Regressions found by an adversarial review of this branch's own diff:
- RoomCycleManager: stop holding the currentBots/currentPets monitor across the
whole bot/pet tick — snapshot under the lock then cycle off-lock. The previous
fix blocked place/pickup and room dispose for the full tick and inverted lock
order vs roomUnitLock->currentBots (latent deadlock for any future cycle code
touching roomUnitLock).
- HabboInfo: complete the currencyLock invariant — getCurrencies() now returns a
snapshot under the lock (UserInfoCommand iterated the live Trove map off-lock,
the exact rehash corruption the lock guards); canBuy() uses the lock-guarded
getCredits()/getCurrencyAmount(); run() reads credits under the lock for save.
- RoomSpecialTypes: synchronize the by-id pet getters (getNest/getPetDrink/
getPetFood/getPetToy/getPetTree) to match their now-synchronized mutators.
- AuthHttpUtil.isTrustedProxy: exact-match trusted IPs; only treat an entry as a
range when it ends with "."/":" so "10.0.0.1" can't also trust "10.0.0.12".
Pre-existing logic bugs found by the deep subsystem analysis:
- RoomUsersComposer: the bulk (room-entry) branch wrote the guild id twice;
the second field must be the 1/-1 group-membership flag (matches the single
branch) — every user showed a wrong group indicator on room entry.
- BotManager.pickUpBot: room owners (and ACC_PLACEFURNI) couldn't pick up bots
placed in their own room — added the room-owner clause that placeBot has.
- PetPickupEvent: compared user id to pet.getId() instead of pet.getUserId(), so
a pet owner who isn't the room owner couldn't pick up their own pet.
- RoomRightsManager.refreshRightsForHabbo: in guild rooms, explicit room_rights
were stripped (overwritten by guild level NONE); now takes the stronger of
explicit rights and guild level.
- RoomRequestBannedUsersEvent: `!hasRights || !ACC_ANYROOMOWNER` required BOTH,
denying legitimate owners the banned-users list — corrected to `&&`.
- InteractionPetBreedingNest.breed: a crafted packet on a not-full nest deleted
the nest furni then NPE'd (furni loss); guard petOne/petTwo/room before the
destructive delete; ConfirmPetBreedingEvent null-checks the room.
- WiredEffectTeleport/UserFurniBase: appended item id instead of sprite id in the
incompatible-triggers list (cosmetic wired-dialog mismatch) — matched the ~10
other effects' getBaseItem().getSpriteId() convention.
Root cause from the CPU audit: every incoming packet handler ran on the Netty
I/O event loop (MULTI_THREADED_PACKET_HANDLING is false by default), so any
blocking handler — login DB + loadHabbo, friends/polls/catalog/guild-forum
JDBC (~48 handlers), synchronous A* per walk — stalled socket I/O for every
other client sharing that I/O thread.
- WebSocketChannelInitializer: register GameMessageHandler on a dedicated
DefaultEventExecutorGroup (max(16, 2x cores), daemon). Netty pins each channel
to one executor in the group, so a client's packets stay strictly ordered (no
new intra-client races) while blocking work moves off the I/O loop. The
cross-client concurrency degree matches the already-multi-threaded I/O group,
and this is strictly safer than the existing (order-losing) shared-pool
MULTI_THREADED_PACKET_HANDLING mode the codebase already supported.
- GameMessageHandler: always run the handler inline (now on the group thread);
drop the shared-pool branch (which would break per-channel ordering and also
removes the rejectable-pool ByteBuf-drop path).
- PathfinderImpl: default the A* execution-time guard ON (25ms) so a pathological
search returns an empty path instead of running unbounded on its thread.
Note: this changes the server's packet-threading model — verified to compile,
unit-test, and assemble the shaded jar, but should be load-tested before prod.
Group size is currently derived from CPU count; can be made a config key if
tuning is needed.
roomUnit.getMoveBlockingTask().get() blocked the Netty event loop with no
timeout; a stuck/delayed move-blocking task would park the worker thread (and
every client on it) indefinitely. Wait at most 2s, then proceed with the walk.
