chore(docs) updated changelog

Cargo.lock

@@ -2200,6 +2200,25 @@ dependencies = [
  "system-deps",
 ]
 
+[[package]]
+name = "h2"
+version = "0.4.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6cb093c84e8bd9b188d4c4a8cb6579fc016968d14c99882163cd3ff402a4f155"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "fnv",
+ "futures-core",
+ "futures-sink",
+ "http",
+ "indexmap",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
 [[package]]
 name = "half"
 version = "2.7.1"
@@ -2367,6 +2386,7 @@ dependencies = [
  "bytes",
  "futures-channel",
  "futures-core",
+ "h2",
  "http",
  "http-body",
  "httparse",
@@ -3518,7 +3538,7 @@ checksum = "35fb2e5f958ec131621fdd531e9fc186ed768cbe395337403ae56c17a74c68ec"
 
 [[package]]
 name = "pdfs-cli"
-version = "1.6.0"
+version = "1.8.0"
 dependencies = [
  "anyhow",
  "clap 4.6.4",
@@ -3533,7 +3553,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-core"
-version = "1.6.0"
+version = "1.8.0"
 dependencies = [
  "async-trait",
  "directories",
@@ -3554,7 +3574,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-fuse"
-version = "1.6.0"
+version = "1.8.0"
 dependencies = [
  "fuser",
  "futures",
@@ -3571,7 +3591,7 @@ dependencies = [
 
 [[package]]
 name = "pdfs-gui"
-version = "1.6.0"
+version = "1.8.0"
 dependencies = [
  "anyhow",
  "async-channel",
@@ -3832,9 +3852,9 @@ dependencies = [
 
 [[package]]
 name = "proton-drive-rs"
-version = "0.5.1"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3abae5e90e6889b120208159eb2c60b332612e5b708fdf3c14287307ce01715d"
+checksum = "fe0e39da5b61ceb862af2f4406870b9c1d9a1554889e8d40a32fff13882ec690"
 dependencies = [
  "async-trait",
  "base64",
@@ -3859,9 +3879,9 @@ dependencies = [
 
 [[package]]
 name = "proton-sdk"
-version = "0.5.1"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3fa729310c1a66e12bb41b9f5f036c86b9857e76536556052203ffecdb0b44c3"
+checksum = "97cde8cf8e3099d1dbb56302462a632a12e0ff818c012b9c9b0a7a5200544fd6"
 dependencies = [
  "aes-gcm 0.11.0",
  "async-trait",
@@ -4252,6 +4272,7 @@ dependencies = [
  "bytes",
  "futures-core",
  "futures-util",
+ "h2",
  "http",
  "http-body",
  "http-body-util",
@@ -5110,6 +5131,7 @@ dependencies = [
  "bytes",
  "futures-core",
  "futures-sink",
+ "libc",
  "pin-project-lite",
  "tokio",
 ]

Cargo.toml

@@ -3,15 +3,15 @@ resolver = "2"
 members = ["crates/pdfs-core", "crates/pdfs-fuse", "crates/pdfs-cli", "crates/pdfs-gui"]
 
 [workspace.package]
-version = "1.6.0"
+version = "1.8.0"
 edition = "2024"
 license = "MIT"
 authors = ["Nils Pukropp <contact@narl.io>"]
 rust-version = "1.96"
 
 [workspace.dependencies]
-proton-sdk = { version = "0.5.1" }
-proton-drive-rs = { version = "0.5.1" }
+proton-sdk = { version = "0.6.0" }
+proton-drive-rs = { version = "0.6.0" }
 
 # Internal crates.
 pdfs-core = { path = "crates/pdfs-core" }

crates/pdfs-cli/src/main.rs

@@ -1336,6 +1336,11 @@ fn cmd_status() -> Result<()> {
             online,
             pending_uploads,
             pending_changes,
+            parked_uploads,
+            failing_ops,
+            failing_error,
+            staged_bytes,
+            staged_oldest_secs,
             ..
         }) => {
             let state = if online { "" } else { ", offline" };
@@ -1344,6 +1349,25 @@ fn cmd_status() -> Result<()> {
                 None => String::new(),
             };
             println!("Mounted at {mountpoint} ({pinned} pinned{state}{queued})");
+            // Everything below is a queue that is not making progress on its
+            // own. Printed only when it exists, because in the ordinary case
+            // there is nothing to say.
+            if staged_bytes > 0 {
+                println!(
+                    "Staged     {} not yet uploaded (oldest {})",
+                    human_bytes(staged_bytes),
+                    human_age(staged_oldest_secs),
+                );
+            }
+            if parked_uploads > 0 {
+                println!("Parked     {parked_uploads} upload(s) waiting for a final filename");
+            }
+            if failing_ops > 0 {
+                println!(
+                    "Stuck      {failing_ops} queued operation(s) keep failing: {}",
+                    failing_error.as_deref().unwrap_or("no error recorded"),
+                );
+            }
         }
         Ok(other) => println!("Mount: unexpected response {other:?}"),
         Err(_) => println!("Mount: not running."),
@@ -1886,6 +1910,18 @@ fn path_arg(path: &Path) -> Result<String> {
         .ok_or_else(|| anyhow!("path is not valid UTF-8"))
 }
 
+/// Format an age in seconds for a report line, coarsely — the point of the
+/// number is whether it is minutes or weeks.
+fn human_age(secs: u64) -> String {
+    const UNITS: [(&str, u64); 4] = [("d", 86_400), ("h", 3_600), ("m", 60), ("s", 1)];
+    for (unit, scale) in UNITS {
+        if secs >= scale {
+            return format!("{}{unit}", secs / scale);
+        }
+    }
+    "just now".to_string()
+}
+
 /// Format a byte count for a report line.
 fn human_bytes(bytes: u64) -> String {
     const UNITS: [(&str, f64); 4] = [
@@ -1925,6 +1961,32 @@ fn cmd_cache(action: CacheCmd) -> Result<()> {
     }
 }
 
+/// Ask for a deep integrity check and wait for it to finish.
+///
+/// The daemon answers the request immediately and runs the check on a thread —
+/// it holds the single database connection for as long as it takes, which would
+/// otherwise stall the mount — so a caller that needs the verdict polls for it.
+/// `pdfs doctor` is that caller: it is a diagnostic the user is watching, so
+/// waiting is right, and the bound is generous enough for a large database
+/// without hanging forever if the daemon never finishes.
+fn await_integrity_check() -> Result<CtlResponse> {
+    const POLL: std::time::Duration = std::time::Duration::from_secs(1);
+    const GIVE_UP_AFTER: std::time::Duration = std::time::Duration::from_secs(120);
+
+    let started = std::time::Instant::now();
+    loop {
+        let response = control_request(CtlRequest::CacheInspect { deep: true })?;
+        match &response {
+            CtlResponse::CacheReport {
+                integrity_running: true,
+                integrity_checked: false,
+                ..
+            } if started.elapsed() < GIVE_UP_AFTER => std::thread::sleep(POLL),
+            _ => return Ok(response),
+        }
+    }
+}
+
 fn cmd_cache_inspect(deep: bool) -> Result<()> {
     let response = control_request(CtlRequest::CacheInspect { deep })?;
     if emit_json(&response)? {
@@ -1940,6 +2002,7 @@ fn cmd_cache_inspect(deep: bool) -> Result<()> {
             cache_budget,
             integrity_problems,
             integrity_checked,
+            integrity_running,
         } => {
             println!(
                 "Database   schema v{schema_version}, {}",
@@ -1972,6 +2035,11 @@ fn cmd_cache_inspect(deep: bool) -> Result<()> {
                         println!("  {problem}");
                     }
                 }
+            } else if integrity_running {
+                // The daemon runs the check on a thread rather than holding its
+                // only database connection for the length of it, so the answer
+                // belongs to a later request.
+                println!("\nIntegrity  checking in the background — run this again for the result");
             } else {
                 println!("\nIntegrity  not checked (pass --deep)");
             }
@@ -2209,14 +2277,24 @@ fn cmd_diagnose() -> Result<()> {
                     format!("{pending} waiting to upload")
                 },
             );
-            match control_request(CtlRequest::CacheInspect { deep: true }) {
+            match await_integrity_check() {
                 Ok(CtlResponse::CacheReport {
-                    integrity_problems, ..
+                    integrity_problems,
+                    integrity_checked: true,
+                    ..
                 }) if integrity_problems.is_empty() => report.finding(
                     DiagnoseLevel::Ok,
                     "  database integrity",
                     "SQLite integrity_check passed",
                 ),
+                Ok(CtlResponse::CacheReport {
+                    integrity_checked: false,
+                    ..
+                }) => report.finding(
+                    DiagnoseLevel::Warn,
+                    "  database integrity",
+                    "still checking — run `pdfs cache inspect --deep` for the result",
+                ),
                 Ok(CtlResponse::CacheReport {
                     integrity_problems, ..
                 }) => report.finding(

crates/pdfs-core/src/cache.rs

@@ -16,17 +16,20 @@
 //! [`AppDirs::cache_dir`]: crate::config::AppDirs::cache_dir
 //! [`Node`]: proton_drive_rs::Node
 
-use std::collections::{BTreeMap, HashSet};
+use std::collections::{BTreeMap, HashMap, HashSet};
 use std::fmt;
 use std::io::{Read, Seek, SeekFrom};
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
-use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::time::{SystemTime, UNIX_EPOCH};
 
 use proton_drive_rs::proton_sdk::ids::NodeUid;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
+use tracing::debug;
+
+use parking_lot::Mutex;
 
 use crate::db::{CacheEntryInput, Db};
 use crate::error::{Error, Result};
@@ -74,6 +77,18 @@ fn now_millis() -> i64 {
 /// `download_range` fetch with no straddling.
 pub const BLOCK_SIZE: u64 = 1 << 22;
 
+/// Buffered LRU touches that trigger a flush. Sized so a sequential read of a
+/// large file flushes a few times rather than per block, while an idle cache
+/// never holds more than a few hundred pending keys.
+const TOUCH_FLUSH_AT: usize = 256;
+
+/// How long block `idx` of a file of `size` bytes must be: a full
+/// [`BLOCK_SIZE`] except for the last one, which holds the remainder.
+fn block_len(size: u64, idx: u64) -> u64 {
+    let start = idx.saturating_mul(BLOCK_SIZE);
+    size.saturating_sub(start).min(BLOCK_SIZE)
+}
+
 /// How many eviction candidates to pull per batch when a pool is over budget.
 /// Large enough that a normal overshoot (a store or two past the cap) is settled
 /// in one query, small enough that a badly-over-budget cache does not read its
@@ -131,6 +146,19 @@ struct PinFile {
 /// in the gaps. Uploading that as-is would silently corrupt the file.
 /// `authored` says which ranges are real; `complete` says whether the file can
 /// be uploaded as it stands.
+/// What `staging/` currently holds, from [`ContentCache::staging_usage`].
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct StagingUsage {
+    /// Bytes across every staged blob.
+    pub bytes: u64,
+    /// Staged blobs, not counting their sidecars.
+    pub files: u64,
+    /// Age of the oldest staged blob, in seconds. Zero when there are none —
+    /// what makes a figure here interesting is it being *large*, which is the
+    /// shape of a write that is never going to drain by itself.
+    pub oldest_secs: u64,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct StagedWrite {
     /// Node the write targeted, in `volume~link` display form.
@@ -263,6 +291,23 @@ pub struct ContentCache {
     /// store/read/evict updates it, and the budget enforcers query it instead of
     /// scanning the cache directories (plan.md P4).
     db: Arc<Db>,
+    /// LRU touches not yet written to `cache_entries`, keyed by cache key.
+    ///
+    /// A cache hit only reorders eviction candidates, so it does not have to be
+    /// durable — but as one write transaction per hit it was 32 transactions per
+    /// 4 MiB block delivered, taken on the read path, under the connection mutex
+    /// every `lookup` also needs. Buffered here and flushed in one transaction
+    /// once [`TOUCH_FLUSH_AT`] have accumulated, and always before an eviction
+    /// pass reads the ordering these describe.
+    touches: Mutex<HashMap<String, i64>>,
+    /// Whether `recovery/` may hold a write the daemon has not queued yet.
+    ///
+    /// The drain asks for recovered writes on every idle pass — a `read_dir` of
+    /// a directory that is empty in every run that did not crash. The directory
+    /// gains entries in exactly two places (the rescue at open, and the release
+    /// path that parks a still-authoritative scratch file), both of which set
+    /// this; a scan that finds nothing actionable clears it.
+    recovery_pending: AtomicBool,
 }
 
 impl ContentCache {
@@ -325,6 +370,9 @@ impl ContentCache {
             block_bytes: AtomicU64::new(0),
             thumb_bytes: AtomicU64::new(0),
             db,
+            touches: Mutex::new(HashMap::new()),
+            // The rescue above has already run, so start by looking once.
+            recovery_pending: AtomicBool::new(true),
         };
         cache.reconcile()?;
         // After reconcile, so the totals describe the index that now describes
@@ -473,7 +521,7 @@ impl ContentCache {
         let blob = self.valid_blob(uid, mtime, size)?;
         // Record the access for LRU in the cache index. Best effort — a failed
         // update only makes eviction order slightly less accurate.
-        let _ = self.db.cache_accessed(&Self::key(uid), now_millis());
+        self.touch(Self::key(uid));
         let mut f = std::fs::File::open(&blob).ok()?;
         if offset >= size {
             return Some(Vec::new());
@@ -563,22 +611,38 @@ impl ContentCache {
             .join(format!("{}.b{idx}.meta", Self::key(uid)))
     }
 
+    /// Whether block `idx` of `uid` is cached and fresh for `(mtime, size)`,
+    /// without reading it. For a caller deciding whether a block is worth
+    /// *fetching* — reading 4 MiB to answer that would defeat the point.
+    pub fn has_block(&self, uid: &NodeUid, mtime: i64, size: u64, idx: u64) -> bool {
+        self.block_tag(uid, idx) == Some(Meta { mtime, size })
+            && std::fs::metadata(self.block_blob(uid, idx))
+                .is_ok_and(|m| m.len() == block_len(size, idx))
+    }
+
+    /// The `(mtime, size)` a cached block claims to belong to, from its sidecar.
+    fn block_tag(&self, uid: &NodeUid, idx: u64) -> Option<Meta> {
+        serde_json::from_slice(&std::fs::read(self.block_meta(uid, idx)).ok()?).ok()
+    }
+
     /// Serve cached block `idx` (a [`BLOCK_SIZE`]-aligned chunk) of `uid`, or
     /// `None` on miss/stale. Validated against `(mtime, size)` like a whole-file
     /// blob, so a new revision (which bumps the mtime) is detected. Bumps the
     /// block's mtime for LRU, best effort.
     pub fn cached_block(&self, uid: &NodeUid, mtime: i64, size: u64, idx: u64) -> Option<Vec<u8>> {
-        let want = Meta { mtime, size };
-        let meta: Meta =
-            serde_json::from_slice(&std::fs::read(self.block_meta(uid, idx)).ok()?).ok()?;
-        if meta != want {
+        if self.block_tag(uid, idx)? != (Meta { mtime, size }) {
             return None;
         }
         let mut f = std::fs::File::open(self.block_blob(uid, idx)).ok()?;
-        // LRU touch in the cache index (best effort).
-        let _ = self
-            .db
-            .cache_accessed(&self.block_key(uid, idx), now_millis());
+        // The blob is not fsynced on the way in (see `store_block`), so a crash
+        // can leave a short one behind the sidecar that describes it. Length is
+        // what makes that detectable, the same check `valid_blob` applies to a
+        // whole-file blob.
+        if f.metadata().ok()?.len() != block_len(size, idx) {
+            return None;
+        }
+        // LRU touch, buffered rather than written (see `touch`).
+        self.touch(self.block_key(uid, idx));
         let mut buf = Vec::new();
         f.read_to_end(&mut buf).ok()?;
         Some(buf)
@@ -603,7 +667,10 @@ impl ContentCache {
             let mut f = std::fs::File::create(&tmp)?;
             use std::io::Write;
             f.write_all(bytes)?;
-            f.sync_all()?;
+            // Deliberately not fsynced. This runs inline on the thread a kernel
+            // read is waiting on, and the block cache is disposable: the worst a
+            // crash can do is leave a short blob, which `cached_block` rejects
+            // on length and refetches.
         }
         std::fs::rename(&tmp, &blob)?;
         std::fs::write(
@@ -622,6 +689,40 @@ impl ContentCache {
         Ok(())
     }
 
+    /// Record a cache hit for LRU ordering, without touching the database.
+    ///
+    /// Later touches of the same key overwrite earlier ones — only the most
+    /// recent matters — so a hot file costs one map entry however many times it
+    /// is read.
+    fn touch(&self, key: String) {
+        let flush = {
+            let mut touches = self.touches.lock();
+            touches.insert(key, now_millis());
+            touches.len() >= TOUCH_FLUSH_AT
+        };
+        if flush {
+            self.flush_touches();
+        }
+    }
+
+    /// Write the buffered LRU touches out. Best effort: losing them costs
+    /// eviction accuracy, nothing else.
+    ///
+    /// Public because eviction ordering is only as good as what has been
+    /// flushed, and the daemon flushes on its idle passes as well as here.
+    pub fn flush_touches(&self) {
+        let batch: Vec<(String, i64)> = {
+            let mut touches = self.touches.lock();
+            if touches.is_empty() {
+                return;
+            }
+            touches.drain().collect()
+        };
+        if let Err(error) = self.db.cache_accessed_batch(&batch) {
+            debug!(?error, "flushing cache LRU touches failed");
+        }
+    }
+
     /// Re-read every pool total from the cache index, which is authoritative.
     fn reseed_totals(&self) {
         for kind in [KIND_BLOB, KIND_BLOCK, KIND_THUMB] {
@@ -689,7 +790,10 @@ impl ContentCache {
             return; // the common case: one atomic load, no query
         }
 
-        // Over budget by the running count — now do the accurate work.
+        // Over budget by the running count — now do the accurate work. The
+        // buffered hits are what say which entries are cold, so they have to
+        // reach the index before it is asked.
+        self.flush_touches();
         let mut running = match self.db.cache_total_bytes(kind) {
             Ok(n) => n,
             Err(_) => return,
@@ -727,6 +831,70 @@ impl ContentCache {
         self.reseed_totals();
     }
 
+    /// Duplicate `src` onto `dst`, preferring a reflink.
+    ///
+    /// The caller is copying a staged revision — a whole file, up to whatever
+    /// the user stores — so the difference between the three strategies is the
+    /// difference between instant and minutes:
+    ///
+    /// 1. `FICLONE` shares the extents copy-on-write. On btrfs, XFS with
+    ///    `reflink=1`, and bcachefs this is O(1) and consumes no extra space.
+    /// 2. `copy_file_range` still copies, but entirely inside the kernel — no
+    ///    round trip through a user-space buffer.
+    /// 3. `std::fs::copy` is the portable fallback, and what this used to
+    ///    always do.
+    ///
+    /// `dst` is truncated. Returns the number of bytes the destination ends up
+    /// holding.
+    pub fn clone_or_copy(src: &Path, dst: &Path) -> Result<u64> {
+        use std::os::fd::AsRawFd as _;
+
+        // `_IOW(0x94, 9, int)` — the whole-file reflink ioctl.
+        const FICLONE: libc::c_ulong = 0x4004_9409;
+
+        let source = std::fs::File::open(src)?;
+        let len = source.metadata()?.len();
+        let target = std::fs::OpenOptions::new()
+            .write(true)
+            .create(true)
+            .truncate(true)
+            .open(dst)?;
+
+        // SAFETY: both descriptors are owned and open; FICLONE takes the source
+        // fd by value and writes nothing back through the pointer-shaped arg.
+        if unsafe { libc::ioctl(target.as_raw_fd(), FICLONE, source.as_raw_fd()) } == 0 {
+            return Ok(len);
+        }
+
+        let mut copied = 0u64;
+        while copied < len {
+            let want = (len - copied) as usize;
+            // SAFETY: owned descriptors; null offsets advance the file offsets,
+            // which is what a whole-file copy wants.
+            let n = unsafe {
+                libc::copy_file_range(
+                    source.as_raw_fd(),
+                    std::ptr::null_mut(),
+                    target.as_raw_fd(),
+                    std::ptr::null_mut(),
+                    want,
+                    0,
+                )
+            };
+            if n <= 0 {
+                break;
+            }
+            copied += n as u64;
+        }
+        if copied == len {
+            return Ok(len);
+        }
+        // A short or refused `copy_file_range` (cross-filesystem on older
+        // kernels, an unsupported fs) leaves a partial destination; the
+        // portable copy truncates and rewrites it from scratch.
+        Ok(std::fs::copy(src, dst)?)
+    }
+
     /// Create a fresh, empty read-write scratch file for a disk-backed write
     /// handle. Returns the open file and its path (for cleanup on release).
     pub fn create_scratch(&self) -> Result<(std::fs::File, PathBuf)> {
@@ -806,8 +974,9 @@ impl ContentCache {
         let _ = std::fs::remove_file(Self::scratch_sidecar(scratch));
     }
 
-    /// Move every scratch file that carries a readable sidecar into `recovery`,
-    /// before the caller empties the scratch directory.
+    /// Move every scratch file that carries a sidecar into `recovery`, before the
+    /// caller empties the scratch directory. A sidecar that does not parse is
+    /// dropped and its blob moved anyway, unaccompanied — see below.
     ///
     /// Blob first, then sidecar: a blob in recovery without its sidecar is
     /// dropped on the next open (indistinguishable from rubbish), whereas a
@@ -828,43 +997,28 @@ impl ContentCache {
                 // Not marked durable, or durability marker was explicitly cleared on release
                 continue;
             }
-            // Parsed if valid; if sidecar exists but is corrupted, build synthetic StagedWrite
+            // A sidecar that does not parse describes nothing, and inventing one
+            // is worse than having none: a fabricated uid parses, so recovery
+            // would queue an op against a node that does not exist and retry it
+            // forever, while a fabricated `complete: true` claims the blob is the
+            // whole file and stops any gap-fill. The blob still moves to
+            // recovery — sidecar-less, which `recovered_writes` skips and so
+            // leaves on disk for a human, the right outcome for bytes we cannot
+            // describe.
             let has_valid_sidecar = std::fs::read(&side)
                 .ok()
                 .and_then(|b| serde_json::from_slice::<StagedWrite>(&b).ok())
                 .is_some();
-            if !has_valid_sidecar {
-                if let Ok(meta) = std::fs::metadata(&blob)
-                    && meta.len() > 0
-                {
-                    let blob_name = blob
-                        .file_name()
-                        .map(|n| n.to_string_lossy().to_string())
-                        .unwrap_or_default();
-                    let synthetic = StagedWrite {
-                        uid: format!("recovered~{blob_name}"),
-                        len: meta.len(),
-                        base_size: meta.len(),
-                        base_mtime: 0,
-                        authored: vec![(0, meta.len())],
-                        complete: true,
-                        based_on: None,
-                    };
-                    if let Ok(data) = serde_json::to_vec(&synthetic) {
-                        let _ = std::fs::write(&side, data);
-                    } else {
-                        continue;
-                    }
-                } else {
-                    continue;
-                }
-            }
             let Some(name) = blob.file_name() else {
                 continue;
             };
             if std::fs::rename(&blob, recovery_dir.join(name)).is_err() {
                 continue;
             }
+            if !has_valid_sidecar {
+                let _ = std::fs::remove_file(&side);
+                continue;
+            }
             let Some(side_name) = side.file_name() else {
                 continue;
             };
@@ -919,7 +1073,14 @@ impl ContentCache {
     /// daemon to hand to the upload queue. A blob whose sidecar did not survive
     /// is not actionable automatically, but is retained for manual recovery:
     /// unexplained bytes are never evidence that user data is disposable.
+    ///
+    /// Asked for on every idle drain pass, so the usual answer — a directory
+    /// that has been empty since the last clean shutdown — costs an atomic load
+    /// rather than a `read_dir` (see [`recovery_pending`](Self::recovery_pending)).
     pub fn recovered_writes(&self) -> Vec<(PathBuf, StagedWrite)> {
+        if !self.recovery_pending.load(Ordering::Relaxed) {
+            return Vec::new();
+        }
         let Ok(entries) = std::fs::read_dir(&self.recovery_dir) else {
             return Vec::new();
         };
@@ -936,6 +1097,39 @@ impl ContentCache {
                 out.push((blob, meta));
             }
         }
+        // Blobs whose sidecar did not survive stay here for a human and are
+        // never actionable, so "nothing to queue" is the condition that stops
+        // the polling — not "the directory is empty".
+        if out.is_empty() {
+            self.recovery_pending.store(false, Ordering::Relaxed);
+        }
+        out
+    }
+
+    /// Every blob in `staging/`, paired with its sidecar when it has a readable
+    /// one.
+    ///
+    /// For the startup reconcile: staging holds the only copy of writes the
+    /// kernel has already been told succeeded, and three failure paths can
+    /// leave one there with no `pending_op` row pointing at it. Nothing walked
+    /// this directory, so those bytes were invisible to the daemon and to the
+    /// user both — `hydrate_pending` walks the op table and
+    /// [`recovered_writes`](Self::recovered_writes) walks `recovery/`.
+    pub fn staged_writes(&self) -> Vec<(PathBuf, Option<StagedWrite>)> {
+        let Ok(entries) = std::fs::read_dir(&self.staging_dir) else {
+            return Vec::new();
+        };
+        let mut out = Vec::new();
+        for entry in entries.flatten() {
+            let blob = entry.path();
+            if blob.extension().is_some_and(|e| e == "json") {
+                continue;
+            }
+            let meta = std::fs::read(Self::scratch_sidecar(&blob))
+                .ok()
+                .and_then(|b| serde_json::from_slice::<StagedWrite>(&b).ok());
+            out.push((blob, meta));
+        }
         out
     }
 
@@ -1106,6 +1300,10 @@ impl ContentCache {
                         if std::fs::rename(scratch, &recovery).is_ok() {
                             let _ = std::fs::File::open(&self.recovery_dir)
                                 .and_then(|dir| dir.sync_all());
+                            // These bytes are unaccompanied, but re-arm the
+                            // scan anyway: the flag exists to skip a directory
+                            // nothing has written to, and something just did.
+                            self.recovery_pending.store(true, Ordering::Relaxed);
                             return Err(retained(
                                 &recovery,
                                 Error::Other(format!(
@@ -1210,9 +1408,7 @@ impl ContentCache {
         // Record the access for LRU, as the blob and block readers do. Without
         // it the pool evicts in insertion order, so scrolling back up a gallery
         // would drop exactly the tiles being looked at.
-        let _ = self
-            .db
-            .cache_accessed(&self.thumb_key(uid, ttype), now_millis());
+        self.touch(self.thumb_key(uid, ttype));
         Some(bytes)
     }
 
@@ -1243,9 +1439,7 @@ impl ContentCache {
         if !blob.exists() {
             return None;
         }
-        let _ = self
-            .db
-            .cache_accessed(&self.thumb_key(uid, ttype), now_millis());
+        self.touch(self.thumb_key(uid, ttype));
         Some(blob)
     }
 
@@ -1388,6 +1582,44 @@ impl ContentCache {
             .sum()
     }
 
+    /// Bytes and file count held in `staging/`, with the age of the oldest
+    /// entry in seconds.
+    ///
+    /// Staging is not part of [`usage`](Self::usage) and is not subject to the
+    /// budget — it holds the *only* copy of writes that have not been uploaded,
+    /// so nothing here may ever be evicted to reclaim space. That is precisely
+    /// why it has to be reportable: a parked transient create (three abandoned
+    /// 4 GiB downloads, say) otherwise occupies disk with nothing in any UI
+    /// accounting for it.
+    ///
+    /// A `read_dir` per call, which is fine for the callers it has: a status
+    /// request, polled every couple of seconds against a directory that is
+    /// empty whenever the queue is drained.
+    pub fn staging_usage(&self) -> StagingUsage {
+        let Ok(entries) = std::fs::read_dir(&self.staging_dir) else {
+            return StagingUsage::default();
+        };
+        let mut usage = StagingUsage::default();
+        let now = std::time::SystemTime::now();
+        for entry in entries.flatten() {
+            // Sidecars describe the blob beside them; counting both would double
+            // the file count and report metadata as user bytes.
+            if entry.path().extension().is_some_and(|e| e == "json") {
+                continue;
+            }
+            let Ok(meta) = entry.metadata() else { continue };
+            usage.files += 1;
+            usage.bytes += meta.len();
+            if let Ok(age) = meta.modified().and_then(|m| {
+                now.duration_since(m)
+                    .map_err(|_| std::io::Error::other("modified in the future"))
+            }) {
+                usage.oldest_secs = usage.oldest_secs.max(age.as_secs());
+            }
+        }
+        usage
+    }
+
     /// Configured soft byte cap (`0` = unlimited), for display alongside
     /// [`usage`](Self::usage). This is the whole-cache figure; internally it is
     /// divided between the pools (see [`SPLIT_BLOB`]), but the user set one
@@ -2004,45 +2236,63 @@ mod tests {
     fn block_store_then_read() {
         let (c, _d) = cache();
         let u = uid("a");
-        c.store_block(&u, 100, 4096, 0, b"block-zero").unwrap();
-        assert_eq!(c.cached_block(&u, 100, 4096, 0).unwrap(), b"block-zero");
+        // A 10-byte file is one block of 10 bytes: the length a reader expects
+        // is derived from the file size, so the fixture has to be geometrically
+        // real or the torn-write check rejects it.
+        c.store_block(&u, 100, 10, 0, b"block-zero").unwrap();
+        assert_eq!(c.cached_block(&u, 100, 10, 0).unwrap(), b"block-zero");
         // A different index is a separate cache entry, absent here.
-        assert!(c.cached_block(&u, 100, 4096, 1).is_none());
+        assert!(c.cached_block(&u, 100, 10, 1).is_none());
         // A new revision (mtime/size bump) invalidates the block.
-        assert!(c.cached_block(&u, 101, 4096, 0).is_none());
+        assert!(c.cached_block(&u, 101, 10, 0).is_none());
         assert!(c.cached_block(&u, 100, 5000, 0).is_none());
     }
 
+    /// The block cache is written without an fsync, so a crash can leave a blob
+    /// shorter than the sidecar beside it claims. Serving that as content would
+    /// hand the reader a hole; it has to read as a miss.
+    #[test]
+    fn a_truncated_block_reads_as_a_miss() {
+        let (c, _d) = cache();
+        let u = uid("a");
+        c.store_block(&u, 100, 10, 0, b"block-zero").unwrap();
+        std::fs::write(c.block_blob(&u, 0), b"block").unwrap();
+        assert!(c.cached_block(&u, 100, 10, 0).is_none());
+    }
+
     #[test]
     fn evict_drops_blocks() {
         let (c, _d) = cache();
         let u = uid("a");
-        c.store_block(&u, 1, 8, 0, b"aaaa").unwrap();
-        c.store_block(&u, 1, 8, 1, b"bbbb").unwrap();
+        // Both blocks are stored at their real geometry (a 4-byte file is one
+        // block), so this fails if `evict` misses them rather than because the
+        // fixture was unreadable to begin with.
+        c.store_block(&u, 1, 4, 0, b"aaaa").unwrap();
+        assert!(c.cached_block(&u, 1, 4, 0).is_some());
         c.evict(&u);
-        assert!(c.cached_block(&u, 1, 8, 0).is_none());
-        assert!(c.cached_block(&u, 1, 8, 1).is_none());
+        assert!(c.cached_block(&u, 1, 4, 0).is_none());
     }
 
     #[test]
     fn block_budget_evicts_lru() {
-        // Block pool fits two 4-byte blocks but not three.
+        // Block pool fits two 4-byte blocks but not three. One block per file,
+        // so each fixture is a geometrically real single-block file.
         let (c, _d) = cache_with_pool_cap(KIND_BLOCK, 8);
-        let u = uid("a");
-        c.store_block(&u, 1, 12, 0, b"aaaa").unwrap();
+        let (a, b, d) = (uid("a"), uid("b"), uid("d"));
+        c.store_block(&a, 1, 4, 0, b"aaaa").unwrap();
         std::thread::sleep(std::time::Duration::from_millis(10));
-        c.store_block(&u, 1, 12, 1, b"bbbb").unwrap();
-        // Touch block 0 so block 1 is the least-recently-used.
+        c.store_block(&b, 1, 4, 0, b"bbbb").unwrap();
+        // Touch a's block so b's is the least-recently-used.
         std::thread::sleep(std::time::Duration::from_millis(10));
-        assert!(c.cached_block(&u, 1, 12, 0).is_some());
+        assert!(c.cached_block(&a, 1, 4, 0).is_some());
         std::thread::sleep(std::time::Duration::from_millis(10));
-        c.store_block(&u, 1, 12, 2, b"cccc").unwrap();
+        c.store_block(&d, 1, 4, 0, b"cccc").unwrap();
         assert!(
-            c.cached_block(&u, 1, 12, 0).is_some(),
+            c.cached_block(&a, 1, 4, 0).is_some(),
             "recently-read survives"
         );
-        assert!(c.cached_block(&u, 1, 12, 1).is_none(), "LRU block evicted");
-        assert!(c.cached_block(&u, 1, 12, 2).is_some(), "newest survives");
+        assert!(c.cached_block(&b, 1, 4, 0).is_none(), "LRU block evicted");
+        assert!(c.cached_block(&d, 1, 4, 0).is_some(), "newest survives");
     }
 
     /// A failed upload must never cost the user their bytes (offline.md Phase 2):
@@ -2185,6 +2435,39 @@ mod tests {
         drop(reopened);
     }
 
+    /// `clone_or_copy` has three strategies and only the last one is portable,
+    /// so what matters is that the destination is byte-identical whichever the
+    /// filesystem under `/tmp` happens to take — including when it already held
+    /// longer content, which a reflink or a short `copy_file_range` must not
+    /// leave a tail of.
+    #[test]
+    fn clone_or_copy_reproduces_the_source_over_any_existing_destination() {
+        let dir = TempDir::new();
+        let src = dir.path().join("src");
+        let dst = dir.path().join("dst");
+        std::fs::write(&src, b"the new content").unwrap();
+        std::fs::write(&dst, vec![b'x'; 4096]).unwrap();
+
+        let len = ContentCache::clone_or_copy(&src, &dst).unwrap();
+
+        assert_eq!(len, 15);
+        assert_eq!(std::fs::read(&dst).unwrap(), b"the new content");
+    }
+
+    /// An empty staged blob is a real case — a file truncated to zero and
+    /// closed — and `FICLONE` of a zero-length source is a documented `EINVAL`.
+    #[test]
+    fn clone_or_copy_handles_an_empty_source() {
+        let dir = TempDir::new();
+        let src = dir.path().join("empty");
+        let dst = dir.path().join("dst");
+        std::fs::write(&src, b"").unwrap();
+        std::fs::write(&dst, b"stale").unwrap();
+
+        assert_eq!(ContentCache::clone_or_copy(&src, &dst).unwrap(), 0);
+        assert!(std::fs::read(&dst).unwrap().is_empty());
+    }
+
     /// A2: `fsync(2)` promises the bytes survive a crash, but queueing happens
     /// at `release`. A scratch file marked durable must therefore outlive the
     /// open that clears the scratch directory — and an unmarked one must not,

crates/pdfs-core/src/control.rs

@@ -1193,6 +1193,28 @@ pub enum Response {
         /// calling a queued `mkdir` an upload is a lie.
         #[serde(default)]
         pending_changes: u64,
+        /// Queued writes held back because the file still wears a transient
+        /// name (a browser's `*.crdownload`, an editor's `*.swp`). They are not
+        /// waiting on the network and will not drain until the finalising
+        /// rename, so they are counted apart from `pending_uploads`.
+        #[serde(default)]
+        parked_uploads: u64,
+        /// Queued ops that have failed enough times to be considered stuck.
+        /// They keep retrying forever — correctly, since their staged bytes are
+        /// the only copy — which is why they have to be visible.
+        #[serde(default)]
+        failing_ops: u64,
+        /// The most recent error from a stuck op, when there is one.
+        #[serde(default)]
+        failing_error: Option<String>,
+        /// Bytes held in `staging/`: writes accepted from the kernel whose only
+        /// copy is on local disk. Never evictable, never counted against the
+        /// cache budget — hence reported separately.
+        #[serde(default)]
+        staged_bytes: u64,
+        /// Age of the oldest staged write, in seconds.
+        #[serde(default)]
+        staged_oldest_secs: u64,
     },
     /// A human-readable success message.
     Ok { message: String },
@@ -1217,6 +1239,12 @@ pub enum Response {
         /// Whether the integrity check actually ran, so an empty problem list
         /// is not mistaken for a clean bill of health it never gave.
         integrity_checked: bool,
+        /// Whether a deep check is running right now. It reads every page of
+        /// the database under the daemon's only connection, so it is answered
+        /// asynchronously: this says "ask again shortly", which is a different
+        /// thing to say than either of the two above.
+        #[serde(default)]
+        integrity_running: bool,
     },
     /// The pin registry.
     Pins { pins: Vec<Pin> },

crates/pdfs-core/src/db/activity.rs

@@ -33,7 +33,7 @@ impl Db {
     /// whose stored `kind` no longer parses (written by an older build) are
     /// skipped rather than failing the whole read.
     pub fn activity_list(&self, limit: usize) -> Result<Vec<ActivityEntry>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT time, kind, target, detail, ok FROM activity
              ORDER BY id DESC LIMIT ?1",

crates/pdfs-core/src/db/albums.rs

@@ -76,7 +76,7 @@ impl Db {
 
     /// The persisted album listing, in stored order.
     pub fn albums_list(&self) -> Result<Vec<StoredAlbum>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT uid, name, photo_count, cover_uid, last_activity, shared
              FROM albums ORDER BY seq",
@@ -98,7 +98,7 @@ impl Db {
 
     /// Number of albums in the persisted listing.
     pub fn albums_count(&self) -> Result<usize> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let n: i64 = conn.query_row("SELECT COUNT(*) FROM albums", [], |r| r.get(0))?;
         Ok(n.max(0) as usize)
     }
@@ -164,7 +164,7 @@ impl Db {
         offset: usize,
         limit: usize,
     ) -> Result<Vec<StoredPhoto>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT uid, capture_time, name, ratio, thumb_state, kind FROM album_photos
              WHERE album_uid = ?1 ORDER BY seq LIMIT ?2 OFFSET ?3",
@@ -190,7 +190,7 @@ impl Db {
 
