fix(downloads): harden lifecycle cancellation races

This commit is contained in:
NimBold
2026-07-15 00:10:56 +03:30
parent e07182fbf2
commit e8487ee71b
4 changed files with 318 additions and 63 deletions
+19
View File
@@ -238,6 +238,12 @@ fn import_legacy_data(
}
}
}
if portable {
// Sanitize before importing as well as sanitizing the legacy
// source afterward. A crash between those two operations must
// not leave raw transfer credentials in the portable database.
sanitize_download_strings(&mut legacy.downloads)?;
}
merge_legacy_data(connection, legacy)?;
if migration_complete {
connection
@@ -779,6 +785,19 @@ fn mark_portable_download_unresumable(object: &mut serde_json::Map<String, Value
}
}
fn sanitize_download_strings(downloads: &mut [String]) -> Result<(), String> {
for data in downloads {
let mut value: Value = serde_json::from_str(data).map_err(|error| {
format!("failed to decode download for portable sanitization: {error}")
})?;
remove_persisted_transfer_secrets(&mut value);
*data = serde_json::to_string(&value).map_err(|error| {
format!("failed to encode download for portable sanitization: {error}")
})?;
}
Ok(())
}
fn sanitize_persisted_downloads(connection: &mut Connection) -> Result<(), String> {
let transaction = connection
.transaction()
+137 -7
View File
@@ -142,7 +142,9 @@ async fn run_coordinator(
mut media_rx: mpsc::Receiver<MediaCmd>,
) {
let mut active_media = HashMap::<String, (u64, watch::Sender<bool>)>::new();
let mut pending_media_acks = HashMap::<String, (u64, tokio::sync::oneshot::Sender<()>)>::new();
let mut cancelled_media_generations = HashMap::<String, u64>::new();
let mut pending_media_acks =
HashMap::<(String, u64), tokio::sync::oneshot::Sender<()>>::new();
let mut pending_captured_urls = Vec::<String>::new();
let mut frontend_ready = false;
@@ -180,24 +182,53 @@ async fn run_coordinator(
};
match command {
MediaCmd::Register { id, lifecycle_generation, cancel_tx } => {
if let Some((_, previous)) = active_media.insert(id, (lifecycle_generation, cancel_tx)) {
if active_media
.get(&id)
.is_some_and(|(generation, _)| *generation > lifecycle_generation)
{
let _ = cancel_tx.send(true);
continue;
}
let pending_cancel = cancelled_media_generations.get(&id).copied();
if pending_cancel.is_some_and(|generation| generation < lifecycle_generation) {
cancelled_media_generations.remove(&id);
}
let cancelled = pending_cancel.is_some_and(|generation| generation >= lifecycle_generation);
if let Some((_, previous)) = active_media.insert(id.clone(), (lifecycle_generation, cancel_tx)) {
let _ = previous.send(true);
}
if cancelled {
if let Some((_, cancel_tx)) = active_media.get(&id) {
let _ = cancel_tx.send(true);
}
if pending_cancel == Some(lifecycle_generation) {
cancelled_media_generations.remove(&id);
}
}
}
MediaCmd::Pause { id, lifecycle_generation } => {
if active_media.get(&id).is_some_and(|(generation, _)| *generation == lifecycle_generation) {
if let Some((_, cancel_tx)) = active_media.remove(&id) {
let _ = cancel_tx.send(true);
}
} else {
cancelled_media_generations
.entry(id)
.and_modify(|generation| *generation = (*generation).max(lifecycle_generation))
.or_insert(lifecycle_generation);
}
}
MediaCmd::PauseWithAck { id, lifecycle_generation, ack } => {
if active_media.get(&id).is_some_and(|(generation, _)| *generation == lifecycle_generation) {
if let Some((_, cancel_tx)) = active_media.remove(&id) {
let _ = cancel_tx.send(true);
pending_media_acks.insert(id, (lifecycle_generation, ack));
pending_media_acks.insert((id, lifecycle_generation), ack);
}
} else {
cancelled_media_generations
.entry(id)
.and_modify(|generation| *generation = (*generation).max(lifecycle_generation))
.or_insert(lifecycle_generation);
let _ = ack.send(());
}
}
@@ -205,10 +236,16 @@ async fn run_coordinator(
if active_media.get(&id).is_some_and(|(generation, _)| *generation == lifecycle_generation) {
active_media.remove(&id);
}
if pending_media_acks.get(&id).is_some_and(|(generation, _)| *generation == lifecycle_generation) {
