// paint as the scan reaches them.
///
/// Short enough to land two or three batches inside the GUI's 250 ms result
/// fade, so the fade reveals a filling list rather than an empty one.
const FLUSH_INTERVAL: Duration = Duration::from_millis(80);
struct Cx<'a> {
conn: &'a Connection,
query: &'a CascadeQuery,
@ -223,6 +249,39 @@ struct Cx<'a> {
sink: &'a mut dyn FnMut(Vec<SearchHit>),
}
/// Drives [`Cx::flush_if_due`]: when this pass last handed hits over, and
/// whether it has handed over anything at all.
struct FlushClock {
last: Instant,
sent_anything: bool,
}
impl FlushClock {
fn new() -> FlushClock {
FlushClock {
last: Instant::now(),
sent_anything: false,
}
}
/// Whether a buffer of `len` hits should go now.
///
/// The first batch of a pass goes the moment there is anything to send, so
/// a single early hit paints immediately instead of waiting out an
/// interval it has no reason to.
fn due(&self, len: usize, batch: usize) -> bool {
if len == 0 {
return false;
}
!self.sent_anything || len >= batch || self.last.elapsed() >= FLUSH_INTERVAL
}
fn mark_sent(&mut self) {
self.last = Instant::now();
self.sent_anything = true;
}
}
impl<'a> Cx<'a> {
fn cancelled(&self) -> bool {
self.generation != self.latest_gen.load(Ordering::Relaxed)
@ -288,6 +347,24 @@ impl<'a> Cx<'a> {
Ok(re.is_match(&String::from_utf8_lossy(&raw)))
}
/// Hand `buf` over mid-scan if it is due, leaving it empty when it goes.
///
/// Partial and final flushes are the same operation — [`flush_pass`] sorts
/// what it is given and emits it — so a pass streams simply by calling this
/// each time round its row loop. Ordering *between* batches is the caller's
/// problem: the GUI keeps its table sorted by whichever column is keyed, so
/// a better-ranked hit found later lands above the ones already shown.
fn flush_if_due(&mut self, buf: &mut Vec<SearchHit>, clock: &mut FlushClock) {
if !clock.due(buf.len(), self.options.batch.max(1)) {
return;
}
let batch = std::mem::take(buf);
// `overflowed` belongs to the pass as a whole, not to one batch; the
// final flush reports it.
self.flush_pass(batch, false);
clock.mark_sent();
}
/// Sort a finished pass buffer, truncate to what's left of the display
/// limit, and stream it out in `options.batch`-sized events.
fn flush_pass(&mut self, mut buf: Vec<SearchHit>, overflowed: bool) {
@ -387,6 +464,7 @@ impl<'a> Cx<'a> {
let mut overflowed = false;
let mut path_overflowed = false;
let mut scanned = 0usize;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
scanned += 1;
if scanned % CANCEL_CHECK_ROWS == 0 && self.cancelled() {
@ -451,6 +529,7 @@ impl<'a> Cx<'a> {
} else {
buf.push(hit);
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
}
drop(rows);
@ -529,6 +608,7 @@ impl<'a> Cx<'a> {
let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
let mut buf: Vec<SearchHit> = Vec::new();
let mut overflowed = false;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
// Decompression dominates: check every row.
if self.cancelled() {
@ -546,14 +626,22 @@ impl<'a> Cx<'a> {
let (rank, stage, snip) = match &text {
Some(text) => {
let (count, stage, ci) = {
// Fold once. The case-insensitive count, the first-match
// search and the snippet extraction all need it, and each
// used to make its own copy of a document that can run to
// `maximum_text_size`. The trigram tokenizer is
// case-insensitive, so nearly every candidate takes this
// path — it is a per-row cost, not a per-hit one.
let mut folded: Option<String> = None;
let (count, stage) = {
let count_cs = pattern.count(text, false);
if count_cs > 0 {
(count_cs, 5, false)
(count_cs, 5)
} else {
let count_ci = pattern.count(text, true);
let lower = folded.insert(text.to_ascii_lowercase());
let count_ci = pattern.count_folded(lower);
if count_ci > 0 {
(count_ci, 6, true)
(count_ci, 6)
} else {
// Folded/unordered FTS candidate: the
// pattern never occurs — drop it.
@ -563,9 +651,15 @@ impl<'a> Cx<'a> {
};
// Literal terms keep the richer multi-occurrence
// extract; a wildcard match marks its own first range.
let lower = folded.unwrap_or_else(|| text.to_ascii_lowercase());
let snip = match pattern.literal() {
Some(term) => Some(snippet::extract(text, &[term], &snippet_opts)),
None => pattern.find_first(text, ci).map(|r| {
Some(term) => Some(snippet::extract_folded(
text,
&lower,
&[term],
&snippet_opts,
)),
None => pattern.find_first_folded(&lower).map(|r| {
let r = clamp_match_range(text, r, SNIPPET_WINDOW_CHARS);
snippet::window_around(text, (r.start, r.end), &snippet_opts)
}),
@ -601,6 +695,7 @@ impl<'a> Cx<'a> {
snippet: snip,
});
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
drop(rows);
if self.cancelled() {
@ -644,6 +739,7 @@ impl<'a> Cx<'a> {
let mut overflowed = false;
let mut path_overflowed = false;
let mut scanned = 0usize;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
scanned += 1;
if scanned % 1024 == 0 && self.cancelled() {
@ -656,14 +752,18 @@ impl<'a> Cx<'a> {
continue;
}
// The name is the better match when both fire, so it wins and
// only a name miss falls through to the path tier.
// only a name miss falls through to the path tier. Distance and
// match range come from one sweep — taking them separately meant a
// second full scan of the same buffer for every hit.
let folded_name = name.to_ascii_lowercase();
let (rank, field, folded_field) = match bitap.best_distance(folded_name.as_bytes()) {
Some(distance) => (7.0 + 0.1 * distance as f64, &name, folded_name),
let (rank, field, range) = match bitap.best_distance_and_first(folded_name.as_bytes()) {
Some((distance, range)) => (7.0 + 0.1 * distance as f64, &name, range),
None if with_paths => {
let folded_path = path.to_ascii_lowercase();
match bitap.best_distance(folded_path.as_bytes()) {
Some(distance) => (11.0 + 0.1 * distance as f64, &path, folded_path),
match bitap.best_distance_and_first(folded_path.as_bytes()) {
Some((distance, range)) => {
(11.0 + 0.1 * distance as f64, &path, range)
}
None => continue,
}
}
@ -675,15 +775,13 @@ impl<'a> Cx<'a> {
// Mark the approximate matched span in the matched field for
// the GUI's [matched field] rendering. window_around clamps
// and aligns.
let snip = bitap.count_and_first(folded_field.as_bytes()).1.map(|range| {
snippet::window_around(
field,
range,
&snippet::Options {
approx_chars: field.len().saturating_mul(2).max(8),
},
)
});
let snip = Some(snippet::window_around(
field,
range,
&snippet::Options {
approx_chars: field.len().saturating_mul(2).max(8),
},
));
let is_path_tier = rank >= 11.0;
let hit = SearchHit {
file_id,
@ -701,6 +799,7 @@ impl<'a> Cx<'a> {
} else {
buf.push(hit);
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
}
drop(rows);
@ -746,6 +845,7 @@ impl<'a> Cx<'a> {
let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
let mut buf: Vec<SearchHit> = Vec::new();
let mut overflowed = false;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
if self.cancelled() {
return Ok(false);
@ -785,6 +885,7 @@ impl<'a> Cx<'a> {
snippet: snip,
});
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
drop(rows);
if self.cancelled() {
@ -816,6 +917,7 @@ impl<'a> Cx<'a> {
let mut overflowed = false;
let mut path_overflowed = false;
let mut scanned = 0usize;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
scanned += 1;
if scanned % 1024 == 0 && self.cancelled() {
@ -858,6 +960,7 @@ impl<'a> Cx<'a> {
} else {
buf.push(hit);
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
}
drop(rows);
@ -890,6 +993,7 @@ impl<'a> Cx<'a> {
let snippet_opts = snippet::Options { approx_chars: SNIPPET_WINDOW_CHARS };
let mut buf: Vec<SearchHit> = Vec::new();
let mut overflowed = false;
let mut clock = FlushClock::new();
while let Some(row) = rows.next().map_err(|e| e.to_string())? {
// Decompression dominates: check every row.
if self.cancelled() {
@ -926,6 +1030,7 @@ impl<'a> Cx<'a> {
snippet: snip,
});
overflowed |= self.enforce_cap(&mut buf);
self.flush_if_due(&mut buf, &mut clock);
}
drop(rows);
if self.cancelled() {

3
crates/quicksearch-core/src/search/duplicates.rs
View file

@ -115,6 +115,9 @@ mod tests {
mime: None,
ftype: FileType::TEXT,
hash,
// No MIME, so nothing claims it — and duplicate detection
// never looks at content anyway.
needs_content: false,
},
)
.unwrap()

69
crates/quicksearch-core/src/search/fuzzy.rs
View file

@ -11,6 +11,15 @@
//! convention). Patterns are limited to 64 bytes by the machine word; the
//! cascade skips fuzzy stages for longer terms.
/// Registers the automaton needs: one per error count `0..=k`.
///
/// `k` is bounded by [`edit_budget`]'s one-edit-per-three-characters ladder
/// against a pattern the machine word caps at 64 bytes, so it never exceeds 21
/// however hostile `fuzzy_max_edits` is (pinned by
/// `edit_budget_stays_within_the_bitap_word_size`). [`Bitap::new`] rejects
/// anything larger, so the array is always big enough.
const MAX_REGISTERS: usize = 22;
pub struct Bitap {
/// `masks[c]` has bit `i` set iff `pattern[i] == c`.
masks: [u64; 256],
@ -22,9 +31,10 @@ pub struct Bitap {
impl Bitap {
/// `None` when the pattern is empty, longer than 64 bytes, or the edit
/// budget reaches the word size (`initial_registers` shifts by `k`).
/// budget does not fit the registers (`reset` shifts by `k`, and the
/// register array holds [`MAX_REGISTERS`]).
pub fn new(pattern: &[u8], k: usize) -> Option<Bitap> {
if pattern.is_empty() || pattern.len() > 64 || k >= 64 {
if pattern.is_empty() || pattern.len() > 64 || k >= MAX_REGISTERS {
return None;
}
let mut masks = [0u64; 256];
@ -38,14 +48,19 @@ impl Bitap {
})
}
/// Fresh per-distance state registers. Bit `i` of `r[d]` set means "a
/// Reset the per-distance state registers. Bit `i` of `r[d]` set means "a
/// match of pattern[..=i] with ≤ d errors ends at the current text
/// position". With d errors the first d pattern bytes can be deleted
/// before any text is read, hence the pre-set low bits.
fn initial_registers(&self) -> Vec<u64> {
(0..=self.k)
.map(|d| if d == 0 { 0 } else { (1u64 << d) - 1 })
.collect()
///
/// Writes into a caller-owned array rather than returning a `Vec`: the
/// fuzzy passes call this once per scanned row — millions of them on a
/// whole-table sweep — and a heap allocation per row is pure overhead for
/// at most [`MAX_REGISTERS`] words.
fn reset(&self, r: &mut [u64; MAX_REGISTERS]) {
for (d, reg) in r.iter_mut().enumerate().take(self.k + 1) {
*reg = if d == 0 { 0 } else { (1u64 << d) - 1 };
}
}
/// Advance all registers by one haystack byte. Returns the smallest
@ -78,15 +93,30 @@ impl Bitap {
/// Minimum edit distance (≤ k) of any occurrence of the pattern in
/// `hay`, or `None` if nothing matches within k edits.
pub fn best_distance(&self, hay: &[u8]) -> Option<usize> {
let mut r = self.initial_registers();
let mut best: Option<usize> = None;
for &b in hay {
self.best_distance_and_first(hay).map(|(d, _)| d)
}
/// [`best_distance`](Self::best_distance) plus the first match's
/// approximate byte range, from one sweep.
///
/// The filename pass needs both — the distance to rank by, the range to
/// mark in the snippet — and taking them separately meant a second full
/// scan of the same buffer for every hit. The range carries the same
/// caveat as [`count_and_first`](Self::count_and_first): it assumes a
/// pattern-length match, so edits can shift the true start by up to `k`.
pub fn best_distance_and_first(&self, hay: &[u8]) -> Option<(usize, (usize, usize))> {
let mut r = [0u64; MAX_REGISTERS];
self.reset(&mut r);
let mut best: Option<(usize, (usize, usize))> = None;
for (i, &b) in hay.iter().enumerate() {
if let Some(d) = self.step(&mut r, b) {
let end = i + 1;
let range = (end.saturating_sub(self.len), end);
if d == 0 {
return Some(0);
return Some((0, range));
}
if best.map_or(true, |cur| d < cur) {
best = Some(d);
if best.is_none_or(|(cur, _)| d < cur) {
best = Some((d, range));
}
}
}
@ -100,7 +130,8 @@ impl Bitap {
/// The reported range assumes pattern-length matches — edits can shift
/// the true start by up to k bytes, which is fine for snippet windows.
pub fn count_and_first(&self, hay: &[u8]) -> (usize, Option<(usize, usize)>) {
let mut r = self.initial_registers();
let mut r = [0u64; MAX_REGISTERS];
self.reset(&mut r);
let mut count = 0usize;
let mut first: Option<(usize, usize)> = None;
for (i, &b) in hay.iter().enumerate() {
@ -110,9 +141,7 @@ impl Bitap {
let end = i + 1;
first = Some((end.saturating_sub(self.len), end));
}
for (d, reg) in r.iter_mut().enumerate() {
*reg = if d == 0 { 0 } else { (1u64 << d) - 1 };
}
self.reset(&mut r);
}
}
(count, first)
@ -183,10 +212,12 @@ mod tests {
#[test]
fn oversized_k_is_rejected_not_shifted() {
// `initial_registers` shifts by k; k >= 64 would overflow u64.
// `reset` shifts by k and writes k+1 registers, so both the u64 shift
// and the fixed array have to be respected.
assert!(Bitap::new(b"abc", MAX_REGISTERS).is_none());
assert!(Bitap::new(b"abc", 64).is_none());
assert!(Bitap::new(b"abc", usize::MAX).is_none());
assert!(Bitap::new(b"abc", 63).is_some());
assert!(Bitap::new(b"abc", MAX_REGISTERS - 1).is_some());
}
#[test]

19
crates/quicksearch-core/src/snippet.rs
View file

@ -63,6 +63,18 @@ impl Snippet {
/// `terms` (ASCII-case-insensitive). With no terms or no matches, returns
/// the head of the text as the window with no ranges.
pub fn extract(text: &str, terms: &[&str], opts: &Options) -> Snippet {
extract_folded(text, &text.to_ascii_lowercase(), terms, opts)
}
/// [`extract`] against a haystack the caller has already ASCII-folded.
///
/// The search cascade folds each document once per row and then counts,
/// searches and extracts from that one buffer; folding again here would copy
/// up to `maximum_text_size` a second time for every hit. `folded` must be
/// `text.to_ascii_lowercase()` — the fold is byte-length preserving, which is
/// what lets offsets found in it slice the original.
pub fn extract_folded(text: &str, folded: &str, terms: &[&str], opts: &Options) -> Snippet {
debug_assert_eq!(folded.len(), text.len(), "ASCII folding preserves length");
if text.is_empty() {
return Snippet::empty();
}
@ -75,10 +87,9 @@ pub fn extract(text: &str, terms: &[&str], opts: &Options) -> Snippet {
return head_window(text, opts.approx_chars);
}
// Case-fold once; we do all positioning on the folded buffer and slice
// from the original. Both buffers have identical byte layout because
// `to_ascii_lowercase` only touches ASCII letters.
let folded = text.to_ascii_lowercase();
// All positioning happens on the folded buffer; slices come from the
// original. Both have identical byte layout because `to_ascii_lowercase`
// only touches ASCII letters.
let folded_bytes = folded.as_bytes();
let mut matches: Vec<(usize, usize)> = Vec::new();