     /// How many photos of `album_uid` are persisted.
     pub fn album_photos_count(&self, album_uid: &str) -> Result<usize> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let n: i64 = conn.query_row(
             "SELECT COUNT(*) FROM album_photos WHERE album_uid = ?1",
             [album_uid],
@@ -207,7 +207,7 @@ impl Db {
         if uids.is_empty() {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
+        let conn = self.read();
         let placeholders = vec!["?"; uids.len()].join(",");
         let mut stmt = conn.prepare(&format!(
             "SELECT uid, capture_time, name, ratio, thumb_state, kind FROM album_photos

crates/pdfs-core/src/db/cache.rs

@@ -42,6 +42,30 @@ impl Db {
         Ok(())
     }
 
+    /// Bump `last_accessed` for many keys in one transaction.
+    ///
+    /// The read path touches a key per 128 KiB the kernel asks for — 32 times
+    /// per 4 MiB block — and each of those was its own write transaction under
+    /// the single connection mutex, contending with every `lookup` and
+    /// `readdir`. Callers accumulate touches in memory and flush them here.
+    pub fn cache_accessed_batch(&self, touches: &[(String, i64)]) -> Result<()> {
+        if touches.is_empty() {
+            return Ok(());
+        }
+        let mut conn = self.conn.lock();
+        let tx = conn.transaction()?;
+        {
+            let mut stmt = tx.prepare_cached(
+                "UPDATE cache_entries SET last_accessed = ?2 WHERE cache_key = ?1",
+            )?;
+            for (key, at) in touches {
+                stmt.execute(params![key, at])?;
+            }
+        }
+        tx.commit()?;
+        Ok(())
+    }
+
     /// Drop a single cache-entry row (one evicted blob or block).
     pub fn cache_remove(&self, key: &str) -> Result<()> {
         let conn = self.conn.lock();
@@ -75,7 +99,7 @@ impl Db {
     /// because `enforce_block_budget` runs on every cached block — once per
     /// 4 MiB of every cold read — while holding the shared connection.
     pub fn cache_total_bytes(&self, kind: &str) -> Result<u64> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let total: i64 = conn.query_row(
             "SELECT COALESCE(SUM(size_bytes), 0) FROM cache_entries WHERE kind = ?1",
             params![kind],
@@ -95,7 +119,7 @@ impl Db {
         kind: &str,
         limit: usize,
     ) -> Result<Vec<(String, u64)>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare_cached(
             "SELECT cache_key, size_bytes FROM cache_entries
              WHERE kind = ?1 ORDER BY last_accessed ASC LIMIT ?2",
@@ -115,7 +139,7 @@ impl Db {
     /// (`clear_unpinned`); incremental eviction wants
     /// [`cache_eviction_candidates`](Self::cache_eviction_candidates).
     pub fn cache_entries_by_kind(&self, kind: &str) -> Result<Vec<(String, u64)>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT cache_key, size_bytes FROM cache_entries
              WHERE kind = ?1 ORDER BY last_accessed ASC",

crates/pdfs-core/src/db/devices.rs

@@ -15,7 +15,7 @@ pub struct StoredDevice {
 
 impl Db {
     pub fn device_get(&self) -> Result<Option<StoredDevice>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         conn.query_row(
             "SELECT uid, share_id, root_uid, name, created FROM device LIMIT 1",
             [],

crates/pdfs-core/src/db/local.rs

@@ -41,32 +41,43 @@ impl Db {
         Ok(next)
     }
 
-    /// Write one batch of walked entries under scan generation `generation`. The
-    /// FTS index is *not* touched here — it is rebuilt once in
-    /// [`local_finish_scan`](Self::local_finish_scan).
+    /// Write one batch of walked entries under scan generation `generation`,
+    /// keeping the FTS index in step with them.
+    ///
+    /// The index is maintained incrementally rather than rebuilt at the end of
+    /// the scan, because a rebuild reads and re-tokenises every row in the table
+    /// — trigrams over a few hundred thousand paths — inside one transaction on
+    /// the single shared connection, which stalls the whole mount for as long as
+    /// it takes. Incremental maintenance is possible here for a reason specific
+    /// to this table: `local_fts` covers `name` and `path`, `path` is the
+    /// table's key, and `name` is derived from it, so an existing row's indexed
+    /// content can never change. Only insertions and deletions touch the index;
+    /// a re-scan that finds the same files does no index work at all.
+    ///
+    /// Hence the update-then-insert shape below rather than an upsert: the
+    /// distinction between "already indexed" and "new row" is exactly what
+    /// decides whether the index needs a write.
     pub fn local_upsert_batch(&self, generation: i64, entries: &[LocalEntry]) -> Result<()> {
         let mut conn = self.conn.lock();
         let tx = conn.transaction()?;
         {
-            let mut stmt = tx.prepare_cached(
-                "INSERT INTO local_files (path, name, is_dir, size, mtime, scan_gen)
-                 VALUES (?1, ?2, ?3, ?4, ?5, ?6)
-                 ON CONFLICT(path) DO UPDATE SET
-                   name     = excluded.name,
-                   is_dir   = excluded.is_dir,
-                   size     = excluded.size,
-                   mtime    = excluded.mtime,
-                   scan_gen = excluded.scan_gen",
+            let mut update = tx.prepare_cached(
+                "UPDATE local_files
+                    SET name = ?2, is_dir = ?3, size = ?4, mtime = ?5, scan_gen = ?6
+                  WHERE path = ?1",
             )?;
+            let mut insert = tx.prepare_cached(
+                "INSERT INTO local_files (path, name, is_dir, size, mtime, scan_gen)
+                 VALUES (?1, ?2, ?3, ?4, ?5, ?6)",
+            )?;
+            let mut index =
+                tx.prepare_cached("INSERT INTO local_fts (rowid, name, path) VALUES (?1, ?2, ?3)")?;
             for e in entries {
-                stmt.execute(params![
-                    e.path,
-                    e.name,
-                    e.is_dir as i64,
-                    e.size,
-                    e.mtime,
-                    generation
-                ])?;
+                let row = params![e.path, e.name, e.is_dir as i64, e.size, e.mtime, generation];
+                if update.execute(row)? == 0 {
+                    insert.execute(row)?;
+                    index.execute(params![tx.last_insert_rowid(), e.name, e.path])?;
+                }
             }
         }
         tx.commit()?;
@@ -74,16 +85,25 @@ impl Db {
     }
 
     /// Close scan generation `generation`: drop every row an older scan wrote
-    /// (those paths are gone from disk), rebuild the FTS index over what remains,
-    /// and stamp the completion time. Returns the number of indexed entries.
+    /// (those paths are gone from disk), drop their FTS entries, and stamp the
+    /// completion time. Returns the number of indexed entries.
+    ///
+    /// The FTS deletes name the old values explicitly, which is how an
+    /// external-content FTS5 table is told to retract a row — see
+    /// [`local_upsert_batch`](Self::local_upsert_batch) for why this is done
+    /// per-row instead of by rebuilding the index.
     pub fn local_finish_scan(&self, generation: i64, finished_at: i64) -> Result<i64> {
         let mut conn = self.conn.lock();
         let tx = conn.transaction()?;
+        tx.execute(
+            "INSERT INTO local_fts (local_fts, rowid, name, path)
+             SELECT 'delete', id, name, path FROM local_files WHERE scan_gen != ?1",
+            params![generation],
+        )?;
         tx.execute(
             "DELETE FROM local_files WHERE scan_gen != ?1",
             params![generation],
         )?;
-        tx.execute_batch("INSERT INTO local_fts(local_fts) VALUES('rebuild');")?;
         tx.execute(
             "INSERT INTO sync_state (key, value) VALUES ('local_indexed_at', ?1)
              ON CONFLICT(key) DO UPDATE SET value = excluded.value",
@@ -98,7 +118,7 @@ impl Db {
     /// has never been built. The daemon uses this to decide whether a fresh mount
     /// needs an immediate rescan or can serve the existing index.
     pub fn local_indexed_at(&self) -> Result<Option<i64>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         Ok(conn
             .query_row(
                 "SELECT value FROM sync_state WHERE key = 'local_indexed_at'",
@@ -118,7 +138,7 @@ impl Db {
         if query.is_empty() {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt;
         let rows = if query.chars().count() < TRIGRAM_MIN {
             let pat = format!("%{}%", like_escape(query));
@@ -157,7 +177,7 @@ impl Db {
         if query.is_empty() || limit == 0 {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
+        let conn = self.read();
         let terms = candidate_trigrams(query);
         let lane_limit = limit.div_ceil(2);
         let mut rows: Vec<LocalFileHit> = if terms.is_empty() {

crates/pdfs-core/src/db/migrations.rs

@@ -9,7 +9,7 @@ use super::Db;
 use crate::Result;
 
 /// Current schema version. Bump on every forward migration added below.
-pub(super) const SCHEMA_VERSION: i64 = 21;
+pub(super) const SCHEMA_VERSION: i64 = 27;
 
 impl Db {
     pub(super) fn migrate(&self) -> Result<()> {
@@ -113,6 +113,94 @@ impl Db {
                 tx.execute_batch(MIGRATION_V21)?;
             }
         }
+        if current < 22 {
+            // Older fixtures (and a database that never grew a table because
+            // its feature was never used) may be missing any of these, and an
+            // index on an absent table is a hard error rather than a no-op.
+            for (table, sql) in MIGRATION_V22 {
+                let present: bool = tx.query_row(
+                    "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = ?1",
+                    [table],
+                    |row| row.get::<_, i64>(0),
+                )? > 0;
+                if present {
+                    tx.execute_batch(sql)?;
+                }
+            }
+            // `activity_time` indexes `time DESC` while both activity queries
+            // order by `id`: it has only ever cost writes.
+            tx.execute_batch("DROP INDEX IF EXISTS activity_time;")?;
+        }
+        if current < 23 {
+            // One last full rebuild of the local index, because from here on the
+            // scan maintains it incrementally (`local_upsert_batch`) and so can
+            // no longer repair a `local_fts` that drifted from `local_files`.
+            // Whatever it costs, it costs once, at startup, before the mount
+            // serves anything.
+            let present: bool = tx.query_row(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'local_fts'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            if present {
+                tx.execute_batch("INSERT INTO local_fts(local_fts) VALUES('rebuild');")?;
+            }
+        }
+        if current < 24 {
+            tx.execute_batch(MIGRATION_V24)?;
+        }
+        if current < 25 {
+            // Same guard as V21: an old fixture may not have `nodes` at all,
+            // and there is no path to materialise without it.
+            let has_nodes: bool = tx.query_row(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'nodes'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            if has_nodes {
+                tx.execute_batch(MIGRATION_V25)?;
+            }
+        }
+        if current < 26 {
+            // Same guard as V22: a fixture that never exercised the queue may
+            // not have `pending_op`, and altering an absent table is an error
+            // rather than a no-op.
+            // A fixture built from the current schema and then rewound to an
+            // older `schema_version` already has the column, and `ADD COLUMN`
+            // has no `IF NOT EXISTS`.
+            let has_column: bool = tx.query_row(
+                "SELECT COUNT(*) FROM pragma_table_info('pending_op') WHERE name = 'claimed_at'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            let has_ops: bool = tx.query_row(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'pending_op'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            if has_ops && !has_column {
+                tx.execute_batch(MIGRATION_V26)?;
+            }
+        }
+        if current < 27 {
+            // Same guards as V26, for the same two reasons: the table may not
+            // exist in an old fixture, and a fixture rewound from the current
+            // schema already has the column.
+            let has_column: bool = tx.query_row(
+                "SELECT COUNT(*) FROM pragma_table_info('pending_op') \
+                 WHERE name = 'access_deferred_since'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            let has_ops: bool = tx.query_row(
+                "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'pending_op'",
+                [],
+                |row| row.get::<_, i64>(0),
+            )? > 0;
+            if has_ops && !has_column {
+                tx.execute_batch(MIGRATION_V27)?;
+            }
+        }
         tx.execute(
             "INSERT INTO sync_state (key, value) VALUES ('schema_version', ?1)
              ON CONFLICT(key) DO UPDATE SET value = excluded.value",
@@ -583,3 +671,140 @@ SELECT n.rowid, n.name, paths.path
   FROM nodes n JOIN paths ON paths.rowid = n.rowid
  WHERE paths.trashed = 0;
 ";
+
+/// V22: indexes for the lookups that were scanning, each paired with the table
+/// it needs to exist for.
+///
+/// Each backs a query on a hot path that had no index to use:
+///
+/// - `sync_entry(remote_uid)` — the reconcile resolves a remote uid to its
+///   local entry once per node, per pass.
+/// - `local_files(scan_gen)` — the sweep that retires the previous generation
+///   after a local index scan.
+/// - `nodes(parent_uid) WHERE trashed = 0` — every listing, with the partial
+///   index keeping it to the live rows a listing actually wants.
+///
+/// Deliberately *not* here: an index on `pending_op(next_attempt_at, id)`.
+/// `next_due_op` orders by `id`, which is the rowid, so it already stops at the
+/// first due row; offering the planner a `next_attempt_at` index makes it scan
+/// that index and then sort the matches by `id` instead —
+/// `next_due_op_does_not_scale_with_queue_length` measures the regression.
+const MIGRATION_V22: [(&str, &str); 3] = [
+    (
+        "sync_entry",
+        "CREATE INDEX IF NOT EXISTS idx_sync_entry_remote_uid ON sync_entry(remote_uid);",
+    ),
+    (
+        "local_files",
+        "CREATE INDEX IF NOT EXISTS idx_local_files_scan_gen ON local_files(scan_gen);",
+    ),
+    (
+        "nodes",
+        "CREATE INDEX IF NOT EXISTS idx_nodes_parent_live ON nodes(parent_uid) WHERE trashed = 0;",
+    ),
+];
+
+/// V24: the revisions this daemon sealed itself, so restarting does not reopen
+/// the conflict hole B70 layer B closed.
+///
+/// `Core::own_sealed_revs` held these in memory: a queued write whose baseline
+/// names an earlier revision is a conflict *unless* the remote sits at a
+/// revision we sealed, in which case it is one writer stalling and resuming and
+/// must supersede rather than fork a `(sync-conflict)` copy. Losing the map at
+/// shutdown meant the first drain after a restart could not tell the two apart.
+///
+/// `sealed_at` is what bounds the table: entries older than the window in which
+/// a stall→resume is plausible are pruned on write.
+const MIGRATION_V24: &str = "
+CREATE TABLE IF NOT EXISTS own_sealed_rev (
+  uid         TEXT PRIMARY KEY,
+  revision_id TEXT NOT NULL,
+  sealed_at   INTEGER NOT NULL
+);
+";
+
+/// V25: `nodes.path` — a node's mountpoint-relative path, stored rather than
+/// walked.
+///
+/// `path_of` was a recursive CTE per node, and search ran it once per candidate
+/// (up to a thousand per keystroke) on top of the one the write path already ran
+/// per descendant of every folder upsert. The value only changes when a node's
+/// name or parent changes, which is exactly when the write path is already
+/// walking the subtree — so it belongs in a column.
+///
+/// The backfill is the same walk V21 used for `nodes_fts`, minus the trashed
+/// filter: a trashed node stays out of the index but still has a path. Roots get
+/// `''` (their name is the mount), and a node whose parent is merely uncached
+/// keeps its own name — both matching what `path_of` returned at runtime.
+///
+/// Materialised into a temp table rather than written by a correlated subquery,
+/// so the recursive walk runs once instead of once per row.
+const MIGRATION_V25: &str = "
+ALTER TABLE nodes ADD COLUMN path TEXT;
+
+CREATE TEMP TABLE node_paths (uid TEXT PRIMARY KEY, path TEXT NOT NULL);
+
+INSERT INTO node_paths (uid, path)
+WITH RECURSIVE paths(uid, path, depth) AS (
+  SELECT n.uid, CASE WHEN n.parent_uid IS NULL THEN '' ELSE n.name END, 0
+    FROM nodes n
+   WHERE n.parent_uid IS NULL
+      OR NOT EXISTS (SELECT 1 FROM nodes p WHERE p.uid = n.parent_uid)
+  UNION ALL
+  SELECT n.uid,
+         CASE WHEN paths.path = '' THEN n.name ELSE paths.path || '/' || n.name END,
+         paths.depth + 1
+    FROM nodes n JOIN paths ON n.parent_uid = paths.uid
+   WHERE paths.depth < 256
+)
+SELECT uid, path FROM paths;
+
+UPDATE nodes
+   SET path = (SELECT p.path FROM node_paths p WHERE p.uid = nodes.uid)
+ WHERE EXISTS (SELECT 1 FROM node_paths p WHERE p.uid = nodes.uid);
+
+DROP TABLE node_paths;
+";
+
+/// V26: `pending_op.claimed_at` — the drain's claim column.
+///
+/// The drain was one thread taking one op at a time, so a 10 GiB upload held
+/// every queued rename and small write behind it for as long as it ran. Several
+/// workers can share the queue only if a claim is visible to the others: this
+/// column is what makes "already being drained by someone" a fact about the row
+/// rather than about one thread's local variable.
+///
+/// `0` is unclaimed; anything else is the ms at which a worker took it. A
+/// claim is process-local state that outlives nothing — the single-writer lock
+/// means a non-zero value found at open belongs to a crashed run, and
+/// [`Db::clear_op_claims`](super::Db::clear_op_claims) drops the lot.
+///
+/// The index is partial so it costs nothing on the ordinary row: only the
+/// handful of claimed rows are in it, which is exactly the set the claim query
+/// has to subtract to keep two workers off one node.
+const MIGRATION_V26: &str = "
+ALTER TABLE pending_op ADD COLUMN claimed_at INTEGER NOT NULL DEFAULT 0;
+
+CREATE INDEX IF NOT EXISTS idx_pending_op_claimed
+    ON pending_op(uid) WHERE claimed_at <> 0;
+";
+
+/// Schema v27: when an op was first deferred by the local access check.
+///
+/// An access deferral consumes no attempt and records no error — the op has not
+/// failed remotely, it has not been sent at all. That is right for a permission
+/// change the user is about to undo, and wrong for a condition that never
+/// clears: the row is re-deferred every
+/// [`DRAIN_ACCESS_RECHECK`](../../../pdfs_fuse/drain/constant.DRAIN_ACCESS_RECHECK.html)
+/// forever while `attempts`, `last_error`, `failing` and the logs all stay
+/// empty, so a queue that can never drain is indistinguishable from a busy one.
+/// Found in the field as 28 revision ops holding 18 GiB of accepted writes for
+/// 30 days, none of them visible anywhere.
+///
+/// `0` means "not currently access-deferred"; anything else is the ms at which
+/// the first consecutive deferral happened, which is what
+/// [`Db::defer_op_for_access`](super::Db::defer_op_for_access) measures the
+/// deferral against.
+const MIGRATION_V27: &str = "
+ALTER TABLE pending_op ADD COLUMN access_deferred_since INTEGER NOT NULL DEFAULT 0;
+";

crates/pdfs-core/src/db/mod.rs

@@ -2,9 +2,17 @@
 //! inode bookkeeping, full-text search, content-cache LRU tracking, and pins.
 //!
 //! Only the daemon (`pdfs-fuse`) opens this for writes; the GUI and CLI reach
-//! the same data through the control socket. The connection is wrapped in a
-//! `Mutex` because the FUSE callbacks are synchronous and already serialize
-//! behind the `State` lock, so a connection pool would be overkill.
+//! the same data through the control socket. There is one write connection
+//! behind a `Mutex` — SQLite allows exactly one writer, and the single-instance
+//! `flock` makes sure it is ours — and a small pool of read-only connections
+//! behind [`Db::read`], which every `SELECT`-only method uses.
+//!
+//! The pool is what makes WAL worth having. The original justification for a
+//! single connection was that "FUSE callbacks already serialize behind the
+//! `State` lock"; that stopped being true when reads moved onto an 11-thread
+//! worker pool, and a lookup then queued behind whatever listing happened to be
+//! committing. Under WAL a reader never blocks the writer and vice versa, so the
+//! only thing left serialising them was this process's own mutex.
 //!
 //! This module is the P0 foundation: it opens the database, enables WAL, and
 //! applies the forward-only schema migrations. Write-through of nodes, the
@@ -18,8 +26,10 @@
 use parking_lot::Mutex;
 use rusqlite::Connection;
 
-use crate::Result;
-use std::path::Path;
+use crate::{Error, Result};
+use std::fs::File;
+use std::os::fd::AsRawFd;
+use std::path::{Path, PathBuf};
 
 mod activity;
 mod albums;
@@ -66,18 +76,208 @@ const WAL_SIZE_LIMIT: i64 = 64 * 1024 * 1024;
 /// genuinely stuck holder surfaces as an error rather than an apparent hang.
 const BUSY_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(5);
 
+/// Page cache, in KiB (applied negated, which is how SQLite reads a size in KiB
+/// rather than in pages). 32 MiB against a database that is typically tens of
+/// MiB: large enough to hold the working set of a listing-heavy session, small
+/// enough to be unremarkable in a desktop daemon's RSS.
+const CACHE_SIZE_KIB: i64 = 32 * 1024;
+
+/// Ceiling on the memory-mapped window over the database file. 256 MiB covers
+/// any realistic index while keeping the mapping bounded.
+const MMAP_SIZE: i64 = 256 * 1024 * 1024;
+
+/// WAL pages between automatic checkpoints (default 1000, i.e. ~4 MiB). Halved
+/// so the synchronous checkpoint a committing thread inherits stays short.
+const WAL_AUTOCHECKPOINT_PAGES: i64 = 500;
+
 /// Handle to the unified metadata database.
 ///
 /// Cheap to wrap in an `Arc`; clone the `Arc`, not this. All access goes through
 /// the inner `Mutex<Connection>`.
 pub struct Db {
     conn: Mutex<Connection>,
+    /// Read-only connections for `SELECT`-only methods. `None` for an in-memory
+    /// database, where a second connection would open a different, empty one —
+    /// those fall back to the write connection.
+    readers: Option<ReadPool>,
+    /// The single-writer lock, held open for as long as this handle lives.
+    ///
+    /// Never read; dropping the `File` is what releases the `flock`, so it has
+    /// to be owned here rather than by `open`'s stack frame. `None` for
+    /// in-memory databases, which nothing else can reach.
+    _single_writer: Option<File>,
+}
+
+/// Idle read-only connections, opened on demand and kept for reuse.
+///
+/// Not a bounded pool: the number in flight is bounded by the number of threads
+/// that can be inside a read at once (the FUSE worker lanes plus the control
+/// handlers), and blocking one of those on a permit would reintroduce exactly
+/// the queueing this exists to remove. Only the *idle* set is capped, so a burst
+/// does not leave a connection per thread parked forever.
+struct ReadPool {
+    path: PathBuf,
+    idle: Mutex<Vec<Connection>>,
+}
+
+/// How many idle read connections to keep. Enough that the steady state never
+/// reopens, small enough to be unremarkable.
+const MAX_IDLE_READERS: usize = 4;
+
+impl ReadPool {
+    fn take(&self) -> Option<Connection> {
+        if let Some(conn) = self.idle.lock().pop() {
+            return Some(conn);
+        }
+        match open_read_only(&self.path) {
+            Ok(conn) => Some(conn),
+            Err(error) => {
+                // Falling back to the write connection is correct, just slower,
+                // so this is a warning and not a failed read.
+                tracing::warn!(%error, "could not open a read-only connection");
+                None
+            }
+        }
+    }
+
+    fn put(&self, conn: Connection) {
+        let mut idle = self.idle.lock();
+        if idle.len() < MAX_IDLE_READERS {
+            idle.push(conn);
+        }
+    }
+}
+
+/// A connection to read through: one borrowed from [`ReadPool`], or the write
+/// connection when there is no pool (in-memory) or opening one failed.
+///
+/// Derefs to `Connection`, so a method converts by swapping `self.conn.lock()`
+/// for `self.read()` and nothing else.
+pub(super) struct Reader<'a> {
+    pooled: Option<Connection>,
+    writer: Option<parking_lot::MutexGuard<'a, Connection>>,
+    pool: Option<&'a ReadPool>,
+}
+
+impl std::ops::Deref for Reader<'_> {
+    type Target = Connection;
+
+    fn deref(&self) -> &Connection {
+        self.pooled
+            .as_ref()
+            .or(self.writer.as_deref())
+            .expect("a Reader always holds one of the two")
+    }
+}
+
+impl Drop for Reader<'_> {
+    fn drop(&mut self) {
+        if let (Some(conn), Some(pool)) = (self.pooled.take(), self.pool) {
+            pool.put(conn);
+        }
+    }
+}
+
+/// Open a second connection to the same file for reads only.
+///
+/// `SQLITE_OPEN_READ_ONLY` plus `query_only` is belt and braces: the flag stops
+/// a routing mistake at the file, the pragma stops it at the statement. The rest
+/// mirror [`Db::open_configured`] — a read connection has its own page cache,
+/// its own temp store and its own busy timeout, and gets nothing from the
+/// writer's.
+///
+/// `journal_mode` is deliberately absent: it is a property of the database file,
+/// which the writer already set to WAL, and a read-only connection cannot change
+/// it anyway.
+fn open_read_only(path: &Path) -> Result<Connection> {
+    use rusqlite::OpenFlags;
+    let conn = Connection::open_with_flags(
+        path,
+        OpenFlags::SQLITE_OPEN_READ_ONLY | OpenFlags::SQLITE_OPEN_NO_MUTEX,
+    )?;
+    conn.busy_timeout(BUSY_TIMEOUT)?;
+    conn.pragma_update(None, "cache_size", -CACHE_SIZE_KIB)?;
+    conn.pragma_update(None, "temp_store", "MEMORY")?;
+    conn.pragma_update(None, "mmap_size", MMAP_SIZE)?;
+    conn.pragma_update(None, "query_only", true)?;
+    Ok(conn)
 }
 
 impl Db {
     /// Open (creating if absent) the database at `path`, enable WAL, and bring
     /// the schema up to [`SCHEMA_VERSION`].
+    ///
+    /// Takes the single-instance lock first (see
+    /// [`acquire_single_writer_lock`]), and rebuilds from empty if what is on
+    /// disk turns out not to be a readable database (see [`is_corrupt`]).
     pub fn open(path: &Path) -> Result<Self> {
+        let single_writer = acquire_single_writer_lock(path)?;
+        let conn = match Self::open_configured(path) {
+            Ok(conn) => conn,
+            Err(e) if is_corrupt(&e) => {
+                // Every byte in here is derived from the account or from the
+                // content cache directory, so the recovery is to start over —
+                // loudly, and keeping the damaged file for a post-mortem rather
+                // than deleting evidence.
+                tracing::error!(
+                    db = %path.display(),
+                    error = %e,
+                    "cache database is corrupt — moving it aside and rebuilding \
+                     (metadata will be re-fetched; no user data is stored here)"
+                );
+                quarantine_corrupt_db(path)?;
+                Self::open_configured(path)?
+            }
+            Err(e) => return Err(e),
+        };
+        let db = Self {
+            conn: Mutex::new(conn),
+            // No pool for `:memory:`. A second connection to that name is a
+            // second, empty database rather than another view of this one, so
+            // every routed read would find a schema that does not exist.
+            readers: (!is_in_memory(path)).then(|| ReadPool {
+                path: path.to_path_buf(),
+                idle: Mutex::new(Vec::new()),
+            }),
+            _single_writer: single_writer,
+        };
+        // A drain claim names a worker thread in the process that took it, and
+        // the single-writer lock above says that process is not running. Any
+        // claim still on disk is therefore a crashed run's, and leaving it would
+        // hide those ops from every worker for good.
+        match db.clear_op_claims() {
+            Ok(0) => {}
+            Ok(n) => tracing::info!(ops = n, "released drain claims left by a previous run"),
+            Err(e) => tracing::warn!(error = %e, "clearing stale drain claims failed"),
+        }
+        Ok(db)
+    }
+
+    /// A connection for a `SELECT`-only method.
+    ///
+    /// Reads see the last committed state, which is the same thing they saw
+    /// through the write connection: every method here commits before it
+    /// returns, and nothing holds a transaction open across calls.
+    pub(super) fn read(&self) -> Reader<'_> {
+        if let Some(pool) = &self.readers
+            && let Some(conn) = pool.take()
+        {
+            return Reader {
+                pooled: Some(conn),
+                writer: None,
+                pool: Some(pool),
+            };
+        }
+        Reader {
+            pooled: None,
+            writer: Some(self.conn.lock()),
+            pool: None,
+        }
+    }
+
+    /// Open the file, apply every PRAGMA, and migrate. Split out of
+    /// [`open`](Self::open) so the corruption path can retry it verbatim.
+    fn open_configured(path: &Path) -> Result<Connection> {
         let conn = Connection::open(path)?;
         // WAL: readers never block the single writer. NORMAL sync is the
         // standard durability/throughput tradeoff for WAL.
@@ -102,12 +302,39 @@ impl Db {
         // (Bare SQLite does default to 0, which is where the belief that this
         // was unset came from. That default is not what we get.)
         conn.busy_timeout(BUSY_TIMEOUT)?;
+        // Page cache. The default is 2 MiB, against a schema whose hot table
+        // stores a JSON blob per node — a listing of a few thousand files
+        // evicts itself while it is being read.
+        conn.pragma_update(None, "cache_size", -CACHE_SIZE_KIB)?;
+        // Sorts and recursive CTEs materialise their intermediate results.
+        // Every `ORDER BY … COLLATE NOCASE` listing and every ancestor walk
+        // spills those to a file in `/tmp` by default; they are small and
+        // short-lived, so memory is both faster and less surprising.
+        conn.pragma_update(None, "temp_store", "MEMORY")?;
+        // Read the database through the page cache of the OS rather than
+        // `pread` per page. Bounded rather than unlimited so the daemon's RSS
+        // does not track the database size.
+        conn.pragma_update(None, "mmap_size", MMAP_SIZE)?;
+        // Checkpointing is synchronous in whichever thread happens to commit
+        // the transaction that crosses the threshold — often a FUSE callback.
+        // A smaller autocheckpoint makes that debt smaller and more frequent
+        // rather than rare and multi-second.
+        conn.pragma_update(None, "wal_autocheckpoint", WAL_AUTOCHECKPOINT_PAGES)?;
 
         let db = Self {
             conn: Mutex::new(conn),
+            readers: None,
+            _single_writer: None,
         };
         db.migrate()?;
-        Ok(db)
+        Ok(db.into_conn())
+    }
+
+    /// Unwrap the connection back out of a temporary handle. Only
+    /// [`open_configured`](Self::open_configured) uses it, to run `migrate`
+    /// (which is written against `&self`) before the real handle exists.
+    fn into_conn(self) -> Connection {
+        self.conn.into_inner()
     }
 
     /// Open an in-memory database. For tests.
@@ -120,6 +347,8 @@ impl Db {
         conn.pragma_update(None, "foreign_keys", "ON")?;
         let db = Self {
             conn: Mutex::new(conn),
+            readers: None,
+            _single_writer: None,
         };
         db.migrate()?;
         Ok(db)
@@ -133,5 +362,102 @@ impl Db {
     }
 }
 
+/// Whether `path` names the in-memory database rather than a file. Both
+/// spellings SQLite accepts are checked, because neither has a lock file or a
+/// corruption story.
+fn is_in_memory(path: &Path) -> bool {
+    let p = path.as_os_str();
+    p == ":memory:" || p.is_empty()
+}
+
+/// Take the single-instance write lock for the database at `path`.
+///
+/// The daemon is the only writer by design — the front-ends go through the
+/// control socket — but nothing stopped a hand-run `pdfs mount` from starting
+/// alongside the systemd unit, at which point two processes own the same inode
+/// space, the same content cache, and the same drain queue. SQLite's own
+/// locking makes each *statement* safe and does nothing about that.
+///
+/// `flock` on a sibling `.lock` file rather than on the database itself:
+/// SQLite uses POSIX record locks, so the two never interact, and a lock file
+/// left behind carries no state — the lock lives in the kernel and dies with
+/// the process, however it exits.
+fn acquire_single_writer_lock(path: &Path) -> Result<Option<File>> {
+    if is_in_memory(path) {
+        return Ok(None);
+    }
+    let lock_path = lock_path_for(path);
+    let file = std::fs::OpenOptions::new()
+        .create(true)
+        .read(true)
+        .truncate(false)
+        .write(true)
+        .open(&lock_path)?;
+    // SAFETY: `file` owns the descriptor and outlives the call.
+    let rc = unsafe { libc::flock(file.as_raw_fd(), libc::LOCK_EX | libc::LOCK_NB) };
+    if rc != 0 {
+        let e = std::io::Error::last_os_error();
+        if e.raw_os_error() == Some(libc::EWOULDBLOCK) {
+            return Err(Error::Other(format!(
+                "another Proton Drive daemon is already using {} — \
+                 stop it first (`systemctl --user stop proton-drive.service`)",
+                path.display()
+            )));
+        }
+        return Err(Error::Io(e));
+    }
+    Ok(Some(file))
+}
+
+fn lock_path_for(path: &Path) -> PathBuf {
+    let mut name = path.as_os_str().to_owned();
+    name.push(".lock");
+    PathBuf::from(name)
+}
+
+/// Whether `e` says the file is not a usable database, as opposed to any other
+/// failure. These are the only two conditions rebuilding fixes; a disk error, a
+/// permission problem or a newer-than-supported schema must all still fail.
+fn is_corrupt(e: &Error) -> bool {
+    matches!(
+        e,
+        Error::Db(rusqlite::Error::SqliteFailure(inner, _))
+            if inner.code == rusqlite::ErrorCode::DatabaseCorrupt
+                || inner.code == rusqlite::ErrorCode::NotADatabase
+    )
+}
+
+/// Move a damaged database (and its WAL sidecars) out of the way so the next
+/// open starts from empty.
+///
+/// Renamed rather than deleted: this is a cache, but it is also the only
+/// evidence of what went wrong, and it costs one file the user can remove.
+fn quarantine_corrupt_db(path: &Path) -> Result<()> {
+    let stamp = std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0);
+    for suffix in ["", "-wal", "-shm"] {
+        let mut from = path.as_os_str().to_owned();
+        from.push(suffix);
+        let from = PathBuf::from(from);
+        if !from.exists() {
+            continue;
+        }
+        let mut to = path.as_os_str().to_owned();
+        to.push(format!(".corrupt-{stamp}{suffix}"));
+        let to = PathBuf::from(to);
+        // A failure to rename the sidecars is survivable — SQLite discards a
+        // WAL whose database header no longer matches — but failing to move the
+        // database itself is not, because the retry would hit the same file.
+        match std::fs::rename(&from, &to) {
+            Ok(()) => tracing::warn!(from = %from.display(), to = %to.display(), "quarantined"),
+            Err(e) if suffix.is_empty() => return Err(Error::Io(e)),
+            Err(e) => tracing::warn!(file = %from.display(), error = %e, "could not quarantine"),
+        }
+    }
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests;

crates/pdfs-core/src/db/mounts.rs

@@ -74,7 +74,7 @@ impl Db {
 
     /// Every local location, with device fields joined from `sync_folder`.
     pub fn mount_list(&self) -> Result<Vec<MountSpec>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT
                  m.id,

crates/pdfs-core/src/db/nodes.rs

@@ -10,7 +10,9 @@ use crate::{Access, Result};
 use proton_drive_rs::proton_sdk::ids::{LinkId, NodeUid, VolumeId};
 use proton_drive_rs::{Node, NodeKind};
 
-use super::utils::{TRIGRAM_MIN, collect_pairs, like_escape, pair, path_of};
+use super::utils::{
+    HitRow, TRIGRAM_MIN, collect_hits, hit_row, join_path, like_escape, path_of, walk_path_of,
+};
 
 pub struct StoredNode {
     pub node: Node,
@@ -59,7 +61,7 @@ impl Db {
     /// flag. Synthetic shared listings use this to serve the last completed
     /// snapshot while offline, including after a live event expired its TTL.
     pub fn visible_children(&self, parent: &NodeUid) -> Result<Vec<Node>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT node_json FROM nodes
               WHERE parent_uid = ?1 AND trashed = 0 AND node_json IS NOT NULL
@@ -285,7 +287,7 @@ impl Db {
     /// Check if a folder node has any non-trashed children in the database.
     pub fn has_children(&self, parent_uid: &NodeUid) -> Result<bool> {
         let uid_str = parent_uid.to_string();
-        let conn = self.conn.lock();
+        let conn = self.read();
         let count: i64 = conn.query_row(
             "SELECT COUNT(*) FROM nodes WHERE parent_uid = ?1 AND trashed = 0",
             params![uid_str],
@@ -294,6 +296,25 @@ impl Db {
         Ok(count > 0)
     }
 
+    /// A node's mountpoint-relative path, or `None` when it is not cached.
+    ///
+    /// One indexed read of the stored `path`. Callers that need to relate many
+    /// nodes to a handful of roots — search decorating each hit with its
+    /// mountpoint — resolve the roots once through this and do the rest as
+    /// string arithmetic, instead of a `path_relative_to` walk per pair.
+    pub fn node_path(&self, uid: &str) -> Result<Option<String>> {
+        let conn = self.read();
+        let exists: Option<i64> = conn
+            .query_row("SELECT 1 FROM nodes WHERE uid = ?1", params![uid], |r| {
+                r.get(0)
+            })
+            .optional()?;
+        if exists.is_none() {
+            return Ok(None);
+        }
+        Ok(Some(path_of(&conn, uid)?))
+    }
+
     /// Resolve `uid` relative to an ancestor node.
     ///
     /// This is used when a remote subtree has its own local sync mount: the
@@ -302,7 +323,7 @@ impl Db {
     /// chain keeps callers from accidentally joining a Drive-wide path onto the
     /// wrong mountpoint. The ancestor itself resolves to the empty path.
     pub fn path_relative_to(&self, ancestor_uid: &str, uid: &str) -> Result<Option<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "WITH RECURSIVE anc(uid, parent_uid, name, depth) AS (
                SELECT uid, parent_uid, name, 0 FROM nodes WHERE uid = ?1
@@ -338,15 +359,15 @@ impl Db {
         if query.is_empty() {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
-        let rows: Vec<(String, String)> = if query.chars().count() < TRIGRAM_MIN {
+        let conn = self.read();
+        let rows: Vec<HitRow> = if query.chars().count() < TRIGRAM_MIN {
             let pat = format!("%{}%", like_escape(query));
             let mut stmt = conn.prepare(
-                "SELECT node_json, uid FROM nodes
+                "SELECT node_json, uid, path FROM nodes
                  WHERE name LIKE ?1 ESCAPE '\\' AND trashed = 0 AND node_json IS NOT NULL
                  ORDER BY name LIMIT ?2",
             )?;
-            collect_pairs(stmt.query_map(params![pat, limit as i64], pair)?)?
+            collect_hits(stmt.query_map(params![pat, limit as i64], hit_row)?)?
         } else {
             // Escape double quotes and quote each term, then combine with AND so
             // all terms must match but can appear in any order or position.
@@ -356,18 +377,21 @@ impl Db {
                 .collect::<Vec<_>>()
                 .join(" AND ");
             let mut stmt = conn.prepare(
-                "SELECT n.node_json, n.uid
+                "SELECT n.node_json, n.uid, n.path
                  FROM nodes_fts f JOIN nodes n ON n.rowid = f.rowid
                  WHERE f.name MATCH ?1 AND n.trashed = 0 AND n.node_json IS NOT NULL
                  ORDER BY f.rank LIMIT ?2",
             )?;
-            collect_pairs(stmt.query_map(params![phrase, limit as i64], pair)?)?
+            collect_hits(stmt.query_map(params![phrase, limit as i64], hit_row)?)?
         };
 
         let mut hits = Vec::with_capacity(rows.len());
-        for (json, uid) in rows {
+        for (json, uid, path) in rows {
             let node: Node = serde_json::from_str(&json)?;
-            let path = path_of(&conn, &uid)?;
+            let path = match path {
+                Some(path) => path,
+                None => walk_path_of(&conn, &uid)?,
+            };
             hits.push(SearchHit { node, path });
         }
         Ok(hits)
@@ -382,26 +406,26 @@ impl Db {
         if query.is_empty() || limit == 0 {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
+        let conn = self.read();
         let terms = candidate_trigrams(query);
         let lane_limit = limit.div_ceil(2);
-        let mut rows: Vec<(String, String)> = if terms.is_empty() {
+        let mut rows: Vec<HitRow> = if terms.is_empty() {
             let pat = format!("%{}%", like_escape(query));
             let mut stmt = conn.prepare(
-                "SELECT node_json, uid FROM nodes
+                "SELECT node_json, uid, path FROM nodes
                  WHERE name LIKE ?1 ESCAPE '\\' AND trashed = 0 AND node_json IS NOT NULL
                  ORDER BY name LIMIT ?2",
             )?;
-            collect_pairs(stmt.query_map(params![pat, limit as i64], pair)?)?
+            collect_hits(stmt.query_map(params![pat, limit as i64], hit_row)?)?
         } else {
             let expression = terms.join(" OR ");
             let mut stmt = conn.prepare(
-                "SELECT n.node_json, n.uid
+                "SELECT n.node_json, n.uid, n.path
                    FROM nodes_fts f JOIN nodes n ON n.rowid = f.rowid
                   WHERE nodes_fts MATCH ?1 AND n.trashed = 0 AND n.node_json IS NOT NULL
                   ORDER BY f.rank LIMIT ?2",
             )?;
-            collect_pairs(stmt.query_map(params![expression, lane_limit as i64], pair)?)?
+            collect_hits(stmt.query_map(params![expression, lane_limit as i64], hit_row)?)?
         };
         // Trigram rank does not know that a name *starts* with the query, and a
         // short substitution can destroy every trigram outright (`vedio` vs
@@ -411,7 +435,7 @@ impl Db {
             && let Some(first) = query.chars().next()
         {
             let mut stmt = conn.prepare(
-                "SELECT node_json, uid FROM nodes
+                "SELECT node_json, uid, path FROM nodes
                  WHERE name COLLATE NOCASE LIKE ?1 ESCAPE '\\'
                    AND trashed = 0 AND node_json IS NOT NULL
                  ORDER BY name COLLATE NOCASE LIMIT ?2",
@@ -421,23 +445,26 @@ impl Db {
                 format!("{}%", like_escape(query)),
                 format!("{}%", like_escape(&first.to_string())),
             ] {
-                extra.extend(collect_pairs(
-                    stmt.query_map(params![pattern, lane_limit as i64], pair)?,
+                extra.extend(collect_hits(
+                    stmt.query_map(params![pattern, lane_limit as i64], hit_row)?,
                 )?);
             }
             let trigram = std::mem::take(&mut rows);
             for row in extra.into_iter().chain(trigram) {
-                if !rows.iter().any(|(_, uid)| uid == &row.1) {
+                if !rows.iter().any(|(_, uid, _)| uid == &row.1) {
                     rows.push(row);
                 }
             }
             rows.truncate(limit);
         }
         rows.into_iter()
-            .map(|(json, uid)| {
+            .map(|(json, uid, path)| {
                 Ok(SearchHit {
                     node: serde_json::from_str(&json)?,
-                    path: path_of(&conn, &uid)?,
+                    path: match path {
+                        Some(path) => path,
+                        None => walk_path_of(&conn, &uid)?,
+                    },
                 })
             })
             .collect()
@@ -457,7 +484,7 @@ impl Db {
 
     /// Load every persisted node for cold-start hydration of the `State` maps.
     pub fn load_all(&self) -> Result<Vec<StoredNode>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt =
             conn.prepare("SELECT node_json, listed FROM nodes WHERE node_json IS NOT NULL")?;
         let rows = stmt.query_map([], |row| {
@@ -478,7 +505,7 @@ impl Db {
     /// when the API is unreachable, so the mount can still serve the cached tree
     /// (offline.md Phase 1).
     pub fn node_by_uid(&self, uid: &str) -> Result<Option<Node>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let json: Option<String> = conn
             .query_row(
                 "SELECT node_json FROM nodes WHERE uid = ?1 AND node_json IS NOT NULL",
@@ -496,7 +523,7 @@ impl Db {
     /// The daemon resumes from this on restart instead of reseeding to the
     /// server head, so changes made while unmounted are still applied (P2).
     pub fn children_if_listed(&self, parent: &NodeUid) -> Result<Option<Vec<Node>>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let listed: Option<i64> = conn
             .query_row(
                 "SELECT listed FROM nodes WHERE uid = ?1",
@@ -529,13 +556,57 @@ impl Db {
     // in sync on every store/read/evict, so it is authoritative for eviction.
 }
 
