fix(files): ending every session and making a token lock the user without holding up rows that refer to it

backend/app/routes/auth.py

@@ -1150,12 +1150,14 @@ def bump_session_epoch(db: Session, user: User) -> None:
     outlive the answer to that theft. Not committed — the caller commits it with its change.
 
     The user's row is locked before the tokens go, the lock `create_token` takes too: a token made at
-    the same moment either commits first and is deleted here, or waits and sees the new epoch."""
+    the same moment either commits first and is deleted here, or waits and sees the new epoch. FOR NO
+    KEY UPDATE (`key_share`): the two still exclude each other, but a row that only refers to the user
+    (a notice, an audit line) is not held up — a token's expiry notice would deadlock with it."""
     if user.id is None:
         user.session_epoch = (user.session_epoch or 0) + 1
         return
     epoch = db.execute(
-        select(User.session_epoch).where(User.id == user.id).with_for_update()
+        select(User.session_epoch).where(User.id == user.id).with_for_update(key_share=True)
     ).scalar_one()
     user.session_epoch = (epoch or 0) + 1
     apitokens.revoke_all(db, user.id)

backend/app/routes/tokens.py

@@ -122,9 +122,10 @@ def create_token(
         raise refused(400, "api_token_name_required")
     # The user's row, locked: two requests at once cannot both slip under the ceiling, and a
     # session ended since `current_user` read the epoch (unlocked) cannot mint a token that outlives
-    # it — `bump_session_epoch` takes this lock before it revokes them all.
+    # it — `bump_session_epoch` takes this lock before it revokes them all. No key update: a row that
+    # only refers to the user (a folder delete's audit line) must not wait on it, or the two deadlock.
     epoch = db.execute(
-        select(User.session_epoch).where(User.id == user.id).with_for_update()
+        select(User.session_epoch).where(User.id == user.id).with_for_update(key_share=True)
     ).scalar_one()
     if (epoch or 0) != (user.session_epoch or 0):
         raise refused(401, "not_authenticated")

backend/tests/test_api_tokens.py

@@ -12,9 +12,9 @@ from datetime import datetime, timedelta, timezone
 import pytest
 from fastapi import Depends, Request
 from fastapi.testclient import TestClient
-from sqlalchemy import select
+from sqlalchemy import select, update
 
-from app import apitokens
+from app import apitokens, audit
 from app.apitokens import files_caller
 from app.auth import current_user
 from app.downloads import folders, service
@@ -24,6 +24,7 @@ from app.downloads.status import JobStatus
 from app.config import settings
 from app.db import SessionLocal, get_db
 from app.main import app
+from app.notifications import create_notification
 from app.models import (
     ApiToken,
     ApiTokenGrant,
@@ -427,6 +428,74 @@ def test_a_token_asked_for_as_every_session_ends_is_refused(db, make_user, clien
     assert db.execute(select(ApiToken.id)).scalars().all() == []
 
 
+def _meanwhile(work):
+    """Run `work(other_session)` in another thread and session, committed; (thread, outcome)."""
+    import threading
+
+    outcome: list[object] = []
+
+    def run():
+        other = SessionLocal()
+        try:
+            outcome.append(work(other))
+            other.commit()
+        except Exception as e:  # noqa: BLE001 — the assertion below names what went wrong
+            outcome.append(e)
+        finally:
+            other.close()
+
+    t = threading.Thread(target=run)
+    t.start()
+    return t, outcome
+
+
+def test_ending_every_session_never_deadlocks_with_a_tokens_expiry_notice(db, make_user):
+    """The notice's sweep holds the token's row and then writes a row naming the user; the sessions'
+    end holds the user and then deletes the token. A lock on the user that a mere reference to it
+    waits on would make that a deadlock — one of the two failing."""
+    alice = make_user()
+    raw = apitokens.mint()
+    soon = datetime.now(timezone.utc) + timedelta(days=1)
+    tok = ApiToken(
+        user_id=alice.id, name="t", prefix=raw[:12], token_hash=hash_token(raw), expires_at=soon
+    )
+    db.add(tok)
+    db.commit()
+    # `warn_expiring`, its two halves held apart: the token row taken…
+    db.execute(update(ApiToken).where(ApiToken.id == tok.id).values(expiry_warned=True))
+    t, outcome = _meanwhile(lambda other: bump_session_epoch(other, other.get(User, alice.id)))
+    t.join(timeout=1)
+    assert t.is_alive()  # the sessions' end waits for the token row
+    # …and then its notice, naming the user.
+    create_notification(db, alice.id, "api_token_expiring", title="t", commit=False)
+    db.commit()
+    t.join(timeout=10)
+    assert outcome == [None]
+    db.expire_all()
+    assert db.execute(select(ApiToken.id)).scalars().all() == []
+
+
+def test_a_token_granted_a_folder_being_deleted_never_deadlocks(db, make_user):
+    """The delete holds the folder and then writes its audit line naming the user; the token holds
+    the user and then waits for the folder."""
+    alice = make_user()
+    fid = _own(db, alice.id, "Mine").id
+    folders.delete_own(db, db.get(FileFolder, fid))  # the folder held, as the delete route does
+    body = tokens_routes.TokenIn(
+        name="x", grants=[tokens_routes.GrantIn(folder_id=fid, rights=["read"])]
+    )
+    t, outcome = _meanwhile(
+        lambda other: tokens_routes.create_token(body, other.get(User, alice.id), other)
+    )
+    t.join(timeout=1)
+    assert t.is_alive()  # the token waits for the folder
+    audit.record(db, alice.id, "file_folder_delete", target_type="file_folder", target_id=fid)
+    db.commit()
+    t.join(timeout=10)
+    [refusal] = outcome
+    assert getattr(refusal, "status_code", None) == 409, refusal  # the folder went first
+
+
 def test_a_token_made_as_every_session_ends_goes_with_them(db, make_user):
     """The other order: the token is on its way in when the sessions end."""
     import threading