171
crates/quicksearch-core/src/walk.rs
View file

@ -178,12 +178,12 @@ struct Shared {
/// observational, for progress display: two relaxed atomic ops per
/// *job* (a directory read plus up to [`FILES_PER_JOB`] files), so it
/// costs nothing the queue mutex didn't already.
busy: AtomicUsize,
stats: WorkerStats,
}
/// Decrements the busy count however the worker leaves its job — including
/// early returns on stop and panics.
struct BusyGuard<'a>(&'a AtomicUsize);
pub(crate) struct BusyGuard<'a>(&'a AtomicUsize);
impl Drop for BusyGuard<'_> {
fn drop(&mut self) {
@ -191,17 +191,32 @@ impl Drop for BusyGuard<'_> {
}
}
/// Lock-free view of walker activity for progress displays.
/// Lock-free view of a worker pool's activity for progress displays.
///
/// Shared by the walk and by [`crate::content`]: a root runs one pool and then
/// the other, and the progress line has to report whichever is live — a pool
/// whose threads have exited reads as busy 0, which is only the truth while
/// that pool is the one running.
#[derive(Clone)]
pub struct WorkerStats {
shared: Arc<Shared>,
busy: Arc<AtomicUsize>,
total: usize,
}
impl WorkerStats {
pub(crate) fn new(total: usize) -> Self {
WorkerStats { busy: Arc::new(AtomicUsize::new(0)), total }
}
/// Count the calling thread as busy until the returned guard drops.
pub(crate) fn enter(&self) -> BusyGuard<'_> {
self.busy.fetch_add(1, Ordering::Relaxed);
BusyGuard(&self.busy)
}
/// Workers doing work right now (the rest are parked).
pub fn active(&self) -> usize {
self.shared.busy.load(Ordering::Relaxed).min(self.total)
self.busy.load(Ordering::Relaxed).min(self.total)
}
pub fn total(&self) -> usize {
@ -378,7 +393,7 @@ struct Ctx {
/// from bytes saves the content pass an open/read/close per file.
registry: Arc<Registry>,
unreadable: UnreadableDirs,
stop_flag: Arc<Mutex<bool>>,
stop_flag: Arc<AtomicBool>,
suspend_flag: Arc<AtomicBool>,
}
@ -458,16 +473,28 @@ fn read_directory(
// directory now *should* lose its row.
Ok(ft) if ft.is_dir() => found.push(Found::Dir(path)),
Ok(ft) if ft.is_symlink() => {
// With links off there is nothing here to index — targets
// included. Returning before `canonicalize` also drops a
// readlink chain and a stat per symlink, so honouring the
// setting costs fewer syscalls than ignoring it, not more.
//
// Both kinds have to be gated together, or the two walkers
// disagree: `filtered_walk` (which the watcher and the
// incremental path use) passes `follow_links` straight to
// walkdir and follows neither. A file target followed here but
// not there is indexed by every full run and never updated
// between them — and, if it resolves outside every configured
// root, indexed despite living somewhere the user never asked
// us to look.
if !ctx.follow_symlinks {
continue;
}
// Resolve aliases where they are found. The target's canonical
// path is what the index stores, and pushing only canonical
// directories is what keeps `seen_dirs` able to break cycles.
if let Ok(target) = path.canonicalize() {
match fs::metadata(&target) {
Ok(m) if m.is_dir() => {
if ctx.follow_symlinks {
found.push(Found::Dir(target));
}
}
Ok(m) if m.is_dir() => found.push(Found::Dir(target)),
// The row for a resolved target belongs to the
// target's own directory, not this one, so it is not
// marked present here and cannot be classified
@ -589,8 +616,7 @@ fn prepare(path: PathBuf, known: Known<'_>, ctx: &Ctx) -> WalkedFile {
fn worker(shared: &Shared, ctx: &Ctx, tx: &mpsc::SyncSender<WalkEvent>) {
while let Some((job, slot)) = shared.take() {
shared.busy.fetch_add(1, Ordering::Relaxed);
let _busy = BusyGuard(&shared.busy);
let _busy = shared.stats.enter();
if should_abort(&ctx.stop_flag, &ctx.suspend_flag) {
shared.shutdown();
return;
@ -728,11 +754,11 @@ impl ParallelWalk {
/// A cheap, cloneable handle for reading worker activity while the
/// walk's iterator is mutably borrowed by a `for` loop.
///
/// Meaningful only while the walk is running: once the workers exit, the
/// pool size stays but the busy count is permanently zero.
pub fn worker_stats(&self) -> WorkerStats {
WorkerStats {
shared: self.shared.clone(),
total: self.handles.len(),
}
self.shared.stats.clone()
}
/// Join the workers and report whether every one of them finished
@ -766,27 +792,37 @@ impl ParallelWalk {
}
}
/// Result of a non-blocking pull from a walk.
pub enum TryNext {
Item(WalkEvent),
/// Nothing ready right now; the walk is still running.
/// Result of a non-blocking pull from a producer pool.
///
/// Shared by every pass the writer loop multiplexes — the walk here and the
/// content pass in [`crate::content`] — so that draining one root's work reads
/// identically whichever pass it came from.
pub enum TryNext<T> {
Item(T),
/// Nothing ready right now; the pass is still running.
Empty,
/// The walk has ended (all workers exited, for any reason).
/// The pass has ended (all workers exited, for any reason).
Finished,
}
/// Translate a non-blocking channel pull into [`TryNext`]. `None` is a
/// receiver the owner already dropped, which reads as finished.
pub(crate) fn try_recv_next<T>(rx: Option<&mpsc::Receiver<T>>) -> TryNext<T> {
match rx {
None => TryNext::Finished,
Some(rx) => match rx.try_recv() {
Ok(item) => TryNext::Item(item),
Err(mpsc::TryRecvError::Empty) => TryNext::Empty,
Err(mpsc::TryRecvError::Disconnected) => TryNext::Finished,
},
}
}
impl ParallelWalk {
/// Non-blocking variant of `next`, for callers multiplexing several
/// walks (the per-root writer loop).
pub fn try_next(&mut self) -> TryNext {
match &self.rx {
None => TryNext::Finished,
Some(rx) => match rx.try_recv() {
Ok(file) => TryNext::Item(file),
Err(mpsc::TryRecvError::Empty) => TryNext::Empty,
Err(mpsc::TryRecvError::Disconnected) => TryNext::Finished,
},
}
pub fn try_next(&mut self) -> TryNext<WalkEvent> {
try_recv_next(self.rx.as_ref())
}
}
@ -823,7 +859,7 @@ pub fn walk_indexable_files(
db_path: &str,
config: Config,
registry: Arc<Registry>,
stop_flag: Arc<Mutex<bool>>,
stop_flag: Arc<AtomicBool>,
suspend_flag: Arc<AtomicBool>,
workers: usize,
) -> ParallelWalk {
@ -862,7 +898,7 @@ pub fn walk_indexable_files(
let shared = Arc::new(Shared {
queue: Mutex::new(queue),
idle: Condvar::new(),
busy: AtomicUsize::new(0),
stats: WorkerStats::new(threads),
});
let ctx = Arc::new(Ctx {
follow_symlinks,
@ -883,7 +919,12 @@ pub fn walk_indexable_files(
let handles = (0..threads)
.map(|_| {
let (shared, ctx, tx) = (shared.clone(), ctx.clone(), tx.clone());
thread::spawn(move || worker(&shared, &ctx, &tx))
thread::spawn(move || {
// Walker threads are the bulk of a run's CPU and I/O; the
// foreground must stay ahead of them.
crate::platform::set_background_priority();
worker(&shared, &ctx, &tx)
})
})
.collect();
// The workers must hold the only senders, or `recv` never reports the end
@ -893,7 +934,10 @@ pub fn walk_indexable_files(
let prefetch = {
let (shared, db_path) = (shared.clone(), db_path.to_string());
thread::spawn(move || prefetcher(&shared, &db_path))
thread::spawn(move || {
crate::platform::set_background_priority();
prefetcher(&shared, &db_path)
})
};
ParallelWalk { rx: Some(rx), handles, prefetch: Some(prefetch), shared, ctx }
@ -999,7 +1043,7 @@ mod tests {
db.to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
))
@ -1181,7 +1225,7 @@ mod tests {
empty_db("unreadable-dir").to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
);
@ -1240,15 +1284,52 @@ mod tests {
)
.unwrap();
let files = walk(&root, &empty_db("symlink-file"));
let files = walk_with(&root, &empty_db("symlink-file"), true, false);
let paths: HashSet<&String> = files.iter().map(|f| &f.path).collect();
assert_eq!(paths.len(), 1, "both routes report one canonical path");
let canonical = path_to_db_string(&root.join("real/target.txt").canonicalize().unwrap());
assert_eq!(*paths.into_iter().next().unwrap(), canonical, "the target, not the alias");
// The alias itself is still reported, so its row is never mistaken for
// deleted — it is reported under the *target's* path.
assert_eq!(files.len(), 2, "seen twice, spelled once");
assert!(files.iter().any(|f| f.aliased), "the link route is marked");
fs::remove_dir_all(&root).ok();
}
/// The counterpart, and the reason both symlink kinds are gated together:
/// `filtered_walk` — which the watcher and the incremental path use —
/// passes `follow_links` to walkdir and follows neither kind. A file link
/// followed only here would be re-indexed by every full run and never
/// updated between them, and its target may sit outside every root.
#[test]
#[cfg(unix)]
fn a_file_symlink_is_not_followed_when_links_are_off() {
let root = tmp_tree("symlink-off");
touch(&root.join("real/target.txt"));
fs::create_dir_all(root.join("links")).unwrap();
std::os::unix::fs::symlink(root.join("real/target.txt"), root.join("links/alias.txt"))
.unwrap();
// A target outside the walked tree: with links off it must not be
// reachable at all.
let outside = tmp_tree("symlink-off-outside");
touch(&outside.join("elsewhere.txt"));
std::os::unix::fs::symlink(outside.join("elsewhere.txt"), root.join("links/out.txt"))
.unwrap();
let files = walk_with(&root, &empty_db("symlink-off"), false, false);
assert_eq!(
names(&files),
vec!["target.txt"],
"only the real file, reached directly"
);
assert!(!files.iter().any(|f| f.aliased), "nothing was resolved");
fs::remove_dir_all(&root).ok();
fs::remove_dir_all(&outside).ok();
}
#[test]
fn hidden_and_ignored_entries_are_pruned() {
let root = tmp_tree("prune");
@ -1272,7 +1353,7 @@ mod tests {
empty_db("prune").to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
));
@ -1311,7 +1392,7 @@ mod tests {
db.to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
));
@ -1348,7 +1429,7 @@ mod tests {
db.to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
));
@ -1377,7 +1458,7 @@ mod tests {
touch(&root.join(format!("f{:04}.txt", i)));
}
let stop = Arc::new(Mutex::new(true));
let stop = Arc::new(AtomicBool::new(true));
let files: Vec<WalkedFile> = files_only(walk_indexable_files(
&[root.to_string_lossy().into_owned()],
false,
@ -1412,7 +1493,7 @@ mod tests {
empty_db("early-drop").to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
);
@ -1438,7 +1519,7 @@ mod tests {
empty_db("overlap").to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
));
@ -1486,7 +1567,7 @@ mod tests {
empty_db("finish").to_str().unwrap(),
Config::default(),
Arc::new(Registry::default_set()),
Arc::new(Mutex::new(false)),
Arc::new(AtomicBool::new(false)),
Arc::new(AtomicBool::new(false)),
4,
);

94
crates/quicksearch-core/src/watcher.rs
View file

@ -157,8 +157,25 @@ fn fmt_cap(cap: usize) -> String {
}
}
/// Tuning for the whole filesystem-event pipeline: the watcher's own
/// per-directory debounce, and the coordinator's queue on the far side of it.
/// Both halves are debounce knobs on one path, so they live together and
/// tests that want a fast pipeline shorten them in one place.
#[derive(Debug, Clone)]
pub struct WatcherConfig {
/// How long the coordinator's queue must go quiet before it is applied.
///
/// An `rm -rf` arrives as a burst; applying it in waves as it lands means
/// no wave sees the whole set to collapse against, and each pays its own
/// costs. Waiting for quiet trades a little latency for one pass over the
/// complete picture. Creates and modifies benefit too — the queue is
/// last-event-wins per path, so a build that rewrites the same files
/// repeatedly collapses for free.
pub pending_settle: Duration,
/// Ceiling on how long [`WatcherConfig::pending_settle`] may hold the
/// queue back. A steady trickle of changes never goes quiet, and must not
/// starve.
pub pending_max_defer: Duration,
/// Per-directory debounce window. Bursts of events in the same directory
/// collapse to one flush after this interval of quiet.
pub throttle_window: Duration,
@ -177,6 +194,8 @@ pub struct WatcherConfig {
impl Default for WatcherConfig {
fn default() -> Self {
Self {
pending_settle: Duration::from_secs(2),
pending_max_defer: Duration::from_secs(30),
throttle_window: Duration::from_secs(30),
tick_interval: Duration::from_millis(500),
max_dirs_per_tick: 64,
@ -244,6 +263,17 @@ impl WatchRegistry {
/// directories on its own; unwatching anyway keeps notify's internal
/// descriptor map from growing across a long session of directory churn.
fn remove_tree(&mut self, dir: &Path) -> usize {
// The scan below is O(watched dirs), and the event loop calls this for
// every Remove — files included. Deleting a directory of 10k files
// would otherwise cost 10k passes over a set capped at 128k entries.
//
// Exact, not a heuristic: registration always walks top-down (see
// `register_tree` and startup), so a watched directory beneath `dir`
// implies `dir` itself is watched. An unwatched `dir` therefore has no
// watched descendants and there is nothing to find.
if !self.dirs.contains(dir) {
return 0;
}
let doomed: Vec<PathBuf> = self
.dirs
.iter()
@ -464,13 +494,8 @@ struct LoopCtx {
degraded: Degraded,
}
/// A queued event, deduplicated per path within a window.
#[derive(Debug, Clone)]
struct QueuedEvent {
op: QueuedOp,
}
#[derive(Debug, Clone, Copy)]
/// A queued operation, deduplicated per path within a window.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum QueuedOp {
Create,
Modify,
@ -484,7 +509,7 @@ struct DirThrottleEntry {
record_time: Instant,
/// Per-path pending op. Same path seen twice in a window keeps only the
/// latest op — coalescing a rename-as-create+modify spam into one event.
queue: HashMap<PathBuf, QueuedEvent>,
queue: HashMap<PathBuf, QueuedOp>,
/// If true, the next tick flushes regardless of window age. Set for the
/// first event in a previously-idle directory so it reacts fast.
immediate: bool,
@ -694,14 +719,12 @@ fn enqueue(
immediate: true,
});
// Coalesce: Remove after Create → drop both. Modify after Modify → one Modify.
match (op, entry.queue.get(&path).map(|q| q.op)) {
match (op, entry.queue.get(&path).copied()) {
(QueuedOp::Remove, Some(QueuedOp::Create)) => {
entry.queue.remove(&path);
}
_ => {
entry
.queue
.insert(path, QueuedEvent { op });
entry.queue.insert(path, op);
}
}
}
@ -726,11 +749,7 @@ fn flush_ready(
}
for dir in ready {
if let Some(entry) = throttle.get_mut(&dir) {
let drained: Vec<(PathBuf, QueuedOp)> = entry
.queue
.drain()
.map(|(p, q)| (p, q.op))
.collect();
let drained: Vec<(PathBuf, QueuedOp)> = entry.queue.drain().collect();
entry.immediate = false;
entry.record_time = now;
for (path, op) in drained {
@ -801,6 +820,42 @@ mod tests {
dir
}
/// A registry with `dirs` seeded directly, so the pure set logic can be
/// tested without registering real kernel watches.
fn registry_with(dirs: &[&str]) -> WatchRegistry {
let raw = RecommendedWatcher::new(|_res| {}, NotifyConfig::default()).unwrap();
WatchRegistry {
raw,
dirs: dirs.iter().map(PathBuf::from).collect(),
cap: 64,
}
}
/// Most `Remove` events name *files*, which are never watched. Returning
/// before the O(watched) scan is what keeps deleting a 10k-file directory
/// off 10k passes over a set capped at 128k entries.
#[test]
fn remove_tree_skips_the_scan_for_a_path_that_is_not_watched() {
let mut reg = registry_with(&["/a/b", "/a/b/c", "/a/bc"]);
// A file inside a watched directory: not watched itself, and nothing
// can live beneath it.
assert_eq!(reg.remove_tree(Path::new("/a/b/file.txt")), 0);
assert_eq!(reg.dirs.len(), 3, "the watch set is untouched");
// An entirely unrelated path is likewise a no-op.
assert_eq!(reg.remove_tree(Path::new("/elsewhere")), 0);
assert_eq!(reg.dirs.len(), 3);
// The watched directory itself still takes its descendants with it —
// and only its descendants: /a/bc is a sibling, not a child.
assert_eq!(reg.remove_tree(Path::new("/a/b")), 2);
assert_eq!(
reg.dirs.iter().collect::<Vec<_>>(),
vec![&PathBuf::from("/a/bc")]
);
}
#[test]
fn enqueue_create_then_remove_cancels() {
let mut map: HashMap<PathBuf, DirThrottleEntry> = HashMap::new();
@ -866,10 +921,7 @@ mod tests {
fn prune_stale_keeps_active_entries() {
let mut map: HashMap<PathBuf, DirThrottleEntry> = HashMap::new();
let mut queue = HashMap::new();
queue.insert(
PathBuf::from("/tmp/a"),
QueuedEvent { op: QueuedOp::Modify },
);
queue.insert(PathBuf::from("/tmp/a"), QueuedOp::Modify);
map.insert(
PathBuf::from("/tmp"),
DirThrottleEntry {