if let Some((_, ack)) = pending_media_acks.remove(&id) {
let _ = ack.send(());
}
if let Some(ack) =
pending_media_acks.remove(&(id.clone(), lifecycle_generation))
{
let _ = ack.send(());
}
if cancelled_media_generations
.get(&id)
.is_some_and(|generation| *generation <= lifecycle_generation)
{
cancelled_media_generations.remove(&id);
}
}
}
@@ -317,4 +354,97 @@ mod tests {
.unwrap();
assert!(*new_cancel.borrow_and_update());
}
#[tokio::test]
async fn pause_before_media_registration_cancels_the_late_lifecycle() {
let (coordinator, _events) = DownloadCoordinator::spawn_headless();
let (ack_tx, ack_rx) = tokio::sync::oneshot::channel();
coordinator
.pause_media_with_ack("late-media".to_string(), 7, ack_tx)
.await
.unwrap();
tokio::time::timeout(Duration::from_secs(1), ack_rx)
.await
.unwrap()
.unwrap();
let mut cancel_rx = coordinator
.register_media("late-media".to_string(), 7)
.await
.unwrap();
tokio::time::timeout(Duration::from_secs(1), async {
while !*cancel_rx.borrow_and_update() {
cancel_rx.changed().await.unwrap();
}
})
.await
.expect("late media registration was not cancelled");
}
#[tokio::test]
async fn stale_media_registration_cannot_replace_a_newer_lifecycle() {
let (coordinator, _events) = DownloadCoordinator::spawn_headless();
let mut new_cancel = coordinator
.register_media("same-id".to_string(), 2)
.await
.unwrap();
let mut old_cancel = coordinator
.register_media("same-id".to_string(), 1)
.await
.unwrap();
tokio::time::timeout(Duration::from_secs(1), old_cancel.changed())
.await
.unwrap()
.unwrap();
assert!(*old_cancel.borrow_and_update());
let (ack_tx, ack_rx) = tokio::sync::oneshot::channel();
coordinator
.pause_media_with_ack("same-id".to_string(), 2, ack_tx)
.await
.unwrap();
coordinator.finish_media("same-id".to_string(), 2).await;
tokio::time::timeout(Duration::from_secs(1), ack_rx)
.await
.unwrap()
.unwrap();
assert!(*new_cancel.borrow_and_update());
}
#[tokio::test]
async fn media_pause_ack_is_preserved_across_lifecycle_replacement() {
let (coordinator, _events) = DownloadCoordinator::spawn_headless();
let _old_cancel = coordinator
.register_media("same-id".to_string(), 1)
.await
.unwrap();
let (old_ack_tx, old_ack_rx) = tokio::sync::oneshot::channel();
coordinator
.pause_media_with_ack("same-id".to_string(), 1, old_ack_tx)
.await
.unwrap();
let _new_cancel = coordinator
.register_media("same-id".to_string(), 2)
.await
.unwrap();
let (new_ack_tx, new_ack_rx) = tokio::sync::oneshot::channel();
coordinator
.pause_media_with_ack("same-id".to_string(), 2, new_ack_tx)
.await
.unwrap();
coordinator.finish_media("same-id".to_string(), 1).await;
tokio::time::timeout(Duration::from_secs(1), old_ack_rx)
.await
.unwrap()
.unwrap();
coordinator.finish_media("same-id".to_string(), 2).await;
tokio::time::timeout(Duration::from_secs(1), new_ack_rx)
.await
.unwrap()
.unwrap();
}
}
+98 -24
View File
@@ -3357,11 +3357,11 @@ async fn pause_download(
let _control_guard = state.queue_manager.acquire_aria2_control(&id).await;
let active_kind = state.queue_manager.active_kind(&id).await;
let media_lifecycle_generation = state
let registered_lifecycle_generation = state
.queue_manager
.registered_lifecycle_generation(&id)
.await
.unwrap_or_default();
.await;
let media_lifecycle_generation = registered_lifecycle_generation.unwrap_or_default();
let removed_pending = state.queue_manager.remove_from_pending(&id).await;
let gid = state.queue_manager.aria2_gid_for_download(&id);
@@ -3379,17 +3379,68 @@ async fn pause_download(
log::info!("aria2 pause [{}]: gid {} was already paused", id, gid);
}
"active" | "waiting" => {
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
let result = rpc_call(
let pause_result = match rpc_call(
state.aria2_port.load(std::sync::atomic::Ordering::Relaxed),
&state.aria2_secret,
"aria2.forcePause",
serde_json::json!([gid]),
)
.await
.map_err(|error| format!("failed to pause aria2 gid {gid}: {error}"))?;
ensure_aria2_gid_result("forcePause", gid, &result)?;