+/// What a node's row looked like before this upsert, for the fields that decide
+/// whether the subtree below it has to be touched at all.
+struct PriorRow {
+    parent_uid: Option<String>,
+    name: String,
+    path: Option<String>,
+    trashed: bool,
+}
+
 fn upsert_node_tx(tx: &Transaction<'_>, node: &Node) -> Result<()> {
     let json = serde_json::to_string(node)?;
     let uid = node.uid.to_string();
+    let parent_uid = node.parent_uid.as_ref().map(|u| u.to_string());
+
+    let prior: Option<PriorRow> = tx
+        .query_row(
+            "SELECT parent_uid, name, path, trashed FROM nodes WHERE uid = ?1",
+            params![uid],
+            |row| {
+                Ok(PriorRow {
+                    parent_uid: row.get(0)?,
+                    name: row.get(1)?,
+                    path: row.get(2)?,
+                    trashed: row.get::<_, i64>(3)? != 0,
+                })
+            },
+        )
+        .optional()?;
+
+    // A node's path is its parent's path plus its own name — one indexed lookup,
+    // where resolving it from scratch is a recursive walk to the root. A node
+    // whose parent is not cached (a device folder's root lives in `device`, never
+    // in `nodes`) starts a path of its own, which is what the walk did too.
+    let path = match &parent_uid {
+        None => String::new(),
+        Some(parent) => {
+            let parent_path: Option<Option<String>> = tx
+                .query_row(
+                    "SELECT path FROM nodes WHERE uid = ?1",
+                    params![parent],
+                    |row| row.get(0),
+                )
+                .optional()?;
+            join_path(parent_path.flatten().as_deref().unwrap_or(""), &node.name)
+        }
+    };
+
     tx.execute(
         "INSERT INTO nodes
-           (uid, parent_uid, name, is_dir, size, mtime, trashed, node_json)
-         VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8)
+           (uid, parent_uid, name, is_dir, size, mtime, trashed, node_json, path)
+         VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9)
          ON CONFLICT(uid) DO UPDATE SET
            parent_uid = excluded.parent_uid,
            name       = excluded.name,
@@ -543,65 +614,116 @@ fn upsert_node_tx(tx: &Transaction<'_>, node: &Node) -> Result<()> {
            size       = excluded.size,
            mtime      = excluded.mtime,
            trashed    = excluded.trashed,
-           node_json  = excluded.node_json",
+           node_json  = excluded.node_json,
+           path       = excluded.path",
         params![
             uid,
-            node.parent_uid.as_ref().map(|u| u.to_string()),
+            parent_uid,
             node.name,
             node.is_folder() as i64,
             node_size(node),
             node.modification_time,
             node.trashed as i64,
             json,
+            path,
         ],
     )?;
-    // FTS5 has no UPSERT. Folder moves also change every descendant path, so
-    // refresh the whole affected subtree while keeping rowid deletes indexed.
     let rowid: i64 = tx.query_row(
         "SELECT rowid FROM nodes WHERE uid = ?1",
         params![uid],
         |row| row.get(0),
     )?;
-    let mut affected = vec![rowid];
-    if node.is_folder() {
-        // The depth cap is a liveness guard, not a policy: `UNION ALL` over a
-        // parent cycle (a's parent is b, b's parent is a — corrupt data, but the
-        // API can hand it to us) never terminates, and this runs inside the
-        // write transaction that holds the daemon's only SQLite connection.
-        // Real Drive trees are nowhere near this deep.
-        let mut stmt = tx.prepare(
-            "WITH RECURSIVE descendants(rowid, uid, depth) AS (
-               SELECT rowid, uid, 0 FROM nodes WHERE parent_uid = ?1
-               UNION ALL
-               SELECT n.rowid, n.uid, d.depth + 1 FROM nodes n
-                 JOIN descendants d ON n.parent_uid = d.uid
-                WHERE d.depth < 256
-             ) SELECT rowid FROM descendants",
-        )?;
-        for descendant in stmt.query_map(params![uid], |row| row.get(0))? {
-            affected.push(descendant?);
-        }
-    }
-    for affected_rowid in affected {
+
+    // FTS5 has no UPSERT, so the node's own row is always replaced.
+    let indexable = node_is_indexable_tx(tx, &uid)?;
+    tx.execute("DELETE FROM nodes_fts WHERE rowid = ?1", params![rowid])?;
+    if indexable {
         tx.execute(
-            "DELETE FROM nodes_fts WHERE rowid = ?1",
-            params![affected_rowid],
+            "INSERT INTO nodes_fts (rowid, name, path) VALUES (?1, ?2, ?3)",
+            params![rowid, node.name, path],
         )?;
-        let indexed: Option<(String, String)> = tx
-            .query_row(
-                "SELECT name, uid FROM nodes WHERE rowid = ?1",
-                params![affected_rowid],
-                |row| Ok((row.get(0)?, row.get(1)?)),
-            )
-            .optional()?;
-        if let Some((name, indexed_uid)) = indexed
-            && node_is_indexable_tx(tx, &indexed_uid)?
-        {
-            let path = path_of(tx, &indexed_uid)?;
-            tx.execute(
-                "INSERT INTO nodes_fts (rowid, name, path) VALUES (?1, ?2, ?3)",
-                params![affected_rowid, name, path],
-            )?;
+    }
+
+    // Everything below only matters when the subtree's paths or reachability
+    // actually moved. Re-listing a folder upserts it unchanged, and paying for a
+    // full descendant walk on each of those was most of what a large tree spent
+    // its time on.
+    let subtree_moved = prior.is_none_or(|prior| {
+        prior.parent_uid != parent_uid
+            || prior.name != node.name
+            || prior.path.as_deref() != Some(path.as_str())
+            || prior.trashed != node.trashed
+    });
+    if node.is_folder() && subtree_moved {
+        reindex_subtree_tx(tx, &uid, &path, indexable)?;
+    }
+    Ok(())
+}
+
+/// Rewrite every descendant's stored path and search-index row after `folder`
+/// moved, was renamed, or changed reachability.
+///
+/// One recursive walk carries both the new path and whether anything on the way
+/// down is trashed, so each descendant costs two trivial statements instead of a
+/// `path_of` walk plus an ancestor walk of its own. `folder_indexable` is the
+/// answer `node_is_indexable_tx` gave for the folder itself: what is above the
+/// subtree is the same for every node in it.
+///
+/// The depth cap is a liveness guard, not a policy: `UNION ALL` over a parent
+/// cycle (a's parent is b, b's parent is a — corrupt data, but the API can hand
+/// it to us) never terminates, and this runs inside the write transaction that
+/// holds the daemon's only SQLite connection. Real Drive trees are nowhere near
+/// this deep.
+fn reindex_subtree_tx(
+    tx: &Transaction<'_>,
+    folder: &str,
+    folder_path: &str,
+    folder_indexable: bool,
+) -> Result<()> {
+    let descendants: Vec<(i64, String, bool)> = {
+        let mut stmt = tx.prepare(
+            "WITH RECURSIVE sub(rowid, uid, name, path, blocked, depth) AS (
+               SELECT n.rowid, n.uid, n.name,
+                      CASE WHEN ?2 = '' THEN n.name ELSE ?2 || '/' || n.name END,
+                      n.trashed, 0
+                 FROM nodes n WHERE n.parent_uid = ?1
+               UNION ALL
+               SELECT n.rowid, n.uid, n.name,
+                      sub.path || '/' || n.name,
+                      MAX(sub.blocked, n.trashed), sub.depth + 1
+                 FROM nodes n JOIN sub ON n.parent_uid = sub.uid
+                WHERE sub.depth < 256
+             )
+             SELECT rowid, path, blocked FROM sub",
+        )?;
+        let rows = stmt.query_map(params![folder, folder_path], |row| {
+            Ok((
+                row.get::<_, i64>(0)?,
+                row.get::<_, String>(1)?,
+                row.get::<_, i64>(2)? != 0,
+            ))
+        })?;
+        let mut out = Vec::new();
+        for row in rows {
+            out.push(row?);
+        }
+        out
+    };
+    if descendants.is_empty() {
+        return Ok(());
+    }
+
+    let mut set_path = tx.prepare("UPDATE nodes SET path = ?2 WHERE rowid = ?1")?;
+    let mut unindex = tx.prepare("DELETE FROM nodes_fts WHERE rowid = ?1")?;
+    let mut index = tx.prepare(
+        "INSERT INTO nodes_fts (rowid, name, path)
+         SELECT rowid, name, ?2 FROM nodes WHERE rowid = ?1",
+    )?;
+    for (rowid, path, blocked) in descendants {
+        set_path.execute(params![rowid, path])?;
+        unindex.execute(params![rowid])?;
+        if folder_indexable && !blocked {
+            index.execute(params![rowid, path])?;
         }
     }
     Ok(())

crates/pdfs-core/src/db/ops.rs

@@ -119,15 +119,32 @@ pub struct PendingOp {
 }
 
 /// How much work the queue owes the server, by whether it carries bytes.
-#[derive(Debug, Clone, Copy, Default)]
+#[derive(Debug, Clone, Default)]
 pub struct PendingCounts {
     /// Queued `create`/`revision` ops: files whose content is not on the remote.
     pub uploads: i64,
     /// Queued `mkdir`/`rename`/`trash` ops: metadata the remote has not been
     /// told about.
     pub changes: i64,
+    /// Ops parked at [`PARK_UNTIL`]: a transient file's bytes, held back until
+    /// the rename that gives it its finished name. They are queued but will
+    /// never drain on their own, so counting them as "waiting to upload" would
+    /// have the front-end promise progress that is not coming.
+    pub parked: i64,
+    /// Ops that have failed at least [`FAILING_ATTEMPTS`] times. A permanently
+    /// failing op never wedges the queue — the backoff sees to that — which is
+    /// exactly why it needs surfacing: it retries forever, invisibly.
+    pub failing: i64,
+    /// The most recent error text from a failing op, for the front-end to show
+    /// instead of the bare count.
+    pub last_error: Option<String>,
 }
 
+/// Failed attempts after which an op is reported as failing rather than merely
+/// retrying. Six is where the backoff has reached its ceiling, so an op past it
+/// is retrying at the slowest rate it ever will and has been for minutes.
+pub const FAILING_ATTEMPTS: i64 = 6;
+
 /// The outcome of folding freshly written bytes into a queued create.
 #[derive(Debug, Clone)]
 pub struct AttachedBlob {
@@ -140,7 +157,7 @@ pub struct AttachedBlob {
 impl Db {
     /// Whether a specific desired-state operation is still queued for a node.
     pub fn has_pending_op(&self, uid: &str, kind: &str) -> Result<bool> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         conn.query_row(
             "SELECT EXISTS(
                 SELECT 1 FROM pending_op WHERE uid = ?1 AND kind = ?2
@@ -153,7 +170,7 @@ impl Db {
 
     /// Read an existing operation's metadata before a superseding enqueue.
     pub fn pending_op_meta(&self, uid: &str, kind: &str) -> Result<Option<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         conn.query_row(
             "SELECT meta_json FROM pending_op WHERE uid = ?1 AND kind = ?2",
             params![uid, kind],
@@ -352,7 +369,7 @@ impl Db {
 
     /// Check if a queued create or mkdir op exists for `uid`.
     pub fn has_create_op(&self, uid: &str) -> Result<bool> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let count: i64 = conn.query_row(
             "SELECT COUNT(*) FROM pending_op WHERE uid = ?1 AND kind IN (?2, ?3)",
             params![uid, OP_CREATE, OP_MKDIR],
@@ -368,7 +385,7 @@ impl Db {
     /// bytes are still owed an upload, and dropping it would discard the staged
     /// blob that holds them (`docs/BUGS.md` B71).
     pub fn has_any_op(&self, uid: &str) -> Result<bool> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let count: i64 = conn.query_row(
             "SELECT COUNT(*) FROM pending_op WHERE uid = ?1",
             params![uid],
@@ -464,7 +481,7 @@ impl Db {
     /// sort, and no separate index to maintain. The common case (something is
     /// due) exits on the first row.
     pub fn next_due_op(&self, now: i64) -> Result<Option<PendingOp>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare_cached(&format!(
             "SELECT id, kind, uid, parent_uid, name, blob_path, meta_json, created_at,
                     attempts, last_error, next_attempt_at
@@ -495,17 +512,115 @@ impl Db {
         Ok(op)
     }
 
+    /// Take the oldest due op for this worker, marking it claimed so no other
+    /// drain worker takes it, or `None` when there is nothing to do.
+    ///
+    /// The drain is several workers over one queue, which is only safe because
+    /// the claim is a column: a row is offered to exactly one worker, and the
+    /// worker that took it is the only one that may retire, defer or fail it.
+    ///
+    /// Two exclusions on top of [`next_due_op`](Self::next_due_op)'s:
+    ///
+    /// * `claimed_at = 0` — not already in flight somewhere else.
+    /// * no *other* claimed op shares this `uid`. Ordering only has to hold per
+    ///   node, but there it has to hold absolutely: a queued rename and a queued
+    ///   revision of one file must land in the order they were made, and two
+    ///   workers on one node would also race over its staged blob and its
+    ///   pending entry. `ORDER BY id` plus this exclusion gives per-uid serial
+    ///   order for free — the older op is claimed first and the newer one is
+    ///   invisible until it retires.
+    ///
+    /// Select and mark are one transaction, so two workers cannot both see the
+    /// row as unclaimed.
+    pub fn claim_next_due_op(&self, now: i64) -> Result<Option<PendingOp>> {
+        let mut conn = self.conn.lock();
+        let tx = conn.transaction()?;
+        let op = {
+            let mut stmt = tx.prepare_cached(&format!(
+                "SELECT id, kind, uid, parent_uid, name, blob_path, meta_json, created_at,
+                        attempts, last_error, next_attempt_at
+                 FROM pending_op
+                 WHERE next_attempt_at <= ?1
+                   AND claimed_at = 0
+                   AND (parent_uid IS NULL OR substr(parent_uid, 1, {n}) <> '{v}~')
+                   AND uid NOT IN (SELECT uid FROM pending_op WHERE claimed_at <> 0)
+                 ORDER BY id LIMIT 1",
+                v = LOCAL_VOLUME,
+                n = LOCAL_VOLUME.len() + 1,
+            ))?;
+            stmt.query_row(params![now], |r| {
+                Ok(PendingOp {
+                    id: r.get(0)?,
+                    kind: r.get(1)?,
+                    uid: r.get(2)?,
+                    parent_uid: r.get(3)?,
+                    name: r.get(4)?,
+                    blob_path: r.get(5)?,
+                    meta_json: r.get(6)?,
+                    created_at: r.get(7)?,
+                    attempts: r.get(8)?,
+                    last_error: r.get(9)?,
+                    next_attempt_at: r.get(10)?,
+                })
+            })
+            .optional()?
+        };
+        if let Some(op) = op.as_ref() {
+            tx.execute(
+                "UPDATE pending_op SET claimed_at = ?2 WHERE id = ?1",
+                params![op.id, now.max(1)],
+            )?;
+        }
+        tx.commit()?;
+        Ok(op)
+    }
+
+    /// Give a claimed op back to the queue, whatever became of the attempt.
+    ///
+    /// Called on every path out of a drain attempt rather than only the failing
+    /// ones: a handler that retires its own row leaves nothing to release (the
+    /// `UPDATE` matches nothing), and a handler that returns without retiring
+    /// *and* without failing would otherwise leave the row claimed by a worker
+    /// that has moved on — invisible to every worker, forever.
+    pub fn release_op_claim(&self, id: i64) -> Result<()> {
+        let conn = self.conn.lock();
+        conn.execute(
+            "UPDATE pending_op SET claimed_at = 0 WHERE id = ?1",
+            params![id],
+        )?;
+        Ok(())
+    }
+
+    /// Drop every claim in the table.
+    ///
+    /// A claim describes a worker in *this* process, so one surviving on disk
+    /// means the previous run died holding it. The single-writer lock
+    /// ([`Db::open`]) is what makes that inference safe: nobody else can be
+    /// draining this queue while we are opening it.
+    pub fn clear_op_claims(&self) -> Result<usize> {
+        let conn = self.conn.lock();
+        Ok(conn.execute(
+            "UPDATE pending_op SET claimed_at = 0 WHERE claimed_at <> 0",
+            [],
+        )?)
+    }
+
     /// The earliest `next_attempt_at` among all queued ops, or `None` if the
     /// queue is empty. Used by the drain loop to sleep exactly until the next
     /// debounced or backed-off op becomes eligible rather than waiting the full
     /// idle-poll interval.
+    ///
+    /// Claimed ops are excluded: another worker is already on them, so they are
+    /// not work this one is waiting for, and counting them would have an idle
+    /// worker spin on a row it cannot have.
     pub fn earliest_due_at(&self) -> Result<Option<i64>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let ts: Option<i64> = conn
             .query_row(
                 &format!(
                     "SELECT MIN(next_attempt_at) FROM pending_op \
-                     WHERE parent_uid IS NULL OR substr(parent_uid, 1, {n}) <> '{v}~'",
+                     WHERE claimed_at = 0 \
+                       AND (parent_uid IS NULL OR substr(parent_uid, 1, {n}) <> '{v}~')",
                     v = LOCAL_VOLUME,
                     n = LOCAL_VOLUME.len() + 1,
                 ),
@@ -520,7 +635,7 @@ impl Db {
     /// Every queued op, oldest first. For status read-outs and the CLI — the
     /// drain wants [`next_due_op`](Self::next_due_op) instead.
     pub fn pending_ops(&self) -> Result<Vec<PendingOp>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT id, kind, uid, parent_uid, name, blob_path, meta_json, created_at,
                     attempts, last_error, next_attempt_at
@@ -553,7 +668,7 @@ impl Db {
     /// mount owes the server, but it carries no bytes and reporting it as an
     /// upload is simply untrue.
     pub fn pending_op_counts(&self) -> Result<PendingCounts> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let uploads = conn.query_row(
             "SELECT COUNT(*) FROM pending_op WHERE kind IN (?1, ?2)",
             params![OP_REVISION, OP_CREATE],
@@ -564,7 +679,33 @@ impl Db {
             params![OP_REVISION, OP_CREATE],
             |r| r.get(0),
         )?;
-        Ok(PendingCounts { uploads, changes })
+        let parked = conn.query_row(
+            "SELECT COUNT(*) FROM pending_op WHERE next_attempt_at >= ?1",
+            params![PARK_UNTIL],
+            |r| r.get(0),
+        )?;
+        let failing = conn.query_row(
+            "SELECT COUNT(*) FROM pending_op WHERE attempts >= ?1",
+            params![FAILING_ATTEMPTS],
+            |r| r.get(0),
+        )?;
+        let last_error = conn
+            .query_row(
+                "SELECT last_error FROM pending_op
+                  WHERE attempts >= ?1 AND last_error IS NOT NULL
+                  ORDER BY attempts DESC LIMIT 1",
+                params![FAILING_ATTEMPTS],
+                |r| r.get::<_, Option<String>>(0),
+            )
+            .optional()?
+            .flatten();
+        Ok(PendingCounts {
+            uploads,
+            changes,
+            parked,
+            failing,
+            last_error,
+        })
     }
 
     /// Nothing bearing this volume may be handed to the API — it would 404. The
@@ -580,11 +721,16 @@ impl Db {
 
     /// Record a failed attempt and when to next try. Leaves the row in place —
     /// the staged bytes are still the only copy of the user's write.
+    ///
+    /// Clears any access-deferral window: whatever the op was waiting for has
+    /// now been reported, so the next uncleared deferral starts its own window
+    /// rather than escalating again on the following recheck.
     pub fn record_op_failure(&self, id: i64, error: &str, next_attempt_at: i64) -> Result<()> {
         let conn = self.conn.lock();
         conn.execute(
             "UPDATE pending_op
-             SET attempts = attempts + 1, last_error = ?2, next_attempt_at = ?3
+             SET attempts = attempts + 1, last_error = ?2, next_attempt_at = ?3,
+                 access_deferred_since = 0
              WHERE id = ?1",
             params![id, error, next_attempt_at],
         )?;
@@ -609,4 +755,132 @@ impl Db {
         )?;
         Ok(())
     }
+
+    /// Defer an op the local access check refused, and report how long it has
+    /// been refused for.
+    ///
+    /// Identical to [`defer_op_without_attempt`](Self::defer_op_without_attempt)
+    /// on the retry clock, plus the one thing that was missing: the first
+    /// deferral of a run stamps `access_deferred_since`, and every deferral
+    /// returns that stamp. A caller can therefore tell a deferral that is one
+    /// recheck old from one that has been repeating unchanged for days, which is
+    /// the difference between a permission the user is mid-way through changing
+    /// and a queue that will never drain.
+    ///
+    /// Returns the ms at which the current run of deferrals began — `now` on the
+    /// first one.
+    pub fn defer_op_for_access(&self, id: i64, now: i64, next_attempt_at: i64) -> Result<i64> {
+        let conn = self.conn.lock();
+        conn.execute(
+            "UPDATE pending_op
+             SET next_attempt_at = CASE
+                 WHEN next_attempt_at >= ?4 THEN next_attempt_at
+                 ELSE ?3
+             END,
+             access_deferred_since = CASE
+                 WHEN access_deferred_since = 0 THEN ?2
+                 ELSE access_deferred_since
+             END
+             WHERE id = ?1",
+            params![id, now, next_attempt_at, PARK_UNTIL],
+        )?;
+        let since = conn
+            .query_row(
+                "SELECT access_deferred_since FROM pending_op WHERE id = ?1",
+                params![id],
+                |r| r.get::<_, i64>(0),
+            )
+            .optional()?;
+        // A row that drained out from under this call leaves nothing to measure;
+        // reporting `now` says "this run just started", which is the reading that
+        // never escalates something that no longer exists.
+        Ok(since.filter(|since| *since != 0).unwrap_or(now))
+    }
+
+    /// Forget an op's access-deferral window, after an attempt the access check
+    /// let through. The op is no longer blocked, so a later deferral is the
+    /// start of a new run rather than the continuation of an old one.
+    pub fn clear_op_access_deferral(&self, id: i64) -> Result<()> {
+        let conn = self.conn.lock();
+        conn.execute(
+            "UPDATE pending_op SET access_deferred_since = 0
+             WHERE id = ?1 AND access_deferred_since <> 0",
+            params![id],
+        )?;
+        Ok(())
+    }
+
+    /// Every staged blob the queue still refers to.
+    ///
+    /// The startup staging reconcile subtracts this from what is actually in
+    /// `staging/`: the difference is bytes the kernel was told were written and
+    /// that nothing is going to upload.
+    pub fn op_blob_paths(&self) -> Result<std::collections::HashSet<String>> {
+        let conn = self.read();
+        let mut stmt =
+            conn.prepare("SELECT blob_path FROM pending_op WHERE blob_path IS NOT NULL")?;
+        let rows = stmt.query_map([], |r| r.get::<_, String>(0))?;
+        let mut out = std::collections::HashSet::new();
+        for row in rows {
+            out.insert(row?);
+        }
+        Ok(out)
+    }
+
+    /// Record that this daemon sealed `revision_id` on `uid`, and prune entries
+    /// past [`OWN_SEALED_TTL_MS`].
+    ///
+    /// One row per node, overwritten as the node advances. See
+    /// [`own_sealed_rev`](Self::own_sealed_rev) for what reads it.
+    pub fn set_own_sealed_rev(&self, uid: &str, revision_id: &str, now_ms: i64) -> Result<()> {
+        let conn = self.conn.lock();
+        conn.execute(
+            "INSERT INTO own_sealed_rev (uid, revision_id, sealed_at)
+             VALUES (?1, ?2, ?3)
+             ON CONFLICT(uid) DO UPDATE SET
+               revision_id = excluded.revision_id,
+               sealed_at   = excluded.sealed_at",
+            params![uid, revision_id, now_ms],
+        )?;
+        conn.execute(
+            "DELETE FROM own_sealed_rev WHERE sealed_at < ?1",
+            params![now_ms - OWN_SEALED_TTL_MS],
+        )?;
+        Ok(())
+    }
+
+    /// The last revision of `uid` this daemon sealed itself, if it is still
+    /// within [`OWN_SEALED_TTL_MS`].
+    ///
+    /// The drain asks before treating a moved-on remote as a conflict: a
+    /// revision we sealed means one writer stalled and resumed, which
+    /// supersedes rather than forking a `(sync-conflict)` copy (docs/BUGS.md
+    /// B70 layer B).
+    pub fn own_sealed_rev(&self, uid: &str, now_ms: i64) -> Result<Option<String>> {
+        let conn = self.read();
+        Ok(conn
+            .query_row(
+                "SELECT revision_id FROM own_sealed_rev
+                  WHERE uid = ?1 AND sealed_at >= ?2",
+                params![uid, now_ms - OWN_SEALED_TTL_MS],
+                |r| r.get(0),
+            )
+            .optional()?)
+    }
+
+    /// Forget what we sealed on `uid` — it has been trashed, so there is no
+    /// revision left to chain onto.
+    pub fn clear_own_sealed_rev(&self, uid: &str) -> Result<()> {
+        let conn = self.conn.lock();
+        conn.execute("DELETE FROM own_sealed_rev WHERE uid = ?1", params![uid])?;
+        Ok(())
+    }
 }
+
+/// How long a revision this daemon sealed stays recognisable as its own.
+///
+/// Bounds a table that would otherwise grow with every write for the life of
+/// the install. A week is far longer than any stall→resume — a browser download
+/// that pauses and continues, an editor that reopens its handle — and short
+/// enough that the table tracks recent work rather than accumulating history.
+pub const OWN_SEALED_TTL_MS: i64 = 7 * 24 * 60 * 60 * 1000;

crates/pdfs-core/src/db/photos.rs

@@ -115,7 +115,7 @@ impl Db {
         range: Option<(i64, i64)>,
         favorites: bool,
     ) -> Result<Vec<StoredPhoto>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         // Built up so any combination of the optional filters is one indexed
         // query rather than a statement per case.
         let mut sql = String::from(
@@ -174,7 +174,7 @@ impl Db {
         &self,
         kind: Option<crate::control::PhotoKind>,
     ) -> Result<Vec<(i32, i32, usize)>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let (filter, binds): (&str, Vec<i64>) = match kind {
             Some(k) => (" WHERE kind = ?1", vec![k.as_i64()]),
             None => ("", Vec::new()),
@@ -201,7 +201,7 @@ impl Db {
     /// Per-tab counts for the Photos page subtitle: `(photos, videos, raw)`.
     pub fn photos_counts(&self) -> Result<(usize, usize, usize)> {
         use crate::control::PhotoKind;
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare("SELECT kind, COUNT(*) FROM photos GROUP BY kind")?;
         let rows = stmt.query_map([], |r| Ok((r.get::<_, i64>(0)?, r.get::<_, i64>(1)?)))?;
         let (mut photos, mut videos, mut raw) = (0usize, 0usize, 0usize);
@@ -223,7 +223,7 @@ impl Db {
         if uids.is_empty() {
             return Ok(Vec::new());
         }
-        let conn = self.conn.lock();
+        let conn = self.read();
         let placeholders = vec!["?"; uids.len()].join(",");
         let mut stmt = conn.prepare(&format!(
             "SELECT uid, capture_time, name, ratio, thumb_state, kind, favorite FROM photos
@@ -273,7 +273,7 @@ impl Db {
 
     /// Number of photos in the persisted timeline.
     pub fn photos_count(&self) -> Result<usize> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let n: i64 = conn.query_row("SELECT COUNT(*) FROM photos", [], |r| r.get(0))?;
         Ok(n.max(0) as usize)
     }

crates/pdfs-core/src/db/pins.rs

@@ -4,6 +4,7 @@
 use rusqlite::{OptionalExtension, params};
 
 use super::Db;
+use super::utils::MAX_PATH_DEPTH;
 use crate::Result;
 
 /// One row of [`Db::pin_list`]. `recursive` is pin *policy* — whether the whole
@@ -39,7 +40,7 @@ impl Db {
 
     /// Every directly-pinned entry, ordered by uid.
     pub fn pin_list(&self) -> Result<Vec<PinRow>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         // `recursive` is pin policy; the node's kind comes from `nodes`, and is
         // NULL for a pin whose node was never cached.
         let mut stmt = conn.prepare(
@@ -66,8 +67,13 @@ impl Db {
     /// is pinned recursively. The ancestor check walks `parent_uid` to the root
     /// via a CTE; a direct pin is honoured even when the node has no `nodes` row
     /// yet (e.g. a CLI that never hydrates the node cache).
+    ///
+    /// The walk is depth-capped like [`super::utils::path_of`]: a `parent_uid`
+    /// cycle is corrupt data the API can hand us, and `UNION ALL` over one never
+    /// terminates — while holding the daemon's only SQLite connection, on a path
+    /// that runs per cached read.
     pub fn is_pinned(&self, uid: &str) -> Result<bool> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let direct: Option<i64> = conn
             .query_row("SELECT 1 FROM pins WHERE uid = ?1", params![uid], |r| {
                 r.get(0)
@@ -78,14 +84,18 @@ impl Db {
         }
         let anc: Option<i64> = conn
             .query_row(
-                "WITH RECURSIVE anc(uid, parent_uid) AS (
-                   SELECT uid, parent_uid FROM nodes WHERE uid = ?1
-                   UNION ALL
-                   SELECT n.uid, n.parent_uid FROM nodes n JOIN anc ON n.uid = anc.parent_uid
-                 )
-                 SELECT 1 FROM anc a JOIN pins p ON p.uid = a.uid
-                 WHERE a.uid != ?1 AND p.recursive = 1
-                 LIMIT 1",
+                &format!(
+                    "WITH RECURSIVE anc(uid, parent_uid, depth) AS (
+                       SELECT uid, parent_uid, 0 FROM nodes WHERE uid = ?1
+                       UNION ALL
+                       SELECT n.uid, n.parent_uid, anc.depth + 1
+                       FROM nodes n JOIN anc ON n.uid = anc.parent_uid
+                       WHERE anc.depth < {MAX_PATH_DEPTH}
+                     )
+                     SELECT 1 FROM anc a JOIN pins p ON p.uid = a.uid
+                     WHERE a.uid != ?1 AND p.recursive = 1
+                     LIMIT 1"
+                ),
                 params![uid],
                 |r| r.get(0),
             )
@@ -97,7 +107,7 @@ impl Db {
     /// every descendant of a recursively-pinned folder. Used to build the
     /// eviction-exempt set (the budget enforcer hashes these into cache keys).
     pub fn pinned_uids(&self) -> Result<Vec<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "WITH RECURSIVE sub(uid) AS (
                SELECT uid FROM pins WHERE recursive = 1
@@ -119,7 +129,7 @@ impl Db {
     /// Every descendant uid of `folder` (all depths), via a `parent_uid` CTE.
     /// Used to evict a recursively-pinned subtree's cached blobs on unpin.
     pub fn descendants(&self, folder: &str) -> Result<Vec<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "WITH RECURSIVE sub(uid) AS (
                SELECT uid FROM nodes WHERE parent_uid = ?1

crates/pdfs-core/src/db/share_access.rs

@@ -16,7 +16,7 @@ impl Db {
     /// remains the restart/offline authority and every explicit role change is
     /// written through before this cache is updated.
     pub fn all_share_access(&self) -> Result<HashMap<NodeUid, Access>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare("SELECT root_uid, access FROM share_access")?;
         let rows = stmt.query_map([], |row| {
             Ok((row.get::<_, String>(0)?, row.get::<_, String>(1)?))
@@ -42,7 +42,7 @@ impl Db {
 
     /// Read the effective access recorded for a shared-tree root.
     pub fn share_access(&self, uid: &NodeUid) -> Result<Option<Access>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let value = conn
             .query_row(
                 "SELECT access FROM share_access WHERE root_uid = ?1",
@@ -96,7 +96,7 @@ impl Db {
     /// share root. `None` means the node itself is not persisted, which callers
     /// must not treat as owned: stale handles need to fail closed.
     pub fn effective_node_access(&self, uid: &NodeUid) -> Result<Option<Access>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let exists: bool = conn.query_row(
             "SELECT EXISTS(SELECT 1 FROM nodes WHERE uid = ?1)",
             [uid.to_string()],

crates/pdfs-core/src/db/state.rs

@@ -8,7 +8,7 @@ use crate::Result;
 
 impl Db {
     pub fn get_event_cursor(&self) -> Result<Option<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let v = conn
             .query_row(
                 "SELECT value FROM sync_state WHERE key = 'event_cursor'",
@@ -32,7 +32,7 @@ impl Db {
 
     /// Read a `sync_state` value as a string.
     pub fn state_str(&self, key: &str) -> Result<Option<String>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let v = conn
             .query_row(
                 "SELECT value FROM sync_state WHERE key = ?1",
@@ -66,7 +66,7 @@ impl Db {
     /// dropping it would leave a file that exists nowhere but this machine with
     /// nothing left to create it. (Writes to a node that is itself still queued
     pub fn state_i64(&self, key: &str) -> Result<Option<i64>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let v: Option<String> = conn
             .query_row(
                 "SELECT value FROM sync_state WHERE key = ?1",

crates/pdfs-core/src/db/sync.rs

@@ -57,7 +57,7 @@ impl Db {
 
     /// Every synced folder, oldest first.
     pub fn sync_folder_list(&self) -> Result<Vec<StoredSyncFolder>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT id, local_path, remote_uid, remote_share_id, mode, pending_mode, state, last_sync
              FROM sync_folder ORDER BY id",
@@ -83,7 +83,7 @@ impl Db {
 
     /// Look up one synced folder by id.
     pub fn sync_folder_get(&self, id: i64) -> Result<Option<StoredSyncFolder>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         conn.query_row(
             "SELECT id, local_path, remote_uid, remote_share_id, mode, pending_mode, state, last_sync
              FROM sync_folder WHERE id = ?1",
@@ -166,7 +166,7 @@ impl Db {
 
     /// The whole per-file sync baseline for a folder, keyed by relative path.
     pub fn sync_entries(&self, folder_id: i64) -> Result<HashMap<String, StoredSyncEntry>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT rel_path, remote_uid, local_mtime, local_size, remote_rev, remote_hash
              FROM sync_entry WHERE folder_id = ?1",
@@ -196,7 +196,7 @@ impl Db {
     /// `sync_folder` and descendants come from the last successful
     /// `sync_entry` baseline.
     pub fn mirror_contains_uid(&self, uid: &str) -> Result<bool> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         conn.query_row(
             "SELECT EXISTS(
                  SELECT 1

crates/pdfs-core/src/db/tests.rs

@@ -867,6 +867,67 @@ fn shared_root_pinned_name_survives_database_reopen() {
     let _ = std::fs::remove_file(path);
 }
 
+/// Two writers on one database is the failure mode the single-instance lock
+/// exists to prevent — a hand-run `pdfs mount` next to the systemd unit gives
+/// two daemons the same inode space, content cache, and drain queue. The second
+/// open has to fail, and it has to stop failing once the first handle is gone.
+#[test]
+fn a_second_writer_is_refused_while_the_first_holds_the_database() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-single-writer-{}-{}.db",
+        std::process::id(),
+        now_test_id()
+    ));
+    let first = Db::open(&path).unwrap();
+    let second = Db::open(&path);
+    assert!(second.is_err(), "a second writer was allowed in");
+
+    drop(first);
+    // The lock lives in the kernel and dies with the descriptor, so nothing has
+    // to be cleaned up for the next daemon to start.
+    Db::open(&path).expect("reopen after the first handle is dropped");
+
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(path.with_extension("db.lock"));
+}
+
+/// `cache.db` is derived state — every row can be re-fetched or re-scanned — so
+/// a file that is not a database at all must not stop the daemon starting. It
+/// is moved aside (not deleted: it is the only evidence) and rebuilt.
+#[test]
+fn a_corrupt_database_is_moved_aside_and_rebuilt() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-corrupt-{}-{}.db",
+        std::process::id(),
+        now_test_id()
+    ));
+    std::fs::write(&path, b"this is not an SQLite database").unwrap();
+
+    let db = Db::open(&path).unwrap();
+    db.set_state_str("shared_with_me_name", "rebuilt").unwrap();
+    assert_eq!(
+        db.state_str("shared_with_me_name").unwrap().as_deref(),
+        Some("rebuilt")
+    );
+    drop(db);
+
+    let dir = path.parent().unwrap();
+    let stem = path.file_name().unwrap().to_string_lossy().to_string();
+    let quarantined: Vec<_> = std::fs::read_dir(dir)
+        .unwrap()
+        .filter_map(|e| e.ok())
+        .map(|e| e.file_name().to_string_lossy().to_string())
+        .filter(|name| name.starts_with(&format!("{stem}.corrupt-")))
+        .collect();
+    assert_eq!(quarantined.len(), 1, "the damaged file was not kept");
+
+    for name in quarantined {
+        let _ = std::fs::remove_file(dir.join(name));
+    }
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(path.with_extension("db.lock"));
+}
+
 fn local(path: &str, name: &str, is_dir: bool) -> LocalEntry {
     LocalEntry {
         path: path.into(),
@@ -911,6 +972,25 @@ fn local_scan_indexes_then_prunes_stale_paths() {
     assert_eq!(db.search_local("notes", 10).unwrap().len(), 1);
 }
 
+/// The index is maintained incrementally rather than rebuilt at the end of a
+/// scan, so the case that used to be covered for free — a file the scan sees
+/// again — needs its own test: a re-scan must neither lose the entry nor
+/// duplicate it.
+#[test]
+fn rescanning_an_unchanged_file_keeps_exactly_one_index_entry() {
+    let db = Db::open_in_memory().unwrap();
+    for (generation, mtime) in [(1_000, 5), (2_000, 9)] {
+        let g = db.local_begin_scan().unwrap();
+        let mut e = local("/home/u/docs/report.pdf", "report.pdf", false);
+        e.mtime = mtime;
+        db.local_upsert_batch(g, &[e]).unwrap();
+        assert_eq!(db.local_finish_scan(g, generation).unwrap(), 1);
+        let hits = db.search_local("report", 10).unwrap();
+        assert_eq!(hits.len(), 1, "generation {generation}");
+        assert_eq!(hits[0].mtime, mtime);
+    }
+}
+
 /// Queries below the trigram minimum still match, via the `LIKE` fallback.
 #[test]
 fn local_search_short_query_like_fallback() {
@@ -1429,6 +1509,7 @@ fn migration_v17_adds_share_access_to_a_v16_fixture() {
              DROP TABLE albums;
              DROP INDEX idx_photos_favorite;
              ALTER TABLE photos DROP COLUMN favorite;
+             ALTER TABLE nodes DROP COLUMN path;
              UPDATE sync_state SET value = '16' WHERE key = 'schema_version';",
         )
         .unwrap();
@@ -2628,6 +2709,26 @@ fn non_recursive_folder_pin_does_not_cover_children() {
     assert!(!db.is_pinned(&du("rep")).unwrap());
 }
 