146
crates/quicksearch-core/tests/cascade.rs
View file

@ -55,6 +55,7 @@ impl Seeder {
mime: Some("text/plain"),
ftype: FileType::TEXT,
hash: None,
needs_content: true,
},
)
.unwrap()
@ -73,20 +74,47 @@ impl Seeder {
/// Run the cascade synchronously, collecting every batch. Returns
/// (flattened hits in emission order, outcome).
/// Run a search and return its hits in rank order.
///
/// The cascade streams batches *while* each pass scans, so arrival order is
/// table order, not rank order — batch two can hold something better than
/// anything in batch one. Ordering across batches belongs to the consumer, and
/// this mirrors what the GUI does with the default sort key, so the ranking
/// assertions below stay about ranking rather than about scan order.
///
/// Use [`run_collect_batches`] to assert on the stream itself.
fn run_collect(
conn: &rusqlite::Connection,
input: &str,
options: &SearchOptions,
) -> (Vec<SearchHit>, cascade::Outcome) {
let (batches, outcome) = run_collect_batches(conn, input, options);
let mut hits: Vec<SearchHit> = batches.into_iter().flatten().collect();
hits.sort_by(|a, b| {
a.rank
.partial_cmp(&b.rank)
.unwrap_or(std::cmp::Ordering::Equal)
.then_with(|| a.name.cmp(&b.name))
.then_with(|| a.path.cmp(&b.path))
});
(hits, outcome)
}
/// The raw batch stream, one `Vec` per sink call.
fn run_collect_batches(
conn: &rusqlite::Connection,
input: &str,
options: &SearchOptions,
) -> (Vec<Vec<SearchHit>>, cascade::Outcome) {
let split = split_for_cascade(input).expect("split");
let latest = AtomicU64::new(7);
let mut hits = Vec::new();
let mut batches = Vec::new();
let outcome = cascade::run(conn, &split, options, 7, &latest, &mut |batch| {
hits.extend(batch)
batches.push(batch)
})
.expect("cascade run")
.expect("not cancelled");
(hits, outcome)
(batches, outcome)
}
fn fuzzy_options() -> SearchOptions {
@ -1013,3 +1041,115 @@ fn service_reports_missing_db_as_error() {
assert!(got_error, "missing index must surface as a search error");
service.shutdown();
}
/// The point of the whole change: a pass hands hits over *while* it scans, so
/// the UI has something to show long before the scan ends.
///
/// Proven by ordering rather than by batch count — `flush_pass` has always
/// chunked its output, so counting sink calls proves nothing. Pass A scans in
/// `files.path` order, so seeding a *worse* match at an early path and a
/// *better* one at a late path separates the two designs: emitting at the end
/// sorts them and leads with rank 1, while streaming hands over the rank-3 hit
/// before the scan has even reached the rank-1 one.
///
/// That is the trade being made deliberately: arrival order is scan order, and
/// ordering across batches belongs to the consumer.
#[test]
fn a_pass_hands_hits_over_before_the_scan_reaches_the_end() {
let p = tmp_db("stream");
let mut s = Seeder::new(&p, true);
let early_worse = s.add("my_zebra_file.txt", "/aaa", 1, None); // rank 3
for i in 0..400 {
s.add(&format!("filler{:04}.txt", i), "/mmm", i as u64 + 2, None);
}
let late_better = s.add("zebra", "/zzz", 999, None); // rank 1
let conn = s.done();
let options = SearchOptions {
batch: 100,
..SearchOptions::default()
};
let (batches, outcome) = run_collect_batches(&conn, "zebra", &options);
assert_eq!(outcome.total, 2, "both matches still reach the sink");
assert_eq!(
batches.first().map(|b| b.as_slice()).and_then(|b| b.first()).map(|h| h.file_id),
Some(early_worse),
"the early hit should have gone out before the scan found the better one; \
batch sizes: {:?}",
batches.iter().map(|b| b.len()).collect::<Vec<_>>()
);
// ...and sorting the stream the way the GUI does still puts rank 1 on top.
let (hits, _) = run_collect(&conn, "zebra", &options);
assert_eq!(
hits.iter().map(|h| h.file_id).collect::<Vec<_>>(),
vec![late_better, early_worse],
"the consumer's sort restores rank order"
);
drop(conn);
std::fs::remove_file(&p).ok();
}
/// The time bound, which a size-only rule would miss: a query matching a
/// handful of rows out of many still paints them as the scan reaches them
/// rather than at the end.
#[test]
fn a_sparse_match_still_streams_before_the_scan_ends() {
let p = tmp_db("sparse");
let mut s = Seeder::new(&p, true);
// Three needles, spread through a haystack far larger than one batch.
for i in 0..3000 {
let name = if i % 1000 == 500 {
format!("needle{:04}.txt", i)
} else {
format!("hay{:04}.txt", i)
};
s.add(&name, "/d", i as u64 + 1, Some("body"));
}
let conn = s.done();
let options = SearchOptions {
batch: 100,
..SearchOptions::default()
};
let (batches, outcome) = run_collect_batches(&conn, "needle", &options);
assert_eq!(outcome.total, 3, "all three needles found");
assert_eq!(batches.iter().map(|b| b.len()).sum::<usize>(), 3);
assert!(
!batches.is_empty() && batches[0].len() < 3,
"the first needle should not have waited for the other two; batches: {:?}",
batches.iter().map(|b| b.len()).collect::<Vec<_>>()
);
}
/// Streaming must not change *what* a search returns, only when. The rank
/// ordering assertions elsewhere in this file are the detailed version; this
/// pins the set and the count against a mixed-tier query.
#[test]
fn streaming_does_not_change_the_result_set() {
let p = tmp_db("stream-set");
let mut s = Seeder::new(&p, true);
let exact = s.add("zebra", "/d", 1, Some("nothing here"));
let sub = s.add("my_zebra_file.txt", "/d", 2, Some("nothing here"));
let content = s.add("unrelated.txt", "/d", 3, Some("a zebra in the text"));
let conn = s.done();
for batch in [1usize, 2, 100] {
let options = SearchOptions {
batch,
..SearchOptions::default()
};
let (hits, outcome) = run_collect(&conn, "zebra", &options);
let ids: Vec<i64> = hits.iter().map(|h| h.file_id).collect();
assert_eq!(
ids,
vec![exact, sub, content],
"batch size {} must not change ranking",
batch
);
assert_eq!(outcome.total, 3, "batch size {}", batch);
}
}

93
crates/quicksearch-core/tests/encrypted.rs
View file

@ -107,9 +107,50 @@ fn encrypted_index_lifecycle() {
"plaintext content leaked into the encrypted file"
);
// --- Optimizing a keyed index: VACUUM keeps it encrypted. ---
//
// VACUUM rewrites the whole file through a temporary database that
// SQLCipher has to key from the main one. If it did not, the rewrite would
// hand back a plaintext index — silently, and only for protected users.
//
// The slack is manufactured: this tree is two files, and `maintain` only
// rewrites a file with something to reclaim.
{
let conn = db::open_existing(&db_path.to_string_lossy(), true).unwrap();
conn.execute_batch(
"INSERT INTO files (name, path, parent, size, mtime, type, basic_state, content_state)
WITH RECURSIVE n(i) AS (
SELECT 1 UNION ALL SELECT i + 1 FROM n WHERE i < 20000
)
SELECT 'p' || i, '/pad/' || i, '/pad', 0, 0, 0, 1, 3 FROM n;
DELETE FROM files WHERE parent = '/pad';",
)
.unwrap();
drop(conn);
let conn = db::open::open_maintenance(&db_path.to_string_lossy()).unwrap();
let dir = data.to_string_lossy().into_owned();
assert!(
quicksearch_core::db::repo::maintain(&conn, &dir).unwrap(),
"that much slack should have been reclaimed"
);
drop(conn);
assert_ne!(
&header(&db_path),
b"SQLite format 3\0",
"the vacuum's replacement file must still be encrypted"
);
assert_eq!(
match_count(&db_path, "zebrapayload"),
1,
"and still searchable under the same key"
);
}
// --- Wrong password / no password: tagged error, file intact. ---
let before = std::fs::read(&db_path).unwrap();
db::set_process_key(Some(wrong_key));
db::set_process_key(Some(wrong_key.clone()));
let err = db::verify_process_key(&db_path.to_string_lossy()).unwrap_err();
assert!(err.starts_with(db::KEY_MISMATCH_PREFIX), "got: {err}");
db::set_process_key(None);
@ -121,6 +162,56 @@ fn encrypted_index_lifecycle() {
"failed unlocks must never modify the index"
);
// --- A stale schema must not read as a locked index. ---
//
// The reported failure: after a schema bump, a password-protected install
// could not start at all — the correct password was rejected with
// "not a compatible QuickSearch index (schema v4 expected)", because the
// unlock gate verified the key by opening the index the way a *consumer*
// does, which also insists the schema be current. An unprotected install
// in the same state starts and rebuilds on its first run; the protected
// one had no way past the gate.
//
// Whether the schema is current belongs to the indexer, which answers it
// by wiping and rebuilding. Unlocking only has to answer "does this key
// open the file?".
{
db::set_process_key(Some(key.clone()));
// Age the stored schema, exactly as a version bump would.
let conn = db::open_existing(&db_path.to_string_lossy(), true).unwrap();
conn.execute(
"UPDATE schema_info SET value = '1' WHERE key = 'version'",
[],
)
.unwrap();
drop(conn);
// The right password still unlocks...
db::verify_process_key(&db_path.to_string_lossy()).expect(
"a stale schema must not make the correct password look wrong",
);
// ...and the wrong one is still refused, with the same tagged error —
// the relaxation must not have turned the check into a rubber stamp.
db::set_process_key(Some(wrong_key.clone()));
let err = db::verify_process_key(&db_path.to_string_lossy()).unwrap_err();
assert!(err.starts_with(db::KEY_MISMATCH_PREFIX), "got: {err}");
// Consumers still refuse a stale index, which is what sends the
// indexer down its rebuild path.
db::set_process_key(Some(key.clone()));
let err = db::open_existing(&db_path.to_string_lossy(), false).unwrap_err();
assert!(err.contains("not a compatible QuickSearch index"), "got: {err}");
// And the rebuild comes back encrypted and searchable under the same
// key, so the whole path a real user walks is covered.
index_once(&root, &db_path, &config);
assert_ne!(&header(&db_path), b"SQLite format 3\0");
assert_eq!(match_count(&db_path, "zebrapayload"), 1);
// Hand the next section the unkeyed state it expects.
db::set_process_key(None);
}
// --- Disable: delete + rebuild produces a plaintext index. ---
let service = IndexingService::new();
service