{
Ok(result) => ensure_aria2_gid_result("forcePause", gid, &result),
Err(error) => Err(format!("failed to pause aria2 gid {gid}: {error}")),
};
if let Err(pause_error) = pause_result {
match aria2_download_status(
state.aria2_port.load(std::sync::atomic::Ordering::Relaxed),
&state.aria2_secret,
gid,
)
.await
{
Ok(status) if status == "paused" => {
log::info!(
"aria2 pause [{}]: forcePause failed but gid {} is paused",
id,
gid
);
}
Ok(status) if status == "complete" => {
state
.queue_manager
.apply_completion_locked(&id, crate::queue::PendingOutcome::Complete)
.await;
return Ok(());
}
Ok(status) if matches!(status.as_str(), "error" | "removed") => {
let terminal_error = format!(
"cannot pause aria2 gid {gid}: {pause_error}; daemon reports terminal state {status}"
);
state
.queue_manager
.apply_completion_locked(
&id,
crate::queue::PendingOutcome::Error(terminal_error.clone()),
)
.await;
return Err(terminal_error);
}
Ok(status) => {
state.queue_manager.allow_aria2_retries(&id).await;
return Err(format!(
"{pause_error}; aria2 gid {gid} is still {status}"
));
}
Err(status_error) => {
state.queue_manager.allow_aria2_retries(&id).await;
return Err(format!(
"{pause_error}; failed to verify aria2 gid {gid}: {status_error}"
));
}
}
}
state.queue_manager.next_aria2_control_epoch(&id).await;
log::info!("aria2 pause [{}]: gid {} paused", id, gid);
}
"complete" => {
@@ -3442,6 +3493,13 @@ async fn pause_download(
return Ok(());
}
// A terminal runner may have already released its lifecycle while this
// command was waiting for the per-download control lock. Treat pause as
// an idempotent no-op instead of emitting a stale Paused state.
if active_kind.is_none() && !removed_pending && registered_lifecycle_generation.is_none() {
return Ok(());
}
if matches!(active_kind, Some(crate::queue::TaskKind::Aria2)) {
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
@@ -3460,11 +3518,19 @@ async fn pause_download(
rx.await
.map_err(|_| "download worker stopped without acknowledging pause".to_string())?;
if !matches!(active_kind, Some(crate::queue::TaskKind::Media)) {
state.queue_manager.release_permit(&id).await;
}
if removed_pending || matches!(active_kind, Some(crate::queue::TaskKind::Aria2)) {
state.queue_manager.release_registered_id(&id).await;
state.queue_manager.release_permit(&id).await;
if registered_lifecycle_generation.is_some()
|| removed_pending
|| matches!(active_kind, Some(crate::queue::TaskKind::Aria2))
{
if let Some(generation) = registered_lifecycle_generation {
state
.queue_manager
.release_registered_id_for_generation(&id, generation)
.await;
} else {
state.queue_manager.release_registered_id(&id).await;
}
}
use tauri::Emitter;
let _ = app_handle.emit(
@@ -3726,11 +3792,12 @@ async fn remove_download(
.unwrap_or_default();
state.queue_manager.remove_from_pending(&id).await;
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
let gid = state.queue_manager.aria2_gid_for_download(&id);
if let Some(gid) = gid.as_deref() {
// Keep the current epoch and mapping alive until daemon removal is
// confirmed. If removal fails, terminal events from this lifecycle
// must still be accepted so the permit and mapping can be cleaned up.
state.queue_manager.cancel_aria2_retries(&id).await;
let removal_result = async {
force_remove_aria2_gid(
state.aria2_port.load(std::sync::atomic::Ordering::Relaxed),
@@ -3750,11 +3817,17 @@ async fn remove_download(
state.queue_manager.allow_aria2_retries(&id).await;
return Err(error);
}
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.clear_aria2_retry_state(&id).await;
state.queue_manager.forget_aria2_gid(&id).await;
state.queue_manager.release_permit(&id).await;
log::info!("aria2 remove [{}]: gid {} stopped and forgotten", id, gid);
} else {
// There may be an addUri or retry handoff in flight with no mapped
// GID yet. Invalidate that pending lifecycle before acknowledging the
// removal so its late result cannot resurrect the download.