The /api/auth/* handlers ran inline on the Netty worker event loop, so their
blocking work — BCrypt (cost 12 ~tens of ms), JDBC, the Turnstile HTTPS
round-trip and SMTP — stalled every other client multiplexed on that thread; a
burst of logins/registers could freeze game traffic.
Dispatch each auth request to a dedicated bounded pool (4..16 daemon threads,
bounded queue, 503 on saturation) instead. It is deliberately SEPARATE from the
shared game ThreadPooling so auth load can't starve room cycles either. Netty
writes are thread-safe, so the endpoints' sendJson calls work unchanged from the
worker; the FullHttpRequest is released when the task finishes.
Caveat: this allows concurrent handling of pipelined requests on a single
keep-alive connection (out-of-order responses) — not a concern for the Nitro
client which is strictly request/response, but worth load-testing before prod.
Security (S3):
- AuthHttpUtil/WebSocketHttpHandler: only honour the configured ws.ip.header
forwarded-IP header when the DIRECT peer is a trusted reverse proxy, instead
of trusting it unconditionally. Loopback is always trusted; extra proxies can
be allow-listed (exact IP or string prefix, comma-separated) via the new
`ws.ip.header.trusted` config key — default-deny so the header can't be
spoofed from the open internet to evade per-IP rate limiting and IP bans.
Also take only the first comma token when setting the game-session WS_IP.
Leak cleanup (C4):
- WiredVariableReferenceSupport.invalidateRoom(): drop a room's shared
wired-variable assignment caches; called from Room.dispose so the static
USER/ROOM_ASSIGNMENT_CACHE maps don't retain entries for the JVM lifetime.
- SessionResumeManager.parkHabbo: if the scheduler refuses the grace-expiry
task (future == null), disconnect immediately instead of parking an
un-reapable GhostSession that would pin the Habbo + room refs forever.
Note: ws.ip.header.trusted defaults to loopback-only; deployments whose proxy
is on another host must add its IP/prefix to that key or client IPs will
collapse to the proxy address.
Continuation of the concurrency hardening from the audit:
- InteractionWired/WiredHandler (E4): add an atomic per-box processing guard so
one trigger box is handled by a single thread at a time, making the cooldown
check-and-set effectively atomic; mark `cooldown` volatile. Prevents a packet
thread and the room cycle thread from double-firing the same wired stack
(double teleport/reward).
- RoomUnit (C1): the walk path is now a volatile ConcurrentLinkedDeque instead of
a plain LinkedList, so the room cycle popping steps can't corrupt it while a
walk packet rebuilds it via findPath/setPath.
- RoomItemManager (C2): iterate roomItems under its own monitor in getFloorItems/
getWallItems/getPostItNotes/getUserUniqueFurniCount/getItemsAt, matching the
existing put/remove sync sites — stops place/pickup from corrupting the
traversal into a silently-incomplete item set.
- pom.xml (S4): bump netty-all 4.1.115 -> 4.1.118.Final (CVE-2025-24970 SslHandler
pre-auth DoS, CVE-2025-25193).
From the full-codebase audit. Economy/security (Batch A):
- CatalogBuyItemEvent: clamp client `count` to 1..100 — the club-offer branch
accumulated cost in plain ints, so a huge count overflowed to a negative
total, bypassed the affordability checks and CREDITED the buyer (free
currency/subscription exploit).
- HousekeepingGiveCredits/GiveCurrency: bound `amount` to +/-1e9 to stop
overflow/negative-balance grants via the privileged path.
- RoomTrade: synchronize accept/confirm/offer/remove and add a `completed`
re-entry guard so two simultaneous confirms can't run tradeItems() twice
(item/credit duplication).
- HabboInfo: serialize credits + currencies read-modify-write and the
saveCurrencies snapshot on a dedicated lock (never held across DB I/O) —
fixes lost updates and Trove rehash-during-iteration corruption between the
credit-roller thread and purchase/trade handlers.
- AchievementManager/HabboStats: atomic incrementProgress() so concurrent
progress sources don't lose updates.
Resource/stability (Batch B):
- GameMessageRateLimit: release the wrapped ByteBuf on every drop path
(ClientMessage isn't ReferenceCounted, so the decoder's auto-release is a
no-op) — fixes a refcount leak on pre-auth/rate-limited packets.