+/// A `parent_uid` cycle is corrupt data the API can hand us, and `is_pinned`'s
+/// ancestor walk is `UNION ALL` — uncapped it never terminates, while holding
+/// the daemon's only SQLite connection, on a path that runs per cached read.
+/// The answer here matters less than the fact that there is one.
+#[test]
+fn is_pinned_terminates_on_a_parent_cycle() {
+    let db = Db::open_in_memory().unwrap();
+    let du = |l: &str| uid(l).to_string();
+    // a -> b -> a, with nothing pinned above either.
+    db.upsert_node(&folder("a", Some("b"), "A")).unwrap();
+    db.upsert_node(&folder("b", Some("a"), "B")).unwrap();
+    db.upsert_node(&file("leaf", "a", "leaf.txt", 1)).unwrap();
+
+    assert!(!db.is_pinned(&du("leaf")).unwrap());
+
+    // And a pin *inside* the cycle is still found rather than walked past.
+    db.pin_add(&du("a"), "A", true).unwrap();
+    assert!(db.is_pinned(&du("leaf")).unwrap());
+}
+
 #[test]
 fn schema_objects_exist() {
     let db = Db::open_in_memory().unwrap();
@@ -3182,3 +3283,430 @@ fn clear_state_prefix_takes_only_the_prefixed_keys() {
         "the listing's own stamp is not a per-album one"
     );
 }
+
+// --- `nodes.path`, and the subtree work it replaces (DB3) -------------------
+
+/// Read the stored path column directly: the point of these tests is that the
+/// column itself is right, not that `path_of` can still fall back to a walk.
+fn stored_path(db: &Db, link: &str) -> Option<String> {
+    use rusqlite::OptionalExtension;
+    db.with_conn(|conn| {
+        Ok(conn
+            .query_row(
+                "SELECT path FROM nodes WHERE uid = ?1",
+                [uid(link).to_string()],
+                |row| row.get::<_, Option<String>>(0),
+            )
+            .optional()?
+            .flatten())
+    })
+    .unwrap()
+}
+
+#[test]
+fn a_node_stores_its_path_when_it_is_written() {
+    let db = Db::open_in_memory().unwrap();
+    db.upsert_node(&folder("root", None, "My Files")).unwrap();
+    db.upsert_node(&folder("docs", Some("root"), "Documents"))
+        .unwrap();
+    db.upsert_node(&file("f1", "docs", "report.pdf", 1))
+        .unwrap();
+
+    // The root's name is the mount, so it contributes nothing.
+    assert_eq!(stored_path(&db, "root").as_deref(), Some(""));
+    assert_eq!(stored_path(&db, "docs").as_deref(), Some("Documents"));
+    assert_eq!(
+        stored_path(&db, "f1").as_deref(),
+        Some("Documents/report.pdf")
+    );
+}
+
+#[test]
+fn a_node_below_an_uncached_parent_starts_its_own_path() {
+    // A device folder's root lives in `device`/`sync_folder` and never gets a
+    // `nodes` row, so its children are the top of the cached chain.
+    let db = Db::open_in_memory().unwrap();
+    db.upsert_node(&folder("orphan", Some("never-cached"), "Work"))
+        .unwrap();
+    db.upsert_node(&file("f1", "orphan", "notes.txt", 1))
+        .unwrap();
+
+    assert_eq!(stored_path(&db, "orphan").as_deref(), Some("Work"));
+    assert_eq!(stored_path(&db, "f1").as_deref(), Some("Work/notes.txt"));
+}
+
+#[test]
+fn renaming_a_folder_rewrites_every_descendant_path() {
+    let db = Db::open_in_memory().unwrap();
+    db.upsert_node(&folder("root", None, "My Files")).unwrap();
+    db.upsert_node(&folder("a", Some("root"), "Before"))
+        .unwrap();
+    db.upsert_node(&folder("b", Some("a"), "Inner")).unwrap();
+    db.upsert_node(&file("f1", "b", "deep.txt", 1)).unwrap();
+    assert_eq!(
+        stored_path(&db, "f1").as_deref(),
+        Some("Before/Inner/deep.txt")
+    );
+
+    db.upsert_node(&folder("a", Some("root"), "After")).unwrap();
+
+    assert_eq!(stored_path(&db, "a").as_deref(), Some("After"));
+    assert_eq!(stored_path(&db, "b").as_deref(), Some("After/Inner"));
+    assert_eq!(
+        stored_path(&db, "f1").as_deref(),
+        Some("After/Inner/deep.txt")
+    );
+    let hit = &db.search("deep.txt", 10).unwrap()[0];
+    assert_eq!(
+        hit.path, "After/Inner/deep.txt",
+        "the search index must carry the rewritten path too"
+    );
+}
+
+#[test]
+fn moving_a_folder_rewrites_every_descendant_path() {
+    let db = Db::open_in_memory().unwrap();
+    db.upsert_node(&folder("root", None, "My Files")).unwrap();
+    db.upsert_node(&folder("old", Some("root"), "Old")).unwrap();
+    db.upsert_node(&folder("new", Some("root"), "New")).unwrap();
+    db.upsert_node(&folder("moved", Some("old"), "Moved"))
+        .unwrap();
+    db.upsert_node(&file("f1", "moved", "deep.txt", 1)).unwrap();
+
+    db.upsert_node(&folder("moved", Some("new"), "Moved"))
+        .unwrap();
+
+    assert_eq!(stored_path(&db, "moved").as_deref(), Some("New/Moved"));
+    assert_eq!(
+        stored_path(&db, "f1").as_deref(),
+        Some("New/Moved/deep.txt")
+    );
+}
+
+#[test]
+fn trashing_a_folder_drops_its_subtree_from_the_index() {
+    let db = Db::open_in_memory().unwrap();
+    db.upsert_node(&folder("root", None, "My Files")).unwrap();
+    db.upsert_node(&folder("a", Some("root"), "Holder"))
+        .unwrap();
+    db.upsert_node(&file("f1", "a", "hidden.txt", 1)).unwrap();
+    assert_eq!(db.search("hidden", 10).unwrap().len(), 1);
+
+    let mut trashed = folder("a", Some("root"), "Holder");
+    trashed.trashed = true;
+    db.upsert_node(&trashed).unwrap();
+
+    assert_eq!(
+        db.search("hidden", 10).unwrap().len(),
+        0,
+        "a file under a trashed folder is not reachable and must not be findable"
+    );
+
+    // …and comes back when the folder does.
+    db.upsert_node(&folder("a", Some("root"), "Holder"))
+        .unwrap();
+    assert_eq!(db.search("hidden", 10).unwrap().len(), 1);
+}
+
+/// The re-listing case, which is most of what the write path actually does: a
+/// folder upserted with the same name and parent must not touch its subtree.
+///
+/// Written as a ratio between two subtree sizes, so it fails on the scaling
+/// regression rather than on a slow machine. Before this, every re-listing paid
+/// a `path_of` walk plus an ancestor walk per descendant.
+#[test]
+fn re_listing_a_folder_does_not_scale_with_its_subtree() {
+    fn elapsed_for(children: usize) -> std::time::Duration {
+        let db = Db::open_in_memory().unwrap();
+        db.upsert_node(&folder("root", None, "My Files")).unwrap();
+        db.upsert_node(&folder("big", Some("root"), "Big")).unwrap();
+        let kids: Vec<Node> = (0..children)
+            .map(|i| file(&format!("k{i}"), "big", &format!("child-{i}.txt"), 1))
+            .collect();
+        db.upsert_nodes(&kids).unwrap();
+
+        let unchanged = folder("big", Some("root"), "Big");
+        let start = std::time::Instant::now();
+        for _ in 0..20 {
+            db.upsert_node(&unchanged).unwrap();
+        }
+        start.elapsed()
+    }
+
+    let small = elapsed_for(100);
+    let large = elapsed_for(2_000);
+    assert!(
+        large < small * 8,
+        "re-listing a folder scaled with its subtree: {small:?} for 100 children, \
+         {large:?} for 2000 (20x the descendants)"
+    );
+}
+
+#[test]
+fn migration_v25_backfills_paths_for_an_existing_tree() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-db-v24-fixture-{}-{}.db",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos()
+    ));
+    {
+        let db = Db::open(&path).unwrap();
+        db.upsert_node(&folder("root", None, "My Files")).unwrap();
+        db.upsert_node(&folder("docs", Some("root"), "Documents"))
+            .unwrap();
+        db.upsert_node(&file("f1", "docs", "report.pdf", 1))
+            .unwrap();
+        db.upsert_node(&folder("orphan", Some("never-cached"), "Work"))
+            .unwrap();
+        db.upsert_node(&file("f2", "orphan", "notes.txt", 1))
+            .unwrap();
+    }
+    {
+        // Put the file back in the state a released V24 database was in.
+        let conn = rusqlite::Connection::open(&path).unwrap();
+        conn.execute_batch(
+            "ALTER TABLE nodes DROP COLUMN path;
+             UPDATE sync_state SET value = '24' WHERE key = 'schema_version';",
+        )
+        .unwrap();
+    }
+
+    let db = Db::open(&path).unwrap();
+    assert_eq!(stored_path(&db, "root").as_deref(), Some(""));
+    assert_eq!(stored_path(&db, "docs").as_deref(), Some("Documents"));
+    assert_eq!(
+        stored_path(&db, "f1").as_deref(),
+        Some("Documents/report.pdf"),
+        "an existing tree must come out of the migration with the same paths \
+         the walk used to produce"
+    );
+    // A subtree whose top has an uncached parent is ordinary data, not an error.
+    assert_eq!(stored_path(&db, "orphan").as_deref(), Some("Work"));
+    assert_eq!(stored_path(&db, "f2").as_deref(), Some("Work/notes.txt"));
+
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(format!("{}.lock", path.display()));
+}
+
+/// A queue carried over from V26 keeps its rows, and every one of them starts
+/// out not access-deferred — a migration must not invent a deferral window that
+/// would report an ordinary queued upload as blocked.
+#[test]
+fn migration_v27_adds_the_access_deferral_column_to_a_v26_queue() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-db-v26-fixture-{}-{}.db",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos()
+    ));
+    {
+        let db = Db::open(&path).unwrap();
+        db.enqueue_op(&crate::db::PendingOp {
+            id: 0,
+            kind: crate::db::OP_REVISION.to_string(),
+            uid: "v~l".to_string(),
+            parent_uid: None,
+            name: None,
+            blob_path: Some("/tmp/blob".to_string()),
+            meta_json: Some("{}".to_string()),
+            created_at: 1,
+            attempts: 0,
+            last_error: None,
+            next_attempt_at: 0,
+        })
+        .unwrap();
+    }
+    {
+        // Put the file back in the state a released V26 database was in.
+        let conn = rusqlite::Connection::open(&path).unwrap();
+        conn.execute_batch(
+            "ALTER TABLE pending_op DROP COLUMN access_deferred_since;
+             UPDATE sync_state SET value = '26' WHERE key = 'schema_version';",
+        )
+        .unwrap();
+    }
+
+    let db = Db::open(&path).unwrap();
+    let ops = db.pending_ops().unwrap();
+    assert_eq!(ops.len(), 1, "the migration keeps the queued write");
+    let now = 5_000i64;
+    assert_eq!(
+        db.defer_op_for_access(ops[0].id, now, now + 5_000).unwrap(),
+        now,
+        "a carried-over op is not already deferred"
+    );
+
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(format!("{}.lock", path.display()));
+}
+
+// --- read-only connection pool (DB2) ---------------------------------------
+
+/// A read must not queue behind whatever happens to be writing.
+///
+/// The write connection is held for the whole measurement — which is what a
+/// large listing's commit does to it — while a reader asks for a search. Before
+/// the pool this was the same mutex, so the read waited for the writer;
+/// `busy_timeout` never entered into it, because the contention was this
+/// process's own lock rather than SQLite's.
+#[test]
+fn a_read_does_not_wait_for_the_write_connection() {
+    let path = std::env::temp_dir().join(format!(
+        "pdfs-db-readpool-{}-{}.db",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos()
+    ));
+    let db = std::sync::Arc::new(Db::open(&path).unwrap());
+    db.upsert_node(&folder("root", None, "My Files")).unwrap();
+    db.upsert_node(&file("f1", "root", "findme.txt", 1))
+        .unwrap();
+
+    let held = std::time::Duration::from_millis(500);
+    let (tx, rx) = std::sync::mpsc::channel();
+    let writer = {
+        let db = db.clone();
+        std::thread::spawn(move || {
+            db.with_conn(|_conn| {
+                tx.send(()).unwrap();
+                std::thread::sleep(held);
+                Ok(())
+            })
+            .unwrap();
+        })
+    };
+    rx.recv().unwrap();
+
+    let start = std::time::Instant::now();
+    let hits = db.search("findme", 10).unwrap();
+    let waited = start.elapsed();
+    writer.join().unwrap();
+
+    assert_eq!(hits.len(), 1, "the read must still see committed data");
+    assert!(
+        waited < held / 2,
+        "the read waited {waited:?} for a write connection held {held:?}"
+    );
+
+    drop(db);
+    let _ = std::fs::remove_file(&path);
+    let _ = std::fs::remove_file(format!("{}.lock", path.display()));
+}
+
+/// §5: the drain is several workers over one queue, and the claim column is the
+/// only thing keeping them off each other's work. Two properties matter — no
+/// row is ever handed to two workers, and no *node* is ever worked by two
+/// workers even though its ops are separate rows.
+#[test]
+fn a_claimed_op_is_invisible_to_every_other_worker() {
+    let db = Db::open_in_memory().unwrap();
+    let root = uid("root").to_string();
+    for i in 0..3 {
+        db.enqueue_op(&bulk_op(i, &root)).unwrap();
+    }
+
+    let first = db.claim_next_due_op(10).unwrap().expect("an op is due");
+    let second = db
+        .claim_next_due_op(10)
+        .unwrap()
+        .expect("another op is due");
+    let third = db
+        .claim_next_due_op(10)
+        .unwrap()
+        .expect("a third op is due");
+    assert!(
+        db.claim_next_due_op(10).unwrap().is_none(),
+        "a fourth claim must find nothing left: every queued op is taken"
+    );
+
+    let mut ids = [first.id, second.id, third.id];
+    ids.sort_unstable();
+    assert_eq!(
+        ids,
+        [1, 2, 3],
+        "each worker must get a distinct row, oldest first"
+    );
+
+    // Releasing puts one back, and it is the one that comes out next.
+    db.release_op_claim(second.id).unwrap();
+    let again = db.claim_next_due_op(10).unwrap().expect("the released op");
+    assert_eq!(again.id, second.id);
+}
+
+#[test]
+fn two_workers_never_share_a_node() {
+    let db = Db::open_in_memory().unwrap();
+    let root = uid("root").to_string();
+    // Two ops of different kinds against one node — a queued rename and a
+    // queued revision, which is an ordinary `mv` plus a save.
+    let mut revision = bulk_op(0, &root);
+    revision.parent_uid = None;
+    db.enqueue_op(&revision).unwrap();
+    let rename = PendingOp {
+        kind: OP_RENAME.to_string(),
+        parent_uid: Some(root.clone()),
+        blob_path: None,
+        meta_json: None,
+        ..bulk_op(0, &root)
+    };
+    db.enqueue_op(&rename).unwrap();
+
+    let claimed = db.claim_next_due_op(10).unwrap().expect("an op is due");
+    assert_eq!(claimed.kind, OP_REVISION, "the older op goes first");
+    assert!(
+        db.claim_next_due_op(10).unwrap().is_none(),
+        "the second op targets the same node, so it must wait for the first: \
+         two workers on one uid would race over its staged blob and land its \
+         changes out of order"
+    );
+
+    db.release_op_claim(claimed.id).unwrap();
+    db.delete_op(claimed.id).unwrap();
+    assert_eq!(
+        db.claim_next_due_op(10).unwrap().map(|o| o.kind),
+        Some(OP_RENAME.to_string()),
+        "once the revision retires, the rename is claimable"
+    );
+}
+
+#[test]
+fn claims_do_not_survive_the_process_that_took_them() {
+    let db = Db::open_in_memory().unwrap();
+    let root = uid("root").to_string();
+    for i in 0..2 {
+        db.enqueue_op(&bulk_op(i, &root)).unwrap();
+    }
+    db.claim_next_due_op(10).unwrap().unwrap();
+    db.claim_next_due_op(10).unwrap().unwrap();
+    assert!(db.claim_next_due_op(10).unwrap().is_none());
+
+    // What `Db::open` does after the single-writer lock says the previous run
+    // is gone. Without it those ops are invisible to every worker forever.
+    assert_eq!(db.clear_op_claims().unwrap(), 2);
+    assert!(
+        db.claim_next_due_op(10).unwrap().is_some(),
+        "a crashed run's claims must not park its queue"
+    );
+}
+
+#[test]
+fn an_idle_worker_does_not_wait_on_an_op_another_worker_has() {
+    let db = Db::open_in_memory().unwrap();
+    let root = uid("root").to_string();
+    db.enqueue_op(&bulk_op(0, &root)).unwrap();
+
+    assert_eq!(db.earliest_due_at().unwrap(), Some(0));
+    db.claim_next_due_op(10).unwrap().unwrap();
+    assert_eq!(
+        db.earliest_due_at().unwrap(),
+        None,
+        "an op somebody else is draining is not work this worker is waiting for"
+    );
+}

crates/pdfs-core/src/db/trash.rs

@@ -39,7 +39,7 @@ impl Db {
     /// The persisted trash listing, folders first then by name — the order the
     /// Trash page shows it in.
     pub fn trash_list(&self) -> Result<Vec<StoredTrash>> {
-        let conn = self.conn.lock();
+        let conn = self.read();
         let mut stmt = conn.prepare(
             "SELECT uid, name, is_dir, size, mtime FROM trash
              ORDER BY is_dir DESC, name COLLATE NOCASE",

crates/pdfs-core/src/db/utils.rs

@@ -1,6 +1,6 @@
 //! Small query helpers shared across the table modules.
 
-use rusqlite::{Connection, params};
+use rusqlite::{Connection, OptionalExtension, params};
 
 use crate::Result;
 
@@ -8,14 +8,19 @@ use crate::Result;
 /// grams), so short queries fall back to a `LIKE` scan over `nodes.name`.
 pub(super) const TRIGRAM_MIN: usize = 3;
 
-pub(super) fn pair(row: &rusqlite::Row) -> rusqlite::Result<(String, String)> {
-    Ok((row.get(0)?, row.get(1)?))
+/// One raw search row: the node's JSON, its uid, and its stored path — `None`
+/// only for a row written before the `path` column existed, which
+/// [`path_of`] repairs by walking.
+pub(super) type HitRow = (String, String, Option<String>);
+
+pub(super) fn hit_row(row: &rusqlite::Row) -> rusqlite::Result<HitRow> {
+    Ok((row.get(0)?, row.get(1)?, row.get(2)?))
 }
 
-/// Drain a `query_map` of [`pair`] rows into a `Vec`, propagating row errors.
-pub(super) fn collect_pairs(
-    rows: impl Iterator<Item = rusqlite::Result<(String, String)>>,
-) -> Result<Vec<(String, String)>> {
+/// Drain a `query_map` of [`hit_row`] rows into a `Vec`, propagating row errors.
+pub(super) fn collect_hits(
+    rows: impl Iterator<Item = rusqlite::Result<HitRow>>,
+) -> Result<Vec<HitRow>> {
     let mut out = Vec::new();
     for r in rows {
         out.push(r?);
@@ -35,18 +40,44 @@ pub(super) fn like_escape(s: &str) -> String {
 /// nowhere near this; a parent *cycle* is unbounded, and this query is on the
 /// path of every search result, including the `LIKE` lane that reads `nodes`
 /// directly and so never passed through the index's cycle check.
-const MAX_PATH_DEPTH: usize = 256;
+pub(super) const MAX_PATH_DEPTH: usize = 256;
 
-/// Resolve a node's mountpoint-relative path by walking `parent_uid` to the
-/// root via a recursive CTE. The root (the node with no parent) is excluded, so
-/// a top-level file `report.pdf` yields `"report.pdf"`, not `"My Files/report.pdf"`.
+/// Append `name` to a parent's stored path. An empty parent path is the root,
+/// whose name is the mount and so contributes nothing.
+pub(super) fn join_path(parent_path: &str, name: &str) -> String {
+    if parent_path.is_empty() {
+        name.to_string()
+    } else {
+        format!("{parent_path}/{name}")
+    }
+}
+
+/// Resolve a node's mountpoint-relative path. The root (the node with no
+/// parent) is excluded, so a top-level file `report.pdf` yields
+/// `"report.pdf"`, not `"My Files/report.pdf"`.
+///
+/// Reads the stored `nodes.path`, which the write path maintains (see
+/// `upsert_node_tx`). Falls back to [`walk_path_of`] for a row that predates the
+/// column or was never written through — a search serving a keystroke used to
+/// run that walk once per candidate, which is the reason the column exists.
+pub(super) fn path_of(conn: &Connection, uid: &str) -> Result<String> {
+    let stored: Option<Option<String>> = conn
+        .query_row("SELECT path FROM nodes WHERE uid = ?1", params![uid], |r| {
+            r.get(0)
+        })
+        .optional()?;
+    match stored.flatten() {
+        Some(path) => Ok(path),
+        None => walk_path_of(conn, uid),
+    }
+}
+
+/// Resolve a path by walking `parent_uid` to the root via a recursive CTE.
 ///
 /// A cycle in `parent_uid` is corrupt data the API can still hand us, and
 /// `UNION ALL` over one never terminates — the walk is therefore depth-capped
-/// and returns the truncated path rather than hanging the caller. That caller is
-/// usually a search serving a keystroke, holding the daemon's only SQLite
-/// connection.
-pub(super) fn path_of(conn: &Connection, uid: &str) -> Result<String> {
+/// and returns the truncated path rather than hanging the caller.
+pub(super) fn walk_path_of(conn: &Connection, uid: &str) -> Result<String> {
     let mut stmt = conn.prepare(&format!(
         "WITH RECURSIVE anc(uid, parent_uid, name, depth) AS (
            SELECT uid, parent_uid, name, 0 FROM nodes WHERE uid = ?1

crates/pdfs-core/src/error.rs

@@ -45,6 +45,27 @@ pub enum Error {
     Other(String),
 }
 
+impl Error {
+    /// Whether this is the local disk being full (`ENOSPC`), as opposed to any
+    /// other I/O failure.
+    ///
+    /// Worth distinguishing because it is the one storage failure the user can
+    /// act on, and because the caller's response differs: a full disk is met by
+    /// evicting cached blobs and answering `ENOSPC`, where `EIO` would read as a
+    /// broken file. `std::io::ErrorKind::StorageFull` is still unstable, so this
+    /// compares the raw code.
+    pub fn is_disk_full(&self) -> bool {
+        match self {
+            Error::Io(e) => e.raw_os_error() == Some(libc::ENOSPC),
+            // SQLite reports it as its own code rather than surfacing the errno.
+            Error::Db(rusqlite::Error::SqliteFailure(e, _)) => {
+                e.code == rusqlite::ErrorCode::DiskFull
+            }
+            _ => false,
+        }
+    }
+}
+
 pub type Result<T> = std::result::Result<T, Error>;
 
 /// A failure on its way to a front-end: prose for the user, plus the

crates/pdfs-fuse/src/background.rs

@@ -8,35 +8,20 @@ use super::*;
 ///
 /// Returns whether this mount actually held the node, so the caller can evict
 /// its content blob once if *any* mount did.
-fn forget_and_notify(st: &mut State, notifier: Option<&Notifier>, uid: &NodeUid) -> bool {
+fn forget_and_notify(st: &mut State, notify: &mut NotifyBatch, uid: &NodeUid) -> bool {
     // Capture the inode before `forget` clears the uid mapping.
     let child = st.by_uid.get(uid).copied();
     let Some((parent, name)) = st.forget(uid) else {
         return false;
     };
-    if let Some(notifier) = notifier {
-        match child {
-            Some(child) => {
-                let _ = notifier.delete(INodeNo(parent), INodeNo(child), OsStr::new(&name));
-            }
-            None => {
-                let _ = notifier.inval_entry(INodeNo(parent), OsStr::new(&name));
-            }
-        }
+    match child {
+        Some(child) => notify.delete(parent, child, name),
+        None => notify.inval_entry(parent, name),
     }
     true
 }
 
-fn invalidate_access_changes(notifier: Option<&Notifier>, changed: &[u64]) {
-    let Some(notifier) = notifier else {
-        return;
-    };
-    for &ino in changed {
-        let _ = notifier.inval_inode(INodeNo(ino), 0, 0);
-    }
-}
-
-fn hide_foreign_deleted_and_notify(st: &mut State, notifier: Option<&Notifier>, uid: &NodeUid) {
+fn hide_foreign_deleted_and_notify(st: &mut State, notify: &mut NotifyBatch, uid: &NodeUid) {
     let child = st.by_uid.get(uid).copied();
     let dentry = child.and_then(|ino| {
         st.entries
@@ -45,9 +30,9 @@ fn hide_foreign_deleted_and_notify(st: &mut State, notifier: Option<&Notifier>,
     });
     let changed = st.downgrade_shared_subtree(uid);
     st.hide_foreign_subtree(uid);
-    invalidate_access_changes(notifier, &changed);
-    if let (Some(notifier), Some((parent, name)), Some(child)) = (notifier, dentry, child) {
-        let _ = notifier.delete(INodeNo(parent), INodeNo(child), OsStr::new(&name));
+    notify.extend_inodes(&changed);
+    if let (Some((parent, name)), Some(child)) = (dentry, child) {
+        notify.delete(parent, child, name);
     }
 }
 
@@ -132,18 +117,18 @@ fn apply_event(core: &Core, event: &DriveEvent, dirty: &mut DirtyParents) -> pdf
         && foreign_delete
     {
         apply_foreign_delete(&core.db, node_uid, |uid| {
-            core.for_each_mount(|st, notifier| {
-                hide_foreign_deleted_and_notify(st, notifier, uid);
+            core.for_each_mount(|st, notify| {
+                hide_foreign_deleted_and_notify(st, notify, uid);
             });
         })
     } else {
         apply_access_downgrade(&core.db, event, |root| {
-            core.for_each_mount(|st, notifier| {
+            core.for_each_mount(|st, notify| {
                 let changed = match root {
                     Some(uid) => st.downgrade_shared_subtree(uid),
                     None => st.downgrade_known_shared_access(),
                 };
-                invalidate_access_changes(notifier, &changed);
+                notify.extend_inodes(&changed);
             });
         })
     };
@@ -173,19 +158,17 @@ fn apply_event(core: &Core, event: &DriveEvent, dirty: &mut DirtyParents) -> pdf
                 debug!(uid = %node_uid, "ignoring remote event for a node with a queued write");
             }
             let mut had_node = false;
-            core.for_each_mount(|st, notifier| {
+            core.for_each_mount(|st, notify| {
                 if *is_trashed {
                     // Trashing makes a node vanish from its parent listing.
-                    had_node |= forget_and_notify(st, notifier, node_uid);
+                    had_node |= forget_and_notify(st, notify, node_uid);
                 } else if !ours && let Some(&ino) = st.by_uid.get(node_uid) {
                     // Known node changed: drop its cached attrs/data (and
                     // listing if it is a directory) so the next access
                     // re-fetches. Its content blob may now be stale too.
                     had_node = true;
                     st.invalidate_listing(ino);
-                    if let Some(notifier) = notifier {
-                        let _ = notifier.inval_inode(INodeNo(ino), 0, 0);
-                    }
+                    notify.inval_inode(ino);
                 }
             });
             // Once, not once per mount: the content cache is per-daemon.
@@ -209,8 +192,8 @@ fn apply_event(core: &Core, event: &DriveEvent, dirty: &mut DirtyParents) -> pdf
         DriveEvent::NodeDeleted { node_uid, .. } => {
             core.cache.evict(node_uid);
             if !foreign_delete {
-                core.for_each_mount(|st, notifier| {
-                    forget_and_notify(st, notifier, node_uid);
+                core.for_each_mount(|st, notify| {
+                    forget_and_notify(st, notify, node_uid);
                 });
             }
         }
@@ -223,13 +206,11 @@ fn apply_event(core: &Core, event: &DriveEvent, dirty: &mut DirtyParents) -> pdf
         | DriveEvent::SharedWithMeUpdated { .. } => {
             warn!("event access continuity lost; downgrading shared trees and resyncing lazily");
             apply_result = apply_result.and(core.db.clear_state(SHARED_WITH_ME_SYNCED_MS));
-            core.for_each_mount(|st, notifier| {
+            core.for_each_mount(|st, notify| {
                 let dirs: Vec<u64> = st.children.keys().copied().collect();
                 for &ino in &dirs {
                     st.invalidate_listing(ino);
-                    if let Some(notifier) = notifier {
-                        let _ = notifier.inval_inode(INodeNo(ino), 0, 0);
-                    }
+                    notify.inval_inode(ino);
                 }
             });
         }
@@ -303,18 +284,16 @@ fn flush_dirty_parents(core: &Core, dirty: &DirtyParents) {
     if dirty.foreign.is_empty() && dirty.ours.is_empty() {
         return;
     }
-    core.for_each_mount(|st, notifier| {
-        let drop_listing = |st: &mut State, parent_uid: &NodeUid| {
-            let Some(&parent) = st.by_uid.get(parent_uid) else {
-                return;
-            };
-            st.invalidate_listing(parent);
-            if let Some(notifier) = notifier {
-                let _ = notifier.inval_inode(INodeNo(parent), 0, 0);
-            }
+    fn drop_listing(st: &mut State, notify: &mut NotifyBatch, parent_uid: &NodeUid) {
+        let Some(&parent) = st.by_uid.get(parent_uid) else {
+            return;
         };
+        st.invalidate_listing(parent);
+        notify.inval_inode(parent);
+    }
+    core.for_each_mount(|st, notify| {
         for parent_uid in &dirty.foreign {
-            drop_listing(st, parent_uid);
+            drop_listing(st, notify, parent_uid);
         }
         for (parent_uid, children) in &dirty.ours {
             if dirty.foreign.contains(parent_uid) {
@@ -328,7 +307,7 @@ fn flush_dirty_parents(core: &Core, dirty: &DirtyParents) {
             {
                 continue;
             }
-            drop_listing(st, parent_uid);
+            drop_listing(st, notify, parent_uid);
         }
     });
 }

crates/pdfs-fuse/src/control.rs

@@ -29,7 +29,23 @@ use super::{Core, count_noun, human_bytes, human_duration, parse_uid};
 
 const MAX_CONTROL_REQUEST_BYTES: u64 = 1024 * 1024;
 const CONTROL_READ_TIMEOUT: Duration = Duration::from_secs(10);
+/// A client that connects, sends a request and then never reads its reply used
+/// to park a handler thread forever in `write_all`; 64 of those exhaust
+/// [`MAX_CONTROL_HANDLERS`] permanently. Generous enough for the largest
+/// responses (photo bytes) over a local socket, finite so a wedged peer can't
+/// hold a permit.
+const CONTROL_WRITE_TIMEOUT: Duration = Duration::from_secs(30);
 const MAX_CONTROL_HANDLERS: usize = 64;
+/// Ceiling on any client-supplied row `limit`. Every in-tree caller asks for far
+/// less (200 for activity, 60 for a photo page), but the value goes straight
+/// into `SQL LIMIT ?` under the daemon's single connection — an unclamped one
+/// materialises the whole account, and then serialises it into one JSON line.
+const MAX_ROW_LIMIT: usize = 1000;
+
+/// Clamp a client-supplied row limit to [`MAX_ROW_LIMIT`].
+fn clamp_limit(limit: usize) -> usize {
+    limit.min(MAX_ROW_LIMIT)
+}
 static ACTIVE_CONTROL_HANDLERS: AtomicUsize = AtomicUsize::new(0);
 
 struct ControlHandlerPermit;
@@ -109,6 +125,10 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
         warn!(error = %e, "control: could not set request timeout");
         return;
     }
+    if let Err(e) = stream.set_write_timeout(Some(CONTROL_WRITE_TIMEOUT)) {
+        warn!(error = %e, "control: could not set response timeout");
+        return;
+    }
     let mut reader = BufReader::new(match stream.try_clone() {
         Ok(s) => s,
         Err(e) => {
@@ -128,7 +148,13 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
         Ok(CtlRequest::Status) => {
             let pins = core.cache.list_pins();
             let queued = core.db.pending_op_counts().unwrap_or_default();
+            let staging = core.cache.staging_usage();
             CtlResponse::Status {
+                parked_uploads: queued.parked.max(0) as u64,
+                failing_ops: queued.failing.max(0) as u64,
+                failing_error: queued.last_error.clone(),
+                staged_bytes: staging.bytes,
+                staged_oldest_secs: staging.oldest_secs,
                 username: username.to_string(),
                 mountpoint: mountpoint.display().to_string(),
                 pinned: pins.len(),
@@ -204,7 +230,7 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
             kind,
             range,
             favorites,
-        }) => match core.photos_timeline(offset, limit, kind, range, favorites) {
+        }) => match core.photos_timeline(offset, clamp_limit(limit), kind, range, favorites) {
             Ok(Some(items)) => CtlResponse::Photos {
                 available: true,
                 items,
@@ -245,7 +271,7 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
             // An album page is a timeline page as far as the front-end is
             // concerned, so it comes back in the same reply — with `counts`
             // unset, which describe the whole timeline rather than this album.
-            Some(album) => match core.album_photos(&album, offset, limit) {
+            Some(album) => match core.album_photos(&album, offset, clamp_limit(limit)) {
                 Ok(items) => CtlResponse::Photos {
                     available: true,
                     items,
@@ -346,22 +372,24 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
                 Err(e) => CtlResponse::error(e),
             }
         }
-        Ok(CtlRequest::Search { query, limit }) => match core.search(&query, limit) {
+        Ok(CtlRequest::Search { query, limit }) => match core.search(&query, clamp_limit(limit)) {
             Ok(hits) => CtlResponse::SearchResults { hits },
             Err(e) => CtlResponse::error(e),
         },
-        Ok(CtlRequest::SearchLocal { query, limit }) => match core.search_local(&query, limit) {
-            Ok(hits) => CtlResponse::LocalResults {
-                hits,
-                indexing: core.indexing.load(Ordering::Relaxed),
-            },
-            Err(e) => CtlResponse::error(e),
-        },
+        Ok(CtlRequest::SearchLocal { query, limit }) => {
+            match core.search_local(&query, clamp_limit(limit)) {
+                Ok(hits) => CtlResponse::LocalResults {
+                    hits,
+                    indexing: core.indexing.load(Ordering::Relaxed),
+                },
+                Err(e) => CtlResponse::error(e),
+            }
+        }
         Ok(CtlRequest::SearchV2 {
             query,
             limit,
             filters,
-        }) => match core.search_v2(&query, limit, &filters) {
+        }) => match core.search_v2(&query, clamp_limit(limit), &filters) {
             Ok((drive_hits, local_hits)) => CtlResponse::SearchResultsV2 {
                 drive_hits,
                 local_hits,
@@ -540,19 +568,20 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
         }
         Ok(CtlRequest::CacheInspect { deep }) => match core.db.stats() {
             Ok(stats) => {
-                // The deep check reads every page, so it runs only on request;
-                // `integrity_checked` tells the caller which answer it is
-                // holding, since "no problems found" and "never looked" are
-                // both an empty list.
-                let (integrity_problems, integrity_checked) = if deep {
-                    match core.db.integrity_check() {
-                        Ok(problems) => (problems, true),
-                        Err(e) => (vec![format!("integrity check failed: {e:?}")], true),
-                    }
+                // The deep check reads every page — tens of seconds on a real
+                // database, all of it holding the connection every FUSE handler
+                // needs — so it runs on a thread and this request returns
+                // whatever the last completed check found. `integrity_checked`
+                // tells the caller which answer it is holding, since "no
+                // problems found" and "never looked" are both an empty list;
+                // `integrity_running` says a fresh answer is on its way.
+                let (integrity_problems, integrity_checked, integrity_running) = if deep {
+                    core.start_or_read_integrity_check()
                 } else {
-                    (Vec::new(), false)
+                    (Vec::new(), false, false)
                 };
                 CtlResponse::CacheReport {
+                    integrity_running,
                     schema_version: stats.schema_version,
                     db_bytes: stats.total_bytes(),
                     db_reclaimable_bytes: stats.reclaimable_bytes(),
@@ -565,17 +594,16 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
             }
             Err(e) => CtlResponse::error(CoreError::internal(format!("read db stats: {e:?}"))),
         },
-        Ok(CtlRequest::CacheVacuum) => match core.db.vacuum() {
-            Ok(outcome) => CtlResponse::Ok {
-                message: format!(
-                    "vacuumed: {:.1} MiB freed ({:.1} MiB → {:.1} MiB), {} WAL frame(s) checkpointed",
-                    outcome.freed_bytes() as f64 / 1_048_576.0,
-                    outcome.before_bytes as f64 / 1_048_576.0,
-                    outcome.after_bytes as f64 / 1_048_576.0,
-                    outcome.wal_frames_checkpointed,
-                ),
+        // Ack and vacuum on a thread. `VACUUM` rewrites the whole database
+        // under the connection mutex — minutes on a large one — and the
+        // request's own socket timeout is 10 s, so running it inline both froze
+        // the mount and lost its own answer. Progress shows up as a job in
+        // `GetQueueStatus`; the outcome goes to the log.
+        Ok(CtlRequest::CacheVacuum) => match core.start_vacuum() {
+            Ok(()) => CtlResponse::Ok {
+                message: "compacting the cache database in the background".into(),
             },
-            Err(e) => CtlResponse::error(CoreError::internal(format!("vacuum: {e:?}"))),
+            Err(e) => CtlResponse::error(e),
         },
         Ok(CtlRequest::GetQueueStatus) => CtlResponse::Transfers {
             items: core.transfers.snapshot(),
@@ -1023,7 +1051,7 @@ fn handle_control_conn(core: &Core, username: &str, mountpoint: &Path, stream: U
             Err(e) => CtlResponse::error(e),
         },
         Ok(CtlRequest::ListActivity { limit }) => CtlResponse::Activity {
-            items: core.list_activity(limit),
+            items: core.list_activity(clamp_limit(limit)),
         },
         // The request did not parse: the caller sent something malformed.
         Err(e) => CtlResponse::error(CoreError::invalid(format!("bad request: {e}"))),

crates/pdfs-fuse/src/devices.rs

@@ -536,7 +536,14 @@ impl Core {
         // folder the daemon no longer tracks, and nothing would ever unmount it.
         // Tear it down first — including before trashing the remote tree it
         // serves, which would otherwise leave it answering for deleted nodes.
-        if let Some(mount) = self.mounts.lock().remove(&id)
+        // Taken in its own statement so the registry lock is released before the
+        // teardown: an edition-2024 `if let` chain holds the guard for the whole
+        // body, and `teardown` unmounts and joins the session thread — which can
+        // be inside a multi-second transfer. The shutdown path takes the same
+        // lock, so holding it here risks systemd's stop timeout landing mid-
+        // unmount and stranding the endpoint.
+        let taken = self.mounts.lock().remove(&id);
+        if let Some(mount) = taken
             && let Err(error) = commit_after_teardown(mount.teardown(), || ())
         {
             let _ = self.db.sync_folder_set_state(id, "error", now_secs());
@@ -584,19 +591,8 @@ impl Core {
     pub(crate) fn fork_state(&self) -> Core {
         let mut fork = self.clone();
         fork.primary = false;
-        let share_access = self.state.lock().share_access.clone();
-        fork.state = Arc::new(Mutex::new(State {
-            entries: HashMap::new(),
-            by_uid: HashMap::new(),
-            children: HashMap::new(),
-            next_ino: 2,
-            active_writes: HashMap::new(),
-            handles: HashMap::new(),
-            next_fh: 1,
-            access_changes: std::collections::HashSet::new(),
-            share_access,
-            db: self.db.clone(),
-        }));
+        let share_access = self.state().share_access.clone();
+        fork.state = Arc::new(Mutex::new(State::new(self.db.clone(), share_access, 2)));
         // A fresh inode space needs a fresh notification channel: this fork's
         // session is the only one that knows these inodes, so it must be the one
         // notified about them. Filled in by `spawn_ondemand_mount`.
@@ -866,8 +862,11 @@ impl Core {
                 Ok(format!("{} is now on-demand", local.display()))
             }
             ApplicableMountMode::Mirror => {
-                // ondemand→mirror: tear down the secondary mount first.
-                if let Some(mount) = self.mounts.lock().remove(&id)
+                // ondemand→mirror: tear down the secondary mount first. Taken
+                // before the teardown so the registry lock is not held across
+                // the unmount and thread join (see `remove_sync_folder`).
+                let taken = self.mounts.lock().remove(&id);
+                if let Some(mount) = taken
                     && let Err(error) = commit_after_teardown(mount.teardown(), || ())
                 {
                     let _ = self.db.sync_folder_set_state(id, "error", now_secs());

crates/pdfs-fuse/src/drain.rs

@@ -20,10 +20,14 @@
 //! are kept as a conflict copy rather than discarded — see
 //! [`Core::revision_conflict`].
 