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
let (tx, rx) = tokio::sync::oneshot::channel();
if matches!(active_kind, Some(crate::queue::TaskKind::Media)) {
state
@@ -3766,9 +3839,7 @@ async fn remove_download(
}
let _ = rx.await;
if !matches!(active_kind, Some(crate::queue::TaskKind::Media)) {
state.queue_manager.release_permit(&id).await;
}
state.queue_manager.release_permit(&id).await;
state.queue_manager.clear_aria2_retry_state(&id).await;
state.queue_manager.forget_aria2_gid(&id).await;
}
@@ -3881,12 +3952,13 @@ async fn detach_download_for_reconfigure(
.await
.unwrap_or_default();
state.queue_manager.remove_from_pending(&id).await;
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
let gid = state.queue_manager.aria2_gid_for_download(&id);
if let Some(gid) = gid.as_deref() {
// Do not invalidate the mapped lifecycle until the daemon confirms
// the detach. A failed pause must leave terminal-event reconciliation
// and permit ownership intact.
state.queue_manager.cancel_aria2_retries(&id).await;
let removal_result = async {
let pause_res = rpc_call(
state.aria2_port.load(std::sync::atomic::Ordering::Relaxed),
@@ -3914,12 +3986,16 @@ async fn detach_download_for_reconfigure(
state.queue_manager.allow_aria2_retries(&id).await;
return Err(error);
}
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.clear_aria2_retry_state(&id).await;
state.queue_manager.forget_aria2_gid(&id).await;
state.queue_manager.release_permit(&id).await;
state.queue_manager.release_registered_id(&id).await;
log::info!("aria2 detach [{}]: gid {} stopped and forgotten", id, gid);
} else {
// Invalidate a queued/addUri lifecycle that has not published a GID.
state.queue_manager.next_aria2_control_epoch(&id).await;
state.queue_manager.cancel_aria2_retries(&id).await;
let (tx, rx) = tokio::sync::oneshot::channel();
if matches!(active_kind, Some(crate::queue::TaskKind::Media)) {
state
@@ -3931,9 +4007,7 @@ async fn detach_download_for_reconfigure(
}
let _ = rx.await; // Wait for the writer to stop
if !matches!(active_kind, Some(crate::queue::TaskKind::Media)) {
state.queue_manager.release_permit(&id).await;
}
state.queue_manager.release_permit(&id).await;
state.queue_manager.clear_aria2_retry_state(&id).await;
state.queue_manager.forget_aria2_gid(&id).await;
state.queue_manager.release_registered_id(&id).await;
+64 -32
View File
@@ -277,7 +277,7 @@ impl<R: tauri::Runtime> QueueManager<R> {
== Some(generation)
}
async fn release_registered_id_for_generation(&self, id: &str, generation: u64) {
pub(crate) async fn release_registered_id_for_generation(&self, id: &str, generation: u64) {
let released = {
let mut registered = self.registered_ids.lock().await;
let mut generations = self.registered_lifecycle_generations.lock().await;
@@ -362,7 +362,7 @@ impl<R: tauri::Runtime> QueueManager<R> {
pub async fn commit_reserved_enqueue(
&self,
task: QueuedTask,
mut task: QueuedTask,
generation: u64,
) -> Result<(), String> {
let id = task.id.clone();
@@ -373,6 +373,7 @@ impl<R: tauri::Runtime> QueueManager<R> {
{
return Err("Download enqueue was superseded by a newer user action".to_string());
}
task.lifecycle_generation = generation;
self.pending.lock().await.push_back(task);
self.emit_state(id, DownloadStatus::Queued);
self.notify.notify_one();
@@ -685,6 +686,20 @@ impl<R: tauri::Runtime> QueueManager<R> {
async fn dispatch_one(self: Arc<Self>, permit: OwnedSemaphorePermit, task: QueuedTask) {
let id = task.id.clone();
let lifecycle_generation = task.lifecycle_generation;
// Serialize activation with pause/remove/detach. The dispatcher can
// pop a task just as a control command arrives; registering the
// permit and active kind under the same guard prevents that command
// from observing a half-started lifecycle.
let control_guard = self.acquire_aria2_control(&id).await;
if !self
.is_registered_generation(&id, lifecycle_generation)
.await
{
drop(control_guard);
drop(permit);
return;
}
// Park the permit BEFORE spawning. Uniform parking:
// aria2's RPC returns instantly, so the permit must outlive the
// dispatch_one call. Media runners release on exit.
