fix(server): transactional sync push; single-use refresh rotation
Two concurrency fixes from the 2026-07-06 review (findings M1, M2): sync push (M1): the load→merge→store cycle ran as three separate DB operations. Two devices pushing concurrently both read the same stored payload, merged independently, and the second store overwrote the first merge — the server visibly regressed until the losing device pushed again. The whole cycle (including the leaderboard update) now runs in one transaction; SQLite serialises the writers. The leaderboard helper's stale docstring (claiming a single conditional UPDATE that was actually two statements) is corrected — the transaction now provides the atomicity it described. refresh rotation (M2): SELECT-then-DELETE let two concurrent refreshes with the same token both pass the liveness check and both mint fresh token pairs. The SELECT is gone; rotation now gates on the DELETE's rows_affected — whoever removes the jti row wins, everyone else gets 401. Sequential reuse was already covered by consumed_refresh_token_is_rejected, which still passes unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
@@ -283,23 +283,20 @@ pub async fn refresh(
|
||||
// Tokens without jti predate rotation — require re-login.
|
||||
let jti = claims.jti.ok_or(AppError::Unauthorized)?;
|
||||
|
||||
// Verify this jti is still live (not yet consumed or from a deleted account).
|
||||
// SQLite TEXT columns are always nullable in sqlx; flatten the double-Option.
|
||||
let exists: Option<String> =
|
||||
sqlx::query_scalar!("SELECT jti FROM refresh_tokens WHERE jti = ?", jti)
|
||||
.fetch_optional(&state.pool)
|
||||
.await?
|
||||
.flatten();
|
||||
|
||||
if exists.is_none() {
|
||||
return Err(AppError::Unauthorized);
|
||||
}
|
||||
|
||||
// Consume the old token before issuing new ones. If the insert below
|
||||
// fails, the user loses this session (must re-login) — safe by design.
|
||||
sqlx::query!("DELETE FROM refresh_tokens WHERE jti = ?", jti)
|
||||
// Consume the old token before issuing new ones, gating on the DELETE
|
||||
// actually removing a row. rows_affected == 0 covers both "jti never
|
||||
// existed / account deleted" and "a concurrent refresh already consumed
|
||||
// it" — the previous SELECT-then-DELETE let two concurrent refreshes
|
||||
// both pass the check and both mint fresh token pairs. The DELETE is
|
||||
// the mutex: whoever removes the row wins; everyone else gets 401.
|
||||
// If the insert below fails, the user loses this session (must
|
||||
// re-login) — safe by design.
|
||||
let deleted = sqlx::query!("DELETE FROM refresh_tokens WHERE jti = ?", jti)
|
||||
.execute(&state.pool)
|
||||
.await?;
|
||||
if deleted.rows_affected() != 1 {
|
||||
return Err(AppError::Unauthorized);
|
||||
}
|
||||
|
||||
let new_access = make_access_token(&claims.sub, &state.jwt_secret)?;
|
||||
let (new_refresh, new_jti) = make_refresh_token(&claims.sub, &state.jwt_secret)?;
|
||||
|
||||
@@ -5,7 +5,6 @@
|
||||
|
||||
use axum::{Json, extract::State};
|
||||
use chrono::Utc;
|
||||
use sqlx::SqlitePool;
|
||||
use uuid::Uuid;
|
||||
|
||||
use solitaire_sync::{
|
||||
@@ -26,13 +25,19 @@ struct SyncRow {
|
||||
|
||||
/// Load the stored `SyncPayload` for `user_id` from the database.
|
||||
/// Returns `None` if this user has not pushed any data yet.
|
||||
async fn load_sync_row(pool: &SqlitePool, user_id: &str) -> Result<Option<SyncRow>, AppError> {
|
||||
///
|
||||
/// Executor-generic so `push` can run it inside its transaction while
|
||||
/// `pull` keeps passing the pool directly.
|
||||
async fn load_sync_row(
|
||||
exec: impl sqlx::Executor<'_, Database = sqlx::Sqlite>,
|
||||
user_id: &str,
|
||||
) -> Result<Option<SyncRow>, AppError> {
|
||||
let row = sqlx::query_as!(
|
||||
SyncRow,
|
||||
"SELECT stats_json, achievements_json, progress_json FROM sync_state WHERE user_id = ?",
|
||||
user_id
|
||||
)
|
||||
.fetch_optional(pool)
|
||||
.fetch_optional(exec)
|
||||
.await?;
|
||||
Ok(row)
|
||||
}
|
||||
@@ -69,7 +74,7 @@ fn row_to_payload(row: &SyncRow, user_id: &str) -> Result<SyncPayload, AppError>
|
||||
|
||||
/// Persist a `SyncPayload` for `user_id` using an upsert.