+use std::collections::HashMap;
 use std::fs::File;
 use std::path::{Path, PathBuf};
-use std::sync::atomic::Ordering;
-use std::time::Duration;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::{Duration, Instant};
+
+use parking_lot::Mutex;
 
 use pdfs_core::batch;
 use pdfs_core::cache::{Baseline, StagedWrite};
@@ -36,17 +40,132 @@ use tracing::{debug, error, info, warn};
 use super::state::Intervals;
 use super::transfers::CountingReader;
 use super::{
-    Core, DRAIN_BACKOFF_MAX, DRAIN_BACKOFF_MIN, DRAIN_IDLE_POLL, ROOT_INO, conflict_name,
+    Core, DRAIN_BACKOFF_MAX, DRAIN_BACKOFF_MIN, DRAIN_IDLE_POLL, DRAIN_REVISION_DEBOUNCE,
+    DRAIN_REVISION_DEBOUNCE_MAX, ROOT_INO, UPLOAD_TIME_MEMORY, WriteAuthority, conflict_name,
     is_already_exists, is_gone, is_local_uid_str, media_type_for, node_revision_id, node_size,
     now_millis, now_secs, parse_node_uid,
 };
 
 const DRAIN_ACCESS_RECHECK: Duration = Duration::from_secs(5);
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+/// How long an op may stay access-deferred before the deferral is reported as a
+/// failure instead of retried in silence.
+///
+/// An access deferral costs no attempt and records no error, which is right for
+/// the case it was written for: a share downgraded while the queue was moving,
+/// undone a moment later. It is wrong for anything that does not clear. Until
+/// this window existed such an op was re-deferred every
+/// [`DRAIN_ACCESS_RECHECK`] indefinitely with `attempts` at zero, `last_error`
+/// null and only a `debug!` line to show for it, so `pdfs status` counted it
+/// among ordinary queued uploads and nothing ever said otherwise.
+///
+/// The condition that produced that in practice — a node absent from the local
+/// tree, which [`Db::effective_node_access`] used to report indistinguishably
+/// from a revoked permission — is now answered directly by
+/// [`Core::resolve_unknown_authority`]. This window remains the backstop for
+/// every deferral that does not resolve, including a refetch that keeps not
+/// helping.
+///
+/// Five minutes is far longer than a permission change takes to settle and far
+/// shorter than a user's patience with bytes that are not moving. Past it each
+/// recheck records a failure, so the op enters the ordinary backoff and, at
+/// [`FAILING_ATTEMPTS`], is counted and named by `Response::Status`.
+///
+/// [`Db::effective_node_access`]: pdfs_core::db::Db::effective_node_access
+/// [`FAILING_ATTEMPTS`]: pdfs_core::db::FAILING_ATTEMPTS
+const DRAIN_ACCESS_DEFER_LIMIT: Duration = Duration::from_secs(300);
+
+/// A registered cancellation flag for one in-flight upload, deregistered when
+/// dropped. See [`Core::begin_cancellable_upload`].
+struct UploadCancel {
+    registry: Arc<Mutex<HashMap<NodeUid, Arc<AtomicBool>>>>,
+    uid: NodeUid,
+    flag: Arc<AtomicBool>,
+}
+
+impl UploadCancel {
+    /// Whether this upload was told to stop. Distinguishes an upload we
+    /// abandoned on purpose from one the network broke.
+    fn cancelled(&self) -> bool {
+        self.flag.load(Ordering::Relaxed)
+    }
+
+    fn flag(&self) -> Arc<AtomicBool> {
+        self.flag.clone()
+    }
+}
+
+impl Drop for UploadCancel {
+    fn drop(&mut self) {
+        let mut registry = self.registry.lock();
+        // By identity: a supersede racing this upload's last bytes may already
+        // have registered the next one under the same uid.
+        if registry
+            .get(&self.uid)
+            .is_some_and(|current| Arc::ptr_eq(current, &self.flag))
+        {
+            registry.remove(&self.uid);
+        }
+    }
+}
+
+/// The debounce arithmetic of [`Core::revision_debounce`], without the map:
+/// the last measured upload time for this node, bounded on both sides.
+///
+/// A node nobody has uploaded yet gets the fixed grace period, which is the
+/// same answer the fixed debounce always gave.
+fn adaptive_debounce(measured: Option<Duration>) -> Duration {
+    measured
+        .unwrap_or(DRAIN_REVISION_DEBOUNCE)
+        .clamp(DRAIN_REVISION_DEBOUNCE, DRAIN_REVISION_DEBOUNCE_MAX)
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
 enum DrainDisposition {
     Applied,
     AccessDeferred,
+    /// An authority this op needs is missing from the local tree, so no access
+    /// answer exists for it yet. Carries the uid to ask the remote about.
+    AuthorityUnknown(NodeUid),
+}
+
+/// Why an op went back on the clock without consuming an attempt. Only affects
+/// what the user is told — every variant is deferred and reported identically.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum AccessDeferral {
+    /// An authority the tree knows about refuses the write.
+    NotWritable,
+    /// An authority missing from the tree was just re-read from the remote and
+    /// re-interned; the access check gets another look on the next pass.
+    AuthorityRefetched,
+    /// An authority missing from the tree could not be asked about.
+    RemoteUnreachable,
+}
+
+impl AccessDeferral {
+    fn log_reason(self) -> &'static str {
+        match self {
+            Self::NotWritable => "the node is not writable",
+            Self::AuthorityRefetched => "the node was missing locally and has been re-read",
+            Self::RemoteUnreachable => "the node is missing locally and the remote is unreachable",
+        }
+    }
+
+    /// Phrased to complete "<reason> for 300s." in a recorded `last_error`, so
+    /// the text the user sees names the actual condition rather than the
+    /// permission question it used to be flattened into.
+    fn failure_reason(self) -> &'static str {
+        match self {
+            Self::NotWritable => "the share has not been writable",
+            Self::AuthorityRefetched => {
+                "the node has been missing from the local tree, and re-reading it from the remote \
+                 has not restored it"
+            }
+            Self::RemoteUnreachable => {
+                "the node has been missing from the local tree and the remote has been unreachable"
+            }
+        }
+    }
 }
 
 fn pending_op_authorities(op: &PendingOp) -> Result<Vec<NodeUid>, Box<dyn std::error::Error>> {
@@ -89,7 +208,7 @@ fn pending_op_authorities(op: &PendingOp) -> Result<Vec<NodeUid>, Box<dyn std::e
 
 fn run_authorized_drain(
     op: &PendingOp,
-    mut require: impl FnMut(&NodeUid) -> Result<(), fuser::Errno>,
+    mut authority_of: impl FnMut(&NodeUid) -> WriteAuthority,
     apply: impl FnOnce() -> Result<(), Box<dyn std::error::Error>>,
 ) -> Result<DrainDisposition, Box<dyn std::error::Error>> {
     // This check is the admission/linearization point. A downgrade that lands
@@ -97,15 +216,14 @@ fn run_authorized_drain(
     // attempt. Every later queued operation rechecks. No permission lock is
     // held across remote latency.
     for authority in pending_op_authorities(op)? {
-        match require(&authority) {
-            Ok(()) => {}
-            Err(error) if error.code() == libc::EACCES => {
-                return Ok(DrainDisposition::AccessDeferred);
-            }
-            Err(error) => {
-                return Err(
-                    format!("drain access check failed with errno {}", error.code()).into(),
-                );
+        match authority_of(&authority) {
+            WriteAuthority::Writable => {}
+            WriteAuthority::Denied => return Ok(DrainDisposition::AccessDeferred),
+            // Nothing in the tree can answer for this node. Deferring on it
+            // waits for a permission change that has no reason to come, so the
+            // caller asks the remote instead (B83).
+            WriteAuthority::Unknown => {
+                return Ok(DrainDisposition::AuthorityUnknown(authority));
             }
         }
     }
@@ -127,44 +245,78 @@ impl Core {
     /// that is due — one file wedged against a folder that no longer exists must
     /// not hold up an unrelated upload behind it. The worker only blocks once
     /// nothing is due at all.
-    pub(crate) fn run_pending_drain(&self) {
+    ///
+    /// Several of these run at once ([`DRAIN_WORKERS`]), sharing the queue
+    /// through [`Db::claim_next_due_op`]: the alternative was that a 10 GiB
+    /// upload held every queued rename, trash and small write behind it for as
+    /// long as it took. Ordering only has to hold *per node*, and the claim
+    /// query is what guarantees it — no two workers are ever on one uid.
+    ///
+    /// `primary` marks the one worker that also runs the queue's idle chores.
+    /// They are cheap, but they are not per-op work and doing them in every
+    /// worker would just multiply them.
+    ///
+    /// [`DRAIN_WORKERS`]: super::DRAIN_WORKERS
+    /// [`Db::claim_next_due_op`]: pdfs_core::db::Db::claim_next_due_op
+    pub(crate) fn run_pending_drain(&self, primary: bool) {
         loop {
             let now = now_millis();
-            // One row, chosen by the database. Reading the whole queue to pick
-            // one op made a long queue quadratic to drain, and held the shared
-            // connection — and so every FUSE metadata call — for the duration.
-            let due = self.db.next_due_op(now).unwrap_or_default();
-
-            let Some(op) = due.filter(|_| self.online.load(Ordering::Relaxed)) else {
-                self.recover_fsynced_writes();
-                // A debounced or backed-off op may be waiting: sleep only
-                // until it becomes due rather than the full idle-poll.
-                self.wait_for_drain_work_or_due();
-                continue;
+            // One row, chosen and *claimed* by the database. Reading the whole
+            // queue to pick one op made a long queue quadratic to drain, and
+            // held the shared connection — and so every FUSE metadata call —
+            // for the duration.
+            let due = match self.db.claim_next_due_op(now) {
+                Ok(due) => due,
+                Err(error) => {
+                    error!(%error, "claiming the next pending operation failed");
+                    self.wait_for_drain_work();
+                    continue;
+                }
+            };
+            let online = self.online.load(Ordering::Relaxed);
+            let op = match due {
+                Some(op) if online => op,
+                idle => {
+                    // Claimed but offline: hand it straight back, or it stays
+                    // invisible to every worker until this one happens to be
+                    // the next to notice the reconnect.
+                    if let Some(op) = idle {
+                        self.release_claim(&op);
+                    }
+                    if primary {
+                        // Nothing to upload is exactly when the connection is
+                        // free, so this is where the read path's buffered LRU
+                        // touches get written (`ContentCache::flush_touches`).
+                        // Cheap and a no-op when empty.
+                        self.cache.flush_touches();
+                        self.recover_fsynced_writes();
+                    }
+                    // A debounced or backed-off op may be waiting: sleep only
+                    // until it becomes due rather than the full idle-poll.
+                    self.wait_for_drain_work_or_due();
+                    continue;
+                }
             };
             let outcome = run_authorized_drain(
                 &op,
-                |uid| self.require_uid_writable(uid),
+                |uid| self.uid_write_authority(uid),
                 || self.drain_op(&op),
             );
             if matches!(outcome, Ok(DrainDisposition::AccessDeferred)) {
-                let next_attempt_at = now_millis() + DRAIN_ACCESS_RECHECK.as_millis() as i64;
-                if let Err(error) = self.db.defer_op_without_attempt(op.id, next_attempt_at) {
-                    error!(
-                        uid = %op.uid,
-                        %error,
-                        "deferring access-blocked pending operation failed"
-                    );
-                    self.wait_for_drain_work();
-                } else {
-                    debug!(
-                        uid = %op.uid,
-                        next_attempt_at,
-                        "pending operation deferred until access is writable"
-                    );
-                }
+                self.defer_for_access(&op, AccessDeferral::NotWritable);
                 continue;
             }
+            if let Ok(DrainDisposition::AuthorityUnknown(ref authority)) = outcome {
+                self.resolve_unknown_authority(&op, authority);
+                continue;
+            }
+            // The access check let this attempt through, so any earlier run of
+            // deferrals is over and must not be counted against the next one.
+            if matches!(outcome, Ok(DrainDisposition::Applied))
+                && let Err(error) = self.db.clear_op_access_deferral(op.id)
+            {
+                debug!(uid = %op.uid, %error, "clearing an access-deferral window failed");
+            }
             if let Err(e) = outcome {
                 let attempts = op.attempts + 1;
                 let err_str = e.to_string();
@@ -184,9 +336,234 @@ impl Core {
                     now_millis() + backoff.as_millis() as i64,
                 ) {
                     error!(uid = %op.uid, error = %e, "recording a drain failure failed");
+                    self.release_claim(&op);
                     self.wait_for_drain_work();
+                    continue;
                 }
             }
+            // Unconditionally, on every path: a handler that retired its own row
+            // leaves nothing to release, and one that returned without retiring
+            // must not leave the row claimed by a worker that has moved on.
+            self.release_claim(&op);
+        }
+    }
+
+    /// Put an op the access check refused back on the clock, and decide whether
+    /// its refusal is still news.
+    ///
+    /// Inside [`DRAIN_ACCESS_DEFER_LIMIT`] a deferral is what it always was: no
+    /// attempt consumed, no error recorded, retried in five seconds. Past it the
+    /// deferral is reported through [`Db::record_op_failure`] like any other
+    /// stuck op, which is what puts it in `attempts`, in `last_error`, and — at
+    /// [`FAILING_ATTEMPTS`] — in the `failing_ops` count and error text that
+    /// `pdfs status` and the GUI show. The row and its staged blob stay exactly
+    /// where they are either way; this changes only whether the user can see
+    /// them.
+    ///
+    /// [`Db::record_op_failure`]: pdfs_core::db::Db::record_op_failure
+    /// [`FAILING_ATTEMPTS`]: pdfs_core::db::FAILING_ATTEMPTS
+    fn defer_for_access(&self, op: &PendingOp, reason: AccessDeferral) {
+        let now = now_millis();
+        let next_attempt_at = now + DRAIN_ACCESS_RECHECK.as_millis() as i64;
+        let since = match self.db.defer_op_for_access(op.id, now, next_attempt_at) {
+            Ok(since) => since,
+            Err(error) => {
+                error!(
+                    uid = %op.uid,
+                    %error,
+                    "deferring access-blocked pending operation failed"
+                );
+                self.release_claim(op);
+                self.wait_for_drain_work();
+                return;
+            }
+        };
+        let blocked = Duration::from_millis(now.saturating_sub(since).max(0) as u64);
+        if since == now {
+            // The first deferral of a run, at `warn!`: a write the mount accepted
+            // is not going anywhere for now, and that is worth a line even when
+            // the next recheck clears it.
+            warn!(
+                uid = %op.uid,
+                kind = %op.kind,
+                reason = reason.log_reason(),
+                "pending operation deferred"
+            );
+        }
+        if blocked >= DRAIN_ACCESS_DEFER_LIMIT {
+            let attempts = op.attempts + 1;
+            let backoff = DRAIN_BACKOFF_MIN
+                .saturating_mul(1u32 << attempts.min(6))
+                .min(DRAIN_BACKOFF_MAX);
+            let error = format!(
+                "{} for {}s. The staged bytes are kept.",
+                reason.failure_reason(),
+                blocked.as_secs()
+            );
+            warn!(uid = %op.uid, attempts, blocked_secs = blocked.as_secs(),
+                  "pending operation has been access-blocked past the limit; reporting it");
+            if let Err(error) =
+                self.db
+                    .record_op_failure(op.id, &error, now_millis() + backoff.as_millis() as i64)
+            {
+                error!(uid = %op.uid, %error, "recording an access-deferral failure failed");
+            }
+        } else {
+            debug!(
+                uid = %op.uid,
+                next_attempt_at,
+                blocked_secs = blocked.as_secs(),
+                reason = reason.log_reason(),
+                "pending operation deferred until access is writable"
+            );
+        }
+        self.release_claim(op);
+    }
+
+    /// Answer an authority the local tree has no row for by asking the remote,
+    /// rather than deferring on a permission question nobody asked (B83).
+    ///
+    /// Three outcomes, and the point of the whole exercise is that they are
+    /// three:
+    ///
+    /// * the node is still there — the tree simply lost it (a pruned subtree, a
+    ///   mount whose registration went away). Re-intern it and the next pass has
+    ///   a real access answer to work with.
+    /// * the node is gone remotely. That is a permanent, *known* condition, so
+    ///   it is recorded as a failure immediately instead of waiting out
+    ///   [`DRAIN_ACCESS_DEFER_LIMIT`] to say so. The row and its staged blob are
+    ///   kept, as on every other failure path — the bytes are still the user's,
+    ///   and `pdfs recover` is how they get them back.
+    /// * we could not ask. Ordinary deferral; the network is the network.
+    fn resolve_unknown_authority(&self, op: &PendingOp, authority: &NodeUid) {
+        match self.fetch_node_remote(authority) {
+            Ok(Some(node)) => {
+                if let Err(error) = self.db.upsert_node(&node) {
+                    error!(uid = %op.uid, %authority, %error,
+                           "re-interning a queued operation's missing authority failed");
+                } else {
+                    info!(uid = %op.uid, %authority,
+                          "re-interned an authority missing from the local tree; retrying");
+                }
+                // Deferred, not applied: the re-read has to go through the same
+                // access check as everything else, on the next pass. The window
+                // keeps running, so a refetch that never actually helps still
+                // gets reported instead of looping in silence.
+                self.defer_for_access(op, AccessDeferral::AuthorityRefetched);
+            }
+            Ok(None) => {
+                let attempts = op.attempts + 1;
+                let backoff = DRAIN_BACKOFF_MIN
+                    .saturating_mul(1u32 << attempts.min(6))
+                    .min(DRAIN_BACKOFF_MAX);
+                let error = format!(
+                    "{authority} no longer exists remotely, so this operation has nothing to \
+                     apply to. The staged bytes are kept."
+                );
+                warn!(uid = %op.uid, %authority, attempts,
+                      "a queued operation's authority is gone remotely; reporting it");
+                if let Err(error) = self.db.record_op_failure(
+                    op.id,
+                    &error,
+                    now_millis() + backoff.as_millis() as i64,
+                ) {
+                    error!(uid = %op.uid, %error, "recording a missing-authority failure failed");
+                }
+                self.release_claim(op);
+            }
+            Err(error) => {
+                debug!(uid = %op.uid, %authority, %error,
+                       "could not ask the remote about a missing authority");
+                self.defer_for_access(op, AccessDeferral::RemoteUnreachable);
+            }
+        }
+    }
+
+    /// Hand a claimed op back to the queue, logging rather than propagating —
+    /// there is no caller left to propagate to, and the drain must not stop.
+    fn release_claim(&self, op: &PendingOp) {
+        if let Err(error) = self.db.release_op_claim(op.id) {
+            error!(uid = %op.uid, %error, "releasing a drain claim failed");
+        }
+    }
+
+    /// How long a freshly queued revision of `uid` should wait before a worker
+    /// may pick it up.
+    ///
+    /// [`DRAIN_REVISION_DEBOUNCE`] exists for a tool that preallocates a file
+    /// and then writes it, where two seconds is the right scale. It is the wrong
+    /// scale for a large file on a slow link: an editor saving a 200 MB project
+    /// every ten seconds queues a write, the drain starts sending it, the next
+    /// save supersedes it mid-flight, and the file never finishes uploading
+    /// while the user keeps working. Waiting roughly as long as the last upload
+    /// of *this* node took moves that supersede into the queue, where replacing
+    /// a row costs nothing.
+    ///
+    /// Never shorter than the fixed debounce and never longer than
+    /// [`DRAIN_REVISION_DEBOUNCE_MAX`]: a node whose upload is genuinely slow
+    /// still has to reach Drive, and a single pathological stall must not park
+    /// it.
+    ///
+    /// [`DRAIN_REVISION_DEBOUNCE`]: super::DRAIN_REVISION_DEBOUNCE
+    /// [`DRAIN_REVISION_DEBOUNCE_MAX`]: super::DRAIN_REVISION_DEBOUNCE_MAX
+    pub(crate) fn revision_debounce(&self, uid: &NodeUid) -> Duration {
+        let measured = self
+            .upload_times
+            .lock()
+            .get(&uid.to_string())
+            .map(|&(ms, _)| Duration::from_millis(ms));
+        adaptive_debounce(measured)
+    }
+
+    /// Remember how long an upload of `uid` took, for the next write's
+    /// debounce.
+    ///
+    /// Only a *completed* upload is a measurement. A cancelled or failed one
+    /// says how long it ran before it stopped, which is not how long the file
+    /// takes to send, and feeding that back would let one network fault widen
+    /// the debounce for good.
+    fn record_upload_time(&self, uid: &NodeUid, took: Duration) {
+        let mut times = self.upload_times.lock();
+        let now = now_millis();
+        if times.len() >= UPLOAD_TIME_MEMORY
+            && !times.contains_key(&uid.to_string())
+            && let Some(oldest) = times
+                .iter()
+                .min_by_key(|(_, (_, at))| *at)
+                .map(|(k, _)| k.clone())
+        {
+            times.remove(&oldest);
+        }
+        times.insert(uid.to_string(), (took.as_millis() as u64, now));
+    }
+
+    /// Register a cancellation flag for an upload of `uid` that is about to
+    /// start, returning a guard that deregisters it.
+    ///
+    /// Deregistration is by identity rather than by key: a supersede that
+    /// arrives just as this upload finishes may already have registered the
+    /// *next* upload's flag under the same uid, and removing that one would
+    /// leave the newer upload uncancellable.
+    fn begin_cancellable_upload(&self, uid: &NodeUid) -> UploadCancel {
+        let flag = Arc::new(AtomicBool::new(false));
+        self.upload_cancel.lock().insert(uid.clone(), flag.clone());
+        UploadCancel {
+            registry: self.upload_cancel.clone(),
+            uid: uid.clone(),
+            flag,
+        }
+    }
+
+    /// Tell an upload of `uid` that is on the wire to stop, if there is one.
+    ///
+    /// Called when a newer revision of the file is queued. The upload does not
+    /// stop instantly — it stops when the SDK next asks its reader for bytes,
+    /// which for a 4 MiB block is soon enough to matter and late enough to be
+    /// safe: the staged blob it is reading stays open until it lets go.
+    pub(crate) fn cancel_upload(&self, uid: &NodeUid) {
+        if let Some(flag) = self.upload_cancel.lock().get(uid) {
+            flag.store(true, Ordering::Relaxed);
+            debug!(%uid, "cancelling a superseded upload");
         }
     }
 
@@ -351,6 +728,9 @@ impl Core {
         }
         self.db.complete_trash_op(op.id, &uid)?;
         self.own_sealed_revs.lock().remove(&uid);
+        if let Err(e) = self.db.clear_own_sealed_rev(&uid.to_string()) {
+            debug!(%uid, error = %e, "clearing the sealed-revision record failed");
+        }
         self.invalidate_trash();
         self.log_activity(ActivityKind::Trash, &name, "trashed", true);
         info!(%uid, name, "pending trash landed");
@@ -625,7 +1005,16 @@ impl Core {
         // instead of forking — so the finished bytes win the name rather than
         // being exiled to a `(sync-conflict)` copy.
         let remote_rev = node_revision_id(&node);
-        let own_rev = self.own_sealed_revs.lock().get(uid).cloned();
+        // In memory first, then the table it is written through to: the map is
+        // this process's cache of it, and after a restart only the table has
+        // the answer (which is the whole reason it is persisted).
+        let own_rev = match self.own_sealed_revs.lock().get(uid).cloned() {
+            Some(rev) => Some(rev),
+            None => self
+                .db
+                .own_sealed_rev(&uid.to_string(), now_millis())
+                .unwrap_or_default(),
+        };
         if is_own_self_supersede(meta.complete, remote_rev.as_deref(), own_rev.as_deref()) {
             debug!(%uid, ?remote_rev,
                    "queued write chains onto our own sealed revision; not a conflict");
@@ -709,8 +1098,9 @@ impl Core {
         self.db.delete_op(op.id)?;
         // Dropping the pending entry hands the node back to the remote's truth:
         // reads stop coming from the staged blob, and the event sync stops
-        // skipping it as "ahead of the server" (offline.md Phase 3a).
-        self.pending.lock().remove(uid);
+        // skipping it as "ahead of the server" (offline.md Phase 3a). Only if it
+        // is still *this* write's entry — see [`Core::release_pending`].
+        self.release_pending(uid, blob);
         self.cache.discard_staged(blob);
         self.cache.evict(uid);
         self.evict_reader(uid);
@@ -748,7 +1138,7 @@ impl Core {
         error!(%uid, name, reason, staged = %blob.display(),
                "cannot place a conflicted write; bytes kept in staging");
         self.db.delete_op(op.id)?;
-        self.pending.lock().remove(uid);
+        self.release_pending(uid, blob);
         self.cache.evict(uid);
         self.evict_reader(uid);
         self.log_activity(
@@ -760,6 +1150,24 @@ impl Core {
         Ok(())
     }
 
+    /// Drop `uid`'s pending entry, but only while it still names `blob`.
+    ///
+    /// A supersede can replace the entry while the op that owned it is in
+    /// flight, so an unconditional `remove` retires a *newer* write's staged
+    /// blob: reads fall back to a stale remote, the next `remote_baseline` forks
+    /// a spurious `(sync-conflict)` copy of bytes that were never in conflict,
+    /// and the incomplete-write interlock — which keys off the pending entry —
+    /// is disarmed, letting a gap-fill read from a revision that no longer
+    /// describes the file.
+    pub(crate) fn release_pending(&self, uid: &NodeUid, blob: &Path) {
+        let mut pending = self.pending.lock();
+        if let Some(current) = pending.get(uid)
+            && current.path == blob
+        {
+            pending.remove(uid);
+        }
+    }
+
     /// Whether a folder a queued op targets has stopped being a place a node can
     /// go. A failure to ask is not an answer: only a definite "trashed" or "not
     /// there" counts, so a network fault leaves the op to retry normally.
@@ -826,7 +1234,7 @@ impl Core {
     /// root yet, which is not where the drain runs; the drain must not panic
     /// over it regardless, since that would stop the queue for good.
     pub(crate) fn root_uid(&self) -> Option<NodeUid> {
-        let st = self.state.lock();
+        let st = self.state();
         st.entries.get(&ROOT_INO).map(|e| e.uid.clone())
     }
 
@@ -872,28 +1280,43 @@ impl Core {
         let guard =
             self.transfers
                 .begin(&name, meta.uid.clone(), TransferDirection::Upload, meta.len);
-        let reader = CountingReader::new(File::open(&blob)?, &guard);
-        self.rt.block_on(self.client.upload_new_revision_from(
+        // Registered before the first byte moves, so a write that lands while
+        // this is on the wire can stop it rather than paying for a revision the
+        // op it queued will immediately replace.
+        let cancel = self.begin_cancellable_upload(&uid);
+        let reader = CountingReader::new(File::open(&blob)?, &guard).with_cancel(cancel.flag());
+        let started = Instant::now();
+        let sent = self.rt.block_on(self.client.upload_new_revision_from(
             &uid,
             reader,
             meta.len as i64,
             Vec::new(),
             None,
-        ))?;
+        ));
+        let took = started.elapsed();
+        if let Err(e) = sent {
+            if cancel.cancelled() {
+                // Not a failure, and not this op's row to retire: the write that
+                // cancelled us deleted it and took ownership of the staged blob
+                // when it superseded this revision. Leaving the blob alone here
+                // is the whole point — it is either already unlinked or it
+                // belongs to the newer op.
+                info!(%uid, "upload abandoned; a newer write superseded it");
+                return Ok(());
+            }
+            return Err(e.into());
+        }
+        drop(cancel);
         drop(guard);
+        // Only a completed upload times how long this file takes to send, which
+        // is what the next write's debounce is scaled from.
+        self.record_upload_time(&uid, took);
 
         // Retire the op before dropping the blob: a crash between the two leaves
         // an orphaned file (harmless), whereas the reverse would leave a queued
         // op pointing at nothing.
         self.db.delete_op(op.id)?;
-        {
-            let mut p = self.pending.lock();
-            if let Some(op_in_mem) = p.get(&uid)
-                && op_in_mem.path == blob
-            {
-                p.remove(&uid);
-            }
-        }
+        self.release_pending(&uid, &blob);
 
         // The staged blob now matches the sealed revision, so a pinned file keeps
         // it as its cached content rather than re-downloading what we just sent.
@@ -969,7 +1392,16 @@ impl Core {
             // change and does not fork (B70 layer B; consulted in
             // `revision_conflict`).
             if let Some(rev) = sealed_rev.clone() {
-                self.own_sealed_revs.lock().insert(uid.clone(), rev);
+                self.own_sealed_revs.lock().insert(uid.clone(), rev.clone());
+                // Write through: losing this at shutdown reopened the fork
+                // window B70 layer B closed, for the first drain after a
+                // restart.
+                if let Err(e) = self
+                    .db
+                    .set_own_sealed_rev(&uid.to_string(), &rev, now_millis())
+                {
+                    warn!(%uid, error = %e, "persisting our sealed revision failed");
+                }
             }
             self.for_each_state(|st| {
                 for aw in st.active_writes.values_mut() {
@@ -1190,12 +1622,10 @@ mod tests {
         let applied = Cell::new(false);
         let disposition = run_authorized_drain(
             &op,
-            |uid| {
-                db.effective_node_access(uid)
-                    .unwrap()
-                    .filter(|access| access.writable())
-                    .map(|_| ())
-                    .ok_or(fuser::Errno::EACCES)
+            |uid| match db.effective_node_access(uid).unwrap() {
+                Some(access) if access.writable() => WriteAuthority::Writable,
+                Some(_) => WriteAuthority::Denied,
+                None => WriteAuthority::Unknown,
             },
             || {
                 applied.set(true);
@@ -1235,7 +1665,7 @@ mod tests {
 
         let disposition = run_authorized_drain(
             &op,
-            |_| Err(fuser::Errno::EACCES),
+            |_| WriteAuthority::Denied,
             || {
                 applied.set(true);
                 Ok(())
@@ -1256,13 +1686,157 @@ mod tests {
         let _ = std::fs::remove_dir_all(dir);
     }
 
+    /// A deferral that never clears has to stop being silent. Found in the field
+    /// as revision ops whose node had left the local tree: `require_uid_access`
+    /// reads an unknown node as `EACCES`, so they were re-deferred every five
+    /// seconds for thirty days with `attempts` at zero, `last_error` null and
+    /// `failing_ops` at zero — 18 GiB of accepted writes that `pdfs status`
+    /// reported as ordinary queued uploads.
+    #[test]
+    fn an_access_deferral_that_outlives_the_limit_is_reported_as_failing() {
+        use std::sync::atomic::{AtomicU64, Ordering};
+
+        static N: AtomicU64 = AtomicU64::new(0);
+        let dir = std::env::temp_dir().join(format!(
+            "pdfs-drain-defer-limit-{}-{}",
+            std::process::id(),
+            N.fetch_add(1, Ordering::Relaxed)
+        ));
+        std::fs::create_dir_all(&dir).unwrap();
+        let db = Db::open(&dir.join("cache.db")).unwrap();
+        db.enqueue_op(&pending(OP_REVISION)).unwrap();
+        let op = db.pending_ops().unwrap().remove(0);
+
+        // Two deferrals a recheck apart: the window opens on the first and the
+        // second is measured against it, not against itself.
+        let now = 1_000_000i64;
+        let since = db.defer_op_for_access(op.id, now, now + 5_000).unwrap();
+        assert_eq!(since, now, "the first deferral opens the window");
+        let later = now + DRAIN_ACCESS_RECHECK.as_millis() as i64;
+        assert_eq!(
+            db.defer_op_for_access(op.id, later, later + 5_000).unwrap(),
+            now,
+            "a later deferral keeps the original stamp"
+        );
+        let queued = db.pending_ops().unwrap();
+        assert_eq!(queued[0].attempts, 0, "deferrals consume no attempt");
+        assert!(queued[0].last_error.is_none());
+
+        // Past the limit the same deferral is reported like any other stuck op.
+        let past = now + DRAIN_ACCESS_DEFER_LIMIT.as_millis() as i64;
+        assert_eq!(
+            db.defer_op_for_access(op.id, past, past + 5_000).unwrap(),
+            now
+        );
+        db.record_op_failure(op.id, "not writable locally", past + 10_000)
+            .unwrap();
+        let failed = db.pending_ops().unwrap();
+        assert_eq!(failed.len(), 1, "the row and its staged bytes stay queued");
+        assert_eq!(failed[0].attempts, 1);
+        assert_eq!(
+            failed[0].last_error.as_deref(),
+            Some("not writable locally")
+        );
+
+        // Reporting closes the window, so the next deferral starts a fresh one
+        // rather than escalating on every recheck from here on.
+        let after = past + 20_000;
+        assert_eq!(
+            db.defer_op_for_access(op.id, after, after + 5_000).unwrap(),
+            after,
+            "recording the failure re-armed the window"
+        );
+        drop(db);
+        let _ = std::fs::remove_dir_all(dir);
+    }
+
+    /// An op that gets through the access check is no longer blocked, so a
+    /// deferral after it must not inherit the earlier run's age and escalate
+    /// immediately.
+    #[test]
+    fn clearing_the_window_makes_the_next_deferral_a_fresh_run() {
+        use std::sync::atomic::{AtomicU64, Ordering};
+
+        static N: AtomicU64 = AtomicU64::new(0);
+        let dir = std::env::temp_dir().join(format!(
+            "pdfs-drain-defer-clear-{}-{}",
+            std::process::id(),
+            N.fetch_add(1, Ordering::Relaxed)
+        ));
+        std::fs::create_dir_all(&dir).unwrap();
+        let db = Db::open(&dir.join("cache.db")).unwrap();
+        db.enqueue_op(&pending(OP_REVISION)).unwrap();
+        let op = db.pending_ops().unwrap().remove(0);
+
+        let now = 2_000_000i64;
+        assert_eq!(
+            db.defer_op_for_access(op.id, now, now + 5_000).unwrap(),
+            now
+        );
+        db.clear_op_access_deferral(op.id).unwrap();
+        let later = now + DRAIN_ACCESS_DEFER_LIMIT.as_millis() as i64;
+        assert_eq!(
+            db.defer_op_for_access(op.id, later, later + 5_000).unwrap(),
+            later,
+            "the window reopens at the new deferral, not the old one"
+        );
+        drop(db);
+        let _ = std::fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn an_authority_the_tree_does_not_have_is_reported_as_unknown_not_deferred() {
+        // The B83 distinction, at the point it is made: a uid with no row must
+        // not come back as `AccessDeferred`, because deferring waits on a
+        // permission change that nothing is going to make. The caller answers
+        // an `AuthorityUnknown` by asking the remote instead.
+        let op = pending(OP_REVISION);
+        let applied = Cell::new(false);
+        let disposition = run_authorized_drain(
+            &op,
+            |_| WriteAuthority::Unknown,
+            || {
+                applied.set(true);
+                Ok(())
+            },
+        )
+        .unwrap();
+        assert_eq!(
+            disposition,
+            DrainDisposition::AuthorityUnknown(parse_node_uid(&op.uid).unwrap())
+        );
+        assert!(!applied.get(), "an unresolved authority never applies");
+    }
+
+    #[test]
+    fn a_missing_parent_authority_names_the_parent_not_the_node() {
+        // A create's authority is its parent, so that is the uid the remote has
+        // to be asked about — reporting the node instead would refetch the
+        // thing that does not exist yet.
+        let mut op = pending(OP_CREATE);
+        let parent = folder_node("vanished-parent", None);
+        op.parent_uid = Some(parent.uid.to_string());
+        let disposition =
+            run_authorized_drain(&op, |_| WriteAuthority::Unknown, || Ok(())).unwrap();
+        assert_eq!(
+            disposition,
+            DrainDisposition::AuthorityUnknown(parent.uid.clone())
+        );
+    }
+
     #[test]
     fn owner_and_editor_drain_authorities_apply_normally() {
         for access in [Access::Owner, Access::Editor] {
             let applied = Cell::new(false);
             let disposition = run_authorized_drain(
                 &pending(OP_REVISION),
-                |_| access.writable().then_some(()).ok_or(fuser::Errno::EACCES),
+                |_| {
+                    if access.writable() {
+                        WriteAuthority::Writable
+                    } else {
+                        WriteAuthority::Denied
+                    }
+                },
                 || {
                     applied.set(true);
                     Ok(())
@@ -1376,3 +1950,42 @@ mod tests {
         assert!(!is_own_self_supersede(true, None, None));
     }
 }
+
+#[cfg(test)]
+mod debounce_tests {
+    use super::*;
+
+    #[test]
+    fn a_node_never_uploaded_keeps_the_fixed_grace_period() {
+        assert_eq!(adaptive_debounce(None), DRAIN_REVISION_DEBOUNCE);
+    }
+
+    #[test]
+    fn a_fast_upload_does_not_shorten_the_grace_period() {
+        // The 2 seconds are not about upload time at all — they are the window
+        // in which a preallocate-then-write tool supersedes its own first
+        // close. A file that sends in 50 ms must still get them.
+        assert_eq!(
+            adaptive_debounce(Some(Duration::from_millis(50))),
+            DRAIN_REVISION_DEBOUNCE
+        );
+    }
+
+    #[test]
+    fn a_slow_upload_widens_the_grace_period_to_match() {
+        // The case the fixed debounce got wrong: saves arriving faster than the
+        // upload completes, each superseding one already on the wire.
+        assert_eq!(
+            adaptive_debounce(Some(Duration::from_secs(25))),
+            Duration::from_secs(25)
+        );
+    }
+
+    #[test]
+    fn a_pathological_upload_cannot_park_a_node() {
+        assert_eq!(
+            adaptive_debounce(Some(Duration::from_secs(6000))),
+            DRAIN_REVISION_DEBOUNCE_MAX
+        );
+    }
+}

crates/pdfs-fuse/src/filesystem.rs

@@ -3,6 +3,73 @@
 use super::*;
 
 impl Filesystem for ProtonFs {
+    /// Negotiate kernel-side behaviour once, before any other callback.
+    ///
+    /// Everything here is a hint the kernel is free to decline, so each is
+    /// requested and its refusal logged rather than treated as fatal — an older
+    /// kernel that lacks a capability must still mount.
+    ///
+    /// Deliberately *not* requested: `FUSE_ATOMIC_O_TRUNC` and
+    /// `FUSE_DO_READDIRPLUS` change what the kernel sends (truncation folded
+    /// into `open`, `readdirplus` instead of `readdir` + N lookups), so they
+    /// need the matching handlers first; and `FOPEN_KEEP_CACHE` needs the write
+    /// path to invalidate page cache on every locally staged revision, not just
+    /// on remote events.
+    fn init(&mut self, _req: &Request, config: &mut KernelConfig) -> std::io::Result<()> {
+        // Readahead the kernel issues per file. Matching the block size means a
+        // sequential reader's readahead lands inside the block a demand read is
+        // already fetching, instead of one 128 KiB slice at a time.
+        if let Err(nearest) = config.set_max_readahead(BLOCK_SIZE as u32) {
+            debug!(nearest, "kernel capped max_readahead");
+            let _ = config.set_max_readahead(nearest);
+        }
+        if let Err(nearest) = config.set_max_background(MAX_BACKGROUND) {
+            debug!(nearest, "kernel capped max_background");
+        }
+        // Without this the kernel serialises lookups within one directory, so a
+        // cold `ls` of a folder is a chain of round trips rather than a fan-out.
+        if let Err(unsupported) = config.add_capabilities(InitFlags::FUSE_PARALLEL_DIROPS) {
+            debug!(?unsupported, "kernel does not support parallel dirops");
+        }
+        if let Err(unsupported) = config.add_capabilities(InitFlags::FUSE_CACHE_SYMLINKS) {
+            debug!(?unsupported, "kernel does not support symlink caching");
+        }
+        Ok(())
+    }
+
+    /// Free and total space, as `df` and every "will this fit?" preflight ask.
+    ///
+    /// Reported in [`REPORTED_BLKSIZE`] units against the *account* quota, not
+    /// the local cache: what a copy into the mount consumes is Drive storage.
+    /// The default implementation answers all-zeroes, which reads as a full
+    /// filesystem and makes installers and file managers refuse to write.
+    ///
+    /// On a worker, because a cold answer is a round trip
+    /// ([`Core::account_quota_cached`] serves a warm one from memory).
+    fn statfs(&self, _req: &Request, _ino: INodeNo, reply: ReplyStatfs) {
+        let core = self.core.clone();
+        self.core.workers.run(Lane::Meta, move || {
+            let (max_space, used_space) = core.account_quota_cached().unwrap_or((0, 0));
+            let bsize = u64::from(REPORTED_BLKSIZE);
+            let blocks = (max_space.max(0) as u64).div_ceil(bsize);
+            let used = (used_space.max(0) as u64).div_ceil(bsize);
+            let free = blocks.saturating_sub(used);
+            // `files`/`ffree` are unknowable — Drive has no inode budget — so
+            // they are reported as "plenty" rather than as zero, which some
+            // tools read as "no more files may be created".
+            reply.statfs(
+                blocks,
+                free,
+                free,
+                0,
+                u64::MAX,
+                REPORTED_BLKSIZE,
+                MAX_NAME_LEN,
+                REPORTED_BLKSIZE,
+            );
+        });
+    }
+
     fn lookup(&self, _req: &Request, parent: INodeNo, name: &OsStr, reply: ReplyEntry) {
         let parent = parent.0;
         let name = match fuse_name(name) {
@@ -27,13 +94,13 @@ impl Filesystem for ProtonFs {
     }
 
     fn forget(&self, _req: &Request, ino: INodeNo, nlookup: u64) {
-        let mut st = self.core.state.lock();
+        let mut st = self.core.state();
         st.forget_lookup(ino.0, nlookup);
     }
 
     fn getattr(&self, _req: &Request, ino: INodeNo, _fh: Option<FileHandle>, reply: ReplyAttr) {
         let (attr, provisional_in) = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.entries.get(&ino.0) {
                 Some(e) => {
                     // A file still carrying the cheap enumeration's placeholder
@@ -69,23 +136,30 @@ impl Filesystem for ProtonFs {
         // single upgrade. This is why B12's split is still worth having — the
         // plain listing never comes here at all.
         //
-        // Off the dispatch loop, because it goes to the network: blocking there
-        // would stall every other operation on the mount (the B5 lesson).
+        // Starting the batch is local work — a `state` read and a thread spawn —
+        // so it happens here; what goes to the network is the batch itself, and
+        // this reply is parked on it rather than waiting on any thread of the
+        // mount's (audit F4).
+        let ino = ino.0;
         let fs = self.clone();
-        self.core.workers.run(Lane::Meta, move || {
-            fs.core.upgrade_sizes_for_parent(ino.0, &uid, parent);
+        let respond = move || {
             // Re-read: the upgrade writes through `state`, and on timeout or
             // failure this is simply the provisional attr we already had.
             let attr = {
-                let st = fs.core.state.lock();
-                match st.entries.get(&ino.0) {
-                    Some(e) => fs.attr(ino.0, &e.node, e.access),
+                let st = fs.core.state();
+                match st.entries.get(&ino) {
+                    Some(e) => fs.attr(ino, &e.node, e.access),
                     // Forgotten while we waited.
                     None => attr,
                 }
             };
             reply.attr(&TTL, &attr);
-        });
+        };
+        match self.core.upgrade_sizes_for_parent(ino, &uid, parent) {
+            Some(slot) => self.core.await_size(ino, slot, respond),
+            // Nothing to upgrade, so nothing to wait for.
+            None => respond(),
+        }
     }
 
     fn readdir(
@@ -109,109 +183,19 @@ impl Filesystem for ProtonFs {
             .run(Lane::Meta, move || fs.serve_readdir(ino, offset, reply));
     }
 
