GHSA-xj3h-wwxq-gfcj: Race Condition

Summary

Cloudreve v4 splits the storage-quota check (reading the user's used bytes and comparing them to MaxStorage) and the charge (incrementing users.storage) into two non-atomic steps in the PrepareUpload code path. This creates a Time-of-Check to Time-of-Use (TOCTOU) race condition. Any authenticated user — including an unprivileged account in the default User group — can concurrently issue several upload-session requests that all read the same stale used snapshot, each pass the check, and then each contribute their declared size to users.storage. The end result is that the total approved capacity exceeds the group's MaxStorage many times over.

The same primitive is trivially amplifiable into a storage-based denial of service. During PrepareUpload, Cloudreve reserves the declared size against users.storage before any bytes are written, so an attacker can push the reserved amount far beyond the host's physical disk (tested: a 1 GiB-quota account reserved 17 GiB in a single 20-way burst), and can then materialise the reservation by completing chunked uploads to actually write the excess bytes to disk. Amplification to the host's free space fills the disk and denies uploads for every user of the instance.

Exploitation requires only a valid session with Files.Write permission. No administrator configuration, no non-default storage policy, and no elevated privileges are needed. The default deployment (local storage policy, default User group) is affected.

Technical details

PrepareUpload in pkg/filemanager/fs/dbfs/upload.go splits quota enforcement across two stages:

Stage A — the check (snapshot compare, no lock) — pkg/filemanager/fs/dbfs/validator.go:

go func (f DBFS) validateUserCapacity(ctx context.Context, size int64, u ent.User) error { capacity, err := f.Capacity(ctx, u) // reads "used" if err != nil { return ... } return f.validateUserCapacityRaw(ctx, size, capacity) }

func (f DBFS) validateUserCapacityRaw(ctx context.Context, size int64, capacity fs.Capacity) error { if capacity.Used + size > capacity.Total { // snapshot compare only; no lock, no reservation return fs.ErrInsufficientCapacity } return nil }

capacity.Used comes from the in-memory ent.User that was hydrated once at the beginning of the request — pkg/filemanager/fs/dbfs/dbfs.go:

go func (f DBFS) Capacity(ctx context.Context, u ent.User) (fs.Capacity, error) { ... res.Used = f.user.Storage // captured at request start res.Total = requesterGroup.MaxStorage return res, nil }

No SELECT is issued at check time, no row lock is taken on the user row, and pending upload sessions are not counted.

Stage B — the charge (single unconditional UPDATE, outside the tx) — pkg/filemanager/fs/dbfs/upload.go → inventory/tx.go → inventory/user.go:

go // PrepareUpload: check (A) ... then, many statements later ... if err := f.validateUserCapacity(ctx, req.Props.Size, ancestor.Owner()); err != nil { return nil, err } ... if err := inventory.CommitWithStorageDiff(ctx, dbTx, f.l, f.userClient); err != nil { ... } // charge (B)

// inventory/user.go c.client.User.Update().Where(user.ID(uid)).AddStorage(diff).Exec(ctx) // SQL: storage = storage + diff

Between A and B the code performs storage-policy load-balancing, save-path generation, encryption-metadata generation, transaction start, placeholder-file/entity creation, and metadata upsert. This leaves a wide race window. N concurrent PrepareUpload requests each read the same stale Used snapshot at A, each pass their independent quota check, and then each add their own size to users.storage at B. The committed total is up to N × size above MaxStorage.

Three defensive controls that would each independently close the race are missing:

1. The check and the charge are not enclosed in the same transaction with a SELECT ... FOR UPDATE on the user row. 2. The charge is a plain storage = storage + :size UPDATE, not an atomic conditional update of the form UPDATE users SET storage = storage + :size WHERE id = :uid AND storage + :size <= :maxstorage. 3. Used is computed only from committed entities. Concurrent pending upload sessions (which have already been reserved by the accounting model) are not counted, so races among sessions that haven't yet completed are invisible to each other.