|
||||
async fn store_payload(
|
||||
pool: &SqlitePool,
|
||||
exec: impl sqlx::Executor<'_, Database = sqlx::Sqlite>,
|
||||
user_id: &str,
|
||||
payload: &SyncPayload,
|
||||
) -> Result<(), AppError> {
|
||||
@@ -92,7 +97,7 @@ async fn store_payload(
|
||||
progress_json,
|
||||
now
|
||||
)
|
||||
.execute(pool)
|
||||
.execute(exec)
|
||||
.await?;
|
||||
|
||||
Ok(())
|
||||
@@ -159,12 +164,20 @@ pub async fn push(
|
||||
return Err(AppError::BadRequest("user_id mismatch".into()));
|
||||
}
|
||||
|
||||
let server_payload = match load_sync_row(&state.pool, &user.user_id).await? {
|
||||
// The whole read-merge-write cycle runs in ONE transaction. Without it,
|
||||
// two devices pushing concurrently both read the same stored payload,
|
||||
// merge independently, and the second store overwrites the first merge —
|
||||
// the server visibly regresses until the losing device pushes again.
|
||||
// SQLite serialises writers, so the second transaction simply waits.
|
||||
let mut tx = state.pool.begin().await?;
|
||||
|
||||
let server_payload = match load_sync_row(&mut *tx, &user.user_id).await? {
|
||||
Some(row) => row_to_payload(&row, &user.user_id)?,
|
||||
None => {
|
||||
// First push — nothing to merge against; store directly.
|
||||
store_payload(&state.pool, &user.user_id, &client_payload).await?;
|
||||
update_leaderboard_if_opted_in(&state.pool, &user.user_id, &client_payload).await?;
|
||||
store_payload(&mut *tx, &user.user_id, &client_payload).await?;
|
||||
update_leaderboard_if_opted_in(&mut tx, &user.user_id, &client_payload).await?;
|
||||
tx.commit().await?;
|
||||
return Ok(Json(SyncResponse {
|
||||
merged: client_payload,
|
||||
server_time: Utc::now(),
|
||||
@@ -175,8 +188,9 @@ pub async fn push(
|
||||
|
||||
let (merged, conflicts) = merge(&client_payload, &server_payload);
|
||||
|
||||
store_payload(&state.pool, &user.user_id, &merged).await?;
|
||||
update_leaderboard_if_opted_in(&state.pool, &user.user_id, &merged).await?;
|
||||
store_payload(&mut *tx, &user.user_id, &merged).await?;
|
||||
update_leaderboard_if_opted_in(&mut tx, &user.user_id, &merged).await?;
|
||||
tx.commit().await?;
|
||||
|
||||
Ok(Json(SyncResponse {
|
||||
merged,
|
||||
@@ -188,16 +202,18 @@ pub async fn push(
|
||||
/// If the user is opted in to the leaderboard, update their row with the
|
||||
/// better of the stored and incoming `best_single_score` / `fastest_win_seconds`.
|
||||
///
|
||||
/// The opt-in check and the update are performed atomically in a single
|
||||
/// conditional UPDATE (WHERE EXISTS subquery) to avoid a TOCTOU race where
|
||||
/// the user opts out between the check and the write.
|
||||
/// Runs on the caller's transaction connection, so the opt-in check and the
|
||||
/// update are atomic with the surrounding push — an opt-out between the check
|
||||
/// and the write can no longer interleave. (An earlier doc comment claimed
|
||||
/// this was a single conditional UPDATE; it has always been two statements —
|
||||
/// the enclosing transaction is what actually provides the atomicity.)
|
||||
async fn update_leaderboard_if_opted_in(
|
||||
pool: &SqlitePool,
|
||||
conn: &mut sqlx::SqliteConnection,
|
||||
user_id: &str,
|
||||
payload: &SyncPayload,
|
||||
) -> Result<(), AppError> {
|
||||
let opted_in = sqlx::query!("SELECT leaderboard_opt_in FROM users WHERE id = ?", user_id)
|
||||
.fetch_optional(pool)
|
||||
.fetch_optional(&mut *conn)
|
||||
.await?
|
||||
.map(|r| r.leaderboard_opt_in)
|
||||
.unwrap_or(0);
|
||||
@@ -231,7 +247,7 @@ async fn update_leaderboard_if_opted_in(
|
||||
now,
|
||||
user_id
|
||||
)
|
||||
.execute(pool)
|
||||
.execute(&mut *conn)
|
||||
.await?;
|
||||
|
||||
Ok(())
|
||||
|
||||
Reference in New Issue
Block a user