+    /// A read-only open is a map lookup and answers inline; a write open has to
+    /// allocate a scratch file and, when a queued revision is the base, duplicate
+    /// a blob that is the whole file — so it is served from a worker rather than
+    /// stalling every other request on the mount behind it.
     fn open(&self, _req: &Request, ino: INodeNo, flags: OpenFlags, reply: ReplyOpen) {
-        let write_requested = flags.acc_mode() != OpenAccMode::O_RDONLY;
-        let (uid, base_mtime, base_size, base_revision_id) = {
-            let mut st = self.core.state.lock();
-            match st.entries.get_mut(&ino.0) {
-                Some(e) if e.node.is_file() => {
-                    if write_requested && !e.writable() {
-                        reply.error(Errno::EACCES);
-                        return;
-                    }
-                    e.open_count = e.open_count.saturating_add(1);
-                    (
-                        e.uid.clone(),
-                        e.node.modification_time,
-                        node_size(&e.node),
-                        node_revision_id(&e.node),
-                    )
-                }
-                Some(_) => {
-                    reply.error(Errno::EISDIR);
-                    return;
-                }
-                None => {
-                    reply.error(Errno::ENOENT);
-                    return;
-                }
-            }
-        };
-        debug!(ino = ino.0, ?flags, base_size, base_mtime, "open");
-        // Read-only opens stay stateless (fh 0). A write open allocates a
-        // disk-backed handle; bytes are authored into a scratch file and the
-        // untouched remainder is pulled from the base lazily (on read / commit).
         if flags.acc_mode() == OpenAccMode::O_RDONLY {
-            reply.opened(FileHandle(0), FopenFlags::empty());
+            self.serve_open(ino, flags, reply);
             return;
         }
-        // A previous close may already have newer bytes queued locally while
-        // the server still holds the older revision. That staged blob, not the
-        // remote, is the base for this handle. In particular, `truncate file`
-        // opens for writing and then shrinks via setattr; starting its scratch
-        // as a sparse zero file would turn the preserved prefix into zeros.
-        //
-        // Complete pending blobs are the common case and can be copied without
-        // any network access. An incomplete blob still has gaps referring to
-        // the remote base; stacking another write on it cannot be represented
-        // safely by WriteHandle, so fail the open rather than risk corruption.
-        let pending_base = self.core.pending.lock().get(&uid).cloned();
-        if pending_base.as_ref().is_some_and(|p| !p.meta.complete) {
-            if let Some(entry) = self.core.state.lock().entries.get_mut(&ino.0) {
-                entry.open_count = entry.open_count.saturating_sub(1);
-            }
-            error!(%uid, "refusing write over incomplete queued revision");
-            reply.error(Errno::EIO);
-            return;
-        }
-        let (file, path) = match self.core.cache.create_scratch() {
-            Ok(x) => x,
-            Err(e) => {
-                if let Some(entry) = self.core.state.lock().entries.get_mut(&ino.0) {
-                    entry.open_count = entry.open_count.saturating_sub(1);
-                }
-                error!(%uid, error = %e, "create scratch file failed");
-                reply.error(Errno::EIO);
-                return;
-            }
-        };
-        let mut initial_written = Intervals::default();
-        if let Some(pending) = &pending_base {
-            if let Err(e) = std::fs::copy(&pending.path, &path) {
-                if let Some(entry) = self.core.state.lock().entries.get_mut(&ino.0) {
-                    entry.open_count = entry.open_count.saturating_sub(1);
-                }
-                let _ = std::fs::remove_file(&path);
-                error!(%uid, source = %pending.path.display(), error = %e,
-                    "copy queued revision into write scratch failed");
-                reply.error(Errno::EIO);
-                return;
-            }
-            initial_written.add(0, pending.meta.len);
-        }
-        let mut st = self.core.state.lock();
-        let fh = st.next_fh;
-        st.next_fh += 1;
-        let aw = st.active_writes.entry(ino.0).or_insert_with(|| {
-            WriteHandle {
-                ino: ino.0,
-                uid,
-                file: Arc::new(file),
-                path,
-                written: initial_written,
-                // Starts at the current size; reads in [0, base_size) come from
-                // the base until overwritten.
-                len: base_size,
-                base_size,
-                base_mtime,
-                base_revision_id,
-                dirty: false,
-                open_count: 0,
-            }
-        });
-        aw.open_count += 1;
-        st.handles.insert(fh, ino.0);
-        reply.opened(FileHandle(fh), FopenFlags::empty());
+        let fs = self.clone();
+        self.core
+            .workers
+            .run(Lane::Transfer, move || fs.serve_open(ino, flags, reply));
     }
 
     fn read(
@@ -234,7 +218,7 @@ impl Filesystem for ProtonFs {
         // has always had between a write and a read of the same handle; a mostly
         // local read is not worth reasoning about that for.
         let handle = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             st.active_writes.get(&ino.0).map(|h| {
                 (
                     h.file.clone(),
@@ -263,7 +247,7 @@ impl Filesystem for ProtonFs {
             return;
         }
         let (uid, mtime, fsize, is_video) = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.entries.get(&ino.0) {
                 Some(e) if e.node.is_file() => {
                     let media_type = match &e.node.kind {
@@ -363,7 +347,7 @@ impl Filesystem for ProtonFs {
         // Stage the bytes straight into the scratch file (no base download): only
         // the untouched remainder is pulled from the remote, and only at commit.
         let file = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.handles.get(&fh).and_then(|&i| st.active_writes.get(&i)) {
                 Some(aw) => aw.file.clone(),
                 None => {
@@ -374,11 +358,20 @@ impl Filesystem for ProtonFs {
         };
         if let Err(e) = file.write_all_at(data, offset) {
             error!(ino = ino.0, fh, error = %e, "scratch write failed");
-            reply.error(Errno::EIO);
+            // A full local disk is not a broken file: answer `ENOSPC` so the
+            // application can say so (and so `cp`/editors take their own
+            // out-of-space path), and free what the cache is holding — the
+            // scratch file the next write goes to shares that filesystem.
+            if e.raw_os_error() == Some(libc::ENOSPC) {
+                self.core.cache.emergency_evict();
+                reply.error(Errno::ENOSPC);
+            } else {
+                reply.error(Errno::EIO);
+            }
             return;
         }
         let new_len = {
-            let mut st = self.core.state.lock();
+            let mut st = self.core.state();
             let Some(aw) = st
                 .handles
                 .get(&fh)
@@ -435,7 +428,7 @@ impl Filesystem for ProtonFs {
         if let Some(size) = size {
             let fh = fh.map(|f| f.0).filter(|&f| f != 0);
             let handled = {
-                let mut st = self.core.state.lock();
+                let mut st = self.core.state();
                 match st.active_writes.get_mut(&ino.0) {
                     Some(aw) => {
                         if size < aw.len {
@@ -478,7 +471,7 @@ impl Filesystem for ProtonFs {
                 });
                 return;
             }
-            self.core.state.lock().set_size(ino.0, size);
+            self.core.state().set_size(ino.0, size);
         }
         self.reply_attr(ino.0, reply);
     }
@@ -503,7 +496,7 @@ impl Filesystem for ProtonFs {
         let is_punch_hole = (mode & 0x02) != 0; // FALLOC_FL_PUNCH_HOLE
 
         let res = {
-            let mut st = self.core.state.lock();
+            let mut st = self.core.state();
             match st
                 .handles
                 .get(&_fh.0)
@@ -566,6 +559,12 @@ impl Filesystem for ProtonFs {
     /// for fsync wants its bytes to survive a crash, and blocking it on an upload
     /// (which offline would never finish) buys nothing the queue does not already
     /// guarantee.
+    ///
+    /// Served on a worker: `sync_all` is a disk barrier and the sidecar write is
+    /// two more file operations, and an application that fsyncs per record would
+    /// otherwise stall every request on the mount at its barrier rate. Nothing
+    /// orders this against other callbacks — the kernel waits for the reply
+    /// either way, which is the whole contract.
     fn fsync(
         &self,
         _req: &Request,
@@ -574,76 +573,10 @@ impl Filesystem for ProtonFs {
         _datasync: bool,
         reply: ReplyEmpty,
     ) {
-        // Everything the durability sidecar needs, copied out so no lock is held
-        // across the sync.
-        let handle = {
-            let st = self.core.state.lock();
-            let ino = st.handles.get(&fh.0).copied();
-            ino.and_then(|i| st.active_writes.get(&i)).map(|aw| {
-                (
-                    aw.file.clone(),
-                    aw.path.clone(),
-                    aw.uid.clone(),
-                    aw.dirty,
-                    aw.written.clone(),
-                    aw.len,
-                    aw.base_size,
-                    aw.base_mtime,
-                    aw.base_revision_id.clone(),
-                )
-            })
-        };
-        let Some((f, path, uid, dirty, written, len, base_size, base_mtime, base_revision_id)) =
-            handle
-        else {
-            reply.error(Errno::EBADF);
-            return;
-        };
-        if let Err(e) = f.sync_all() {
-            error!(fh = fh.0, error = %e, "fsync of scratch file failed");
-            reply.error(Errno::EIO);
-            return;
-        }
-        // A clean handle has nothing on it that the remote does not already
-        // hold, so there is nothing for a crash to lose.
-        if !dirty {
-            reply.ok();
-            return;
-        }
-        // The bytes are on stable storage, but only the scratch file knows that,
-        // and the scratch directory is cleared at open. Record what the blob is
-        // so a restart can hand it to the upload queue instead of deleting it.
-        //
-        // Deliberately *not* the full release path: staging the blob here would
-        // move the file out from under a handle the application still has open,
-        // and queueing an upload per `fsync` would push a revision for every
-        // barrier in a write loop. This only makes the existing file findable.
-        let authored: Vec<(u64, u64)> = written
-            .segments(0, len)
-            .into_iter()
-            .filter(|&(_, _, authored)| authored)
-            .map(|(s, e, _)| (s, e))
-            .collect();
-        let meta = StagedWrite {
-            uid: uid.to_string(),
-            len,
-            base_size,
-            base_mtime,
-            complete: authored == [(0, len)],
-            authored,
-            based_on: self
-                .core
-                .remote_baseline(&uid, base_mtime, base_size, base_revision_id),
-        };
-        // A failed sidecar means the write is not durable, and `fsync` promising
-        // otherwise is the defect this exists to fix — so it is an error, not a
-        // warning.
-        if let Err(e) = self.core.cache.mark_scratch_durable(&path, &meta) {
-            error!(%uid, error = %e, "recording fsync durability failed");
-            reply.error(Errno::EIO);
-            return;
-        }
-        reply.ok();
+        let fs = self.clone();
+        self.core
+            .workers
+            .run(Lane::Transfer, move || fs.serve_fsync(fh, reply));
     }
 
     fn release(
@@ -657,7 +590,7 @@ impl Filesystem for ProtonFs {
         reply: ReplyEmpty,
     ) {
         let (handle, unlinked_uid) = {
-            let mut st = self.core.state.lock();
+            let mut st = self.core.state();
             let unlinked_uid = release_unlinked_entry(&mut st, _ino.0);
             let ino = st.handles.remove(&fh.0);
             let h = ino.and_then(|i| {
@@ -680,10 +613,11 @@ impl Filesystem for ProtonFs {
         // acceptance suite's concurrency case failed with "concurrent file 1
         // mismatch" on all three runs the moment this moved.
         //
-        // The cost is real — `queue_revision` gap-fills a partial write from the
-        // remote — so moving it off still wants doing, but it needs the ordering
-        // made explicit first (a per-node "staging in flight" barrier that reads
-        // wait on) rather than inherited from the dispatch loop.
+        // That inherited ordering is now affordable because the network came out
+        // of this path: `queue_revision` no longer downloads the ranges an edit
+        // did not touch, it stages an incomplete blob and lets the drain thread
+        // finish it (audit F1). What is left is a rename, two fsyncs, a database
+        // insert, and at most a bounded copy out of the local cache.
         self.finish_release(handle, unlinked_uid, reply);
     }
 
@@ -802,7 +736,7 @@ impl Filesystem for ProtonFs {
     }
 
     fn access(&self, _req: &Request, ino: INodeNo, mask: AccessFlags, reply: ReplyEmpty) {
-        let state = self.core.state.lock();
+        let state = self.core.state();
         let Some(entry) = state.entries.get(&ino.0) else {
             reply.error(Errno::ENOENT);
             return;
@@ -987,7 +921,7 @@ impl ProtonFs {
     /// Build the filesystem rooted at `root` (the user's My Files folder).
     pub(super) fn new(core: Core, root: Node) -> Self {
         {
-            let mut st = core.state.lock();
+            let mut st = core.state();
             if let Err(e) = st.db.upsert_node(&root) {
                 warn!(uid = %root.uid, error = %e, "db upsert root failed");
             }
@@ -1029,7 +963,7 @@ impl ProtonFs {
             return;
         }
         let hit = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             st.children.get(&parent).and_then(|kids| {
                 kids.iter().copied().find_map(|ino| {
                     st.entries
@@ -1074,16 +1008,22 @@ impl ProtonFs {
             });
             return;
         }
-        self.core.upgrade_sizes_for_parent(ino, &uid, grandparent);
-        // Re-read: on timeout or failure this is the provisional attr again,
-        // which is no worse than what this path used to always send.
-        let attr = {
-            let st = self.core.state.lock();
-            st.entries
-                .get(&ino)
-                .map_or(attr, |e| self.attr(ino, &e.node, e.access))
+        let fs = self.clone();
+        let respond = move || {
+            // Re-read: on timeout or failure this is the provisional attr again,
+            // which is no worse than what this path used to always send.
+            let attr = {
+                let st = fs.core.state();
+                st.entries
+                    .get(&ino)
+                    .map_or(attr, |e| fs.attr(ino, &e.node, e.access))
+            };
+            reply.entry(&TTL, &attr, Generation(0));
         };
-        reply.entry(&TTL, &attr, Generation(0));
+        match self.core.upgrade_sizes_for_parent(ino, &uid, grandparent) {
+            Some(slot) => self.core.await_size(ino, slot, respond),
+            None => respond(),
+        }
     }
 
     /// The body of [`Filesystem::readdir`], on whichever thread ends up serving it.
@@ -1092,7 +1032,7 @@ impl ProtonFs {
             reply.error(e);
             return;
         }
-        let st = self.core.state.lock();
+        let st = self.core.state();
         let parent = st.entries.get(&ino).map_or(ROOT_INO, |e| e.parent);
 
         // "." and ".." occupy offsets 0 and 1; real children follow.
@@ -1146,7 +1086,7 @@ impl ProtonFs {
             uid: self.uid,
             gid: self.gid,
             rdev: 0,
-            blksize: 512,
+            blksize: REPORTED_BLKSIZE,
             flags: 0,
         }
     }
@@ -1154,6 +1094,202 @@ impl ProtonFs {
     /// Trash the child `name` under `parent` on the remote, then drop it from the
     /// local cache. Backs both `unlink` and `rmdir` (Proton trashes whole
     /// subtrees, so an `rmdir` of a non-empty dir behaves the same).
+    /// The body of [`Filesystem::open`], run off the dispatch loop for a write
+    /// open.
+    fn serve_open(&self, ino: INodeNo, flags: OpenFlags, reply: ReplyOpen) {
+        let write_requested = flags.acc_mode() != OpenAccMode::O_RDONLY;
+        let (uid, base_mtime, base_size, base_revision_id) = {
+            let mut st = self.core.state();
+            match st.entries.get_mut(&ino.0) {
+                Some(e) if e.node.is_file() => {
+                    if write_requested && !e.writable() {
+                        reply.error(Errno::EACCES);
+                        return;
+                    }
+                    e.open_count = e.open_count.saturating_add(1);
+                    (
+                        e.uid.clone(),
+                        e.node.modification_time,
+                        node_size(&e.node),
+                        node_revision_id(&e.node),
+                    )
+                }
+                Some(_) => {
+                    reply.error(Errno::EISDIR);
+                    return;
+                }
+                None => {
+                    reply.error(Errno::ENOENT);
+                    return;
+                }
+            }
+        };
+        debug!(ino = ino.0, ?flags, base_size, base_mtime, "open");
+        // Read-only opens stay stateless (fh 0). A write open allocates a
+        // disk-backed handle; bytes are authored into a scratch file and the
+        // untouched remainder is pulled from the base lazily (on read / commit).
+        if flags.acc_mode() == OpenAccMode::O_RDONLY {
+            reply.opened(FileHandle(0), FopenFlags::empty());
+            return;
+        }
+        // A previous close may already have newer bytes queued locally while
+        // the server still holds the older revision. That staged blob, not the
+        // remote, is the base for this handle. In particular, `truncate file`
+        // opens for writing and then shrinks via setattr; starting its scratch
+        // as a sparse zero file would turn the preserved prefix into zeros.
+        //
+        // Complete pending blobs are the common case and can be copied without
+        // any network access. An incomplete blob still has gaps referring to
+        // the remote base; stacking another write on it cannot be represented
+        // safely by WriteHandle, so fail the open rather than risk corruption.
+        let pending_base = self.core.pending.lock().get(&uid).cloned();
+        if pending_base.as_ref().is_some_and(|p| !p.meta.complete) {
+            if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
+                entry.open_count = entry.open_count.saturating_sub(1);
+            }
+            error!(%uid, "refusing write over incomplete queued revision");
+            reply.error(Errno::EIO);
+            return;
+        }
+        let (file, path) = match self.core.cache.create_scratch() {
+            Ok(x) => x,
+            Err(e) => {
+                if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
+                    entry.open_count = entry.open_count.saturating_sub(1);
+                }
+                error!(%uid, error = %e, "create scratch file failed");
+                reply.error(Errno::EIO);
+                return;
+            }
+        };
+        let mut initial_written = Intervals::default();
+        if let Some(pending) = &pending_base {
+            // Reflink where the filesystem can: this blob is the whole file, and
+            // a byte-for-byte copy of a multi-GB pending revision is the kind of
+            // work `open` must not do (`ContentCache::clone_or_copy`).
+            if let Err(e) = ContentCache::clone_or_copy(&pending.path, &path) {
+                if let Some(entry) = self.core.state().entries.get_mut(&ino.0) {
+                    entry.open_count = entry.open_count.saturating_sub(1);
+                }
+                let _ = std::fs::remove_file(&path);
+                error!(%uid, source = %pending.path.display(), error = %e,
+                    "copy queued revision into write scratch failed");
+                reply.error(Errno::EIO);
+                return;
+            }
+            initial_written.add(0, pending.meta.len);
+        }
+        let mut st = self.core.state();
+        let fh = st.next_fh;
+        st.next_fh += 1;
+        let aw = st.active_writes.entry(ino.0).or_insert_with(|| {
+            WriteHandle {
+                ino: ino.0,
+                uid,
+                file: Arc::new(file),
+                path,
+                written: initial_written,
+                // Starts at the current size; reads in [0, base_size) come from
+                // the base until overwritten.
+                len: base_size,
+                base_size,
+                base_mtime,
+                base_revision_id,
+                dirty: false,
+                open_count: 0,
+            }
+        });
+        aw.open_count += 1;
+        st.handles.insert(fh, ino.0);
+        reply.opened(FileHandle(fh), FopenFlags::empty());
+    }
+
+    /// The body of [`Filesystem::fsync`], run off the dispatch loop.
+    fn serve_fsync(&self, fh: FileHandle, reply: ReplyEmpty) {
+        // Everything the durability sidecar needs, copied out so no lock is held
+        // across the sync.
+        let handle = {
+            let st = self.core.state();
+            let ino = st.handles.get(&fh.0).copied();
+            ino.and_then(|i| st.active_writes.get(&i)).map(|aw| {
+                (
+                    aw.file.clone(),
+                    aw.path.clone(),
+                    aw.uid.clone(),
+                    aw.dirty,
+                    aw.written.clone(),
+                    aw.len,
+                    aw.base_size,
+                    aw.base_mtime,
+                    aw.base_revision_id.clone(),
+                )
+            })
+        };
+        let Some((f, path, uid, dirty, written, len, base_size, base_mtime, base_revision_id)) =
+            handle
+        else {
+            reply.error(Errno::EBADF);
+            return;
+        };
+        if let Err(e) = f.sync_all() {
+            error!(fh = fh.0, error = %e, "fsync of scratch file failed");
+            // Delayed allocation means a full disk usually surfaces here rather
+            // than at `write(2)`; `ENOSPC` is what an application checks for.
+            if e.raw_os_error() == Some(libc::ENOSPC) {
+                self.core.cache.emergency_evict();
+                reply.error(Errno::ENOSPC);
+            } else {
+                reply.error(Errno::EIO);
+            }
+            return;
+        }
+        // A clean handle has nothing on it that the remote does not already
+        // hold, so there is nothing for a crash to lose.
+        if !dirty {
+            reply.ok();
+            return;
+        }
+        // The bytes are on stable storage, but only the scratch file knows that,
+        // and the scratch directory is cleared at open. Record what the blob is
+        // so a restart can hand it to the upload queue instead of deleting it.
+        //
+        // Deliberately *not* the full release path: staging the blob here would
+        // move the file out from under a handle the application still has open,
+        // and queueing an upload per `fsync` would push a revision for every
+        // barrier in a write loop. This only makes the existing file findable.
+        let authored: Vec<(u64, u64)> = written
+            .segments(0, len)
+            .into_iter()
+            .filter(|&(_, _, authored)| authored)
+            .map(|(s, e, _)| (s, e))
+            .collect();
+        let meta = StagedWrite {
+            uid: uid.to_string(),
+            len,
+            base_size,
+            base_mtime,
+            complete: authored == [(0, len)],
+            authored,
+            based_on: self
+                .core
+                .remote_baseline(&uid, base_mtime, base_size, base_revision_id),
+        };
+        // A failed sidecar means the write is not durable, and `fsync` promising
+        // otherwise is the defect this exists to fix — so it is an error, not a
+        // warning.
+        if let Err(e) = self.core.cache.mark_scratch_durable(&path, &meta) {
+            error!(%uid, error = %e, "recording fsync durability failed");
+            if e.is_disk_full() {
+                self.core.cache.emergency_evict();
+                reply.error(Errno::ENOSPC);
+            } else {
+                reply.error(Errno::EIO);
+            }
+            return;
+        }
+        reply.ok();
+    }
+
     /// The body of [`Filesystem::create`], run off the dispatch loop.
     fn serve_create(&self, parent: u64, name: &str, reply: ReplyCreate) {
         if let Err(error) = self.core.require_writable(parent) {
@@ -1165,7 +1301,7 @@ impl ProtonFs {
             return;
         }
         let parent_uid = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.entries.get(&parent) {
                 Some(e) => e.uid.clone(),
                 None => {
@@ -1247,7 +1383,7 @@ impl ProtonFs {
                 return;
             }
         };
-        let mut st = self.core.state.lock();
+        let mut st = self.core.state();
         let ino = st.intern(parent, node);
         if let Some(entry) = st.entries.get_mut(&ino) {
             entry.open_count = entry.open_count.saturating_add(1);
@@ -1302,7 +1438,7 @@ impl ProtonFs {
             return;
         }
         let parent_uid = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.entries.get(&parent) {
                 Some(e) => e.uid.clone(),
                 None => {
@@ -1352,7 +1488,7 @@ impl ProtonFs {
                 }
             }
         };
-        let mut st = self.core.state.lock();
+        let mut st = self.core.state();
         let local = is_local_uid(&node.uid);
         let uid = node.uid.clone();
         let ino = st.intern(parent, node);
@@ -1384,7 +1520,7 @@ impl ProtonFs {
             return;
         }
         if let Ok((ino, _)) = self.core.lookup_child(parent, name_str) {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             if let Some(entry) = st.entries.get(&ino)
                 && entry.node.is_folder()
             {
@@ -1403,7 +1539,7 @@ impl ProtonFs {
         }
         if let Ok((ino, _)) = self.core.lookup_child(parent, name_str) {
             {
-                let st = self.core.state.lock();
+                let st = self.core.state();
                 if st
                     .entries
                     .get(&ino)
@@ -1420,7 +1556,7 @@ impl ProtonFs {
                 reply.error(e);
                 return;
             }
-            if self.core.state.lock().has_children(ino) {
+            if self.core.state().has_children(ino) {
                 reply.error(Errno::ENOTEMPTY);
                 return;
             }
@@ -1501,7 +1637,7 @@ impl ProtonFs {
         };
         // Ancestor cycle check: moving a directory into itself or one of its descendants is forbidden by POSIX (EINVAL).
         {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             if let Some(entry) = st.entries.get(&ino)
                 && entry.node.is_folder()
                 && st.is_ancestor_of(ino, newparent)
@@ -1520,7 +1656,7 @@ impl ProtonFs {
             return;
         }
         let (old_parent_uid, new_parent_uid) = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match (st.entries.get(&parent), st.entries.get(&newparent)) {
                 (Some(old), Some(new)) => (old.uid.clone(), new.uid.clone()),
                 _ => {
@@ -1561,7 +1697,7 @@ impl ProtonFs {
             // check has to happen before anything is trashed: getting it wrong
             // turns a refusal into the destruction of the destination.
             let (src_dir, dst_dir) = {
-                let st = self.core.state.lock();
+                let st = self.core.state();
                 let is_dir = |i: &u64| {
                     st.entries
                         .get(i)
@@ -1624,12 +1760,9 @@ impl ProtonFs {
                 newname,
             ) {
                 Ok(true) => {
-                    self.core.state.lock().rename_in_place(
-                        ino,
-                        newparent,
-                        &new_parent_uid,
-                        newname,
-                    );
+                    self.core
+                        .state()
+                        .rename_in_place(ino, newparent, &new_parent_uid, newname);
                     // Finalize: a transient scratch file (its create parked, bytes
                     // held back) renamed to a finished name is the moment the
                     // completed file is meant to reach Drive. Un-park the create so
@@ -1770,7 +1903,7 @@ impl ProtonFs {
             if let Err(e) = moved {
                 error!(%uid, attempts, error = %e, "move after rename failed");
                 // The rename half landed in the source directory.
-                let mut state = self.core.state.lock();
+                let mut state = self.core.state();
                 let old_parent_uid = state
                     .entries
                     .get(&parent)
@@ -1796,7 +1929,7 @@ impl ProtonFs {
     /// Answer with an inode's current attributes, or `ENOENT` if it went away
     /// while the caller was doing something slower.
     fn reply_attr(&self, ino: u64, reply: ReplyAttr) {
-        let st = self.core.state.lock();
+        let st = self.core.state();
         match st.entries.get(&ino) {
             Some(e) => {
                 let attr = self.attr(ino, &e.node, e.access);
@@ -1833,7 +1966,7 @@ impl ProtonFs {
                 reply.error(error);
                 return;
             }
-            self.core.state.lock().forget_or_unlink(&uid);
+            self.core.state().forget_or_unlink(&uid);
             debug!(%uid, name, "deleted a node that had not been created remotely yet");
             reply.ok();
             return;
@@ -1867,7 +2000,7 @@ impl ProtonFs {
         // Only remove the namespace entry after Drive accepted the mutation.
         // Otherwise an EIO response still made the path disappear locally.
         self.core.hidden.lock().insert(uid.clone());
-        self.core.state.lock().forget_or_unlink(&uid);
+        self.core.state().forget_or_unlink(&uid);
         if !open_now {
             if let Err(error) = self.core.discard_queued_ops(&uid) {
                 warn!(%uid, ?error, "remote trash landed but queued-op cleanup failed");
@@ -1931,7 +2064,7 @@ impl ProtonFs {
     /// request for an unadvertised name, so nothing becomes unreachable.
     fn serve_listxattr(&self, ino: INodeNo, size: u32, reply: ReplyXattr) {
         let thumbnailable = {
-            let st = self.core.state.lock();
+            let st = self.core.state();
             match st.entries.get(&ino.0) {
                 Some(e) => match &e.node.kind {
                     NodeKind::File { media_type, .. } => {

crates/pdfs-fuse/src/mount.rs

@@ -255,6 +255,14 @@ pub(super) fn spawn_session(
     root: Node,
 ) -> std::io::Result<BackgroundSession> {
     let mut config = Config::default();
+    // `n_threads` stays at fuser's default of 1 (with `clone_fd` off) on
+    // purpose, for now. A concurrent dispatch loop is worth having, but
+    // `release` currently relies on being serialised against the `open`+`read`
+    // an application issues right after `close(2)` — see the comment on
+    // `Filesystem::release`. Raising it before that ordering is made explicit
+    // (a per-node staging-in-flight barrier) reintroduces the acceptance
+    // suite's "concurrent file 1 mismatch". Kernel-side caps that do *not*
+    // depend on that ordering are negotiated in `ProtonFs::init`.
     config.mount_options = vec![
         MountOption::FSName("protondrive".to_string()),
         MountOption::Subtype("protondrive".to_string()),
@@ -349,23 +357,16 @@ pub fn mount(
     let core = Core {
         client: client.clone(),
         rt: rt.clone(),
+        maintenance: Arc::new(Mutex::new(Default::default())),
         primary_root_uid: root.uid.clone(),
         primary: true,
-        state: Arc::new(Mutex::new(State {
-            entries: HashMap::new(),
-            by_uid: HashMap::new(),
-            children: HashMap::new(),
-            next_ino: 2,
-            active_writes: HashMap::new(),
-            handles: HashMap::new(),
-            next_fh: 1,
-            access_changes: HashSet::new(),
-            share_access,
-            db: db.clone(),
-        })),
+        state: Arc::new(Mutex::new(State::new(db.clone(), share_access, 2))),
         cache: Arc::new(cache),
         readers: Arc::new(Mutex::new(HashMap::new())),
-        stream_ring: Arc::new(Mutex::new(StreamRing::default())),
+        block_ring: Arc::new(Mutex::new(BlockRing::default())),
+        block_flight: Arc::new(BlockFlight::default()),
+        prefetch: Arc::new(Prefetch::default()),
+        prefetch_budget: Arc::new(tokio::sync::Semaphore::new(PREFETCH_BUDGET)),
         workers: Arc::new(Workers::new(FUSE_WORKERS)?),
         db,
         shared_publication: Arc::new(Mutex::new(())),
@@ -375,6 +376,8 @@ pub fn mount(
         pending: Arc::new(Mutex::new(HashMap::new())),
         hidden: Arc::new(Mutex::new(HashSet::new())),
         drain_wake: Arc::new((Mutex::new(false), Condvar::new())),
+        upload_times: Arc::new(Mutex::new(HashMap::new())),
+        upload_cancel: Arc::new(Mutex::new(HashMap::new())),
         timeline_refreshing: Arc::new(AtomicBool::new(false)),
         albums_refreshing: Arc::new(AtomicBool::new(false)),
         trash_refreshing: Arc::new(AtomicBool::new(false)),
@@ -385,7 +388,9 @@ pub fn mount(
         self_changes: Arc::new(Mutex::new(HashMap::new())),
         thumb_gen: Arc::new(Mutex::new(HashSet::new())),
         no_thumbnail: Arc::new(Mutex::new(HashMap::new())),
+        quota: Arc::new(Mutex::new(None)),
         size_upgrades: Arc::new(Mutex::new(HashMap::new())),
+        size_waiters: Arc::new(Default::default()),
         notifier: Arc::new(OnceLock::new()),
         session_live: Arc::new(AtomicBool::new(false)),
         transfers: TransferRegistry::new(),
@@ -407,11 +412,19 @@ pub fn mount(
     // aside at open. After `hydrate_pending`, so a recovered partial write can
     // see an already-queued write to the same node.
     core.recover_fsynced_writes();
-    {
+    // And finally the staged blobs neither of those accounts for: bytes a
+    // failed release left behind with nothing pointing at them.
+    core.reconcile_staging();
+    // Several drain workers over one queue. One thread meant a single large
+    // upload held every queued rename, trash and small write behind it for as
+    // long as it ran; the claim column (`Db::claim_next_due_op`) keeps the
+    // workers off each other's nodes, which is the only ordering the queue
+    // actually needs. Worker 0 additionally runs the queue's idle chores.
+    for worker in 0..DRAIN_WORKERS {
         let core = core.clone();
         std::thread::Builder::new()
-            .name("pdfs-drain".into())
-            .spawn(move || core.run_pending_drain())?;
+            .name(format!("pdfs-drain-{worker}"))
+            .spawn(move || core.run_pending_drain(worker == 0))?;
     }
 
     // Start the folder-sync engine. It watches every mirror folder, polls the

crates/pdfs-fuse/src/reads.rs

@@ -10,18 +10,30 @@ use super::*;
 /// usual.
 pub(super) const STREAM_BYPASS_MIN: u64 = 256 * 1024 * 1024;
 
-/// How many bytes of streamed blocks the in-memory ring keeps. Bypassing the
-/// on-disk cache must not mean re-fetching: the kernel asks for a streamed file
-/// in reads far smaller than [`BLOCK_SIZE`], so without a ring every 128 KiB the
-/// player consumes would download and decrypt a whole 4 MiB block again. Sized
-/// for a handful of blocks per concurrently-streamed file.
-const STREAM_RING_BYTES: u64 = 128 * 1024 * 1024;
+/// How many bytes of decrypted blocks the in-memory ring keeps.
+///
+/// The kernel asks for a file in reads far smaller than [`BLOCK_SIZE`], so
+/// without a ring every 128 KiB it consumes costs a whole 4 MiB block again —
+/// downloaded and decrypted for a streaming file, or re-read off disk and
+/// re-allocated for a cached one (audit T3: 128 MiB of I/O to deliver 4 MiB).
+/// Sized for a handful of blocks per concurrently-read file. Bounded and only
+/// ever filled by active reads.
+const RING_BYTES: u64 = 128 * 1024 * 1024;
 
-/// Blocks fetched *past* the one a sequential streaming read needed, warmed into
-/// the ring in the background so the player's next read is already in memory
-/// instead of waiting a round-trip. Kept small so it cannot crowd out the
-/// in-flight block governor.
-const STREAM_READAHEAD: u64 = 4;
+/// Deepest read-ahead window, in blocks, for a read that has proved sequential.
+const PREFETCH_MAX: u64 = 8;
+
+/// Window a sequential read starts at, and returns to after a seek.
+const PREFETCH_MIN: u64 = 2;
+
+/// Blocks that may be prefetching at once, across every file. Prefetch takes a
+/// permit and gives up rather than queueing, so it can never sit in front of a
+/// demand read in the SDK's flat block semaphore.
+pub(super) const PREFETCH_BUDGET: usize = 8;
+
+/// Cap on tracked sequential streams. Only reads in progress matter, so the map
+/// is cleared wholesale rather than aged — losing the ramp costs one read.
+const MAX_PREFETCH_STREAMS: usize = 256;
 
 /// How many [`RevisionReader`]s stay open at once.
 ///
@@ -76,35 +88,39 @@ pub(super) struct PendingOpen {
 /// does; ids are never compared for order.
 static NEXT_OPEN_ID: AtomicU64 = AtomicU64::new(0);
 
-/// Drive client, a Tokio handle to bridge the synchronous FUSE/socket threads
-/// to the async SDK, the inode bookkeeping, and the on-disk content cache.
-///
-/// Cheaply cloneable (every field is a handle/`Arc`), so the control-socket task
-/// gets its own copy while the FUSE session keeps another.
-/// In-memory LRU of blocks belonging to a file streaming past the on-disk cache
-/// (see [`STREAM_BYPASS_MIN`]). Bypassing the disk stops one film evicting the
-/// block LRU; it must not also mean the same 4 MiB block is downloaded once per
-/// 128 KiB the player reads out of it.
+/// In-memory LRU of decrypted 4 MiB blocks, in front of both block paths: the
+/// on-disk block cache, and the disk-bypassing stream of a large unpinned video
+/// (see [`STREAM_BYPASS_MIN`], which stops one film evicting the whole disk
+/// LRU). Either way the point is the same — the kernel reads a block out in ~32
+/// pieces, and none of them should cost the block again.
 ///
 /// Validated by the same `(mtime, size)` pair the on-disk caches use: a node
 /// whose tag no longer matches has its blocks dropped rather than served, so a
 /// new revision can never be stitched together from the old one.
 #[derive(Default)]
-pub(super) struct StreamRing {
+pub(super) struct BlockRing {
     blocks: HashMap<(NodeUid, u64), Arc<Vec<u8>>>,
-    /// Keys oldest-first, for eviction.
+    /// Keys least-recently-used first, for eviction.
     order: VecDeque<(NodeUid, u64)>,
     /// Per-node validity tag, `(mtime, size)`.
     tags: HashMap<NodeUid, (i64, u64)>,
     bytes: u64,
 }
 
-impl StreamRing {
+impl BlockRing {
     fn get(&mut self, uid: &NodeUid, mtime: i64, size: u64, idx: u64) -> Option<Arc<Vec<u8>>> {
         if self.tags.get(uid) != Some(&(mtime, size)) {
             return None;
         }
-        self.blocks.get(&(uid.clone(), idx)).cloned()
+        let key = (uid.clone(), idx);
+        let hit = self.blocks.get(&key).cloned()?;
+        // Move to the young end: a block being read repeatedly (the whole reason
+        // this exists) must not age out under a sequential read passing through.
+        if let Some(at) = self.order.iter().position(|k| *k == key) {
+            self.order.remove(at);
+            self.order.push_back(key);
+        }
+        Some(hit)
     }
 
     fn insert(&mut self, uid: &NodeUid, mtime: i64, size: u64, idx: u64, bytes: Arc<Vec<u8>>) {
@@ -121,13 +137,18 @@ impl StreamRing {
         self.bytes += bytes.len() as u64;
         self.blocks.insert(key.clone(), bytes);
         self.order.push_back(key);
-        while self.bytes > STREAM_RING_BYTES {
+        while self.bytes > RING_BYTES {
             let Some(victim) = self.order.pop_front() else {
                 break;
             };
             if let Some(dropped) = self.blocks.remove(&victim) {
                 self.bytes -= dropped.len() as u64;
             }
+            // A node whose last block just aged out keeps no tag: otherwise the
+            // tag map is the one thing here that grows with every file ever read.
+            if !self.blocks.keys().any(|(u, _)| *u == victim.0) {
+                self.tags.remove(&victim.0);
+            }
         }
     }
 
@@ -147,6 +168,87 @@ impl StreamRing {
     }
 }
 
+/// The result of fetching one block, shared between everyone who wanted it.
+type BlockResult = Result<Arc<Vec<u8>>, Errno>;
+
+/// Identifies a block *of a revision*: node, validity tag, index.
+type BlockKey = (NodeUid, i64, u64, u64);
+
+/// A claim on a block key: who owns the fetch, and where its result is published.
+type Claim = (u64, tokio::sync::watch::Receiver<Option<BlockResult>>);
+
+/// Block fetches in flight, so a block is downloaded and decrypted once however
+/// many readers want it at that moment (audit T1).
+///
+/// With `FUSE_ASYNC_READ` the demand read and the kernel's read-ahead land in
+/// the same 4 MiB block for 31 of every 32 reads, and read-ahead of our own
+/// (see [`Core::prefetch`]) aims at blocks a demand read may be about to ask
+/// for. Without this each of them pays the full round-trip and decrypt and all
+/// but one result is discarded — roughly double the bandwidth and CPU in steady
+/// state. Same shape as [`ReaderSlot::Pending`] one level up.
+#[derive(Default)]
+pub(super) struct BlockFlight {
+    inflight: Mutex<HashMap<BlockKey, Claim>>,
+}
+
+/// Distinguishes attempts at the same key, so a leader only retires its own.
+static NEXT_FLIGHT_ID: AtomicU64 = AtomicU64::new(0);
+
+/// The leader's claim on a key. Retires it on drop — including when the task is
+/// cancelled mid-fetch, which is what keeps a dropped `JoinSet` from leaving a
+/// key nobody will ever publish.
+struct Flight {
+    flight: Arc<BlockFlight>,
+    key: BlockKey,
+    id: u64,
+}
+
+impl Drop for Flight {
+    fn drop(&mut self) {
+        let mut inflight = self.flight.inflight.lock();
+        if inflight
+            .get(&self.key)
+            .is_some_and(|(id, _)| *id == self.id)
+        {
+            inflight.remove(&self.key);
+        }
+    }
+}
+
+/// Where each file's reader is expected to read next, so sequential reads can be
+/// told from random ones and only the former prefetched (audit T2).
+#[derive(Default)]
+pub(super) struct Prefetch {
+    /// Per node: the block index a sequential reader should ask for next, and
+    /// the window depth it has earned.
+    streams: Mutex<HashMap<NodeUid, (u64, u64)>>,
+}
+
+impl Prefetch {
+    /// How far past `last` to read ahead for a read of blocks `[first, last]`,
+    /// and record where that read leaves the file's cursor.
+    ///
+    /// A read that starts where the last one stopped doubles the window (from
+    /// [`PREFETCH_MIN`], up to [`PREFETCH_MAX`]); anything else is a seek and
+    /// earns nothing, so random access — a database file, a linker — never pays
+    /// for speculative blocks it will not read. The first read of a file counts
+    /// as sequential: `(0, 0)` is both the initial cursor and block zero.
+    fn window(&self, uid: &NodeUid, first: u64, last: u64) -> u64 {
+        let mut streams = self.streams.lock();
+        if streams.len() > MAX_PREFETCH_STREAMS {
+            streams.clear();
+        }
+        let stream = streams.entry(uid.clone()).or_insert((0, 0));
+        let depth = if stream.0 == first {
+            (stream.1 * 2).clamp(PREFETCH_MIN, PREFETCH_MAX)
+        } else {
+            0
+        };
+        *stream = (last + 1, depth);
+        depth
+    }
+}
+
 impl Core {
     /// An open [`RevisionReader`] for `uid`, reusing the cached one when it is
     /// still valid for `(mtime, fsize)` and opening a fresh one otherwise.
@@ -304,6 +406,10 @@ impl Core {
     /// revision is no longer the truth.
     pub(super) fn evict_reader(&self, uid: &NodeUid) {
         self.readers.lock().remove(uid);
+        // The ring holds decrypted bytes of the same revision, and an eviction is
+        // exactly the statement that those are no longer the file's content.
+        self.block_ring.lock().drop_node(uid);
+        self.prefetch.streams.lock().remove(uid);
     }
 