- AuthRateLimiter: opportunistically evict window-expired STATE/PROBE_STATE
entries — previously grew unbounded, one entry per unique client IP.
- ForumThread/ForumThreadComment: close getGeneratedKeys() ResultSets via
try-with-resources, and create the first comment after the thread's
connection is released (was holding two pooled connections at once).
- DatabasePool: add socketTimeout/connectTimeout/tcpKeepAlive so a stalled
MariaDB can't pin a pooled connection (and its thread) indefinitely.
Concurrency visibility (Batch C, partial):
- Room: mark allowBotsWalk/allowPets/allowPetsEat volatile (read every cycle,
written from settings handlers on another thread).
- InteractionGift/OpenRecycleBoxEvent: add an atomic open-once guard so two
near-simultaneous OpenRecycleBox packets can't both schedule the async,
delayed OpenGift before the wrapper is removed (redundant double-process).
- ClientMessage.readString: treat the length prefix as unsigned (mask 0xFFFF)
and clamp to the buffered bytes, so a bogus/oversized length no longer
throws mid-read and desyncs the remaining fields of the packet.
Security
- HousekeepingSetUserRankEvent: add rank-ceiling guard (reject granting a
rank above the operator's own and modifying a higher-ranked target),
mirroring GiveRankCommand — closes a privilege-escalation path.
Trade integrity
- RoomTrade.clearAccepted now also resets confirmed; a stale confirmed=true
let a user strip their side and still complete once the partner re-confirms.
Concurrency
- RoomCycleManager: iterate the synchronized bot/pet maps under their own
monitor (lock order stays one-directional vs addBot/addPet — no deadlock).
- RoomSpecialTypes: synchronize nest/petDrink/petFood/petToy/petTree writers
on the same monitor their getters already use.
- HabboStats: synchronize achievement-progress map accessors.
- RebugKickBallAction: drop redundant direct mutation of the shared tile-cache
sets (updateTile invalidates them right after) — removes a data race.
Robustness
- Wired legacy parsers (HabboCount, NotHabboCount, MatchStatePosition,
MoveRotateFurni): guard length/format so one malformed row no longer aborts
the whole room's wired load.
- RoomLayout: fill malformed/short heightmap rows with INVALID tiles instead
of leaving nulls, and bounds-check door coordinates.
- FurnidataWatcher: defer (instead of drop) a throttled delta so furni-name
changes are never lost between broadcasts.
- GuildManager.getGuildMembers: fix LIMIT row-count (page size 14, not
offset+14) so member pages no longer overlap from page 1 on.
FriendsComposer only serialized a buddy's look when online, sending an
empty string for offline friends. The look is already loaded from the DB
for every friend in Messenger.loadFriends (SELECT users.look), so the
gate just discarded valid data: offline friends rendered with the
anonymous/standard avatar in the friend list and messenger, while their
profile (fetched separately) showed the real figure.
Always serialize row.getLook(). StaffChatBuddy keeps a non-null look
("ADM") so there is no NPE risk, and UpdateFriendComposer already sent
the look unconditionally, so this only aligns the initial friend list.
FurniEditorImportTextEvent (incoming 10049, ACC_CATALOGFURNI): resolves
the classname, fetches the admin-configured furnidata URL via HttpClient
with a TTL cache (furni.editor.import.url / .cache.ms, default habbo.it),
finds name/description by classname and returns them via
FurniEditorImportTextResultComposer (outgoing 10049). URL is DB-configured
only (no client-supplied URL -> no SSRF); serves stale cache on failure.
Read sortField/sortDir from the search packet and ORDER BY a whitelisted
items_base column (id/sprite_id/item_name/public_name/type/interaction_type)
with a stable id tie-break, so sorting orders the whole result set instead
of just the page the client received. Column names come from a fixed
whitelist (never raw input) so the dynamic ORDER BY stays injection-safe.
On a successful furnidata name update (10046), after the JSON write +
10047 broadcast, also UPDATE items_base.public_name to the new
(sanitized) name and refresh the in-memory Item cache via loadItems()
so Item.getFullName() stays consistent without a restart. Guarded by
name != null (description-only edits never blank the column), runs only
on the success path, outside FurnidataLock, with a parameterized
statement.
simoleo89