471
crates/quicksearch-core/tests/full_index.rs
View file

@ -7,10 +7,12 @@
//! stale — and only appears on the second run.
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
use quicksearch_core::config::Config;
use quicksearch_core::indexing::{IndexingService, IndexingStatus};
use quicksearch_core::file_handling::{extract_scope_prepare, ExtractCursor};
use quicksearch_core::indexing::{IndexingService, IndexingStatus, RootPhase};
fn tmp_dir(tag: &str) -> PathBuf {
let mut p = std::env::temp_dir();
@ -270,6 +272,137 @@ fn stopping_mid_run_deletes_nothing() {
std::fs::remove_dir_all(&db_dir).ok();
}
#[test]
fn a_stamped_run_has_finished_its_stale_cleanup() {
// `last_full_index` is what the coordinator schedules the next periodic
// reindex from. Stamping it for a run that was cut short suppresses
// reindexing for the whole interval (24 h by default) — and the damage is
// concrete: stale cleanup is skipped when the run is stopped, so rows for
// files that no longer exist stay in the index and keep turning up in
// search results until something else forces a rebuild.
//
// The hole this guards: the writer loop set `aborted` only at the *top* of
// an iteration, while the "every root is Done" exit sits at the bottom and
// breaks directly. A stop landing inside the pass — or inside stale cleanup
// itself, which returns early and leaves rows behind — reached that bottom
// break with `aborted` still false and stamped the run as complete.
//
// The assertion is one-sided on purpose, so timing can never make it fail
// spuriously: a stamp *always* has to mean cleanup finished, whether the
// stop landed inside the window or never landed at all.
let root = tmp_dir("stop-stamp");
let db_dir = tmp_dir("stop-stamp-db");
let db = db_dir.join("index.sqlite");
let mut config = test_config();
// Nothing to extract, so a root goes Walking → Done in one pass and the
// run's whole tail is the stale cleanup this test wants to interrupt.
config.processing.maximum_text_file_size = 0;
const FILES: usize = 8000;
for i in 0..FILES {
touch(&root.join(format!("d{}/f{:05}.txt", i % 25, i)), b"body");
}
index_once(&root, &db, &config);
assert_eq!(rows(&db).len(), FILES);
// Every file vanishes, so the next run has FILES stale rows to delete —
// a tail long enough for a stop to land inside it.
for i in 0..FILES {
std::fs::remove_file(root.join(format!("d{}/f{:05}.txt", i % 25, i))).unwrap();
}
let marker = |db: &Path| -> Option<u64> {
let conn = rusqlite::Connection::open(db).ok()?;
quicksearch_core::db::repo::get_last_full_index(&conn)
};
for delay_ms in [2u64, 5, 10, 20, 35, 60, 100, 200] {
{
let conn = rusqlite::Connection::open(&db).unwrap();
conn.execute("DELETE FROM schema_info WHERE key = 'last_full_index'", [])
.unwrap();
}
assert_eq!(marker(&db), None, "stamp cleared before the run");
let service = IndexingService::new();
service
.start_indexing(
vec![root.to_string_lossy().into_owned()],
db.to_string_lossy().into_owned(),
config.clone(),
)
.unwrap();
std::thread::sleep(Duration::from_millis(delay_ms));
service.stop_indexing().unwrap();
drop(service);
std::thread::sleep(Duration::from_millis(300));
if marker(&db).is_some() {
assert_eq!(
rows(&db).len(),
0,
"delay {}ms: the run stamped itself complete but left stale rows behind",
delay_ms
);
// Cleanup finished, so there is nothing left for later delays to
// interrupt; the rest of the sweep would be vacuous.
break;
}
}
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
#[test]
fn starting_a_run_claims_the_status_before_it_returns() {
// The coordinator enforces the single-writer rule by polling
// `get_status()`. That is only sound if the Running transition has already
// happened when `start_indexing` returns — it used to be performed by the
// service's command thread, *after* it joined the previous run's handle,
// so a caller could see Idle and start writing to the database this run is
// about to reopen (and possibly wipe).
let root = tmp_dir("start-claims");
let db_dir = tmp_dir("start-claims-db");
let db = db_dir.join("index.sqlite");
let config = test_config();
touch(&root.join("a.txt"), b"body");
let service = IndexingService::new();
service
.start_indexing(
vec![root.to_string_lossy().into_owned()],
db.to_string_lossy().into_owned(),
config.clone(),
)
.unwrap();
// No sleep, no poll: the very next observation must already be Running.
assert!(
matches!(service.get_status(), IndexingStatus::Running { .. }),
"status must be claimed synchronously, got {:?}",
service.get_status()
);
// And a second start is a reportable error rather than a silently
// dropped command.
let err = service
.start_indexing(
vec![root.to_string_lossy().into_owned()],
db.to_string_lossy().into_owned(),
config.clone(),
)
.unwrap_err();
assert!(err.contains("already running"), "got: {}", err);
service.stop_indexing().unwrap();
drop(service);
std::thread::sleep(Duration::from_millis(250));
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
#[test]
fn a_wide_tree_indexes_every_file_exactly_once() {
// Exercises the parallel walk's chunking and termination against a real
@ -444,7 +577,10 @@ fn a_symlink_target_in_an_unwalked_directory_survives_reindexing() {
let outside = tmp_dir("alias-outside");
let db_dir = tmp_dir("alias-db");
let db = db_dir.join("index.sqlite");
let config = test_config();
// Aliases only exist when links are followed; with the default (off) a
// symlink is not resolved at all, which the tail of this test checks.
let mut config = test_config();
config.indexing.follow_symlinks = true;
touch(&root.join("normal.txt"), b"inside the root");
@ -471,6 +607,16 @@ fn a_symlink_target_in_an_unwalked_directory_survives_reindexing() {
index_once(&root, &db, &config);
assert_eq!(rows(&db), first, "an aliased row must survive a re-index");
// And the other half of the setting: with links off, neither target is
// indexed — including the one outside the root, which the user never asked
// us to look at. This is also what keeps the full run in agreement with
// `filtered_walk`, which the watcher uses and which follows neither kind.
let db2 = db_dir.join("links-off.sqlite");
index_once(&root, &db2, &test_config());
let off: Vec<String> = rows(&db2).into_iter().map(|(p, _, _)| p).collect();
assert_eq!(off.len(), 1, "only the ordinary file: {:?}", off);
assert!(off[0].ends_with("normal.txt"));
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&outside).ok();
std::fs::remove_dir_all(&db_dir).ok();
@ -486,7 +632,8 @@ fn a_modified_symlink_target_is_updated_not_silently_ignored() {
let outside = tmp_dir("alias-mod-outside");
let db_dir = tmp_dir("alias-mod-db");
let db = db_dir.join("index.sqlite");
let config = test_config();
let mut config = test_config();
config.indexing.follow_symlinks = true;
let target = outside.join("target.txt");
touch(&target, b"first body");
@ -735,6 +882,69 @@ fn undecodable_small_files_are_reported_as_failures_not_silently_skipped() {
std::fs::remove_dir_all(&db_dir).ok();
}
/// End-to-end version of the fix: the extraction denominator the manage-index
/// tab renders is `extract_total`, and it must count files that need text —
/// not every indexed file. Asserted through a real `IndexingService` run so it
/// covers the walk, the batch writers and `extract_scope_prepare` together.
#[test]
fn the_extraction_denominator_counts_only_files_that_need_text() {
let root = tmp_dir("denominator");
let db_dir = tmp_dir("denominator-db");
let db = db_dir.join("index.sqlite");
// Three files an extractor claims, seven it never will. `big.txt` is the
// interesting one: larger than `hash_length`, so the walk cannot finish it
// inline and it is the only row the content pass actually opens.
for name in ["a.txt", "b.json"] {
touch(&root.join(name), b"body bytes with no magic");
}
touch(&root.join("big.txt"), &vec![b'z'; 32 * 1024]);
for name in ["d.mp4", "e.zip", "f.bin", "g.exe", "h.iso", "i.so", "j"] {
touch(&root.join(name), b"body bytes with no magic");
}
let config = Config::default();
index_once(&root, &db, &config);
let conn = rusqlite::Connection::open(&db).unwrap();
let count = |state: i64| -> i64 {
conn.query_row(
"SELECT COUNT(*) FROM files WHERE content_state = ?1",
[state],
|r| r.get(0),
)
.unwrap()
};
assert_eq!(count(0), 0, "a finished run leaves nothing pending");
assert_eq!(count(1), 3, "the claimed files have text");
assert_eq!(count(3), 7, "the rest are NA, and were NA from the walk on");
drop(conn);
// The exact call `indexing.rs` makes to fill `RootProgress::extract_total`,
// run against the index the full pass just produced. Asserted here rather
// than by sampling the live status, which cannot be observed reliably: a
// ten-file tree finishes between two polls.
let conn = Arc::new(Mutex::new(
// Writable: the scope call's first act is the idempotent oversize sweep.
quicksearch_core::db::open_existing(db.to_str().unwrap(), true).unwrap(),
));
let cursor = ExtractCursor::for_root(root.to_str().unwrap());
let scope = extract_scope_prepare(&conn, &cursor, &config).unwrap();
assert_eq!(
(scope.pending, scope.already_done),
(0, 3),
"extract_total is the searchable set, not the file count"
);
// Which is what the row renders: "3 / 3" on an unchanged re-run. Before
// this was decided at walk time it read "10 / 10", seven of them files
// with nothing to extract.
assert_eq!(scope.pending + scope.already_done, 3);
drop(conn);
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
#[test]
fn an_empty_file_is_done_with_no_snippet_sidecar() {
let root = tmp_dir("inline-empty");
@ -822,3 +1032,258 @@ fn contentless_mode_still_indexes_inlined_files_without_storing_bodies() {
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
/// A slow root must not stall the others.
///
/// This is the complaint stated directly: one root doing heavy extraction used
/// to occupy the single writer thread — and the database connection — for a
/// whole batch of files at a time, during which no other root's walk was
/// drained at all. Their walker threads filled their channels and blocked.
///
/// So the assertion is about *stalls*, not throughput. Throughput would be the
/// wrong measure: writing is serial by construction (one SQLite connection),
/// so on a local disk the writer, not extraction, is the bottleneck and a
/// wall-clock comparison would mostly measure the machine.
#[test]
fn a_heavy_root_does_not_stall_a_light_one() {
// HEAVY: few files, each big enough that reading it is real work, with a
// small `maximum_text_size` so the cost lands in extraction rather than in
// the writer's tokenising.
let heavy = tmp_dir("stall-heavy");
let body: Vec<u8> = "sphinx of black quartz judge my vow ".repeat(40_000).into_bytes();
for i in 0..200 {
touch(&heavy.join(format!("d{}/big{:04}.txt", i % 8, i)), &body);
}
// LIGHT: a wide tree of tiny files, so its walk runs long enough to sample
// and its progress counter moves finely.
let light = tmp_dir("stall-light");
for i in 0..6000 {
touch(&light.join(format!("d{}/f{:05}.txt", i % 60, i)), b"x");
}
let db_dir = tmp_dir("stall-db");
let db = db_dir.join("index.sqlite");
let mut config = test_config();
config.processing.maximum_text_size = 1024;
config.processing.maximum_text_file_size = 8 * 1024 * 1024;
let service = IndexingService::new();
service
.start_indexing(
vec![
heavy.to_string_lossy().into_owned(),
light.to_string_lossy().into_owned(),
],
db.to_string_lossy().into_owned(),
config.clone(),
)
.unwrap();
// Sample the light root's progress while the heavy one is extracting, and
// keep the longest interval over which it did not move.
let mut worst = Duration::ZERO;
let mut last_change = Instant::now();
let mut last_seen = 0usize;
let mut sampled_together = false;
let deadline = Instant::now() + Duration::from_secs(120);
while Instant::now() < deadline {
match service.get_status() {
IndexingStatus::Running { roots, .. } => {
let heavy_p = roots.iter().find(|r| r.root.contains("stall-heavy"));
let light_p = roots.iter().find(|r| r.root.contains("stall-light"));
if let (Some(h), Some(l)) = (heavy_p, light_p) {
let light_busy = l.phase != RootPhase::Done;
if h.phase == RootPhase::Extracting && light_busy {
sampled_together = true;
let now = l.walked + l.extracted;
if now != last_seen {
last_seen = now;
last_change = Instant::now();
} else {
worst = worst.max(last_change.elapsed());
}
}
}
}
IndexingStatus::Error(e) => panic!("indexing failed: {}", e),
_ => break,
}
std::thread::sleep(Duration::from_millis(2));
}
service.stop_indexing().unwrap();
drop(service);
assert!(
sampled_together,
"never observed the two roots overlapping; the fixture is not exercising the case"
);
// Measured on this fixture: ~20 ms with the extraction pools, ~130 ms when
// the file reading is forced back onto the writer thread (and unbounded in
// the real failure, where the heavy root is on a network share). The bound
// sits between, with several times the observed headroom.
//
// The fixed design's stall does not grow with the heavy root's cost — it is
// one round-robin pass plus one commit — so making that root heavier only
// widens the margin.
eprintln!("longest light-root stall while heavy extracted: {:?}", worst);
assert!(
worst < Duration::from_millis(100),
"the light root stalled for {:?} while the heavy root extracted",
worst
);
std::fs::remove_dir_all(&heavy).ok();
std::fs::remove_dir_all(&light).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
/// The write-ahead log must not grow for the length of a run.
///
/// SQLite's autocheckpoint copies committed frames into the index but can only
/// *reset* the log at an instant no reader holds a read mark — a lock it tries
/// once, without retrying. A run keeps a reader per root querying continuously,
/// so that instant does not come and the log appends until the run ends: the
/// case that prompted this was a 12.5 GiB index carrying a 21.6 GiB log.
///
/// So the assertion is about the *peak while running*. It has to be sampled
/// in flight — `stop_indexing` and the post-run maintenance both truncate the
/// log on the way out, so a reading taken afterwards proves nothing about what
/// happened during.
#[test]
fn the_wal_stays_bounded_during_a_run() {
let root = tmp_dir("wal-bound");
// Wide and text-heavy: every file lands in the FTS index, which is what
// actually fills the log.
let body: Vec<u8> = "sphinx of black quartz judge my vow ".repeat(200).into_bytes();
for i in 0..4000 {
touch(&root.join(format!("d{}/f{:05}.txt", i % 40, i)), &body);
}
let db_dir = tmp_dir("wal-bound-db");
let db = db_dir.join("index.sqlite");
let wal = db_dir.join("index.sqlite-wal");
let mut config = test_config();
// The floor `MINIMUM_WAL_SIZE` clamps to, so the cap is exercised many
// times over a fixture this size rather than once at the very end.
config.processing.maximum_wal_size = 16 * 1024 * 1024;
let service = IndexingService::new();
service
.start_indexing(
vec![root.to_string_lossy().into_owned()],
db.to_string_lossy().into_owned(),
config.clone(),
)
.unwrap();
let mut peak = 0u64;
let mut checkpointed = false;
let mut last = 0u64;
let deadline = Instant::now() + Duration::from_secs(120);
while Instant::now() < deadline {
let len = std::fs::metadata(&wal).map(|m| m.len()).unwrap_or(0);
peak = peak.max(len);
// A drop in length is a checkpoint that ran mid-run; without one the
// bound below could be met simply by the fixture being too small.
if len + 1024 * 1024 < last {
checkpointed = true;
}
last = len;
match service.get_status() {
IndexingStatus::Running { .. } => {}
IndexingStatus::Error(e) => panic!("indexing failed: {}", e),
_ => break,
}
std::thread::sleep(Duration::from_millis(2));
}
// Let the maintenance pass finish before tearing the service down.
let idle_by = Instant::now() + Duration::from_secs(120);
while Instant::now() < idle_by && !matches!(service.get_status(), IndexingStatus::Idle) {
std::thread::sleep(Duration::from_millis(10));
}
let after = std::fs::metadata(&wal).map(|m| m.len()).unwrap_or(0);
drop(service);
eprintln!("peak WAL during the run: {} bytes", peak);
assert!(
checkpointed,
"the log never shrank mid-run; the fixture is not exercising the cap"
);
// Generously above the 16 MiB cap: the check runs between round-robin
// rounds, so a round's worth of commits can land on top of it, and a
// checkpoint that loses a lock race defers to the next cap of growth.
assert!(
peak < 96 * 1024 * 1024,
"the log peaked at {} bytes against a 16 MiB cap",
peak
);
assert_eq!(after, 0, "the optimize pass leaves an empty log behind");
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}
/// Stopping a run does not skip the optimize pass.
///
/// A run cut short is exactly when the log is at its largest and nothing else
/// will come along to land it: the writer connection closes, and the next run
/// may be hours away. So Stop ends the *indexing*, and the pass that follows
/// runs either way — visible as `Optimizing` until it is done.
#[test]
fn a_stopped_run_is_still_optimized() {
let root = tmp_dir("stop-optimize");
let body: Vec<u8> = "sphinx of black quartz judge my vow ".repeat(200).into_bytes();
for i in 0..4000 {
touch(&root.join(format!("d{}/f{:05}.txt", i % 40, i)), &body);
}
let db_dir = tmp_dir("stop-optimize-db");
let db = db_dir.join("index.sqlite");
let wal = db_dir.join("index.sqlite-wal");
let service = IndexingService::new();
service
.start_indexing(
vec![root.to_string_lossy().into_owned()],
db.to_string_lossy().into_owned(),
test_config(),
)
.unwrap();
// Let it get far enough in to have written something worth landing.
let deadline = Instant::now() + Duration::from_secs(120);
while Instant::now() < deadline {
if std::fs::metadata(&wal).map(|m| m.len()).unwrap_or(0) > 512 * 1024 {
break;
}
if let IndexingStatus::Error(e) = service.get_status() {
panic!("indexing failed: {}", e);
}
std::thread::sleep(Duration::from_millis(2));
}
service.request_stop();
let mut saw_optimizing = false;
let idle_by = Instant::now() + Duration::from_secs(120);
loop {
match service.get_status() {
IndexingStatus::Optimizing => saw_optimizing = true,
IndexingStatus::Idle => break,
IndexingStatus::Error(e) => panic!("indexing failed: {}", e),
_ => {}
}
assert!(Instant::now() < idle_by, "the stopped run never reached Idle");
std::thread::sleep(Duration::from_millis(1));
}
assert!(saw_optimizing, "a stopped run must still publish Optimizing");
assert_eq!(
std::fs::metadata(&wal).map(|m| m.len()).unwrap_or(0),
0,
"the optimize pass must land the stopped run's log"
);
drop(service);
std::fs::remove_dir_all(&root).ok();
std::fs::remove_dir_all(&db_dir).ok();
}

