fix(core): retry the database lock briefly before calling it another daemon (B290)

A child spawned by another thread holds a copy of the lock descriptor
between its fork and its exec, so a reopen right after a drop could be
refused. Seen once in CI in
what_a_create_sent_outlives_a_newer_blob_its_trash_and_a_restart.
This commit is contained in:
nvrl committed 2026-11-15 19:42:22 +01:00
1 parent fd6b2d3e85
commit 593df5c401
2 files changed
+48 -6

No files matched your search

+23 -6
View File
@@ -445,6 +445,13 @@ fn is_in_memory(path: &Path) -> bool {
/// 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.
///
/// A refusal is retried for [`LOCK_RETRY`] before it counts as another
/// daemon. A child this process spawns holds a copy of every descriptor from
/// its fork until its exec closes them, and `flock` belongs to the open file,
/// not the descriptor. So a lock the process has just let go of can stay held
/// for that moment, and a reopen right after a drop is refused
/// (`docs/BUGS.md` B290).
fn acquire_single_writer_lock(path: &Path) -> Result<Option<File>> {
if is_in_memory(path) {
return Ok(None);
@@ -456,22 +463,32 @@ fn acquire_single_writer_lock(path: &Path) -> Result<Option<File>> {
.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 deadline = std::time::Instant::now() + LOCK_RETRY;
loop {
// 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 {
return Ok(Some(file));
}
let e = std::io::Error::last_os_error();
if e.raw_os_error() == Some(libc::EWOULDBLOCK) {
if e.raw_os_error() != Some(libc::EWOULDBLOCK) {
return Err(Error::Io(e));
}
if std::time::Instant::now() >= deadline {
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));
std::thread::sleep(std::time::Duration::from_millis(20));
}
Ok(Some(file))
}
/// How long a refused single-writer lock is retried; far longer than a fork
/// takes to reach its exec, far shorter than anyone notices.
const LOCK_RETRY: std::time::Duration = std::time::Duration::from_millis(500);
fn lock_path_for(path: &Path) -> PathBuf {
let mut name = path.as_os_str().to_owned();
name.push(".lock");