@@ -694,21 +709,30 @@ impl<R: tauri::Runtime> QueueManager<R> {
.lock()
.await
.insert(id.clone(), task.kind.clone());
let aria2_lifecycle_epoch = if matches!(&task.kind, TaskKind::Aria2) {
// Every backend aria2 dispatch starts a new control lifecycle.
// This invalidates retry workers left behind by a previous
// failed or cancelled lifecycle before retry cancellation is
// made reusable for the new task.
let lifecycle_epoch = self.next_aria2_control_epoch(&id).await;
self.aria2_retry_cancelled.lock().await.remove(&id);
self.aria2_payloads
.lock()
.await
.insert(id.clone(), task.payload.clone());
self.aria2_retry_strikes.lock().await.remove(&id);
Some(lifecycle_epoch)
} else {
None
};
self.emit_state(&id, DownloadStatus::Downloading);
drop(control_guard);
match task.kind {
TaskKind::Aria2 => {
// Every backend aria2 dispatch starts a new control lifecycle.
// This invalidates retry workers left behind by a previous
// failed or cancelled lifecycle before retry cancellation is
// made reusable for the new task.
let lifecycle_epoch = self.next_aria2_control_epoch(&id).await;
self.aria2_retry_cancelled.lock().await.remove(&id);
self.aria2_payloads
.lock()
.await
.insert(id.clone(), task.payload.clone());
self.aria2_retry_strikes.lock().await.remove(&id);
let lifecycle_epoch = aria2_lifecycle_epoch
.expect("aria2 dispatch must initialize a control epoch");
match self.spawner.add_uri(&id, &task.payload).await {
Ok(gid) => {
let control_guard = self.acquire_aria2_control(&id).await;
@@ -793,6 +817,10 @@ impl<R: tauri::Runtime> QueueManager<R> {
lifecycle_generation: u64,
outcome: Result<(), String>,
) {
// Completion and pause/remove/detach must observe one ordered
// lifecycle. Without this guard a terminal media outcome could emit
// Completed, then a racing pause could emit Paused afterward.
let _control_guard = self.acquire_aria2_control(id).await;
if !self.is_registered_generation(id, lifecycle_generation).await {
log::info!(
"media runner [{}]: ignoring stale terminal outcome for lifecycle {}",
@@ -1900,25 +1928,29 @@ impl SidecarSpawner for ProductionSpawner {
.download_coordinator
.register_media(id.to_string(), lifecycle_generation)
.await?;
let outcome = crate::start_media_download_internal(
self.app_handle.clone(),
id,
payload.url.clone(),
payload.destination.clone(),
payload.filename.clone(),
payload.format_selector.clone(),
payload.cookie_source.clone(),
payload.speed_limit.clone(),
payload.username.clone(),
payload.password.clone(),
payload.headers.clone(),
payload.cookies.clone(),
payload.proxy.clone(),
payload.user_agent.clone(),
payload.max_tries,
&mut cancel_rx,
)
.await;
let outcome = if *cancel_rx.borrow() {
Err(crate::queue::MEDIA_RUN_CANCELLED.to_string())
} else {
crate::start_media_download_internal(
self.app_handle.clone(),
id,
payload.url.clone(),
payload.destination.clone(),
payload.filename.clone(),
payload.format_selector.clone(),
payload.cookie_source.clone(),
payload.speed_limit.clone(),
payload.username.clone(),
payload.password.clone(),
payload.headers.clone(),
payload.cookies.clone(),
payload.proxy.clone(),
payload.user_agent.clone(),
payload.max_tries,
&mut cancel_rx,
)
.await
};
if let Ok(path) = outcome.as_ref() {
let _ = crate::download_ownership::set_primary_path(&self.app_handle, id, path);
if let Some(file_name) = path.file_name().and_then(|name| name.to_str()) {