235
crates/quicksearch-gui/src/app.rs
View file

@ -4,11 +4,11 @@
use std::sync::mpsc;
use std::time::{Duration, Instant};
use quicksearch_core::cli::{index_counts, IndexCounts};
use quicksearch_core::cli::IndexCounts;
use quicksearch_core::config::{diff_actions, nested_roots, Config, SecurityConfig};
use quicksearch_core::coordinator::{IndexMode, IndexerState, WatcherStatus};
use quicksearch_core::db;
use quicksearch_core::indexing::{ConfigChange, IndexingStatus, RootPhase};
use quicksearch_core::indexing::{overall_progress, ConfigChange, IndexingStatus, RootPhase};
use quicksearch_core::search::SearchOptions;
use quicksearch_core::security::{derive_key, generate_salt, salt_to_hex, IndexKey};
use quicksearch_core::watcher::WatchError;
@ -22,6 +22,7 @@ use crate::logs_tab::LogsTab;
use crate::manage_tab::ManageTab;
use crate::options::{OptionsWindow, SecurityAction};
use crate::search_tab::SearchTab;
use crate::unlock::KeySource;
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Tab {
@ -50,6 +51,12 @@ pub struct QuickSearchApp {
/// Nested roots found in the loaded config (startup validation); shown
/// as a modal over the Manage tab until dismissed.
nested_prompt: Option<Vec<(String, String)>>,
/// How this session's key was obtained, for wording that refers to it.
key_source: KeySource,
/// Set when the index on disk was written by a different schema version
/// and the next run will replace it; see
/// [`QuickSearchApp::stale_index_prompt_ui`].
stale_index_prompt: bool,
/// Set when the watcher gave up on the directory budget and live
/// updates are off; see [`QuickSearchApp::check_watch_cap_warning`].
watch_cap_prompt: Option<WatchError>,
@ -100,6 +107,7 @@ impl QuickSearchApp {
cfg: Config,
config_error: Option<String>,
initial_query: Option<String>,
key_source: KeySource,
) -> Result<QuickSearchApp, String> {
// Compact styling: results density is the whole point.
ctx.style_mut(|style| {
@ -108,6 +116,13 @@ impl QuickSearchApp {
});
ctx.set_zoom_factor(clamp_scale(cfg.ui.scale));
// Probed *before* the backend exists. The coordinator can begin a full
// run — and with it the wipe — the moment it starts, and afterwards
// there is nothing left on disk to tell an upgrade apart from a fresh
// install.
let stale_index_prompt =
db::index_needs_rebuild(&cfg.resolved_database_path().to_string_lossy());
let backend = Backend::start(&cfg, ctx.clone())?;
let fuzzy = cfg.search.fuzzy_default;
// Startup validation: a hand-edited config can nest roots, which
@ -137,6 +152,8 @@ impl QuickSearchApp {
rebuild_prompt: None,
clear_prompt: false,
nested_prompt,
key_source,
stale_index_prompt,
watch_cap_prompt: None,
security_prompt: None,
config_error,
@ -259,6 +276,28 @@ impl QuickSearchApp {
Err(_) => break,
}
}
// Status-bar counts worker.
if let Some(rx) = &self.backend.counts_job {
match rx.try_recv() {
Ok(counts) => {
self.counts = Some((Instant::now(), counts));
self.backend.counts_job = None;
}
Err(mpsc::TryRecvError::Empty) => {}
// The worker gave up (missing or unreadable index). Keep the
// last known figures but restamp them, so the next attempt
// waits its turn instead of respawning a thread every frame.
Err(mpsc::TryRecvError::Disconnected) => {
let last = self.counts.map(|(_, c)| c).unwrap_or(IndexCounts {
files: 0,
content_done: 0,
content_pending: 0,
});
self.counts = Some((Instant::now(), last));
self.backend.counts_job = None;
}
}
}
// Duplicates worker.
if let Some(rx) = &self.backend.dup_job {
match rx.try_recv() {
@ -349,14 +388,10 @@ impl QuickSearchApp {
.map(|(at, _)| at.elapsed() > Duration::from_secs(5))
.unwrap_or(true);
if stale {
let db = self.cfg.resolved_database_path();
let counts =
index_counts(&db.to_string_lossy()).unwrap_or(IndexCounts {
files: 0,
content_done: 0,
content_pending: 0,
});
self.counts = Some((Instant::now(), counts));
// Kicked off, not awaited: the result lands through
// `drain_events` on a later frame.
let cfg = self.cfg.clone();
self.backend.start_index_counts(&cfg, ctx.clone());
}
let files = self.counts.map(|(_, c)| c.files).unwrap_or(0);
ui.label(
@ -377,25 +412,25 @@ impl QuickSearchApp {
IndexingStatus::Stopping => {
ui.label(egui::RichText::new("Stopping indexing…").small());
}
IndexingStatus::Optimizing => {
ui.label(egui::RichText::new("Optimizing index…").small());
}
IndexingStatus::Running { roots, .. } => {
let done = roots.iter().filter(|r| r.phase == RootPhase::Done).count();
let processed: usize = roots.iter().map(|r| r.walked + r.extracted).sum();
let totals_known = roots.iter().all(|r| r.walk_total.is_some());
let denominator: usize = roots
.iter()
.map(|r| r.walk_total.unwrap_or(0) + r.extract_total)
.sum();
let progress = overall_progress(roots);
let frac = progress.fraction();
let mut text = if totals_known && denominator > 0 {
let frac = (processed as f64 / denominator as f64).min(1.0);
format!(
let mut text = match (progress.total, frac) {
(Some(total), Some(frac)) => format!(
"Indexing {} / {} ({:.0}%)",
group_thousands(processed as u64),
group_thousands(denominator as u64),
group_thousands(progress.processed as u64),
group_thousands(total as u64),
frac * 100.0
)
} else {
format!("Indexing · {} files", group_thousands(processed as u64))
),
_ => format!(
"Indexing · {} files",
group_thousands(progress.processed as u64)
),
};
if roots.len() > 1 {
text.push_str(&format!(" · {}/{} roots done", done, roots.len()));
@ -409,11 +444,13 @@ impl QuickSearchApp {
text.push_str(&format!(" · {}/{} workers", active, total_workers));
}
ui.label(egui::RichText::new(text).small());
if totals_known && denominator > 0 {
let frac = (processed as f32 / denominator as f32).clamp(0.0, 1.0);
ui.add(egui::ProgressBar::new(frac).desired_width(120.0));
} else {
ui.add(egui::Spinner::new().size(12.0));
match frac {
Some(frac) => {
ui.add(egui::ProgressBar::new(frac as f32).desired_width(120.0));
}
None => {
ui.add(egui::Spinner::new().size(12.0));
}
}
}
}
@ -781,6 +818,26 @@ impl QuickSearchApp {
}
}
/// Tell the user their index is being replaced because it belongs to an
/// older version — rather than letting them discover it as a run of failed
/// searches.
///
/// This is the one modal that is not a question. The old index genuinely
/// cannot be read by this build, so there is no "keep it" branch to offer;
/// the button starts the rebuild rather than merely dismissing, which is
/// what makes the promise true in manual mode as well as automatic.
fn stale_index_prompt_ui(&mut self, ctx: &egui::Context) {
if !self.stale_index_prompt {
return;
}
if stale_index_window(ctx, self.key_source) {
self.stale_index_prompt = false;
self.backend.coordinator.rebuild_index();
self.counts = None;
self.dups.state = DupState::NotLoaded;
}
}
fn watch_cap_prompt_ui(&mut self, ctx: &egui::Context) {
let Some(reason) = &self.watch_cap_prompt else {
return;
@ -904,6 +961,60 @@ fn pin_live_fields(new: &mut Config, live: &Config) {
new.indexing.auto_index = live.indexing.auto_index;
}
/// The stale-index window's body. Returns whether the user asked for the
/// rebuild.
///
/// Split out from the method so it can be rendered, and its button clicked, in
/// a headless egui context — building a whole [`QuickSearchApp`] would start a
/// coordinator and a watcher.
fn stale_index_window(ctx: &egui::Context, key_source: KeySource) -> bool {
let mut rebuild = false;
egui::Window::new("Index reset for this version")
.collapsible(false)
.resizable(false)
.anchor(egui::Align2::CENTER_CENTER, [0.0, 0.0])
.show(ctx, |ui| {
ui.set_max_width(440.0);
ui.label(
"Your search index was created by an older version of QuickSearch, \
which this version cannot read. It is being reset and rebuilt from \
scratch. Don't worry, this is very Quick!",
);
// Only says "you just entered" when that actually happened: with
// the key remembered on this device there was no prompt to type
// into, and naming one the user never saw would just confuse.
let reassurance = match key_source {
KeySource::Unprotected => None,
KeySource::Prompt => Some(
"The rebuilt index is encrypted with the password you just \
entered, so it stays password protected.",
),
KeySource::Keychain => Some(
"The rebuilt index is encrypted with the key remembered on \
this device, so it stays password protected.",
),
};
if let Some(text) = reassurance {
ui.add_space(4.0);
ui.label(text);
}
ui.add_space(4.0);
ui.label(
egui::RichText::new(
"Your files are not touched. Searches return incomplete results \
until the rebuild finishes; progress is on the Manage Index tab.",
)
.small()
.weak(),
);
ui.add_space(4.0);
if ui.button("Rebuild now").clicked() {
rebuild = true;
}
});
rebuild
}
/// A stored/current config value for the rebuild prompt; list values are
/// already newline-joined and render as-is, empty means unset.
fn display_value(value: &str) -> String {
@ -1034,6 +1145,10 @@ impl eframe::App for QuickSearchApp {
self.security_prompt_ui(ctx);
self.clear_prompt_ui(ctx);
self.nested_prompt_ui(ctx);
// Ahead of the watch-cap warning: on a fresh upgrade both can be true,
// and "your index is being rebuilt" is the one that explains what the
// user is actually looking at.
self.stale_index_prompt_ui(ctx);
self.watch_cap_prompt_ui(ctx);
}
@ -1046,6 +1161,68 @@ impl eframe::App for QuickSearchApp {
mod tests {
use super::*;
fn frame(ctx: &egui::Context, source: KeySource, events: Vec<egui::Event>) -> bool {
let input = egui::RawInput {
screen_rect: Some(egui::Rect::from_min_size(
egui::Pos2::ZERO,
egui::vec2(1000.0, 700.0),
)),
events,
..Default::default()
};
let mut clicked = false;
let _ = ctx.run(input, |ctx| clicked = stale_index_window(ctx, source));
clicked
}
fn click_at(pos: egui::Pos2) -> Vec<egui::Event> {
[true, false]
.into_iter()
.map(|pressed| egui::Event::PointerButton {
pos,
button: egui::PointerButton::Primary,
pressed,
modifiers: egui::Modifiers::default(),
})
.collect()
}
/// The modal the user sees after unlocking onto an index from an older
/// version. It must render under every key source — each produces a
/// different height — and its one button must actually report the click,
/// since that click is what starts the rebuild. A dead button would leave
/// the promise ("it is being rebuilt") unkept in manual mode.
#[test]
fn the_stale_index_modal_renders_and_its_button_fires() {
for source in [
KeySource::Unprotected,
KeySource::Prompt,
KeySource::Keychain,
] {
let ctx = egui::Context::default();
assert!(
!frame(&ctx, source, Vec::new()),
"an untouched frame must not request a rebuild"
);
// Sweep the window for the button rather than hard-coding its
// position: it is left-aligned at the bottom of a centre-anchored
// window whose height depends on which sentence is shown, and
// pinning coordinates would make this a layout test.
let mut fired = None;
'sweep: for y in (230..480).step_by(3) {
for x in (250..760).step_by(6) {
let pos = egui::pos2(x as f32, y as f32);
if frame(&ctx, source, click_at(pos)) {
fired = Some(pos);
break 'sweep;
}
}
}
assert!(fired.is_some(), "no clickable Rebuild button for {source:?}");
}
}
#[test]
fn a_stale_draft_cannot_revert_the_indexing_mode_or_security() {
// The draft as it was when the editor last synced: automatic

29
crates/quicksearch-gui/src/backend.rs
View file

@ -7,6 +7,7 @@
use std::sync::{mpsc, Arc};
use quicksearch_core::cli::IndexCounts;
use quicksearch_core::config::Config;
use quicksearch_core::coordinator::IndexCoordinator;
use quicksearch_core::search::{DuplicateGroup, SearchService, SearchUpdate};
@ -17,6 +18,8 @@ pub struct Backend {
pub search: Option<SearchService>,
pub search_rx: mpsc::Receiver<SearchUpdate>,
pub dup_job: Option<mpsc::Receiver<Result<Vec<DuplicateGroup>, String>>>,
/// In-flight status-bar count; see [`Backend::start_index_counts`].
pub counts_job: Option<mpsc::Receiver<IndexCounts>>,
}
impl Backend {
@ -37,9 +40,35 @@ impl Backend {
search: Some(search),
search_rx,
dup_job: None,
counts_job: None,
})
}
/// Refresh the status bar's "N files indexed" on a worker thread.
///
/// Three `COUNT(*)` scans, and on a multi-million-row index the unfiltered
/// one alone reads the whole primary key. Running it inline in `update()`
/// froze a frame every refresh for a number that is purely decorative — so
/// it goes the same way the duplicates scan does. No-op while one is
/// already in flight.
pub fn start_index_counts(&mut self, config: &Config, ctx: egui::Context) {
if self.counts_job.is_some() {
return;
}
let (tx, rx) = mpsc::channel();
let db = config.resolved_database_path();
std::thread::spawn(move || {
// A missing or unreadable index is not worth reporting here: the
// status bar has nothing useful to say about it that the indexing
// state does not already say.
if let Ok(counts) = quicksearch_core::cli::index_counts(&db.to_string_lossy()) {
let _ = tx.send(counts);
}
ctx.request_repaint();
});
self.counts_job = Some(rx);
}
pub fn search(&self) -> &SearchService {
self.search.as_ref().expect("search service alive")
}

24
crates/quicksearch-gui/src/main.rs
View file

@ -77,8 +77,17 @@ fn main() {
// verified key means no prompt at all. Anything else starts locked —
// the unlock screen owns password entry, bad-salt reporting, and the
// forgot-password escape hatch. No index is touched until unlocked.
let start_unlocked =
!config.security.password_protected || unlock::try_keychain_unlock(&config);
//
// `None` means "start locked". The two unlocked cases stay distinct
// rather than collapsing to a bool, because anything the app later says
// *about* the key has to know whether the user typed one.
let key_source = if !config.security.password_protected {
Some(unlock::KeySource::Unprotected)
} else if unlock::try_keychain_unlock(&config) {
Some(unlock::KeySource::Keychain)
} else {
None
};
let native_options = eframe::NativeOptions {
viewport: egui::ViewportBuilder::default()
@ -94,11 +103,12 @@ fn main() {
"QuickSearch",
native_options,
Box::new(move |cc| {
let gate = if start_unlocked {
unlock::Gate::running(&cc.egui_ctx, config, config_error, initial_query)
.map_err(Box::<dyn std::error::Error + Send + Sync>::from)?
} else {
unlock::Gate::locked(config, config_error, initial_query)
let gate = match key_source {
Some(source) => {
unlock::Gate::running(&cc.egui_ctx, config, config_error, initial_query, source)
.map_err(Box::<dyn std::error::Error + Send + Sync>::from)?
}
None => unlock::Gate::locked(config, config_error, initial_query),
};
Ok(Box::new(gate) as Box<dyn eframe::App>)
}),