     /// Serve bytes `[offset, offset + len)` of `uid`'s active revision, hitting
@@ -389,20 +495,141 @@ impl Core {
         Ok(out)
     }
 
-    /// Warm the [`STREAM_READAHEAD`] blocks after `last` into the ring, in the
-    /// background, so a player reading forward finds its next block already in
-    /// memory rather than paying a fetch + decrypt round-trip at each boundary.
+    /// Fetch block `bidx`, joining a fetch already in flight for it rather than
+    /// issuing a second one, and publish the result to everyone who joined.
     ///
-    /// Fire-and-forget: a failure here is not an error, it only means the read
-    /// that needs those bytes will fetch them itself.
-    fn stream_readahead(&self, uid: &NodeUid, mtime: i64, fsize: u64, last: u64) {
-        let last_block = (fsize.saturating_sub(1)) / BLOCK_SIZE;
-        let wanted: Vec<u64> = ((last + 1)..=(last + STREAM_READAHEAD).min(last_block))
+    /// See [`BlockFlight`]. A follower whose leader disappears (a panic, or a
+    /// cancelled `JoinSet`) fetches the block itself rather than failing — the
+    /// leader's claim is gone by then, so this cannot recurse.
+    async fn fetch_block(
+        &self,
+        reader: &Arc<RevisionReader>,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+        bidx: u64,
+        cache_blocks: bool,
+    ) -> BlockResult {
+        let key: BlockKey = (uid.clone(), mtime, fsize, bidx);
+        let claim = {
+            let mut inflight = self.block_flight.inflight.lock();
+            match inflight.get(&key) {
+                Some((_, rx)) => Err(rx.clone()),
+                None => {
+                    let id = NEXT_FLIGHT_ID.fetch_add(1, Ordering::Relaxed);
+                    let (tx, rx) = tokio::sync::watch::channel(None);
+                    inflight.insert(key.clone(), (id, rx));
+                    Ok((
+                        tx,
+                        Flight {
+                            flight: self.block_flight.clone(),
+                            key: key.clone(),
+                            id,
+                        },
+                    ))
+                }
+            }
+        };
+        let (tx, flight) = match claim {
+            Ok(lead) => lead,
+            Err(mut rx) => {
+                loop {
+                    if let Some(result) = rx.borrow_and_update().clone() {
+                        return result;
+                    }
+                    if rx.changed().await.is_err() {
+                        break;
+                    }
+                }
+                return self
+                    .read_block(reader, uid, mtime, fsize, bidx, cache_blocks)
+                    .await;
+            }
+        };
+        let result = self
+            .read_block(reader, uid, mtime, fsize, bidx, cache_blocks)
+            .await;
+        // Retire the claim before publishing, so a racer that wakes and re-enters
+        // finds no stale entry to join.
+        drop(flight);
+        let _ = tx.send(Some(result.clone()));
+        result
+    }
+
+    /// Download and decrypt one block, then put it where the next reader will
+    /// find it: the in-memory ring always, the on-disk block cache too unless
+    /// this is a disk-bypassing stream.
+    ///
+    /// The disk store is handed to a blocking thread — it is a 4 MiB write plus
+    /// a cache-index update, and neither the reactor nor the kernel read waiting
+    /// on these bytes should be held up by it.
+    async fn read_block(
+        &self,
+        reader: &Arc<RevisionReader>,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+        bidx: u64,
+        cache_blocks: bool,
+    ) -> BlockResult {
+        let bstart = bidx * BLOCK_SIZE;
+        let blen = BLOCK_SIZE.min(fsize - bstart);
+        let bytes = reader.read_at(bstart, blen).await.map_err(|e| {
+            warn!(%uid, bstart, blen, error = %e, "block read failed");
+            Errno::EIO
+        })?;
+        let bytes = Arc::new(bytes);
+        if cache_blocks {
+            let cache = self.cache.clone();
+            let uid = uid.clone();
+            let bytes = bytes.clone();
+            tokio::task::spawn_blocking(move || {
+                let _ = cache.store_block(&uid, mtime, fsize, bidx, &bytes);
+            });
+        }
+        self.block_ring
+            .lock()
+            .insert(uid, mtime, fsize, bidx, bytes.clone());
+        Ok(bytes)
+    }
+
+    /// Warm the blocks after a sequential read into the caches, in the
+    /// background, so the next read finds them instead of paying a round-trip.
+    ///
+    /// A read that continues where the last one stopped doubles the window, up
+    /// to [`PREFETCH_MAX`]; a seek drops it back to nothing and one more
+    /// sequential read is needed to earn it again. Without this every read is a
+    /// serial fetch/serve/stall chain: on a 100 ms link that ceilings a file at
+    /// ~24 MB/s of the ~125 MB/s available, and the SDK's block concurrency is
+    /// never exercised at all.
+    ///
+    /// Fire-and-forget, and permit-bounded ([`PREFETCH_BUDGET`]) so speculative
+    /// work never sits in front of a read someone is waiting on. A failure here
+    /// is not an error: the read that needs those bytes will fetch them itself.
+    fn prefetch_ahead(
+        &self,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+        first: u64,
+        last: u64,
+        cache_blocks: bool,
+    ) {
+        let depth = self.prefetch.window(uid, first, last);
+        if depth == 0 {
+            return;
+        }
+        let last_block = fsize.saturating_sub(1) / BLOCK_SIZE;
+        let wanted: Vec<u64> = ((last + 1)..=(last + depth).min(last_block))
             .filter(|&bidx| {
-                self.stream_ring
+                // Blocks already in memory need nothing; blocks already on disk
+                // are cheap enough to serve that spending a round-trip to move
+                // them into the ring would be the wrong trade.
+                self.block_ring
                     .lock()
                     .get(uid, mtime, fsize, bidx)
                     .is_none()
+                    && !(cache_blocks && self.cache.has_block(uid, mtime, fsize, bidx))
             })
             .collect();
         if wanted.is_empty() {
@@ -416,26 +643,25 @@ impl Core {
             };
             // Concurrently, for the same reason the demand path fetches its
             // misses concurrently: serially awaited blocks cost the sum of their
-            // round-trips, and the player catches up with the readahead before it
-            // finishes. The SDK's in-flight block governor bounds the fan-out.
+            // round-trips, and the reader catches up with the window before it
+            // finishes.
             let mut set = tokio::task::JoinSet::new();
             for bidx in wanted {
-                let reader = reader.clone();
-                let bstart = bidx * BLOCK_SIZE;
-                let blen = BLOCK_SIZE.min(fsize - bstart);
-                set.spawn(async move { (bidx, reader.read_at(bstart, blen).await) });
-            }
-            while let Some(joined) = set.join_next().await {
-                let Ok((bidx, read)) = joined else { continue };
-                match read {
-                    Ok(bytes) => {
-                        core.stream_ring
-                            .lock()
-                            .insert(&uid, mtime, fsize, bidx, Arc::new(bytes))
+                let Ok(permit) = core.prefetch_budget.clone().try_acquire_owned() else {
+                    break; // the budget is spent; the demand path can have it
+                };
+                let (core, reader, uid) = (core.clone(), reader.clone(), uid.clone());
+                set.spawn(async move {
+                    let _permit = permit;
+                    if let Err(error) = core
+                        .fetch_block(&reader, &uid, mtime, fsize, bidx, cache_blocks)
+                        .await
+                    {
+                        debug!(%uid, bidx, ?error, "prefetch failed");
                     }
-                    Err(e) => debug!(%uid, bidx, error = %e, "stream readahead failed"),
-                }
+                });
             }
+            while set.join_next().await.is_some() {}
         });
     }
 
@@ -473,17 +699,23 @@ impl Core {
         let mut blocks: Vec<Option<Arc<Vec<u8>>>> = Vec::with_capacity((last - first + 1) as usize);
         let mut misses: Vec<u64> = Vec::new();
         for bidx in first..=last {
-            // A streaming read keeps its blocks only in memory, so the ring is
-            // the only cache it has — check it before the disk (which will not
-            // hold them) and before the network (which would refetch the same
-            // 4 MiB block for every 128 KiB the player consumes).
-            let hit = if cache_blocks {
-                self.cache
-                    .cached_block(uid, mtime, fsize, bidx)
-                    .map(Arc::new)
-            } else {
-                self.stream_ring.lock().get(uid, mtime, fsize, bidx)
-            };
+            // The in-memory ring comes first on both paths: a streaming read has
+            // no other cache at all, and a cached one would otherwise pay two
+            // opens, a JSON parse and a 4 MiB read+alloc for every 128 KiB the
+            // kernel asks for (audit T3).
+            let hit = self.block_ring.lock().get(uid, mtime, fsize, bidx);
+            let hit = hit.or_else(|| {
+                if !cache_blocks {
+                    return None;
+                }
+                let bytes = Arc::new(self.cache.cached_block(uid, mtime, fsize, bidx)?);
+                // Promote it, so the other ~31 reads of this block are answered
+                // from memory.
+                self.block_ring
+                    .lock()
+                    .insert(uid, mtime, fsize, bidx, bytes.clone());
+                Some(bytes)
+            });
             match hit {
                 Some(b) => blocks.push(Some(b)),
                 None => {
@@ -502,19 +734,11 @@ impl Core {
 
                 let mut set = tokio::task::JoinSet::new();
                 for &bidx in &misses {
-                    let reader = reader.clone();
-                    let uid = uid.clone();
-                    let bstart = bidx * BLOCK_SIZE;
-                    let blen = BLOCK_SIZE.min(fsize - bstart);
+                    let (core, reader, uid) = (self.clone(), reader.clone(), uid.clone());
                     set.spawn(async move {
-                        reader
-                            .read_at(bstart, blen)
+                        core.fetch_block(&reader, &uid, mtime, fsize, bidx, cache_blocks)
                             .await
                             .map(|bytes| (bidx, bytes))
-                            .map_err(|e| {
-                                warn!(%uid, bstart, blen, error = %e, "block read failed");
-                                Errno::EIO
-                            })
                     });
                 }
                 let mut out = Vec::with_capacity(misses.len());
@@ -525,23 +749,12 @@ impl Core {
                 Ok::<_, Errno>(out)
             })?;
             for (bidx, bytes) in fetched {
-                // A streaming read (large unpinned video) skips the on-disk cache
-                // so it can't evict the rest of it; it keeps the block in the
-                // in-memory ring instead, which is bounded and dies with the read.
-                let bytes = Arc::new(bytes);
-                if cache_blocks {
-                    let _ = self.cache.store_block(uid, mtime, fsize, bidx, &bytes);
-                } else {
-                    self.stream_ring
-                        .lock()
-                        .insert(uid, mtime, fsize, bidx, bytes.clone());
-                }
                 blocks[(bidx - first) as usize] = Some(bytes);
             }
-            if !cache_blocks {
-                self.stream_readahead(uid, mtime, fsize, last);
-            }
         }
+        // Also after a wholly-cached read: that is what keeps the window sliding
+        // ahead of a sequential reader instead of stalling once it catches up.
+        self.prefetch_ahead(uid, mtime, fsize, first, last, cache_blocks);
 
         for (i, block) in blocks.into_iter().enumerate() {
             let bidx = first + i as u64;
@@ -605,6 +818,106 @@ impl Core {
         Ok(out)
     }
 
+    /// The largest amount [`Core::fill_gaps_cached`] will copy out of the local
+    /// caches while a `release(2)` is in progress.
+    ///
+    /// The point of the cached fill is that editing a file the mount already
+    /// holds costs nothing extra; it is not to reconstruct arbitrarily large
+    /// files on the dispatch loop. Past this budget the blob stays incomplete
+    /// and the drain finishes it, which is the same work in a place where it
+    /// blocks nobody.
+    const CACHED_FILL_BUDGET: u64 = 64 * 1024 * 1024;
+
+    /// Bytes `[offset, offset + len)` of `uid`'s base revision, but only if the
+    /// local caches can answer in full — the whole-file blob first, then every
+    /// overlapping block. `None` the moment anything would have to be fetched.
+    fn cached_range(
+        &self,
+        uid: &NodeUid,
+        mtime: i64,
+        fsize: u64,
+        offset: u64,
+        len: u64,
+    ) -> Option<Vec<u8>> {
+        if len == 0 {
+            return Some(Vec::new());
+        }
+        if let Some(bytes) = self.cache.read_range(uid, mtime, fsize, offset, len) {
+            return Some(bytes);
+        }
+        let end = offset.checked_add(len)?.min(fsize);
+        if offset >= end {
+            return None;
+        }
+        let mut out = Vec::with_capacity((end - offset) as usize);
+        for bidx in (offset / BLOCK_SIZE)..=((end - 1) / BLOCK_SIZE) {
+            let ring = self.block_ring.lock().get(uid, mtime, fsize, bidx);
+            let block = match ring {
+                Some(bytes) => bytes,
+                None => Arc::new(self.cache.cached_block(uid, mtime, fsize, bidx)?),
+            };
+            let bstart = bidx * BLOCK_SIZE;
+            let s = (offset.max(bstart) - bstart) as usize;
+            let e = (end.min(bstart + block.len() as u64) - bstart) as usize;
+            if s >= e {
+                return None;
+            }
+            out.extend_from_slice(&block[s..e]);
+        }
+        (out.len() as u64 == end - offset).then_some(out)
+    }
+
+    /// Fill what [`Core::fill_gaps`] would, but only from bytes already on this
+    /// machine, and only up to [`Core::CACHED_FILL_BUDGET`]. Ranges past the
+    /// base are zeros the caller's `set_len` already wrote, so they count as
+    /// filled. `written` is extended with everything that ends up resolved.
+    ///
+    /// This is the release-time half of gap filling. Doing the network half
+    /// there as well is what froze the whole mount for the length of a download
+    /// (audit F1): the closing file's untouched ranges are allowed to stay
+    /// unresolved in the staged blob — `StagedWrite::complete` is false, reads
+    /// resolve the gaps against the base, and `drain_revision` fills them on the
+    /// drain thread before it uploads. What is left here is the case where
+    /// filling is free, which is also the common one: a file that was read
+    /// before it was edited.
+    pub(super) fn fill_gaps_cached(
+        &self,
+        uid: &NodeUid,
+        file: &File,
+        len: u64,
+        base_mtime: i64,
+        base_size: u64,
+        written: &mut Intervals,
+    ) {
+        let mut budget = Self::CACHED_FILL_BUDGET;
+        for (s, e, authored) in written.clone().segments(0, len) {
+            if authored {
+                continue;
+            }
+            let bend = e.min(base_size);
+            if s < bend {
+                let want = bend - s;
+                if want > budget {
+                    continue;
+                }
+                let Some(bytes) = self.cached_range(uid, base_mtime, base_size, s, want) else {
+                    continue;
+                };
+                if let Err(error) = file.write_all_at(&bytes, s) {
+                    warn!(%uid, %error, "cached gap-fill write failed");
+                    continue;
+                }
+                budget -= want;
+                written.add(s, bend);
+            }
+            // Past the base there is nothing to fetch: the resize zeroed it, so
+            // those bytes are already the file's content.
+            if e > base_size {
+                written.add(s.max(base_size), e);
+            }
+        }
+    }
+
     /// Fill every unauthored range of a scratch/staged file that overlaps its
     /// base with the base's bytes, so the file becomes the complete new content.
     ///
@@ -643,3 +956,145 @@ impl Core {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod block_ring_tests {
+    use super::*;
+
+    fn uid(link: &str) -> NodeUid {
+        NodeUid::new(VolumeId::from("vol"), LinkId::from(link))
+    }
+
+    /// One block's worth of bytes, so byte accounting is exercised at the scale
+    /// the ring actually sees.
+    fn block(byte: u8) -> Arc<Vec<u8>> {
+        Arc::new(vec![byte; BLOCK_SIZE as usize])
+    }
+
+    /// How many whole blocks fit before the ring starts evicting.
+    fn capacity() -> u64 {
+        RING_BYTES / BLOCK_SIZE
+    }
+
+    #[test]
+    fn a_stored_block_is_served_back() {
+        let mut ring = BlockRing::default();
+        let node = uid("a");
+        ring.insert(&node, 7, 100, 3, block(0xab));
+        assert_eq!(ring.get(&node, 7, 100, 3).map(|b| b[0]), Some(0xab));
+    }
+
+    #[test]
+    fn a_block_of_another_revision_is_never_served() {
+        let mut ring = BlockRing::default();
+        let node = uid("a");
+        ring.insert(&node, 7, 100, 0, block(1));
+        // Same node, newer revision: the tag moved, so the old bytes go.
+        assert!(ring.get(&node, 8, 100, 0).is_none());
+        assert!(ring.get(&node, 7, 101, 0).is_none());
+    }
+
+    #[test]
+    fn a_new_revision_drops_what_the_old_one_left() {
+        let mut ring = BlockRing::default();
+        let node = uid("a");
+        ring.insert(&node, 7, 100, 0, block(1));
+        ring.insert(&node, 8, 100, 0, block(2));
+        assert_eq!(ring.get(&node, 8, 100, 0).map(|b| b[0]), Some(2));
+        assert_eq!(ring.bytes, BLOCK_SIZE);
+    }
+
+    #[test]
+    fn the_ring_evicts_the_least_recently_used_block() {
+        let mut ring = BlockRing::default();
+        let node = uid("a");
+        for idx in 0..capacity() {
+            ring.insert(&node, 1, u64::MAX, idx, block(idx as u8));
+        }
+        // Re-reading block 0 makes block 1 the oldest instead.
+        assert!(ring.get(&node, 1, u64::MAX, 0).is_some());
+        ring.insert(&node, 1, u64::MAX, capacity(), block(0xff));
+        assert!(ring.get(&node, 1, u64::MAX, 0).is_some());
+        assert!(ring.get(&node, 1, u64::MAX, 1).is_none());
+        assert_eq!(ring.bytes, RING_BYTES);
+    }
+
+    #[test]
+    fn a_node_whose_last_block_ages_out_keeps_no_tag() {
+        let mut ring = BlockRing::default();
+        let (old, new) = (uid("old"), uid("new"));
+        ring.insert(&old, 1, u64::MAX, 0, block(1));
+        for idx in 0..capacity() {
+            ring.insert(&new, 1, u64::MAX, idx, block(2));
+        }
+        assert!(!ring.tags.contains_key(&old));
+        assert!(ring.tags.contains_key(&new));
+    }
+
+    #[test]
+    fn dropping_a_node_frees_only_its_own_bytes() {
+        let mut ring = BlockRing::default();
+        let (a, b) = (uid("a"), uid("b"));
+        ring.insert(&a, 1, 100, 0, block(1));
+        ring.insert(&b, 1, 100, 0, block(2));
+        ring.drop_node(&a);
+        assert!(ring.get(&a, 1, 100, 0).is_none());
+        assert_eq!(ring.get(&b, 1, 100, 0).map(|x| x[0]), Some(2));
+        assert_eq!(ring.bytes, BLOCK_SIZE);
+    }
+}
+
+#[cfg(test)]
+mod prefetch_tests {
+    use super::*;
+
+    fn uid(link: &str) -> NodeUid {
+        NodeUid::new(VolumeId::from("vol"), LinkId::from(link))
+    }
+
+    #[test]
+    fn a_sequential_reader_ramps_up_to_the_cap() {
+        let prefetch = Prefetch::default();
+        let node = uid("a");
+        let depths: Vec<u64> = (0..6)
+            .map(|next| prefetch.window(&node, next, next))
+            .collect();
+        assert_eq!(depths, vec![2, 4, 8, 8, 8, 8]);
+    }
+
+    #[test]
+    fn a_seek_forfeits_the_window() {
+        let prefetch = Prefetch::default();
+        let node = uid("a");
+        assert_eq!(prefetch.window(&node, 0, 0), PREFETCH_MIN);
+        assert_eq!(prefetch.window(&node, 40, 40), 0);
+        // And has to be earned again from the bottom.
+        assert_eq!(prefetch.window(&node, 41, 41), PREFETCH_MIN);
+    }
+
+    #[test]
+    fn a_multi_block_read_advances_the_cursor_past_all_of_it() {
+        let prefetch = Prefetch::default();
+        let node = uid("a");
+        assert_eq!(prefetch.window(&node, 0, 3), PREFETCH_MIN);
+        assert_eq!(prefetch.window(&node, 4, 7), 4);
+    }
+
+    #[test]
+    fn files_are_tracked_independently() {
+        let prefetch = Prefetch::default();
+        let (a, b) = (uid("a"), uid("b"));
+        assert_eq!(prefetch.window(&a, 0, 0), PREFETCH_MIN);
+        assert_eq!(prefetch.window(&b, 9, 9), 0);
+        assert_eq!(prefetch.window(&a, 1, 1), 4);
+    }
+
+    #[test]
+    fn the_stream_map_cannot_grow_without_bound() {
+        let prefetch = Prefetch::default();
+        for n in 0..=MAX_PREFETCH_STREAMS + 1 {
+            prefetch.window(&uid(&n.to_string()), 0, 0);
+        }
+        assert!(prefetch.streams.lock().len() <= MAX_PREFETCH_STREAMS + 1);
+    }
+}

crates/pdfs-fuse/src/sharing.rs

@@ -12,6 +12,9 @@ use pdfs_core::control::{
     BookmarkInfo, DirEntry, InvitationInfo, PublicLinkInfo, ShareEntry, ShareEntryKind, SharedItem,
 };
 use pdfs_core::{CoreError, CoreResult};
+use tracing::debug;
+
+use crate::QUOTA_TTL;
 use proton_drive_rs::proton_sdk::ids::NodeUid;
 use proton_drive_rs::{MemberRole, NodeKind};
 
@@ -533,9 +536,33 @@ impl Core {
             .rt
             .block_on(self.client.quota())
             .map_err(|e| CoreError::from_api(&e, "account quota"))?;
+        *self.quota.lock() = Some((std::time::Instant::now(), q.max_space, q.used_space));
         Ok((q.max_space, q.used_space))
     }
 
+    /// The account quota as `statfs(2)` needs it: cached for [`QUOTA_TTL`],
+    /// refetched after that, and `None` when there is no usable answer at all —
+    /// offline with nothing cached.
+    ///
+    /// Never blocks on the network for a *fresh* value when a recent one exists,
+    /// because the callers are `df` and every "is there room?" preflight, and a
+    /// stale byte count is worth more to them than a stalled syscall. Offline,
+    /// the last known figures keep being served rather than degrading to zero.
+    pub(crate) fn account_quota_cached(&self) -> Option<(i64, i64)> {
+        if let Some((at, max_space, used_space)) = *self.quota.lock()
+            && at.elapsed() < QUOTA_TTL
+        {
+            return Some((max_space, used_space));
+        }
+        match self.account_quota() {
+            Ok(q) => Some(q),
+            Err(error) => {
+                debug!(%error, "account quota refresh failed; serving the last known figures");
+                self.quota.lock().map(|(_, max, used)| (max, used))
+            }
+        }
+    }
+
     // ---- shared by me -----------------------------------------------------
 
     /// List the nodes I have shared with others, each with a summary of its share

crates/pdfs-fuse/src/state.rs

@@ -166,13 +166,163 @@ pub(crate) struct State {
     /// on every explicit role/tombstone change. Descendant inheritance reads
     /// this map instead of issuing one SQLite query per interned inode.
     pub(crate) share_access: HashMap<NodeUid, Access>,
-    /// Unified SQLite metadata cache. Every map mutation below writes through to
-    /// it inside the `State` lock so the DB stays the authoritative copy across
-    /// restarts (see plan.md P1).
+    /// Unified SQLite metadata cache. Read from here (`has_children`) but never
+    /// written: mutations go to [`State::outbox`] and are applied after the lock
+    /// is released. See [`StateGuard`].
     pub(crate) db: Arc<Db>,
+    /// Write-throughs this mutation owes SQLite, applied by [`StateGuard`] once
+    /// the inode lock is released.
+    ///
+    /// The DB stays the authoritative copy across restarts, so every map
+    /// mutation still writes through — it just no longer does so with the
+    /// whole-mount lock held. A single listing's write-through is one
+    /// transaction over up to a few thousand rows, and every other FUSE callback
+    /// used to wait behind it for no reason: the maps were already updated
+    /// before the first row was written.
+    outbox: Vec<DbWrite>,
+}
+
+/// A write-through queued by a `State` mutation, applied after the lock drops.
+///
+/// Deliberately a small closed set rather than a boxed closure: these run
+/// outside the lock, and what they can touch should be readable here.
+pub(crate) enum DbWrite {
+    /// One node or a whole listing. Both go through `upsert_nodes`, which is a
+    /// single transaction either way, so there is nothing to gain by splitting
+    /// them — and a `Node` is large enough that a variant holding one inline
+    /// would set the size of every other.
+    Upsert(Vec<Node>),
+    Delete(NodeUid),
+    SetListed(NodeUid, bool),
+    SetShareAccess(NodeUid, Access),
+}
+
+impl DbWrite {
+    fn apply(self, db: &Db) {
+        let outcome = match &self {
+            Self::Upsert(nodes) => db.upsert_nodes(nodes),
+            Self::Delete(uid) => db.delete_node(uid),
+            Self::SetListed(uid, listed) => db.set_listed(uid, *listed),
+            Self::SetShareAccess(uid, access) => db.set_share_access(uid, *access),
+        };
+        if let Err(error) = outcome {
+            let what = match &self {
+                Self::Upsert(_) => "upsert_nodes",
+                Self::Delete(_) => "delete_node",
+                Self::SetListed(..) => "set_listed",
+                Self::SetShareAccess(..) => "set_share_access",
+            };
+            warn!(%error, write = what, "db write-through failed");
+        }
+    }
+}
+
+/// The inode lock, held for the duration of a mutation, plus the write-throughs
+/// that mutation owed the database.
+///
+/// Dropping it releases the lock *first* and only then writes, which is the
+/// whole point: interning a five-thousand-child listing commits one transaction
+/// — several fsyncs' worth — and nothing about that needs the maps frozen. The
+/// write still happens synchronously on this thread before the caller proceeds,
+/// so a read that follows a mutation still sees it.
+pub(crate) struct StateGuard<'a> {
+    guard: Option<parking_lot::MutexGuard<'a, State>>,
+    db: &'a Arc<Db>,
+}
+
+/// Take a mount's inode lock through a [`StateGuard`], for the loops that walk
+/// every live mount rather than this `Core`'s own.
+pub(crate) fn lock_state<'a>(
+    state: &'a parking_lot::Mutex<State>,
+    db: &'a Arc<Db>,
+) -> StateGuard<'a> {
+    StateGuard::new(state.lock(), db)
+}
+
+impl<'a> StateGuard<'a> {
+    pub(crate) fn new(guard: parking_lot::MutexGuard<'a, State>, db: &'a Arc<Db>) -> Self {
+        Self {
+            guard: Some(guard),
+            db,
+        }
+    }
+}
+
+impl std::ops::Deref for StateGuard<'_> {
+    type Target = State;
+
+    fn deref(&self) -> &State {
+        self.guard.as_ref().expect("held until drop")
+    }
+}
+
+impl std::ops::DerefMut for StateGuard<'_> {
+    fn deref_mut(&mut self) -> &mut State {
+        self.guard.as_mut().expect("held until drop")
+    }
+}
+
+impl Drop for StateGuard<'_> {
+    fn drop(&mut self) {
+        let Some(mut guard) = self.guard.take() else {
+            return;
+        };
+        let writes = std::mem::take(&mut guard.outbox);
+        // The lock goes before the writes do. That ordering is the feature.
+        drop(guard);
+        for write in writes {
+            write.apply(self.db);
+        }
+    }
 }
 
 impl State {
+    /// A fresh, empty inode space over `db`.
+    ///
+    /// `next_ino` differs by caller: a real mount hands out [`crate::ROOT_INO`]
+    /// to its root and starts allocating above it, while a bare state under test
+    /// has no root and starts at 1.
+    /// Apply the write-throughs this state owes SQLite, here and now.
+    ///
+    /// Production code never calls this: [`StateGuard`] releases the inode lock
+    /// first and then applies them, which is the whole point. A test driving a
+    /// bare `State` has no guard, so it calls this where the daemon would have
+    /// dropped one.
+    #[cfg(test)]
+    pub(crate) fn flush_outbox(&mut self) {
+        let db = self.db.clone();
+        for write in std::mem::take(&mut self.outbox) {
+            write.apply(&db);
+        }
+    }
+
+    /// The database, with everything this state owes it already written.
+    ///
+    /// What a test means by `st.db` is almost always "the database as of now",
+    /// which in the daemon is the state after the guard dropped. Handing back an
+    /// owned `Arc` rather than a borrow keeps it usable inside an `assert_eq!`
+    /// that also touches `st`.
+    #[cfg(test)]
+    pub(crate) fn flushed_db(&mut self) -> Arc<Db> {
+        self.flush_outbox();
+        self.db.clone()
+    }
+
+    pub(crate) fn new(db: Arc<Db>, share_access: HashMap<NodeUid, Access>, next_ino: u64) -> Self {
+        Self {
+            entries: HashMap::new(),
+            by_uid: HashMap::new(),
+            children: HashMap::new(),
+            next_ino,
+            active_writes: HashMap::new(),
+            handles: HashMap::new(),
+            next_fh: 1,
+            access_changes: HashSet::new(),
+            share_access,
+            db,
+            outbox: Vec::new(),
+        }
+    }
     /// Whether `uid` names an owned node currently reachable from this mount's
     /// root. Resident entries can outlive their dentries for open-handle and
     /// access-revocation semantics, so `by_uid` membership alone is not proof
@@ -278,11 +428,9 @@ impl State {
                 || stored.unwrap_or(Access::Viewer),
                 |membership| access_for(membership.role_exact(), parent_access, true),
             );
-            if let Err(error) = self.db.set_share_access(&node.uid, access) {
-                warn!(uid = %node.uid, ?access, %error, "db set_share_access failed");
-            } else {
-                self.share_access.insert(node.uid.clone(), access);
-            }
+            self.outbox
+                .push(DbWrite::SetShareAccess(node.uid.clone(), access));
+            self.share_access.insert(node.uid.clone(), access);
             return access;
         }
         parent_access
@@ -313,11 +461,9 @@ impl State {
             None => (None, Access::Unknown),
         };
         let access = access_for(role, parent_access, true);
-        if let Err(error) = self.db.set_share_access(uid, access) {
-            warn!(%uid, ?access, %error, "db set observed share access failed");
-        } else {
-            self.share_access.insert(uid.clone(), access);
-        }
+        self.outbox
+            .push(DbWrite::SetShareAccess(uid.clone(), access));
+        self.share_access.insert(uid.clone(), access);
         let Some(root) = root else {
             return Vec::new();
         };
@@ -602,18 +748,14 @@ impl State {
             })
             .collect();
         for (uid, access) in discovered {
-            if let Err(error) = self.db.set_share_access(&uid, access) {
-                warn!(%uid, ?access, %error, "db set hydrated share access failed");
-            } else {
-                self.share_access.insert(uid, access);
-            }
+            self.outbox
+                .push(DbWrite::SetShareAccess(uid.clone(), access));
+            self.share_access.insert(uid, access);
         }
     }
 
     pub(crate) fn intern(&mut self, parent: u64, node: Node) -> u64 {
-        if let Err(e) = self.db.upsert_node(&node) {
-            warn!(uid = %node.uid, error = %e, "db upsert_node failed");
-        }
+        self.outbox.push(DbWrite::Upsert(vec![node.clone()]));
         self.intern_mem(parent, node)
     }
 
@@ -685,9 +827,7 @@ impl State {
     /// This is what keeps `ls` on a large folder quick: one commit for the
     /// listing, rather than one autocommit — and one fsync — per child.
     pub(crate) fn intern_batch(&mut self, parent: u64, nodes: Vec<Node>) -> Vec<u64> {
-        if let Err(e) = self.db.upsert_nodes(&nodes) {
-            warn!(error = %e, "db upsert_nodes failed");
-        }
+        self.outbox.push(DbWrite::Upsert(nodes.clone()));
         nodes
             .into_iter()
             .map(|node| self.intern_mem(parent, node))
@@ -769,9 +909,7 @@ impl State {
     /// `(parent_ino, name)` when the node was known, so the caller can notify
     /// the kernel.
     pub(crate) fn forget(&mut self, uid: &NodeUid) -> Option<(u64, String)> {
-        if let Err(e) = self.db.delete_node(uid) {
-            warn!(%uid, error = %e, "db delete_node failed");
-        }
+        self.outbox.push(DbWrite::Delete(uid.clone()));
         self.forget_mem(uid)
     }
 
@@ -834,9 +972,7 @@ impl State {
                 kids.push(ino);
             }
         }
-        if let Err(e) = self.db.upsert_node(&node) {
-            warn!(uid = %node.uid, error = %e, "db upsert_node failed for a queued rename");
-        }
+        self.outbox.push(DbWrite::Upsert(vec![node]));
         self.recompute_descendant_access(ino);
     }
 
@@ -851,9 +987,7 @@ impl State {
         self.children.remove(&ino);
         if let Some(e) = self.entries.get(&ino) {
             let uid = e.uid.clone();
-            if let Err(err) = self.db.set_listed(&uid, false) {
-                warn!(%uid, error = %err, "db set_listed(false) failed");
-            }
+            self.outbox.push(DbWrite::SetListed(uid, false));
         }
     }
 
@@ -920,14 +1054,15 @@ impl State {
     /// file is as long as the caller was told it is. Without it a restart serves
     /// the stale size and the file reads as truncated — or empty — while its
     /// bytes sit safely in staging (offline.md Phase 3).
-    pub(crate) fn record_pending_write(&mut self, ino: u64, size: u64, mtime: i64) {
+    /// Returns the node to persist rather than queueing it in the outbox, and is
+    /// the one mutation here whose write-through the caller must do itself:
+    /// failure has to be reported, not logged. Dropping the row silently means
+    /// the daemon acknowledged a write and then serves the old size for it, so
+    /// the caller has to be able to turn it into a failed `release`.
+    pub(crate) fn record_pending_write(&mut self, ino: u64, size: u64, mtime: i64) -> Option<Node> {
         self.set_size(ino, size);
         self.touch_mtime(ino, mtime);
-        if let Some(e) = self.entries.get(&ino)
-            && let Err(err) = self.db.upsert_node(&e.node)
-        {
-            warn!(uid = %e.uid, error = %err, "db upsert_node failed for a queued write");
-        }
+        self.entries.get(&ino).map(|e| e.node.clone())
     }
 }
 
@@ -1008,18 +1143,7 @@ mod tests {
         let dir = TempDir::new();
         let db = Db::open(&dir.0.join("cache.db")).unwrap();
         let share_access = db.all_share_access().unwrap();
-        let st = State {
-            entries: HashMap::new(),
-            by_uid: HashMap::new(),
-            children: HashMap::new(),
-            next_ino: 1,
-            active_writes: HashMap::new(),
-            handles: HashMap::new(),
-            next_fh: 1,
-            access_changes: HashSet::new(),
-            share_access,
-            db: Arc::new(db),
-        };
+        let st = State::new(Arc::new(db), share_access, 1);
         (st, dir)
     }
 
@@ -1032,8 +1156,8 @@ mod tests {
         let f = st.intern(src, node("f", "src", "f.txt", false));
         st.children.insert(src, vec![f]);
         st.children.insert(dst, vec![]);
-        st.db.set_listed(&uid("src"), true).unwrap();
-        st.db.set_listed(&uid("dst"), true).unwrap();
+        st.flushed_db().set_listed(&uid("src"), true).unwrap();
+        st.flushed_db().set_listed(&uid("dst"), true).unwrap();
         (st, dir, src, dst)
     }
 
@@ -1047,11 +1171,17 @@ mod tests {
         st.relocate(f, src, dst, &uid("dst"), "f.txt");
 
         assert!(
-            st.db.children_if_listed(&uid("dst")).unwrap().is_none(),
+            st.flushed_db()
+                .children_if_listed(&uid("dst"))
+                .unwrap()
+                .is_none(),
             "destination re-enumerates: it gained a child whose row was just deleted"
         );
         assert!(
-            st.db.children_if_listed(&uid("src")).unwrap().is_none(),
+            st.flushed_db()
+                .children_if_listed(&uid("src"))
+                .unwrap()
+                .is_none(),
             "source re-enumerates: its listing predates the move, and a stale \
              `listed` flag would hide every later remote change under it"
         );
@@ -1074,12 +1204,41 @@ mod tests {
         let (mut st, _dir, src, _dst) = two_folders();
         let f = st.by_uid[&uid("f")];
         st.relocate(f, src, src, &uid("src"), "renamed.txt");
-        assert!(st.db.children_if_listed(&uid("src")).unwrap().is_none());
+        assert!(
+            st.flushed_db()
+                .children_if_listed(&uid("src"))
+                .unwrap()
+                .is_none()
+        );
         assert!(!st.children.contains_key(&src));
         assert_eq!(st.by_uid.get(&uid("f")), Some(&f), "the inode is stable");
         assert_eq!(st.entries[&f].node.name, "renamed.txt");
     }
 
+    /// DB1: a mutation's write-through must not happen with the inode lock
+    /// held. Interning a listing commits a transaction over every child, and
+    /// every other FUSE callback used to wait behind that for no reason — the
+    /// maps it needs were updated before the first row was written.
+    #[test]
+    fn a_mutation_writes_through_only_after_the_lock_is_released() {
+        let (st, _dir) = state();
+        let db = st.db.clone();
+        let root = uid("root").to_string();
+        let state = parking_lot::Mutex::new(st);
+        {
+            let mut guard = lock_state(&state, &db);
+            guard.intern(0, node("root", "none", "My Files", true));
+            assert!(
+                db.node_by_uid(&root).unwrap().is_none(),
+                "the row must not be written while the lock is held"
+            );
+        }
+        assert!(
+            db.node_by_uid(&root).unwrap().is_some(),
+            "…and must be written by the time the guard is gone"
+        );
+    }
+
     /// The failure mode the fix exists for, stated directly: forgetting the node
     /// prunes it from the resident listing but leaves the DB claiming the source
     /// is fully enumerated. This pins *why* `relocate` cannot just call `forget`.
@@ -1087,7 +1246,7 @@ mod tests {
     fn forget_alone_leaves_the_source_claiming_a_complete_listing() {
         let (mut st, _dir, _src, _dst) = two_folders();
         st.forget(&uid("f"));
-        let listed = st.db.children_if_listed(&uid("src")).unwrap();
+        let listed = st.flushed_db().children_if_listed(&uid("src")).unwrap();
         assert!(
             listed.is_some(),
             "forget does not clear the flag — which is exactly why relocate must"
@@ -1136,7 +1295,7 @@ mod tests {
         assert!(!st.has_children(folder), "cleared memory children");
 
         // DB children
-        st.db
+        st.flushed_db()
             .upsert_node(&node("file2_uid", "dir_uid", "file2.txt", false))
             .unwrap();
         assert!(
@@ -1162,7 +1321,7 @@ mod tests {
         assert_eq!(st.entries[&shared_ino].access, Access::Editor);
         assert_eq!(st.entries[&child].access, Access::Editor);
         assert_eq!(
-            st.db.share_access(&uid("shared")).unwrap(),
+            st.flushed_db().share_access(&uid("shared")).unwrap(),
             Some(Access::Editor)
         );
     }
@@ -1174,8 +1333,8 @@ mod tests {
         let virtual_uid = NodeUid::new(VolumeId::from("virtual"), LinkId::from("sharedwithme"));
         let mut virtual_root = node("placeholder", "root", "Shared with me", true);
         virtual_root.uid = virtual_uid.clone();
-        st.db.upsert_node(&virtual_root).unwrap();
-        st.db
+        st.flushed_db().upsert_node(&virtual_root).unwrap();
+        st.flushed_db()
             .set_share_access(&virtual_uid, Access::Viewer)
             .unwrap();
         let virtual_ino = st.intern_published_share_root(owned_root, virtual_root, Access::Viewer);
@@ -1183,15 +1342,17 @@ mod tests {
         let mut shared = node("shared", "placeholder", "Retained", true);
         shared.parent_uid = Some(virtual_uid.clone());
         shared.membership = Some(membership(6));
-        st.db.upsert_node(&shared).unwrap();
-        st.db.set_share_access(&shared.uid, Access::Editor).unwrap();
+        st.flushed_db().upsert_node(&shared).unwrap();
+        st.flushed_db()
+            .set_share_access(&shared.uid, Access::Editor)
+            .unwrap();
         let shared_ino =
             st.intern_published_share_root(virtual_ino, shared.clone(), Access::Editor);
         let child = st.intern(
             shared_ino,
             node("shared-child", "shared", "pending.txt", false),
         );
-        st.db
+        st.flushed_db()
             .enqueue_op(&pdfs_core::db::PendingOp {
                 id: 0,
                 kind: pdfs_core::db::OP_RENAME.to_string(),
@@ -1209,20 +1370,20 @@ mod tests {
         assert_eq!(st.entries[&shared_ino].access, Access::Editor);
         assert_eq!(st.entries[&child].access, Access::Editor);
 