757
crates/quicksearch-gui/src/manage_tab.rs
View file

@ -1,11 +1,14 @@
//! The Manage Index tab: detailed status, mode controls, indexed roots,
//! and the content/ignore filter editors.
use std::path::Path;
use std::time::{Duration, Instant};
use quicksearch_core::config::Config;
use quicksearch_core::coordinator::{IndexMode, IndexerState, WatcherStatus};
use quicksearch_core::indexing::{IndexingStatus, RootPhase, RootProgress};
use crate::format::{fmt_interval, fmt_rate, group_thousands, middle_truncate};
use crate::format::{fmt_interval, fmt_rate, group_thousands, human_size, middle_truncate};
use crate::tracker::SpeedTracker;
/// What the tab asks the app to do after this frame.
@ -35,6 +38,8 @@ pub struct ManageTab {
baseline: Option<Config>,
/// Draft of the roots/filters edited in-place.
draft: Option<Config>,
/// Cached on-disk footprint of the index, restatted on a timer.
db_size: DbSizeProbe,
}
impl ManageTab {
@ -47,6 +52,7 @@ impl ManageTab {
root_error: None,
baseline: None,
draft: None,
db_size: DbSizeProbe::default(),
}
}
@ -59,7 +65,9 @@ impl ManageTab {
let total: usize = roots.iter().map(|r| r.walked + r.extracted).sum();
self.speed.record(total);
}
IndexingStatus::Idle | IndexingStatus::Error(_) => self.speed.reset(),
IndexingStatus::Idle | IndexingStatus::Error(_) | IndexingStatus::Optimizing => {
self.speed.reset()
}
_ => {}
}
}
@ -121,7 +129,14 @@ impl ManageTab {
.auto_shrink([false; 2])
.show(ui, |ui| {
// --- Status ---------------------------------------------------
ui.heading(egui::RichText::new("Status").strong());
ui.horizontal(|ui| {
ui.heading(egui::RichText::new("Status").strong());
// Right-aligned, clear of the progress text below it —
// that text changes width every frame during a run.
ui.with_layout(egui::Layout::right_to_left(egui::Align::Center), |ui| {
db_size_label(ui, self.db_size.size(config, Instant::now()));
});
});
status_panel(ui, state, &self.speed);
watch_panel(ui, state, config);
ui.add_space(8.0);
@ -232,6 +247,8 @@ impl ManageTab {
storage, 16 on network mounts. Takes effect on \
the next indexing run.",
);
#[cfg(test)]
tests::record_widget("workers", &response);
if response.changed() {
if workers == 0 {
draft.indexing.root_workers.remove(root);
@ -280,9 +297,11 @@ impl ManageTab {
}
}
});
if let Some(err) = &self.root_error {
ui.colored_label(ui.visuals().error_fg_color, err);
}
crate::ui_util::stable_section(ui, |ui| {
if let Some(err) = &self.root_error {
ui.colored_label(ui.visuals().error_fg_color, err);
}
});
ui.separator();
// --- Filters ---------------------------------------------------
@ -297,27 +316,32 @@ impl ManageTab {
);
cols[1].label("Ignore patterns (excluded entirely):");
let mut remove_pat: Option<usize> = None;
for (i, pat) in draft.indexing.ignore_patterns.iter().enumerate() {
cols[1].horizontal(|ui| {
ui.with_layout(
egui::Layout::right_to_left(egui::Align::Center),
|ui| {
if ui.small_button("Remove").clicked() {
remove_pat = Some(i);
}
ui.with_layout(
egui::Layout::left_to_right(egui::Align::Center),
|ui| {
ui.monospace(pat);
},
);
},
);
});
}
if draft.indexing.ignore_patterns.is_empty() {
cols[1].label(egui::RichText::new("No ignore patterns.").small().weak());
}
// The list grows and shrinks — including from outside
// this tab — so it is kept off the id of the editor
// below it (see `ui_util::stable_section`).
crate::ui_util::stable_section(&mut cols[1], |ui| {
for (i, pat) in draft.indexing.ignore_patterns.iter().enumerate() {
ui.horizontal(|ui| {
ui.with_layout(
egui::Layout::right_to_left(egui::Align::Center),
|ui| {
if ui.small_button("Remove").clicked() {
remove_pat = Some(i);
}
ui.with_layout(
egui::Layout::left_to_right(egui::Align::Center),
|ui| {
ui.monospace(pat);
},
);
},
);
});
}
if draft.indexing.ignore_patterns.is_empty() {
ui.label(egui::RichText::new("No ignore patterns.").small().weak());
}
});
if let Some(i) = remove_pat {
draft.indexing.ignore_patterns.remove(i);
}
@ -363,13 +387,13 @@ impl ManageTab {
);
ui.add_space(8.0);
if ui
.add(crate::ui_util::bordered_button(
"Apply & Save",
crate::ui_util::BLUE,
))
.clicked()
{
let apply = ui.add(crate::ui_util::bordered_button(
"Apply & Save",
crate::ui_util::BLUE,
));
#[cfg(test)]
tests::record_widget("apply", &apply);
if apply.clicked() {
let mut new_config = draft.clone();
new_config.indexing.content_extensions = parse_lines(&self.ext_filter_text);
let roots = new_config.paths.indexing_paths.clone();
@ -383,10 +407,111 @@ impl ManageTab {
});
crate::ui_util::more_below_hint(ui, &scroll);
// Nothing else asks for repaints while the app sits idle, so without
// this the size would freeze at whatever it read when the pointer
// last moved. One frame per interval, and only while this tab is the
// one on screen.
ui.ctx().request_repaint_after(DB_SIZE_REFRESH);
actions
}
}
/// How often the index files are re-statted. The number moves slowly even
/// during a run, so this is deliberately lazy: it costs three stats and one
/// repaint, and asking any more often would buy nothing.
const DB_SIZE_REFRESH: Duration = Duration::from_secs(10);
/// Total on-disk footprint of the index: the database plus its `-wal` and
/// `-shm` sidecars. The `-wal` file is why the database alone will not do —
/// mid-run it can hold hundreds of megabytes the database does not show yet.
///
/// A file that is not there counts as zero: there is no database at all
/// before the first run, and the sidecars exist only while a connection is
/// open.
fn measure_db_size(db: &Path) -> u64 {
// Only regular files: a misconfigured path pointing at a directory
// would otherwise report that directory's own inode size as an index.
let len = |path: &Path| {
std::fs::metadata(path)
.map(|m| if m.is_file() { m.len() } else { 0 })
.unwrap_or(0)
};
let name = db.file_name().and_then(|s| s.to_str()).unwrap_or("");
// No `-journal`: the index runs in WAL mode, so a rollback journal is
// not part of a live database.
len(db)
+ ["-wal", "-shm"]
.iter()
.map(|suffix| len(&db.with_file_name(format!("{}{}", name, suffix))))
.sum::<u64>()
}
/// Caches the last measurement so the tab can ask for it every frame.
#[derive(Default)]
struct DbSizeProbe {
/// Configured (unresolved) path the cached size belongs to.
path: String,
bytes: u64,
measured_at: Option<Instant>,
}
impl DbSizeProbe {
/// The cached size, restatted when it has gone stale or the configured
/// database path changed under it. `now` is a parameter so the refresh
/// cadence can be tested without sleeping.
fn size(&mut self, config: &Config, now: Instant) -> u64 {
let expired = self
.measured_at
.is_none_or(|at| now.duration_since(at) >= DB_SIZE_REFRESH);
if expired || self.path != config.paths.database_path {
// Resolving the path is only worth doing on a real refresh.
self.bytes = measure_db_size(&config.resolved_database_path());
self.path = config.paths.database_path.clone();
self.measured_at = Some(now);
}
self.bytes
}
}
/// The index's footprint, with the levers for shrinking it on hover: a user
/// who finds the number too large looks here first, and every lever named
/// is either on this tab or in Options.
fn db_size_label(ui: &mut egui::Ui, bytes: u64) {
let response = ui.label(format!("Index size: {}", human_size(bytes)));
#[cfg(test)]
tests::record_widget("db-size", &response);
response.on_hover_ui(db_size_tooltip);
}
fn db_size_tooltip(ui: &mut egui::Ui) {
ui.set_max_width(440.0);
ui.strong("To reduce the index size");
for lever in [
"Add ignore filters for files and folders you never search — the ignore \
pattern list further down this tab.",
"Remove indexed folders you do not need, in Indexed folders above.",
"Narrow the full-text extension whitelist, so text is only extracted \
from the file types you actually search.",
"Turn off \"Store text for snippets\" in Options: full-text search keeps \
working, but without previews, occurrence ranking or fuzzy matching \
inside file contents.",
"Lower \"Max text file size\" and \"Max stored text\" in Options.",
] {
ui.label(format!("• {}", lever));
}
ui.add_space(6.0);
ui.label(
egui::RichText::new(
"The file does not shrink on its own: freed space is reused by the \
index rather than returned to the disk. To hand it back after \
narrowing the filters, use Clear index… and reindex.",
)
.small()
.weak(),
);
}
/// Append a root to the draft unless it would duplicate or nest with an
/// existing one; the rejection reason lands in `error`.
fn try_add_root(draft: &mut Config, candidate: String, error: &mut Option<String>) -> bool {
@ -419,7 +544,14 @@ fn parse_lines(text: &str) -> Vec<String> {
/// Live-update health. Permanent counterpart to the one-time modal: the
/// modal is dismissed and remembered per root, but "live updates are off"
/// stays true and must remain discoverable.
///
/// Wrapped for the same reason as [`status_panel`]: this panel is empty in
/// manual mode and a line long otherwise.
fn watch_panel(ui: &mut egui::Ui, state: &IndexerState, config: &Config) {
crate::ui_util::stable_section(ui, |ui| watch_contents(ui, state, config));
}
fn watch_contents(ui: &mut egui::Ui, state: &IndexerState, config: &Config) {
match &state.watcher {
// Manual mode already says "stopped" in the controls row; repeating
// it here would be noise.
@ -454,7 +586,15 @@ fn watch_panel(ui: &mut egui::Ui, state: &IndexerState, config: &Config) {
}
}
/// Live progress. Its widget count tracks the run (roots come and go, the
/// rate line and the current file appear and vanish), so it renders inside
/// a [`crate::ui_util::stable_section`]: without one, every id below it —
/// including the per-root worker fields — would move mid-run.
fn status_panel(ui: &mut egui::Ui, state: &IndexerState, speed: &SpeedTracker) {
crate::ui_util::stable_section(ui, |ui| status_contents(ui, state, speed));
}
fn status_contents(ui: &mut egui::Ui, state: &IndexerState, speed: &SpeedTracker) {
match &state.activity {
IndexingStatus::Idle => {
// Relative wording makes even a milliseconds-fast run visibly
@ -471,6 +611,11 @@ fn status_panel(ui: &mut egui::Ui, state: &IndexerState, speed: &SpeedTracker) {
IndexingStatus::Stopping => {
ui.label("Stopping…");
}
IndexingStatus::Optimizing => {
// No per-file progress exists to show: this is one bulk rewrite
// of the whole file, and on a large index it runs for minutes.
ui.label("Optimizing index; reclaiming unused space…");
}
IndexingStatus::Running { roots, .. } => {
for root in roots {
root_row(ui, root);
@ -500,7 +645,7 @@ fn root_row(ui: &mut egui::Ui, r: &RootProgress) {
ui.label("indexing");
divider(ui);
let workers = format!("{}/{} workers", r.active_workers, r.total_workers);
match r.walk_total {
match r.walk_denominator() {
Some(total) if total > 0 => {
let frac = (r.walked as f32 / total as f32).clamp(0.0, 1.0);
ui.label(format!(
@ -535,10 +680,12 @@ fn root_row(ui: &mut egui::Ui, r: &RootProgress) {
1.0
};
ui.label(format!(
"{} / {} ({:.0}%)",
"{} / {} ({:.0}%) · {}/{} workers",
group_thousands(r.extracted as u64),
group_thousands(r.extract_total as u64),
frac * 100.0
frac * 100.0,
r.active_workers,
r.total_workers
));
ui.add(egui::ProgressBar::new(frac).desired_width(160.0));
}
@ -567,6 +714,542 @@ fn root_row(ui: &mut egui::Ui, r: &RootProgress) {
mod tests {
use super::*;
use std::cell::RefCell;
// Driving the real tab through a headless egui context is the only way
// to test widget identity, and identity is exactly what these tests are
// about — so the widgets under test report themselves here.
thread_local! {
static WIDGETS: RefCell<Vec<(&'static str, egui::Id, egui::Rect)>> =
const { RefCell::new(Vec::new()) };
}
pub(super) fn record_widget(tag: &'static str, response: &egui::Response) {
WIDGETS.with(|w| w.borrow_mut().push((tag, response.id, response.rect)));
}
fn widget(tag: &str) -> (egui::Id, egui::Rect) {
WIDGETS.with(|w| {
w.borrow()
.iter()
.find(|(t, _, _)| *t == tag)
.map(|(_, id, rect)| (*id, *rect))
.unwrap_or_else(|| panic!("{} widget was not drawn", tag))
})
}
fn idle_state() -> IndexerState {
IndexerState {
mode: IndexMode::Auto,
activity: IndexingStatus::Idle,
last_full_index: Some(0),
queued_events: 0,
watcher: WatcherStatus::Active { dirs: 10 },
}
}
/// A run in progress. `current_file` and the number of roots are the
/// parts that come and go from frame to frame in a real run.
fn running_state(roots: &[&str], current_file: Option<&str>) -> IndexerState {
state_with(
roots
.iter()
.map(|root| RootProgress {
root: (*root).to_string(),
phase: RootPhase::Walking,
walked: 100,
walk_total: Some(1000),
extracted: 0,
extract_total: 0,
current_file: current_file.map(str::to_string),
active_workers: 4,
total_workers: 4,
})
.collect(),
)
}
/// A run whose roots are described one by one, for the rows whose
/// contents — not just their widget ids — are under test.
fn state_with(roots: Vec<RootProgress>) -> IndexerState {
IndexerState {
mode: IndexMode::Auto,
activity: IndexingStatus::Running {
start_time: std::time::Instant::now(),
roots,
},
last_full_index: Some(0),
queued_events: 0,
watcher: WatcherStatus::Active { dirs: 10 },
}
}
fn raw_input(events: Vec<egui::Event>) -> egui::RawInput {
egui::RawInput {
screen_rect: Some(egui::Rect::from_min_size(
egui::Pos2::ZERO,
egui::vec2(1000.0, 900.0),
)),
events,
..Default::default()
}
}
/// One frame of the real tab, with `events` delivered to it.
fn frame(
ctx: &egui::Context,
tab: &mut ManageTab,
cfg: &Config,
state: &IndexerState,
events: Vec<egui::Event>,
) -> ManageActions {
WIDGETS.with(|w| w.borrow_mut().clear());
let mut actions = ManageActions::default();
let _ = ctx.run(raw_input(events), |ctx| {
egui::CentralPanel::default().show(ctx, |ui| {
actions = tab.ui(ui, state, cfg);
});
});
actions
}
/// Every string the tab actually painted this frame. Labels carry no
/// widget id worth recording, so the rendered text is read back off the
/// shapes — the only place a number the user sees can be checked.
fn frame_text(ctx: &egui::Context, tab: &mut ManageTab, state: &IndexerState) -> Vec<String> {
frame_text_with(ctx, tab, &cfg_with_root(), state)
}
fn frame_text_with(
ctx: &egui::Context,
tab: &mut ManageTab,
cfg: &Config,
state: &IndexerState,
) -> Vec<String> {
WIDGETS.with(|w| w.borrow_mut().clear());
let out = ctx.run(raw_input(vec![]), |ctx| {
egui::CentralPanel::default().show(ctx, |ui| {
tab.ui(ui, state, cfg);
});
});
let mut text = Vec::new();
for clipped in &out.shapes {
collect_text(&clipped.shape, &mut text);
}
text
}
fn collect_text(shape: &egui::epaint::Shape, out: &mut Vec<String>) {
match shape {
egui::epaint::Shape::Text(t) => out.push(t.galley.text().to_string()),
egui::epaint::Shape::Vec(shapes) => {
for s in shapes {
collect_text(s, out);
}
}
_ => {}
}
}
fn pointer(pos: egui::Pos2, pressed: bool) -> egui::Event {
egui::Event::PointerButton {
pos,
button: egui::PointerButton::Primary,
pressed,
modifiers: egui::Modifiers::NONE,
}
}
fn click_at(pos: egui::Pos2) -> Vec<egui::Event> {
vec![
egui::Event::PointerMoved(pos),
pointer(pos, true),
pointer(pos, false),
]
}
fn cfg_with_root() -> Config {
let mut cfg = Config::default();
cfg.paths.indexing_paths = vec!["/data".into()];
cfg
}
fn staged_workers(tab: &ManageTab) -> Option<usize> {
tab.draft
.as_ref()
.unwrap()
.indexing
.root_workers
.get("/data")
.copied()
}
/// The per-root worker field must keep the same widget id however the
/// status above it changes: egui hangs focus and in-progress text off
/// that id, so a field that is renamed mid-run silently drops the edit.
#[test]
fn the_worker_field_keeps_its_identity_as_the_status_changes() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let cfg = cfg_with_root();
frame(&ctx, &mut tab, &cfg, &running_state(&["/data"], None), vec![]);
let (baseline, _) = widget("workers");
for state in [
running_state(&["/data"], Some("/data/file")),
running_state(&["/data", "/other"], None),
idle_state(),
IndexerState {
watcher: WatcherStatus::Off,
..idle_state()
},
IndexerState {
activity: IndexingStatus::Error("boom".into()),
..idle_state()
},
] {
frame(&ctx, &mut tab, &cfg, &state, vec![]);
assert_eq!(
widget("workers").0,
baseline,
"status change moved the worker field"
);
}
}
/// Click the field, type a count, click Apply & Save — while a run is
/// reporting progress the whole time.
#[test]
fn a_typed_worker_count_reaches_the_applied_config() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let cfg = cfg_with_root();
frame(&ctx, &mut tab, &cfg, &running_state(&["/data"], None), vec![]);
let field = widget("workers").1.center();
frame(
&ctx,
&mut tab,
&cfg,
&running_state(&["/data"], None),
click_at(field),
);
// The run starts reporting a file: one more label above the field.
let busy = running_state(&["/data"], Some("/data/file"));
frame(&ctx, &mut tab, &cfg, &busy, vec![]);
frame(
&ctx,
&mut tab,
&cfg,
&busy,
vec![egui::Event::Text("8".into())],
);
assert_eq!(staged_workers(&tab), Some(8), "typed count was not staged");
let apply = widget("apply").1.center();
let mut actions = frame(&ctx, &mut tab, &cfg, &busy, click_at(apply));
if actions.apply_config.is_none() {
// egui fires a click on release; give it the follow-up frame.
actions = frame(&ctx, &mut tab, &cfg, &busy, vec![]);
}
let applied = actions.apply_config.expect("Apply & Save produced a config");
assert_eq!(applied.indexing.root_workers.get("/data"), Some(&8));
}
/// The other way to set the field: drag it.
#[test]
fn a_dragged_worker_count_is_staged() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let cfg = cfg_with_root();
let busy = running_state(&["/data"], None);
frame(&ctx, &mut tab, &cfg, &busy, vec![]);
let field = widget("workers").1.center();
frame(
&ctx,
&mut tab,
&cfg,
&busy,
vec![egui::Event::PointerMoved(field), pointer(field, true)],
);
// Drag right across frames, with the status changing underneath.
let mut pos = field;
for state in [
running_state(&["/data"], Some("/data/a")),
running_state(&["/data"], None),
running_state(&["/data"], Some("/data/b")),
] {
pos.x += 4.0;
frame(
&ctx,
&mut tab,
&cfg,
&state,
vec![egui::Event::PointerMoved(pos)],
);
}
frame(&ctx, &mut tab, &cfg, &busy, vec![pointer(pos, false)]);
assert!(
staged_workers(&tab).is_some_and(|w| w > 0),
"dragging staged nothing: {:?}",
staged_workers(&tab)
);
}
fn root_progress(phase: RootPhase, walked: usize, walk_total: Option<usize>) -> RootProgress {
RootProgress {
root: "/data".to_string(),
phase,
walked,
walk_total,
extracted: 0,
extract_total: 0,
current_file: None,
active_workers: 4,
total_workers: 4,
}
}
/// The `find` count is an estimate of *tree entries*, so it runs far
/// ahead of the files a walk actually emits. Once the walk ends the exact
/// number is in hand, and the row must show that instead — the estimate
/// is what kept the bar short of full for the rest of the run.
#[test]
fn a_finished_root_reports_its_exact_count_not_the_estimate() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let mut done = root_progress(RootPhase::Done, 261_088, Some(6_677_062));
done.extracted = 238_929;
done.active_workers = 0;
done.total_workers = 0;
let text = frame_text(&ctx, &mut tab, &state_with(vec![done])).join(" | ");
assert!(
text.contains("indexed 261,088, extracted 238,929"),
"finished row: {}",
text
);
assert!(
!text.contains("6,677,062"),
"the stale estimate is still on screen: {}",
text
);
}
/// While the walk runs the estimate is all there is, so it is shown —
/// but never below what has already been walked, or the row would sit at
/// 100% with the walk still going and read as a hang.
#[test]
fn a_walking_root_shows_the_estimate_raised_to_what_it_has_walked() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let honest = frame_text(
&ctx,
&mut tab,
&state_with(vec![root_progress(RootPhase::Walking, 100, Some(1000))]),
)
.join(" | ");
assert!(honest.contains("100 / 1,000 (10%)"), "{}", honest);
let overtaken = frame_text(
&ctx,
&mut tab,
&state_with(vec![root_progress(RootPhase::Walking, 1500, Some(1000))]),
)
.join(" | ");
assert!(
overtaken.contains("1,500 / 1,500 (100%)"),
"an overtaken estimate must be raised, not shown: {}",
overtaken
);
}
/// No count has landed yet: an indeterminate row, not a fabricated one.
#[test]
fn a_walking_root_without_a_count_shows_no_denominator() {
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let text = frame_text(
&ctx,
&mut tab,
&state_with(vec![root_progress(RootPhase::Walking, 100, None)]),
)
.join(" | ");
assert!(text.contains("100 files"), "{}", text);
assert!(!text.contains(" / "), "invented a denominator: {}", text);
}
/// A scratch directory of its own for each size test, so two of them
/// running in parallel cannot see each other's files.
fn scratch_dir(tag: &str) -> std::path::PathBuf {
let dir = std::env::temp_dir().join(format!("qs-dbsize-{}-{}", std::process::id(), tag));
let _ = std::fs::remove_dir_all(&dir);
std::fs::create_dir_all(&dir).expect("scratch dir");
dir
}
fn write_bytes(path: &Path, len: usize) {
std::fs::write(path, vec![b'x'; len]).expect("write");
}
/// A config whose database is `path`, for the probe and the rendered row.
fn cfg_with_db(path: &Path) -> Config {
let mut cfg = cfg_with_root();
cfg.paths.database_path = path.to_string_lossy().into_owned();
cfg
}
/// Each of the three files SQLite keeps for one database is added, and
/// nothing else that happens to sit beside them is.
#[test]
fn db_size_counts_the_database_and_both_sidecars() {
let dir = scratch_dir("parts");
let db = dir.join("index.sqlite");
write_bytes(&db, 4096);
assert_eq!(measure_db_size(&db), 4096, "the database itself");
write_bytes(&dir.join("index.sqlite-wal"), 1024);
assert_eq!(measure_db_size(&db), 5120, "-wal was not added");
write_bytes(&dir.join("index.sqlite-shm"), 32);
assert_eq!(measure_db_size(&db), 5152, "-shm was not added");
// Decoys: a rollback journal (never present in WAL mode) and an
// unrelated neighbour.
write_bytes(&dir.join("index.sqlite-journal"), 999);
write_bytes(&dir.join("index.sqlite.bak"), 777);
assert_eq!(measure_db_size(&db), 5152, "a decoy was counted");
let _ = std::fs::remove_dir_all(&dir);
}
/// Before the first indexing run there is no database at all, and a
/// closed database has no sidecars — neither is an error.
#[test]
fn a_missing_database_measures_zero() {
let dir = scratch_dir("missing");
let db = dir.join("index.sqlite");
assert_eq!(measure_db_size(&db), 0);
assert_eq!(measure_db_size(&dir), 0, "a directory");
write_bytes(&db, 100);
assert_eq!(measure_db_size(&db), 100, "sidecars are optional");
let _ = std::fs::remove_dir_all(&dir);
}
/// The probe answers from cache until the interval is up: `ui()` asks it
/// every frame, and three stats per frame is churn nobody needs.
#[test]
fn the_probe_caches_until_the_refresh_interval_is_up() {
let dir = scratch_dir("cache");
let db = dir.join("index.sqlite");
write_bytes(&db, 100);
let cfg = cfg_with_db(&db);
let mut probe = DbSizeProbe::default();
let t0 = Instant::now();
assert_eq!(probe.size(&cfg, t0), 100);
write_bytes(&db, 5000);
assert_eq!(
probe.size(&cfg, t0 + DB_SIZE_REFRESH / 2),
100,
"restatted before the interval was up"
);
assert_eq!(
probe.size(&cfg, t0 + DB_SIZE_REFRESH),
5000,
"the refresh never happened"
);
let _ = std::fs::remove_dir_all(&dir);
}
/// A database path edited in Options must not keep showing the old
/// database's size for the rest of the interval.
#[test]
fn the_probe_follows_a_changed_database_path() {
let dir = scratch_dir("moved");
let first = dir.join("index.sqlite");
let second = dir.join("other.sqlite");
write_bytes(&first, 100);
write_bytes(&second, 7000);
let mut probe = DbSizeProbe::default();
let t0 = Instant::now();
assert_eq!(probe.size(&cfg_with_db(&first), t0), 100);
assert_eq!(
probe.size(&cfg_with_db(&second), t0),
7000,
"a new path must restat at once, not wait out the interval"
);
let _ = std::fs::remove_dir_all(&dir);
}
/// The number reaches the screen, and it is the total of all three files
/// rather than the database on its own.
#[test]
fn the_status_row_shows_the_total_index_size() {
let dir = scratch_dir("render");
let db = dir.join("index.sqlite");
write_bytes(&db, 3_000_000);
write_bytes(&dir.join("index.sqlite-wal"), 200_000);
write_bytes(&dir.join("index.sqlite-shm"), 32_768);
let ctx = egui::Context::default();
let mut tab = ManageTab::new();
let text = frame_text_with(&ctx, &mut tab, &cfg_with_db(&db), &idle_state()).join(" | ");
assert!(
text.contains("Index size: 3.2 MB"),
"the size is not on screen: {}",
text
);
let _ = std::fs::remove_dir_all(&dir);
}
/// Hovering is the whole point of the readout: the number tells the user
/// the index is big, the tooltip tells them what to do about it. Each
/// lever named here has to keep matching a control that exists.
#[test]
fn hovering_the_size_explains_how_to_shrink_the_index() {
let dir = scratch_dir("hover");
let db = dir.join("index.sqlite");
write_bytes(&db, 2048);
let ctx = egui::Context::default();
// egui holds tooltips back for a third of a second, and frames here
// are only 1/60 s of simulated time apart.
ctx.style_mut(|s| s.interaction.tooltip_delay = 0.0);
let mut tab = ManageTab::new();
let cfg = cfg_with_db(&db);
frame_text_with(&ctx, &mut tab, &cfg, &idle_state());
let at = widget("db-size").1.center();
frame(
&ctx,
&mut tab,
&cfg,
&idle_state(),
vec![egui::Event::PointerMoved(at)],
);
// The tooltip is painted in the frame after the pointer lands.
let text = frame_text_with(&ctx, &mut tab, &cfg, &idle_state()).join(" | ");
assert!(
text.contains("To reduce the index size"),
"no tooltip: {}",
text
);
for lever in [
"ignore filters",
"Indexed folders",
"whitelist",
"Store text for snippets",
"Options",
] {
assert!(text.contains(lever), "tooltip never mentions {}", lever);
}
let _ = std::fs::remove_dir_all(&dir);
}
fn synced_tab(config: &Config) -> ManageTab {
let mut tab = ManageTab::new();
tab.sync_editors(config);

23
crates/quicksearch-gui/src/options.rs
View file

@ -308,6 +308,18 @@ pub fn config_editor_ui(ui: &mut egui::Ui, config: &mut Config, section: Section
ui.add(egui::DragValue::new(&mut config.processing.batch_size).range(10..=100_000));
ui.end_row();
ui.label("Max WAL size (bytes)");
ui.add(
egui::DragValue::new(&mut config.processing.maximum_wal_size)
.range(0u64..=8_589_934_592u64),
)
.on_hover_text(
"How large index.sqlite-wal may grow during a run before the \
indexer forces a checkpoint. 0 disables forced checkpoints; \
anything below 16 MiB is raised to it.",
);
ui.end_row();
ui.label("Store text for snippets");
ui.checkbox(&mut config.processing.store_text_for_snippets, "")
.on_hover_text(
@ -365,9 +377,14 @@ pub fn config_editor_ui(ui: &mut egui::Ui, config: &mut Config, section: Section
);
ui.end_row();
});
if let Some(warning) = config.search.fuzzy_edits_warning() {
ui.colored_label(ui.visuals().warn_fg_color, warning);
}
// A warning that comes and goes as the value is edited would
// otherwise move every widget below it in the window; see
// `ui_util::stable_section`.
crate::ui_util::stable_section(ui, |ui| {
if let Some(warning) = config.search.fuzzy_edits_warning() {
ui.colored_label(ui.visuals().warn_fg_color, warning);
}
});
}
}
}

328
crates/quicksearch-gui/src/search_tab.rs
View file

@ -41,6 +41,35 @@ pub struct SearchActions {
pub save_fuzzy_default: Option<bool>,
}
/// Add `incoming` to `set`, keeping at most `limit` of them — the best by
/// **rank**, whatever column the table is currently sorted by.
///
/// Retention and display are separate questions. What to keep is about
/// relevance; what order to show it in is the user's choice. That distinction
/// only started to matter once the cascade began streaming: hits now arrive in
/// table order, so a cap that simply stopped accepting at `limit` would fill
/// the table with whatever the scan happened to reach first and never show the
/// good ones. Dropping the worst-ranked instead means a rank-1 hit found late
/// in a scan still displaces a rank-10 one found early.
fn admit(
set: &mut Vec<SearchHit>,
incoming: Vec<SearchHit>,
limit: usize,
limited: &mut bool,
) {
set.extend(incoming);
if set.len() > limit {
set.sort_by(|a, b| {
a.rank
.partial_cmp(&b.rank)
.unwrap_or(std::cmp::Ordering::Equal)
.then_with(|| a.path.cmp(&b.path))
});
set.truncate(limit);
*limited = true;
}
}
pub struct SearchTab {
pub query: String,
pub fuzzy: bool,
@ -142,31 +171,31 @@ impl SearchTab {
SearchUpdate::Hits { hits, .. } => {
if self.swap_pending {
// Old results are still fading out; hold the new ones.
for hit in hits {
if self.staging.len() >= display_limit {
self.limited = true;
break;
}
self.staging_has_snippets |= hit.snippet.is_some();
self.staging.push(hit);
}
// Nothing of this generation is displayed yet, so there is
// no selection to keep.
self.staging_has_snippets |= hits.iter().any(|h| h.snippet.is_some());
admit(&mut self.staging, hits, display_limit, &mut self.limited);
} else {
// Post-swap stream: later cascade passes append live.
for hit in hits {
if self.results.len() >= display_limit {
self.limited = true;
break;
}
self.has_snippets |= hit.snippet.is_some();
self.results.push(hit);
}
// Arrival order *is* rank order, so the default sort
// needs no work; anything else re-sorts on the set.
if self.sort != (SortKey::Rank, true) {
self.sort_dirty = true;
} else {
self.order = (0..self.results.len() as u32).collect();
}
self.has_snippets |= hits.iter().any(|h| h.snippet.is_some());
// A row can be selected while batches are still arriving,
// and admitting them may reorder or drop rows out from
// under its index, so carry the selection by file id.
let selected_id = self
.selected
.and_then(|i| self.results.get(i as usize))
.map(|h| h.file_id);
admit(&mut self.results, hits, display_limit, &mut self.limited);
self.selected = selected_id.and_then(|id| {
self.results
.iter()
.position(|h| h.file_id == id)
.map(|i| i as u32)
});
// Arrival order is *scan* order — the cascade streams each
// pass as it runs, so a better-ranked hit can turn up after
// a worse one. Re-establish the table's own order on every
// batch, under whichever column is keyed.
self.sort_dirty = true;
}
}
SearchUpdate::Completed { limited, .. } => {
@ -213,11 +242,14 @@ impl SearchTab {
SortKey::Size => a.size.cmp(&b.size),
SortKey::Modified => a.mtime.cmp(&b.mtime),
};
if ascending {
ord
} else {
ord.reverse()
}
let ord = if ascending { ord } else { ord.reverse() };
// Break ties on the path, which is unique, so the order is total.
// Without it a stable sort falls back to insertion order — and
// that is now the order batches happened to stream in, so equal
// keys would shuffle under the pointer on every arrival. The
// tiebreak stays ascending regardless of the key's direction; it
// is there for stability, not as a second sort the user asked for.
ord.then_with(|| a.path.cmp(&b.path))
});
// Selection follows the file, not the visual slot.
self.selected = selected_id.and_then(|id| {
@ -395,11 +427,9 @@ impl SearchTab {
self.has_snippets = self.staging_has_snippets;
self.selected = None;
self.swap_pending = false;
if self.sort == (SortKey::Rank, true) {
self.order = (0..self.results.len() as u32).collect();
} else {
self.sort_dirty = true;
}
// Staged hits arrived in scan order too, so the table has to be
// ordered here as well — including under the default key.
self.sort_dirty = true;
}
if self.sort_dirty {
@ -425,6 +455,10 @@ impl SearchTab {
let text_height = egui::TextStyle::Body.resolve(ui.style()).size + 4.0;
let mut open_ignore_dialog: Option<usize> = None;
let mut hovered_now: Option<usize> = None;
// Moved out of `self` rather than cloned. The row closure needs `&mut
// self` for selection and hover, so a field read would conflict — but
// a plain local does not, and this runs every frame.
let order = std::mem::take(&mut self.order);
let table_scroll = ui
.push_id("results", |ui| {
@ -465,7 +499,6 @@ impl SearchTab {
header.col(|ui| self.sort_header(ui, SortKey::Rank, "Rank"));
})
.body(|body| {
let order = self.order.clone();
body.rows(text_height, order.len(), |mut row| {
let display_ix = row.index();
let result_ix = order[display_ix] as usize;
@ -487,13 +520,17 @@ impl SearchTab {
.push(ui.label(egui::RichText::new(&hit.path).weak()));
});
if self.has_snippets {
let snippet = hit.snippet.clone();
// Borrowed, not cloned: a snippet window runs
// to 600 characters and this is per visible row
// per frame. The hover job is built inside the
// closure, so an un-hovered row builds nothing.
let snippet = hit.snippet.as_ref();
// Name and path matches show a whole field, so
// they render bracketed: [matched field].
let whole_field =
hit.stage <= 4 || hit.stage == 7 || hit.stage >= 9;
row.col(|ui| {
if let Some(snip) = &snippet {
if let Some(snip) = snippet {
let width = ui.available_width();
let job = centered_match_job(ui, snip, width, whole_field);
let mut response = ui
@ -505,10 +542,9 @@ impl SearchTab {
)
.inner;
if !snip.ranges.is_empty() {
let hover = snip.clone();
response = response.on_hover_ui(|ui| {
ui.set_max_width(520.0);
let job = snippet_job(ui, &hover, 10);
let job = snippet_job(ui, snip, 10);
ui.label(job);
});
}
@ -592,6 +628,8 @@ impl SearchTab {
})
})
.inner;
// Put the permutation back for the next frame.
self.order = order;
crate::ui_util::more_below_hint(ui, &table_scroll);
self.hovered_row = hovered_now;
@ -900,8 +938,14 @@ fn centered_match_job(
// centered single-line cell. Flatten them to spaces — a byte-for-byte
// ASCII replacement, so the match ranges stay valid. The mouseover
// renders the original window untouched.
let flattened = snip.window.replace(['\n', '\r', '\t'], " ");
let window = flattened.as_str();
//
// Only copied when there is something to flatten: name and path snippets
// never contain these, and this runs per visible row per frame.
let flattened: Option<String> = snip
.window
.contains(['\n', '\r', '\t'])
.then(|| snip.window.replace(['\n', '\r', '\t'], " "));
let window = flattened.as_deref().unwrap_or(&snip.window);
let (start, end) = match snip.ranges.first().copied() {
Some((a, b)) => {
let match_chars = window[a..b].chars().count();
@ -1086,6 +1130,208 @@ mod tests {
.collect()
}
fn hit(id: i64, name: &str, rank: f64, size: u64) -> SearchHit {
SearchHit {
file_id: id,
name: name.to_string(),
path: format!("/d/{name}"),
size,
mtime: 1_700_000_000,
rank,
stage: rank as u8,
snippet: None,
}
}
fn batch(tab: &mut SearchTab, hits: Vec<SearchHit>) {
tab.apply_update(
SearchUpdate::Hits {
generation: tab.generation,
hits,
},
1000,
);
if tab.sort_dirty {
tab.resort();
}
}
/// A tab already past the fade, so batches land straight in `results`.
fn streaming_tab() -> SearchTab {
let mut tab = SearchTab::new(false);
tab.focus_query = false;
tab.query = "zebra".into();
tab.swap_pending = false;
tab
}
fn displayed(tab: &SearchTab) -> Vec<&str> {
tab.order
.iter()
.map(|&i| tab.results[i as usize].name.as_str())
.collect()
}
/// The cascade streams each pass as it scans, so a better-ranked hit can
/// arrive after a worse one. The table has to stay ordered by its keyed
/// column as batches land, not merely append them.
#[test]
fn later_batches_land_in_the_tables_sort_order() {
let mut tab = streaming_tab();
batch(&mut tab, vec![hit(1, "middling.txt", 4.0, 50)]);
assert_eq!(displayed(&tab), vec!["middling.txt"]);
// A rank-1 hit found later in the scan belongs on top.
batch(&mut tab, vec![hit(2, "best.txt", 1.0, 10)]);
assert_eq!(displayed(&tab), vec!["best.txt", "middling.txt"]);
// And a rank-10 one belongs at the bottom, not wherever it arrived.
batch(&mut tab, vec![hit(3, "worst.txt", 10.0, 90)]);
assert_eq!(
displayed(&tab),
vec!["best.txt", "middling.txt", "worst.txt"]
);
}
/// Rank is only the default. Under any other key, arrivals slot into that
/// key's order.
#[test]
fn batches_respect_a_non_rank_sort_key() {
let mut tab = streaming_tab();
tab.sort = (SortKey::Name, true);
batch(&mut tab, vec![hit(1, "mango.txt", 1.0, 50)]);
batch(&mut tab, vec![hit(2, "apple.txt", 9.0, 10)]);
batch(&mut tab, vec![hit(3, "zucchini.txt", 2.0, 90)]);
assert_eq!(
displayed(&tab),
vec!["apple.txt", "mango.txt", "zucchini.txt"],
"name order, not arrival or rank order"
);
tab.sort = (SortKey::Size, false);
tab.sort_dirty = true;
tab.resort();
assert_eq!(displayed(&tab), vec!["zucchini.txt", "mango.txt", "apple.txt"]);
}
/// The user may re-key the sort at any time, including while results are
/// still arriving: rows already shown must re-order, and later batches
/// must land under the new key.
#[test]
fn re_keying_the_sort_mid_stream_reorders_everything() {
let mut tab = streaming_tab();
batch(&mut tab, vec![hit(1, "delta.txt", 1.0, 30)]);
batch(&mut tab, vec![hit(2, "alpha.txt", 5.0, 10)]);
assert_eq!(displayed(&tab), vec!["delta.txt", "alpha.txt"], "rank order");
// Header click, mid-search.
tab.sort = (SortKey::Name, true);
tab.sort_dirty = true;
tab.resort();
assert_eq!(
displayed(&tab),
vec!["alpha.txt", "delta.txt"],
"rows that already arrived re-order under the new key"
);
batch(&mut tab, vec![hit(3, "bravo.txt", 2.0, 20)]);
assert_eq!(
displayed(&tab),
vec!["alpha.txt", "bravo.txt", "delta.txt"],
"and the next batch lands under it too"
);
}
/// At the display cap, retention stays keyed on rank even when the table
/// is shown in another order — otherwise a broad query fills up with
/// whatever the scan reached first and never shows the good hits.
#[test]
fn a_late_better_hit_displaces_the_worst_at_the_cap() {
let mut tab = streaming_tab();
tab.sort = (SortKey::Name, true);
let limit = 3;
let send = |tab: &mut SearchTab, hits: Vec<SearchHit>| {
tab.apply_update(
SearchUpdate::Hits {
generation: tab.generation,
hits,
},
limit,
);
if tab.sort_dirty {
tab.resort();
}
};
send(
&mut tab,
vec![
hit(1, "aaa.txt", 9.0, 10),
hit(2, "bbb.txt", 8.0, 20),
hit(3, "ccc.txt", 7.0, 30),
],
);
assert_eq!(displayed(&tab), vec!["aaa.txt", "bbb.txt", "ccc.txt"]);
assert!(!tab.limited);
// Full. A rank-1 arrival must still get in, evicting rank 9.
send(&mut tab, vec![hit(4, "zzz.txt", 1.0, 40)]);
assert!(tab.limited, "the cap was hit");
assert_eq!(tab.results.len(), limit);
assert_eq!(
displayed(&tab),
vec!["bbb.txt", "ccc.txt", "zzz.txt"],
"worst rank dropped, display still in name order"
);
}
/// Batches arriving during the fade get the same treatment; the ordering
/// problem must not simply move inside the 250 ms window.
#[test]
fn staged_batches_are_ordered_once_the_fade_swaps() {
let mut tab = SearchTab::new(false);
tab.focus_query = false;
tab.query = "zebra".into();
tab.on_search_started(1);
assert!(tab.swap_pending);
for h in [hit(1, "worst.txt", 9.0, 10), hit(2, "best.txt", 1.0, 20)] {
tab.apply_update(
SearchUpdate::Hits {
generation: 1,
hits: vec![h],
},
1000,
);
}
assert!(tab.results.is_empty(), "still staged behind the fade");
// What the fade does when it reaches zero.
tab.results = std::mem::take(&mut tab.staging);
tab.swap_pending = false;
tab.sort_dirty = true;
tab.resort();
assert_eq!(displayed(&tab), vec!["best.txt", "worst.txt"]);
}
/// A selected row is identified by file id, so it survives both the
/// re-ordering and the eviction that a new batch can cause.
#[test]
fn the_selection_follows_its_file_across_batches() {
let mut tab = streaming_tab();
batch(&mut tab, vec![hit(1, "chosen.txt", 5.0, 10)]);
tab.selected = Some(0);
batch(&mut tab, vec![hit(2, "better.txt", 1.0, 20)]);
let sel = tab.selected.expect("still selected");
assert_eq!(
tab.results[sel as usize].file_id,
1,
"selection follows the file, not the slot"
);
}
#[test]
fn rows_respond_over_selectable_label_text() {
let ctx = egui::Context::default();