-        st.db
+        st.flushed_db()
             .publish_shared_roots(&virtual_uid, std::slice::from_ref(&shared.uid), &[])
             .unwrap();
-        let snapshot = st.db.visible_children(&virtual_uid).unwrap();
+        let snapshot = st.flushed_db().visible_children(&virtual_uid).unwrap();
         assert_eq!(snapshot.len(), 1);
         st.intern_published_share_root(virtual_ino, snapshot[0].clone(), Access::Viewer);
 
         assert_eq!(
-            st.db.share_access(&shared.uid).unwrap(),
+            st.flushed_db().share_access(&shared.uid).unwrap(),
             Some(Access::Viewer)
         );
         assert_eq!(st.entries[&shared_ino].access, Access::Viewer);
         assert_eq!(st.entries[&child].access, Access::Viewer);
-        assert_eq!(st.db.pending_ops().unwrap().len(), 1);
+        assert_eq!(st.flushed_db().pending_ops().unwrap().len(), 1);
     }
 
     #[test]
@@ -1233,12 +1394,12 @@ mod tests {
         let child_node = node("offline-child", "offline-share", "nested folder", true);
         let leaf_node = node("offline-leaf", "offline-child", "child.txt", false);
         for node in [&root_node, &shared_node, &child_node, &leaf_node] {
-            st.db.upsert_node(node).unwrap();
+            st.flushed_db().upsert_node(node).unwrap();
         }
-        st.db
+        st.flushed_db()
             .set_share_access(&uid("offline-share"), Access::Viewer)
             .unwrap();
-        st.share_access = st.db.all_share_access().unwrap();
+        st.share_access = st.flushed_db().all_share_access().unwrap();
 
         // Reconstruct in deliberately non-topological order, matching the
         // database hydration path that first allocates every inode and only
@@ -1296,7 +1457,7 @@ mod tests {
     /// `ORPHAN_INO`. `nodes[0]` becomes the mount root.
     fn hydrate_from(st: &mut State, nodes: &[Node]) -> Vec<u64> {
         for node in nodes {
-            st.db.upsert_node(node).unwrap();
+            st.flushed_db().upsert_node(node).unwrap();
         }
         let inos: Vec<u64> = nodes
             .iter()
@@ -1394,10 +1555,10 @@ mod tests {
     #[test]
     fn persisted_authority_outranks_the_own_volume_fail_open() {
         let (mut st, _dir) = state();
-        st.db
+        st.flushed_db()
             .set_share_access(&uid("device-root"), Access::Viewer)
             .unwrap();
-        st.share_access = st.db.all_share_access().unwrap();
+        st.share_access = st.flushed_db().all_share_access().unwrap();
         let nodes = [
             node("root", "none", "My Files", true),
             node("documents", "device-root", "Documents", true),
@@ -1431,7 +1592,7 @@ mod tests {
             assert_eq!(st.require_writable(ino).unwrap_err().code(), libc::EACCES);
         }
         assert_eq!(
-            st.db.share_access(&uid("shared")).unwrap(),
+            st.flushed_db().share_access(&uid("shared")).unwrap(),
             Some(Access::Viewer)
         );
     }
@@ -1441,8 +1602,10 @@ mod tests {
         for persisted in [Access::Viewer, Access::Editor] {
             let (mut st, _dir) = state();
             let owned_root = st.intern(0, node("root", "none", "My Files", true));
-            st.db.set_share_access(&uid("shared"), persisted).unwrap();
-            st.share_access = st.db.all_share_access().unwrap();
+            st.flushed_db()
+                .set_share_access(&uid("shared"), persisted)
+                .unwrap();
+            st.share_access = st.flushed_db().all_share_access().unwrap();
             let shared = st.intern(
                 owned_root,
                 node("shared", "root", "Shared without membership", true),
@@ -1454,7 +1617,10 @@ mod tests {
 
             assert_eq!(st.entries[&shared].access, persisted);
             assert_eq!(st.entries[&child].access, persisted);
-            assert_eq!(st.db.share_access(&uid("shared")).unwrap(), Some(persisted));
+            assert_eq!(
+                st.flushed_db().share_access(&uid("shared")).unwrap(),
+                Some(persisted)
+            );
         }
     }
 
@@ -1462,7 +1628,7 @@ mod tests {
     fn explicit_missing_share_role_fails_closed_and_marks_kernel_attrs() {
         let (mut st, _dir) = state();
         let owned_root = st.intern(0, node("root", "none", "My Files", true));
-        st.db
+        st.flushed_db()
             .set_share_access(&uid("shared"), Access::Editor)
             .unwrap();
         let shared = st.intern(owned_root, node("shared", "root", "Shared", true));
@@ -1474,7 +1640,7 @@ mod tests {
         assert_eq!(st.entries[&shared].access, Access::Viewer);
         assert_eq!(st.entries[&child].access, Access::Viewer);
         assert_eq!(
-            st.db.share_access(&uid("shared")).unwrap(),
+            st.flushed_db().share_access(&uid("shared")).unwrap(),
             Some(Access::Viewer)
         );
         assert!(changed.contains(&shared));
@@ -1491,7 +1657,7 @@ mod tests {
         let virtual_uid = NodeUid::new(VolumeId::from("virtual"), LinkId::from("sharedwithme"));
         let mut virtual_node = node("sharedwithme", "root", "Shared with me", true);
         virtual_node.uid = virtual_uid.clone();
-        st.db
+        st.flushed_db()
             .set_share_access(&virtual_uid, Access::Viewer)
             .unwrap();
         st.share_access.insert(virtual_uid.clone(), Access::Viewer);
@@ -1558,7 +1724,7 @@ mod tests {
         let child = st.intern(shared, node("child", "shared", "folder", true));
         let leaf = st.intern(child, node("leaf", "child", "leaf.txt", false));
 
-        st.db
+        st.flushed_db()
             .set_share_access(&uid("shared"), Access::Viewer)
             .unwrap();
         let changed = st.downgrade_shared_subtree(&uid("shared"));
@@ -1572,7 +1738,7 @@ mod tests {
         assert_eq!(st.entries[&leaf].access, Access::Viewer);
         assert_eq!(st.require_writable(leaf).unwrap_err().code(), libc::EACCES);
         assert_eq!(
-            st.db.effective_node_access(&uid("leaf")).unwrap(),
+            st.flushed_db().effective_node_access(&uid("leaf")).unwrap(),
             Some(Access::Viewer)
         );
 
@@ -1628,7 +1794,7 @@ mod tests {
         let shared = st.intern(owned_root, shared_node);
         let shared_child = st.intern(shared, node("child", "shared", "child.txt", false));
 
-        st.db.downgrade_all_share_access().unwrap();
+        st.flushed_db().downgrade_all_share_access().unwrap();
         let changed = st.downgrade_known_shared_access();
 
         assert!(changed.contains(&shared));
@@ -1652,7 +1818,7 @@ mod tests {
             let parent_link = st.entries[&parent].uid.link_id.to_string();
             parent = st.intern(parent, node(&link, &parent_link, &link, true));
         }
-        st.db.downgrade_all_share_access().unwrap();
+        st.flushed_db().downgrade_all_share_access().unwrap();
 
         let changed = st.downgrade_known_shared_access();
 

crates/pdfs-fuse/src/sweep.rs

@@ -138,9 +138,9 @@ impl Core {
     /// enumerated every node up front and has been doing network I/O since. Before
     /// touching anything remote this re-reads the node and refuses to act unless
     /// it is *still* the same revision, still untrashed, still unwritten and still
-    /// unopened. Without that, a user edit landing inside the window is destroyed
-    /// — `discard_queued_ops` below drops the queued write *and its staged blob*,
-    /// which may hold the only copy of those bytes (`docs/BUGS.md` B71).
+    /// unopened (`docs/BUGS.md` B71). Nothing on this path discards a queued write
+    /// or its staged blob: a write that arrives after that check is left for the
+    /// drain to resolve against the now-trashed node.
     fn remove_conflict_copy(&self, snapshot: &proton_drive_rs::Node, base: &str) {
         let uid = &snapshot.uid;
         let name = snapshot.name.as_str();
@@ -166,10 +166,13 @@ impl Core {
         self.for_each_state(|st| {
             st.forget_or_unlink(uid);
         });
-        if let Err(error) = self.discard_queued_ops(uid) {
-            warn!(%uid, ?error, "conflict sweep: queued-op cleanup failed");
-            return;
-        }
+        // Deliberately *not* `discard_queued_ops`: `duplicate_still_removable`
+        // proved the queue empty before the trash above, but that proof is stale
+        // the moment `trash_nodes` goes to the network. A `close(2)` landing in
+        // that window stages bytes, and discarding here would unlink the staged
+        // blob holding the only copy of them. Leaving the op queued is the
+        // non-destructive outcome: the drain re-reads the node, sees it trashed,
+        // and `revision_conflict` routes it to `keep_as_conflict_copy`.
         self.cache.evict(uid);
         self.evict_reader(uid);
         if let Err(e) = self.db.delete_node(uid) {
@@ -194,8 +197,8 @@ impl Core {
         let uid = &snapshot.uid;
         let name = snapshot.name.as_str();
 
-        // A queued op means bytes are still owed an upload. `discard_queued_ops`
-        // would throw those away along with the staged blob holding them.
+        // A queued op means bytes are still owed an upload; trashing the node
+        // under them turns an ordinary upload into a conflict fork.
         match self.db.has_any_op(&uid.to_string()) {
             Ok(false) => {}
             Ok(true) => {

crates/pdfs-fuse/src/transfers.rs

@@ -19,7 +19,7 @@ use parking_lot::Mutex;
 use std::collections::HashMap;
 use std::io::{Read, Write};
 use std::sync::Arc;
-use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::time::Instant;
 
 use pdfs_core::control::{JobItem, TransferDirection, TransferItem};
@@ -272,20 +272,50 @@ impl<R: Read> Read for OwnedCountingReader<R> {
     }
 }
 
-/// A [`Read`] that tallies bytes read through a [`TransferGuard`] (upload path).
+/// A [`Read`] that tallies bytes read through a [`TransferGuard`] (upload path),
+/// and can be told to stop.
 pub struct CountingReader<'a, R> {
     inner: R,
     guard: &'a TransferGuard,
+    cancel: Option<Arc<AtomicBool>>,
 }
 
 impl<'a, R: Read> CountingReader<'a, R> {
     pub fn new(inner: R, guard: &'a TransferGuard) -> Self {
-        Self { inner, guard }
+        Self {
+            inner,
+            guard,
+            cancel: None,
+        }
+    }
+
+    /// Abandon the upload as soon as the SDK asks for its next bytes when
+    /// `cancel` is set.
+    ///
+    /// This is the only cancellation point an upload has: the SDK takes a plain
+    /// blocking [`Read`] and pulls from it a block at a time, so refusing to
+    /// hand over the next block is what stops it. Used when a newer revision of
+    /// the same file is queued — the bytes on the wire have already been
+    /// superseded, and finishing them only spends uplink on a revision the next
+    /// op replaces.
+    pub fn with_cancel(mut self, cancel: Arc<AtomicBool>) -> Self {
+        self.cancel = Some(cancel);
+        self
     }
 }
 
 impl<R: Read> Read for CountingReader<'_, R> {
     fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        if let Some(cancel) = &self.cancel
+            && cancel.load(Ordering::Relaxed)
+        {
+            // Not `Ok(0)`: a short read is how this reader says "end of file",
+            // and the SDK would seal a truncated revision from it.
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::Interrupted,
+                "upload superseded by a newer write",
+            ));
+        }
         let n = self.inner.read(buf)?;
         self.guard.add(n as u64);
         Ok(n)

crates/pdfs-fuse/src/upload.rs

@@ -171,7 +171,7 @@ impl Core {
                     stats.bytes += size;
                     match self.fetch_node(&uid) {
                         Ok(node) => {
-                            let mut st = self.state.lock();
+                            let mut st = self.state();
                             let ino = st.intern(parent_ino, node);
                             if let Some(kids) = st.children.get_mut(&parent_ino)
                                 && !kids.contains(&ino)
@@ -216,7 +216,7 @@ impl Core {
         self.ensure_children(pino)
             .map_err(|e| self.errno_error(e, "enumerate"))?;
         let existing = {
-            let st = self.state.lock();
+            let st = self.state();
             st.children.get(&pino).and_then(|kids| {
                 kids.iter().find_map(|ino| {
                     st.entries
@@ -243,7 +243,7 @@ impl Core {
         let node = self
             .fetch_node(&new_uid)
             .map_err(|e| self.errno_error(e, "fetch node"))?;
-        let mut st = self.state.lock();
+        let mut st = self.state();
         let ino = st.intern(pino, node);
         if let Some(kids) = st.children.get_mut(&pino)
             && !kids.contains(&ino)

docs/ARCHITECTURE.md

@@ -153,8 +153,9 @@ Because Proton Drive does not support partial byte writes, modified files must b
 2. **Close/Release (`queue_revision`):** When the application closes the file descriptor, the daemon:
    - Fetches any untouched gaps from the remote base file to compile the full file.
    - Durably publishes the scratch data and authored-range sidecar into `staging` under a `{uid}-{millis}-{counter}` name. Temporary data and metadata are synced before atomic rename, and their directory is synced before the source is removed.
-   - Transactionally queues or supersedes a pending database operation (`PendingOp`), so insertion failure cannot erase the previously acknowledged upload. For `OP_REVISION` ops, execution is debounced by 2 seconds (`DRAIN_REVISION_DEBOUNCE = 2s`) to give rapid follow-up writes (e.g. `aria2c` preallocation followed by writing) time to supersede the staged blob before network transmission.
-3. **Async Drain Thread (`run_pending_drain`):** The background drain worker picks up the database operations queue, handles revision uploads, resolves conflicts, and cleans up staging files. Upon landing a revision upload (`refresh_after_upload`), it rebaselines both still-queued ops (`rebaseline_pending`) and open write handles targeting the same node to prevent false self-conflict copies on subsequent writes.
+   - Transactionally queues or supersedes a pending database operation (`PendingOp`), so insertion failure cannot erase the previously acknowledged upload. For `OP_REVISION` ops, execution is debounced to give rapid follow-up writes (e.g. `aria2c` preallocation followed by writing) time to supersede the staged blob before network transmission. The debounce is adaptive (`Core::revision_debounce`): it starts at `DRAIN_REVISION_DEBOUNCE = 2s` and, once a node has been uploaded, widens toward how long that upload actually took, bounded by `DRAIN_REVISION_DEBOUNCE_MAX = 60s`. A file saved faster than it can be sent therefore supersedes in the queue rather than mid-upload.
+   - Signals `Core::cancel_upload` for the node before touching the queue. A drain worker may already be reading the blob this write supersedes; the flag is read by the upload's `CountingReader`, which refuses the SDK's next block. `queue_trash` and `discard_queued_ops` do the same, because they unlink the blob outright.
+3. **Async Drain Threads (`run_pending_drain`):** `DRAIN_WORKERS = 3` background workers share the operations queue. They pick work up through `Db::claim_next_due_op`, which marks the row `claimed_at` in the same transaction that selects it and excludes any op whose uid another worker already holds — ordering only has to hold *per node*, and that exclusion is what guarantees it. Worker 0 additionally runs the queue's idle chores (LRU touch flush, `recover_fsynced_writes`). A claim is process-local state: the single-writer `flock` means one found at open belongs to a crashed run, and `Db::clear_op_claims` drops the lot so those ops are not invisible forever. Each worker handles revision uploads, resolves conflicts, and cleans up staging files. Upon landing a revision upload (`refresh_after_upload`), it rebaselines both still-queued ops (`rebaseline_pending`) and open write handles targeting the same node to prevent false self-conflict copies on subsequent writes.
 
 ```mermaid
 sequenceDiagram

docs/BUGS.md

@@ -12,6 +12,82 @@ Conventions:
 
 ---
 
+## B83 — An access-deferred op retries forever and is reported as an ordinary queued upload
+
+**Status:** Fixed (partly verified) 2026-08-16 — both halves are in and
+unit-tested. The reporting half was confirmed against the live 28-op backlog on
+1.8.0 (schema migrated to v27, all 28 rows window-stamped, `WARN` lines per op);
+the resolution half has not yet been observed draining that backlog.
+**Found:** 2026-08-16, running `scripts/fuse-acceptance.sh --live` on a
+throwaway CLI-made mount. Four regression cases (B69, B70, B74, B79) failed with
+`TimeoutError: daemon mutation queue did not drain` — the suite's
+`wait_for_queue()` polls the *global* `status.mount.pending_uploads`, which had
+been sitting at 28 for 30 days on the development machine. `pdfs status` called
+them 28 queued uploads holding 18.4 GiB; `pdfs transfers` showed nothing;
+`failing_ops` and `parked_uploads` were both 0.
+**Where:** `run_pending_drain` / `run_authorized_drain`
+(`crates/pdfs-fuse/src/drain.rs`), `require_uid_writable`
+(`crates/pdfs-fuse/src/lib.rs`), `Db::effective_node_access`
+(`crates/pdfs-core/src/db/share_access.rs`),
+`Db::defer_op_without_attempt` (`crates/pdfs-core/src/db/ops.rs`).
+
+**Root cause.** `effective_node_access` returns `Ok(None)` when the uid has no
+row in `nodes`, and `require_uid_access` maps that to `EACCES` — "I have never
+heard of this node" and "you may not write this node" are the same answer. The
+drain reads `EACCES` as `DrainDisposition::AccessDeferred` and calls
+`defer_op_without_attempt`, which by design touches neither `attempts` nor
+`last_error`. So the op was re-deferred every `DRAIN_ACCESS_RECHECK` (5s)
+indefinitely, and every channel that could have surfaced it stayed empty:
+`attempts` 0, `last_error` NULL, `parked` counts only `PARK_UNTIL` rows,
+`failing` counts only `attempts >= FAILING_ATTEMPTS`, and the one log line was
+`debug!`. On the machine that found it, all 28 uids were absent from `nodes`:
+
+```sql
+select (select count(*) from pending_op p
+          where exists(select 1 from nodes n where n.uid = p.uid)) as in_nodes,
+       (select count(*) from pending_op) as total;
+-- 0 | 28
+```
+
+**Fix, part 1 — stop guessing.** `WriteAuthority { Writable, Denied, Unknown }`
+replaces the `Result<(), Errno>` the drain's access check used to return
+(`Core::uid_write_authority`). `require_uid_writable` still collapses both
+refusals to `EACCES`, which is right for a syscall: a stale handle must not be
+admitted because the tree forgot the node. The drain does not collapse them —
+`Unknown` becomes `DrainDisposition::AuthorityUnknown(uid)`, and
+`Core::resolve_unknown_authority` asks the remote:
+
+- node still there → `upsert_node` re-interns it and the op is deferred for one
+  recheck, so the next pass gets a real access answer;
+- node gone (`Ok(None)` / `is_gone`) → `record_op_failure` **immediately**, with
+  a message naming the vanished uid, rather than waiting out the 5-minute window
+  to say something vaguer;
+- could not ask → ordinary deferral.
+
+The authority asked about is the op's own (`pending_op_authorities`), so a
+`create`/`mkdir` refetches its *parent*, not the node it has not made yet. The
+row and its staged blob survive every branch.
+
+**Fix, part 2 — say something.** Schema v27 adds `pending_op.access_deferred_since`.
+`Db::defer_op_for_access` stamps it on the first deferral of a run and returns
+the stamp on every one; `Core::defer_for_access` warns on the first deferral and,
+once the run passes `DRAIN_ACCESS_DEFER_LIMIT` (5 min), reports it through
+`record_op_failure` so it enters the ordinary backoff and shows up in
+`failing_ops` and the status error text. `record_op_failure` clears the window so
+each escalation re-arms it rather than firing every recheck, and an attempt the
+access check admits clears it too (`clear_op_access_deferral`). The row and its
+staged blob are untouched on every path — this changes only what the user can
+see.
+
+The deferral window stays as the backstop for everything that does not resolve,
+including a refetch that keeps not helping, and `AccessDeferral` now carries the
+reason into both the log line and the recorded `last_error`.
+
+**Still open:** the acceptance runner waits on the *global* queue
+(`ManagedSyncPair.wait_for_queue`), so any unrelated stuck op fails those four
+cases on a machine that has one — that is what turned this backlog into four
+apparent FUSE regressions.
+
 ## B82 — A `parent_uid` cycle hangs the daemon on any short search
 
 **Status:** Fixed (unverified) 2026-08-12 — unit-tested, not driven against a

docs/CHANGELOG.md

@@ -0,0 +1,292 @@
+# Changelog
+
+All notable changes to the Linux client since the 1.0.0 release. Format follows
+[Keep a Changelog](https://keepachangelog.com/1.1.0/); versions follow semver, and every
+release is tagged `vX.Y.Z` with `Cargo.toml` and `packaging/PKGBUILD` in agreement.
+
+`Schema:` lines record the SQLite `SCHEMA_VERSION` (`pdfs-core/src/db/migrations.rs`) a
+release ships. Migrations are forward-only — a database written by a newer client is a hard
+refuse-to-open, not a downgrade, so rolling back a release means restoring the cache from
+scratch (user data in `staging/` and `recovery/` is never touched by this).
+
+## [1.8.0] — 2026-08-16
+
+A queued write whose node had left the local tree was retried every five seconds forever and
+counted as an ordinary pending upload — 28 of them, holding 18.4 GiB, had been doing that for
+30 days on the machine that found it (`docs/BUGS.md` B83). Schema: 27 — migration V27 adds
+`pending_op.access_deferred_since`.
+
+### Fixed
+- The drain no longer reads "I have never heard of this node" as "you may not write this
+  node". A missing authority is now resolved against the remote: the node is re-interned if
+  it is still there, and reported as a real failure if it is gone. Ops that had been stuck
+  invisibly for weeks now say what is wrong with them on the first pass.
+- A create or mkdir resolves its *parent* as the authority, so the refetch asks about the
+  folder rather than the node it has not made yet.
+
+### Added
+- Access deferrals are bounded. A deferral still costs no attempt and records no error for
+  the first five minutes — the case it exists for is a share downgraded mid-queue and undone
+  a moment later — but past that window each recheck records a failure, so the op enters the
+  ordinary backoff and appears in `failing_ops` and the `pdfs status` error text like any
+  other stuck operation. The first deferral of a run is logged at `warn` rather than `debug`.
+
+### Notes
+- No staged data is touched by any of this. A reported failure keeps its `pending_op` row and
+  its staged blob exactly as before; the change is in what the user is told, and in the queue
+  no longer waiting on a permission change that was never coming.
+
+## [1.7.1] — 2026-08-16
+
+Write-back throughput follow-up to 1.7.0, same audit
+(`docs/plans/audit-2026-08-15-perf-locks.md`). Schema: 26 — migration V26 adds
+`pending_op.claimed_at`.
+
+### Added
+- Parallel write-back drain: three workers share the queue through a `pending_op.claimed_at`
+  claim column (migration V26), so a large upload no longer holds every queued rename, trash
+  and small write behind it. Ordering is preserved per node — the claim query never offers
+  two workers ops for the same uid — and claims left by a crashed run are cleared at open.
+- An upload already on the wire is cancelled when a newer write, a trash or a discard
+  supersedes it, instead of spending the uplink on a revision the next operation replaces.
+
+### Changed
+- The revision debounce is adaptive: it widens toward how long that node's last upload
+  actually took (bounded to 2–60 s), so a file saved faster than it can be sent supersedes
+  in the queue rather than mid-upload.
+- Workspace deps bumped to `proton-sdk` / `proton-drive-rs` 0.6.0.
+
+## [1.7.0] — 2026-08-16
+
+Performance and robustness pass over the mount, from the architecture audit in
+`docs/plans/audit-2026-08-15-perf-locks.md`. Schema: 25.
+
+### Added
+- Sequential-read prefetch (`Core::prefetch_ahead`): per-file sequential detection, depth
+  doubling 2→8, reset to 0 after a seek, skipping blocks already cached, and bounded by its
+  own permit budget so it can never queue ahead of a demand read.
+- Single-flight block fetching (`BlockFlight`) keyed on `(uid, mtime, size, block)`, so a
+  demand read, kernel read-ahead and prefetch of the same block share one download+decrypt.
+- In-memory block LRU (`BlockRing`) in front of *both* block paths; a block read off disk is
+  promoted into it, so the following ~31 kernel reads of that 4 MiB block cost nothing.
+- Read-only SQLite connection pool (`ReadPool` / `Db::read`) with every `SELECT`-only method
+  routed at it, so a read no longer waits on the write connection.
+- `nodes.path` persisted and backfilled, replacing the recursive `path_of` walk.
+- Single-writer `flock` on the database; a `SQLITE_CORRUPT`/`NOTADB` file is quarantined as
+  `cache.db.corrupt-<ts>` (WAL sidecars too) and rebuilt rather than failing the mount.
+- `VACUUM` and the deep `integrity_check` now ack immediately and run as background jobs;
+  `pdfs doctor` polls for the verdict.
+- Staging reconcile at mount: staged blobs no queued operation refers to are re-queued, or
+  named in the log when they cannot be addressed.
+- Queue/staging visibility — `Status` gained `parked_uploads`, `failing_ops`,
+  `failing_error`, `staged_bytes`, `staged_oldest_secs`; `pdfs status` prints them when
+  non-zero.
+- `statfs` implemented against the account quota; FUSE `init` negotiates `max_readahead`,
+  `max_background`, `PARALLEL_DIROPS` and `CACHE_SYMLINKS`; `blksize` raised to 1 MiB.
+
+### Changed
+- `State` no longer writes to SQLite under the inode lock — mutations queue into an outbox
+  applied after the lock is released, so interning a 5000-child listing no longer commits a
+  transaction with the whole mount frozen.
+- Re-listing an unchanged folder no longer touches its subtree; when a subtree really moved,
+  one recursive walk carries both the new path and the inherited trashed flag.
+- `release(2)` no longer gap-fills from the network: it stages an incomplete blob from local
+  caches only and the drain thread finishes it off the dispatch loop.
+- A partial write over an undrained one is merged instead of refused.
+- Size-upgrade waiters park on a queue instead of occupying a worker thread, capped at 8
+  folders in flight; past the cap `stat` answers provisionally.
+- Block-cache disk writes moved off the read's critical path; no more fsync per cached block.
+- Cache LRU touches buffered in memory and flushed in one transaction.
+- Search: sync-folder list and pin lookups hoisted out of the per-hit loop.
+- SQLite tuned (`cache_size`, `mmap_size`, `temp_store=MEMORY`, `wal_autocheckpoint`) and
+  three new indexes added.
+- Mount registry lock released before FUSE teardown/join; kernel notifications batched under
+  the state lock and flushed after it.
+
+### Fixed
+- **Data loss:** `ENOSPC` is now propagated from `write`/`fsync`/staging instead of being
+  swallowed, with emergency eviction wired up and scratch marked durable before any staging
+  error; a failed `record_pending_write` fails `release` rather than dropping the write.
+- **Data loss:** the conflict sweep no longer discards queued operations after the trash
+  round trip.
+- **Data loss:** `rescue_scratch` moves a blob with an unreadable sidecar to `recovery/`
+  unaccompanied instead of fabricating a sidecar for it.
+- Own-sealed revisions are persisted (7-day TTL), so a restart no longer reopens the
+  transient-download fork window fixed in 1.1.0.
+- Depth-capped `is_pinned` (cycle-safe), daemon write timeout, and every client-supplied
+  `limit` clamped.
+- `recover_fsynced_writes` no longer scans the recovery directory on every idle pass.
+
+## [1.6.0] — 2026-08-12
+
+Search-index correctness and ranking (`docs/plans/search-index.md`). Schema: 21 —
+migration V21 rebuilds `nodes_fts`.
+
+### Fixed
+- **B80** — 29% of the account was missing from the search index. `node_is_indexable_tx`
+  required a null-parent ancestor, which only the My Files root has, so every node under a
+  device folder was dropped from `nodes_fts` on its first upsert after the v16 backfill
+  (2,705 of 9,252 nodes on the account this was diagnosed against). The rule now matches the
+  backfill, migration V21 replays it, nodes under a trashed folder stay out, and the walk is
+  depth-capped.
+- **B81** — the prompt labelled every Drive hit "My files", including device folders; pin
+  rows also took `is_dir` from `Pin::recursive`.
+- **B82** — a `parent_uid` cycle hung the daemon on any short search: the recursive CTEs in
+  `path_of` and the pin walk use `UNION ALL`, which does not deduplicate. `path_of` is now
+  depth-capped at 256 and returns the truncated path.
+
+### Changed
+- Local search ranking: whole-query prefix matches get their own lane (newest first) ahead
+  of trigram rank, which on a home directory full of dependency trees was burying the
+  obvious answer — 73% of the 500 best-ranked candidates for "test" came from one Go module
+  cache. The single-character lane, which recovers a typo that destroys every trigram, stays
+  behind it.
+
+## [1.5.0] — 2026-08-11
+
+Schema: 20.
+
+### Added
+- **`pdfs-prompt --dmenu`** — the same search through an external dmenu-style launcher
+  (fuzzel, rofi, wofi, tofi, bemenu, dmenu) instead of the built-in GTK HUD, so the prompt
+  matches a tiling-WM desktop. A launcher filters a fixed list rather than re-querying per
+  keystroke, so searching is a two-step loop; `--query` skips the first step, fuzzel and rofi
+  get file-type icons. Configurable as `"prompt": { "mode": "dmenu", "menu": [...] }`, with
+  `--gtk` overriding for one invocation.
+- **Configurable open-with rules** (`pdfs-core/src/opener.rs`): an ordered list of name/class
+  patterns in `config.json` under `open_with`, each mapped to argv and optionally wrapped in
+  a terminal, first match wins, falling back to `xdg-open`. Lets "open a text file" mean
+  `nvim` in a terminal without breaking every graphical caller. Absent config reproduces the
+  old behaviour exactly.
+
+### Fixed
+- The prompt and dmenu openers hand applications the cached file while matching rules
+  against the *original* Drive filename, so e.g. `.md` files stop being routed by their
+  cache-path extension.
+
+## [1.4.0] — 2026-08-10
+
+Schema: 20.
+
+### Added
+- **File version history** (`pdfs-fuse/src/revisions.rs` + a GTK versions dialog): list a
+  file's revisions, restore one, delete one, or write an old one out to a local path. A
+  restore re-points the file server-side, so no content crosses the wire and nothing enters
+  the drain queue; the daemon drops its cached content and repopulates on the next read.
+- **Persistent SDK entity cache** (`pdfs-core/src/sdkcache.rs`): the SDK's decrypted node
+  metadata now survives a daemon restart instead of re-walking the tree. It lives in its own
+  `sdk_cache.db`, deliberately off the daemon's hot `cache.db` mutex, holds no key material,
+  needs no migrations, and is cleared on logout.
+- Per-node outcomes for batch operations (`pdfs-core/src/batch.rs`): `trash`/`restore`/
+  `delete`/`move` report one result per node, so a node the server rejects no longer fails
+  the whole batch.
+- Photo viewer page in the GUI and richer photo metadata in the details pane.
+
+### Changed
+- SDK telemetry routed into `tracing`, so its transfer/block-storage/request spans land in
+  the daemon log instead of a no-op sink.
+- A file that fits in one block uploads as a single atomic request rather than the
+  draft/block/commit sequence.
+
+## [1.3.0] — 2026-07-31
+
+Schema: 19.
+
+### Added
+- **Photo albums** — `PhotoAlbums` / `AlbumPhotos` control requests, daemon-side `albums.rs`
+  and a GUI albums page (the former gallery view, renamed). Albums other people share with
+  you are included; their photos are persisted per album because a shared album's photos
+  never appear in your own timeline. Served stale-while-refreshing like the timeline.
+- **GUI "Locations" page** — one row per mount spec: the primary `~/ProtonDrive` mount
+  alongside every folder this computer backs up, mirrored or on-demand. Called *Locations*
+  rather than *Mounts* because a mirror folder is a plain directory with no FUSE session.
+  It absorbs the old Computers page's actions and the mountpoint chooser from Settings.
+- `role` on shared-node metadata (`viewer`/`editor`/`admin`), defaulted for wire-compat with
+  older clients and daemons.
+- FUSE acceptance coverage for the permission-based mounts.
+
+## [1.2.1] — 2026-07-31
+
+Schema: 18.
+
+### Fixed
+- Migration bug that could leave the `ProtonDrive` mount unwritable after upgrade.
+- Permission error on on-demand mounts.
+- B78.
+
+### Changed
+- Release pipeline: `.deb` signing added and `.rpm` signing fixed.
+
+## [1.2.0] — 2026-07-29
+
+Sharing and multi-mount release. Schema: 18.
+
+### Added
+- **Shared with me**: shares others sent you are listed, verified and mountable, including on
+  the primary mount.
+  They appear under a synthetic `Shared with me` directory on a virtual volume
+  (`pdfs-fuse/src/virtual.rs`), with name collisions and over-long components resolved by
+  its own naming rules.
+- Per-share access model (`db/share_access.rs`): the effective role of each shared root is
+  persisted and kept resident in mount state, so interning an inode never costs a query. The
+  database stays the restart/offline authority — every role change is written through before
+  the cache updates. A share you cannot write is presented POSIX read-only.
+- Unified location model (`pdfs-core/src/mounts.rs`, `db/mounts.rs`): one presentation shape
+  for every local Proton Drive location, joining device sync state with a new `mount` table.
+  `MountMode` is `Mirror` (full local copy, reconciled both ways) or `OnDemand` (FUSE session
+  fetching on access), with an `Unknown` fallback so a mode written by a newer client does
+  not break an older one.
+
+## [1.1.1] — 2026-07-27
+
+Schema: 16.
+
+### Fixed
+- Trash refresh did not pull updates.
+- Packaging: `PKGBUILD` was left at 1.1.0 by the 1.1.1 version bump, failing the release
+  workflow's tag/manifest/PKGBUILD agreement check.
+
+## [1.1.0] — 2026-07-26
+
+Conflict-handling and concurrency release. Schema: 16.
+
+### Added
+- Conflict sweep (`pdfs-fuse/src/sweep.rs`, 5-minute cadence): removes conflict copies proven
+  identical by size + `content_sha1`, surfaces divergent ones once as an activity entry, and
+  never removes what it cannot prove.
+- Sweep mode and `.pdfsignore`-driven conflict handling; ignore rules are rebuilt per sync
+  pass so edits take effect without a restart.
+- Substantially extended FUSE acceptance suite.
+
+### Fixed
+- **B69** — spurious `(sync-conflict)` copies. Revision identity now uses the server's
+  `revision_id` rather than `(mtime, size)`, which the server re-stamps on seal; the planner
+  gained `AdoptBaseline` so an equal-size no-baseline pair is recorded instead of forked on
+  first reconcile.
+- **B70** — browser downloads into an on-demand mount had their in-flight temporaries sealed
+  as canonical revisions. Transient names are recognised and held until the name is finalized.
+- Assorted race conditions and concurrency bugs on the write/drain paths.
+
+### Changed
+- Workspace deps bumped to `proton-sdk` / `proton-drive-rs` 0.2.2.
+
+## [1.0.0] — 2026-07-22
+
+First stable release: FUSE files-on-demand mount, sync daemon under `proton-drive.service`,
+`pdfs` CLI and the GTK4 app, on top of the pure-Rust Proton Drive SDK. Schema: 16.
+
+### Fixed
+- The outstanding FUSE defects tracked in `docs/BUGS.md`, plus a truncate defect, validated
+  by a new POSIX compliance suite for the filesystem.
+
+[1.7.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.1
+[1.7.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.0
+[1.6.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.6.0
+[1.5.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.5.0
+[1.4.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.4.0
+[1.3.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.3.0
+[1.2.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.2.1
+[1.2.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.2.0
+[1.1.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.1.1
+[1.1.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.1.0
+[1.0.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.0.0

docs/plans/search-index.md

@@ -1,110 +0,0 @@
-# Search index + prompt display fixes
-
-Working plan. Findings come from an audit of a real `cache.db`
-(9,252 Drive nodes, 110,187 local files) on 2026-08-12.
-
-## Evidence
-
-```
-nodes                     9252
-nodes_fts                 6547     ← 2705 rows never indexed
-nodes in orphan subtrees  2705     ← exactly the missing set
-local_files              110187
-  /home/narl/go/pkg/mod   63399    (58% of the local index)
-query "test"             46405 local matches, candidate pool capped at 1000
-  top 500 by fts rank      367 (73%) from go/pkg/mod
-```
-
-`pdfs search tickets` returns only `Documents/Tickets` (My Files); the
-`tickets.pdf` sitting in the device folder `Downloads` is invisible.
-
-## P1 — 29% of Drive is unsearchable
-
-`node_is_indexable_tx` (`crates/pdfs-core/src/db/nodes.rs:606`) indexes a node
-only when the ancestor walk terminates at a row with `parent_uid IS NULL`. The
-DB holds exactly one such row (the My Files root). Device roots live in
-`device`/`sync_folder`, never in `nodes`, so every device-folder subtree
-(`Downloads`, `Pictures`, `Documents`, `Music`, `Videos` under device
-*nthinkpad*) walks up to a `parent_uid` with no row and is judged unindexable.
-Stale `pdfs-acceptance-*` roots hit the same path.
-
-- [x] Relax the rule to "the walk terminates and no ancestor is trashed",
-      keeping the existing cycle guard. A subtree whose top row's parent is
-      simply not cached is still real, reachable data.
-- [x] `MIGRATION_V21`: rebuild `nodes_fts` under the new rule, so existing
-      installs recover the missing rows instead of waiting for each node to be
-      upserted again.
-- [x] Tests: a node under an uncached parent is indexed; a trashed ancestor
-      still excludes; the migration backfills.
-
-### Found while testing: a parent cycle hangs the daemon
-
-Writing the cycle test surfaced a live defect, not a test artefact. A query
-below the trigram minimum takes the `LIKE` lane, which reads `nodes` directly
-and never consults the index, and every row it returns goes through `path_of` —
-whose parent walk is an unguarded `UNION ALL`. On a `parent_uid` cycle it never
-terminates, while holding the daemon's only SQLite connection.
-
-- [x] Depth-cap `path_of` (`db/utils.rs`), matching the 1024 cap
-      `path_relative_to` already carried.
-- [x] Depth-cap the `descendants` CTE in `upsert_node_tx`, which has the same
-      exposure on the write side.
-- [x] Keep cycles out of the index (part of the P1 rule above).
-
-Still unguarded, not touched here: `pin_is_pinned`'s ancestor walk.
-
-## P2 — local index noise crowds real files out of the candidate pool
-
-`SKIP_DIRS` (`crates/pdfs-core/src/localindex.rs:23`) misses the Go module
-cache and build output dirs. `search_v2` clamps candidates to ≤1000 and scores
-*after* SQLite, so for any common term the pool is exhausted by vendored files
-before a real one is considered.
-
-- [x] Skip `go/pkg/mod` (path-relative, not a bare name), `dist`, `build`,
-      `.cache`-style artefact dirs.
-- [x] Give `search_local_candidates` a name-prefix lane, mirroring what
-      `search_candidates` already does for Drive, so an exact-prefix match
-      cannot be starved by trigram rank.
-- [x] Tests for both.
-
-## P3 — display lies about where a hit lives
-
-- [x] `Hit::location()` (`crates/pdfs-gui/src/prompt.rs:150`) hardcodes
-      `My files / …` for every Drive hit. Derive the label from
-      `mounted_path` when the daemon resolved one (device folder, sync folder)
-      and fall back to `My files` only for the primary mount.
-- [x] Pin rows set `is_dir: pin.recursive` (`dmenu.rs:133`, `prompt.rs:872`) —
-      a non-recursively pinned folder renders and activates as a file.
-- [ ] `pdfs locations` prints `Device folder 1..5` rather than the device and
-      folder name. (Cosmetic; deferred.)
-
-## Verification
-
-Done, 2026-08-12:
-
-```
-cargo fmt --all -- --check                                  clean
-cargo clippy --workspace --all-targets -- -D warnings       clean
-cargo test --workspace --locked                             439 passed, 0 failed
-```
-
-The V21 backfill was also replayed against a read-only `.backup` of the live
-`cache.db`:
-
-```
-live_nodes         9276
-indexed            9276     (was 6547)
-downloads_subtree   300
-tickets.pdf ->     Downloads/tickets.pdf
-```
-
-That run is what caught the last defect: the backfill originally seeded an
-uncached-parent root with an empty path, so a node indexed by the migration
-(`tickets.pdf`) and the same node after a later upsert (`Downloads/tickets.pdf`)
-disagreed. The seed now keeps the name for any row that has a parent, matching
-`path_of`.
-
-Still to do, on the live system: rebuild, restart `proton-drive.service`, and
-confirm `pdfs search tickets` returns the device folder's `tickets.pdf`. The
-schema move to 21 is one-way — an older `pdfs` refuses a v21 database. The
-`local_files` shrink (~110k → ~47k) lands with the next background scan.

packaging/PKGBUILD

@@ -9,7 +9,7 @@
 # Build + install:   cd packaging && makepkg -fi
 
 pkgname=proton-drive-linux
-pkgver=1.6.0
+pkgver=1.8.0
 pkgrel=1
 pkgdesc="Proton Drive client for Linux: FUSE files-on-demand mount, CLI, GTK4 app + tray"
 arch=('x86_64')