18
crates/quicksearch-gui/src/tracker.rs
View file

@ -8,18 +8,20 @@
//! computable), and measures against `now` so the estimate decays during
//! stalls instead of freezing at the last burst.
use std::collections::VecDeque;
use std::time::{Duration, Instant};
const HISTORY: Duration = Duration::from_secs(60);
pub struct SpeedTracker {
/// (when, counter value) — appended only on counter change.
points: Vec<(Instant, usize)>,
/// (when, counter value) — appended only on counter change. A deque
/// because pruning drops from the front, which is O(n) on a `Vec`.
points: VecDeque<(Instant, usize)>,
}
impl SpeedTracker {
pub fn new() -> SpeedTracker {
SpeedTracker { points: Vec::new() }
SpeedTracker { points: VecDeque::new() }
}
/// Reset between phases (each phase restarts its counter).
@ -32,18 +34,18 @@ impl SpeedTracker {
}
fn record_at(&mut self, now: Instant, files_processed: usize) {
match self.points.last() {
match self.points.back() {
Some(&(_, last)) if last == files_processed => return,
// Counter went backwards — a new phase started without an
// explicit reset.
Some(&(_, last)) if files_processed < last => self.points.clear(),
_ => {}
}
self.points.push((now, files_processed));
self.points.push_back((now, files_processed));
// Prune old points, but always keep at least two so a slow but
// steady rate never becomes unmeasurable.
while self.points.len() > 2 && now.duration_since(self.points[0].0) > HISTORY {
self.points.remove(0);
self.points.pop_front();
}
}
@ -54,8 +56,8 @@ impl SpeedTracker {
}
fn files_per_sec_at(&self, now: Instant) -> Option<f64> {
let (t0, c0) = *self.points.first()?;
let (_, c1) = *self.points.last()?;
let (t0, c0) = *self.points.front()?;
let (_, c1) = *self.points.back()?;
if self.points.len() < 2 {
return None;
}

19
crates/quicksearch-gui/src/ui_util.rs
View file

@ -20,6 +20,25 @@ pub fn bordered_button(
egui::Button::new(text).stroke(egui::Stroke::new(1.5, color))
}
/// Render a section whose widget count changes from frame to frame inside
/// its own child `Ui`.
///
/// egui derives a widget's id from how many widgets precede it in the same
/// `Ui`. A section that emits, say, one label while idle and four rows plus
/// a progress line while working therefore *renames* every widget below it
/// the moment its content changes — and a `DragValue` or `TextEdit` whose
/// id changes loses keyboard focus and whatever the user was typing. A
/// child `Ui` costs the parent exactly one id no matter what goes inside
/// it, so everything below keeps its identity.
///
/// Wrapping does not help the section's *own* widgets: a child `Ui` mixes
/// the parent's counter into its id, so an unstable section cannot be made
/// stable from the inside. Live text and progress belong in a wrapped
/// section; editable fields belong outside one.
pub fn stable_section<R>(ui: &mut egui::Ui, contents: impl FnOnce(&mut egui::Ui) -> R) -> R {
ui.vertical(contents).inner
}
/// Whether `pattern` is usable as an ignore pattern. `IgnoreSet::compile`
/// silently *skips* patterns that trim to nothing, so emptiness is checked
/// here with the same trimming rules compile applies.

152
crates/quicksearch-gui/src/unlock.rs
View file

@ -21,6 +21,21 @@ use zeroize::{Zeroize, Zeroizing};
use crate::app::QuickSearchApp;
use crate::keychain;
/// Where this session's index key came from.
///
/// The app needs it for anything that *refers* to the key: telling someone
/// "the password you just entered" is wrong when they never typed one, because
/// the keychain answered before the window opened.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum KeySource {
/// The index is not password-protected.
Unprotected,
/// Typed at the unlock prompt during this session.
Prompt,
/// Supplied by the OS keychain, with no prompt shown.
Keychain,
}
/// The application shell handed to eframe: locked (unlock screen) or
/// running (the real app).
pub enum Gate {
@ -36,8 +51,9 @@ impl Gate {
cfg: Config,
config_error: Option<String>,
initial_query: Option<String>,
key_source: KeySource,
) -> Result<Gate, String> {
QuickSearchApp::new(ctx, cfg, config_error, initial_query)
QuickSearchApp::new(ctx, cfg, config_error, initial_query, key_source)
.map(|app| Gate::Running(Box::new(app)))
}
@ -99,8 +115,13 @@ pub fn try_keychain_unlock(cfg: &Config) -> bool {
enum Mode {
/// An index exists: the password must open it.
Unlock,
/// Protection is on but no index file exists yet — the typed password
/// (with confirmation) becomes the one the new index is built under.
/// Protection is on but no index file exists yet, so the typed password
/// becomes the one the new index is built under.
///
/// Still not a *new* password: this mode is only reachable with a salt
/// already in the config, so one was chosen previously and this is the
/// user re-entering it. Choosing a genuinely new password happens in the
/// Options window, which does its own confirmation.
Create,
/// `password_protected = true` but the salt is missing or corrupt; no
/// password can help. Only the reset escape hatch applies.
@ -113,9 +134,14 @@ pub struct UnlockScreen {
initial_query: Option<String>,
mode: Mode,
password: String,
confirm: String,
remember: bool,
error: Option<String>,
/// Put the caret in the password field on the next frame. Set once at
/// startup and again after a failed attempt, so the user can retype
/// straight away — but *not* every frame: re-focusing unconditionally
/// traps the caret, and nothing else on the screen can be tabbed to or
/// clicked into.
focus_password: bool,
/// In-flight Argon2 derivation (+ verification) on a worker thread.
job: Option<mpsc::Receiver<Result<IndexKey, String>>>,
forgot_confirm: bool,
@ -140,9 +166,9 @@ impl UnlockScreen {
initial_query,
mode,
password: String::new(),
confirm: String::new(),
remember,
error: None,
focus_password: true,
job: None,
forgot_confirm: false,
}
@ -159,6 +185,9 @@ impl UnlockScreen {
} else {
e
});
// The field was cleared on submit, so put the caret back
// in it rather than making the user click before retrying.
self.focus_password = true;
}
}
}
@ -206,15 +235,6 @@ impl UnlockScreen {
.hint_text("Password")
.desired_width(240.0),
);
if matches!(self.mode, Mode::Create) {
ui.add(
egui::TextEdit::singleline(&mut self.confirm)
.id(confirm_field_id())
.password(true)
.hint_text("Confirm password")
.desired_width(240.0),
);
}
ui.add_space(4.0);
ui.checkbox(&mut self.remember, "Remember on this device")
.on_hover_text(
@ -231,8 +251,9 @@ impl UnlockScreen {
let entered = field.lost_focus()
&& ui.input(|i| i.key_pressed(egui::Key::Enter));
submitted = clicked || entered;
if !busy && !field.has_focus() && !submitted {
if self.focus_password && !busy {
field.request_focus();
self.focus_password = false;
}
});
if busy {
@ -292,19 +313,12 @@ impl UnlockScreen {
self.error = Some("The password may not be empty.".to_string());
return;
}
if self.password != self.confirm {
self.error = Some("Passwords do not match.".to_string());
return;
}
}
let Ok(salt) = self.cfg.security.salt_bytes() else {
return; // BrokenSalt mode never reaches submit
};
let password = Zeroizing::new(std::mem::take(&mut self.password));
self.confirm.zeroize();
self.confirm.clear();
purge_text_state(ctx, pw_field_id());
purge_text_state(ctx, confirm_field_id());
let verify_against = match self.mode {
Mode::Unlock => Some(self.cfg.resolved_database_path()),
@ -352,11 +366,20 @@ impl UnlockScreen {
/// Construct the real app; on failure stay locked and show why.
fn launch(&mut self, ctx: &egui::Context) -> Option<QuickSearchApp> {
// Reaching here means either the password was just typed, or the
// "forgot password" path disabled protection on the way. The config
// says which, and no caller has to remember to pass it.
let key_source = if self.cfg.security.password_protected {
KeySource::Prompt
} else {
KeySource::Unprotected
};
match QuickSearchApp::new(
ctx,
self.cfg.clone(),
self.config_error.take(),
self.initial_query.take(),
key_source,
) {
Ok(app) => Some(app),
Err(e) => {
@ -420,7 +443,6 @@ impl UnlockScreen {
impl Drop for UnlockScreen {
fn drop(&mut self) {
self.password.zeroize();
self.confirm.zeroize();
}
}
@ -428,10 +450,6 @@ fn pw_field_id() -> egui::Id {
egui::Id::new("unlock-password")
}
fn confirm_field_id() -> egui::Id {
egui::Id::new("unlock-confirm")
}
/// Drop egui's retained state for a password field — its text buffer and
/// undo history — so the plaintext doesn't outlive the submit.
fn purge_text_state(ctx: &egui::Context, id: egui::Id) {
@ -456,3 +474,83 @@ fn delete_index_files(db_path: &std::path::Path) -> Result<(), String> {
}
Ok(())
}
#[cfg(test)]
mod tests {
use super::*;
/// A protected config whose salt parses, so the screen lands in a real
/// password mode rather than `BrokenSalt`.
fn locked_config() -> Config {
let mut cfg = Config::default();
cfg.security.password_protected = true;
cfg.security.salt = Some("0f1e2d3c4b5a69788796a5b4c3d2e1f0".to_string());
cfg.paths.database_path = std::env::temp_dir()
.join(format!("qs-unlock-test-{}.sqlite", std::process::id()))
.to_string_lossy()
.into_owned();
cfg
}
fn frame(ctx: &egui::Context, screen: &mut UnlockScreen) {
let input = egui::RawInput {
screen_rect: Some(egui::Rect::from_min_size(
egui::Pos2::ZERO,
egui::vec2(900.0, 600.0),
)),
..Default::default()
};
let _ = ctx.run(input, |ctx| {
// `update` only builds the app on a successful unlock, which needs
// a derived key — so with nothing submitted this stays on screen.
assert!(screen.update(ctx).is_none());
});
}
/// The caret starts in the password field, and — the regression — can then
/// leave it.
///
/// The screen used to call `request_focus()` on every frame the field did
/// not have focus, which yanked the caret back the instant anything else
/// took it. Nothing else on the screen could be tabbed to or clicked into.
#[test]
fn the_password_field_takes_focus_once_and_then_releases_it() {
let ctx = egui::Context::default();
let mut screen = UnlockScreen::new(locked_config(), None, None);
frame(&ctx, &mut screen);
assert!(
ctx.memory(|m| m.has_focus(pw_field_id())),
"the caret should start in the password field"
);
// Whatever the user clicks or tabs to next takes focus away.
ctx.memory_mut(|m| m.surrender_focus(pw_field_id()));
frame(&ctx, &mut screen);
assert!(
!ctx.memory(|m| m.has_focus(pw_field_id())),
"focus was stolen back; the caret is trapped in the password field"
);
// And it stays released across further frames.
frame(&ctx, &mut screen);
assert!(!ctx.memory(|m| m.has_focus(pw_field_id())));
}
/// A failed attempt is the one case that *should* re-focus: the field was
/// cleared on submit, so the user would otherwise have to click before
/// retyping.
#[test]
fn a_failed_attempt_puts_the_caret_back() {
let ctx = egui::Context::default();
let mut screen = UnlockScreen::new(locked_config(), None, None);
frame(&ctx, &mut screen);
ctx.memory_mut(|m| m.surrender_focus(pw_field_id()));
frame(&ctx, &mut screen);
assert!(!ctx.memory(|m| m.has_focus(pw_field_id())));
screen.focus_password = true; // what the error path sets
frame(&ctx, &mut screen);
assert!(ctx.memory(|m| m.has_focus(pw_field_id())));
}
}