relayUDP started a goroutine for every domain-addressed datagram and
each lookup opened its own netstack UDP endpoint, so the lookups in
flight were bounded only by the client's send rate: one proxy user could
grow the panel by gigabytes and exhaust the outbound stack's ports,
stalling every user of that outbound.
One association may now have 32 datagrams waiting on DNS and drops the
rest, a cached name is delivered from the reader loop without a lookup,
and resolveTunnelVia takes one of 1024 process-wide slots before it
opens an endpoint, which also covers the TCP CONNECT path.
stackTun.WriteNotify runs on whichever goroutine made the stack emit a
packet; for echo replies and RSTs that is amneziawg-go's per-peer
receiver. It waited without bound for room in the outbound queue, whose
reader waits on that same receiver while the device rebinds, so a peer
that flooded the stack while nothing drained the queue deadlocked the
next reconfiguration of the interface.
The handoff now tries once, then waits at most 5 ms and drops the
packet, as channel.Endpoint already does when its own queue is full.
The embedded interface's TCP forwarder admits 65535 connections in
flight for all peers together, and a forwarded request holds its slot
until the handshake ends or gVisor gives up on it. One client sending
bare SYNs from fresh source ports therefore took every slot and left
the other clients' connections unanswered. The UDP relay likewise
opened a SOCKS5 association and a host UDP socket for every fresh
tunnel source port with no per-peer bound.
The forwarder now names the peer on the SYN, refuses a source no peer
owns and caps one peer at 4096 connections; the UDP relay caps one peer
at 1024 sessions and drops the datagrams that would need more.
joinHostPort wrote an endpoint address into the authority of every URL
share link unencoded, and genVlessLink alone skipped encodeUserinfo for
the client id. Both can come from a node's sync (an externalProxy dest,
an adopted host row, the client id), the finished link is never
re-parsed, and the raw subscription is one link per line: a line break
in either added lines of the node's choosing to the master's
subscription, and an id carrying "@" moved the client's VLESS links on
other inbounds to another server.
Whitespace, control characters and URI delimiters in the host are now
percent-encoded, which leaves host names, IP literals and zone ids
unchanged, and the VLESS id is encoded as userinfo like every other
protocol's credential.
The tunnel overlap guard treated every /0 allowedIPs entry as claiming
no address, while the entry was stored with its host bits and the
consumers read them back: the AmneziaWG port forwarder resolves a peer's
target from them and the device routes the /0 by longest prefix. An
entry such as 10.8.1.2/0 therefore passed next to another client's
10.8.1.2/32, and forwarded ports reached that client's device.
A written entry now counts with its whole range, so a /0 collides with
every held address of its family. On the stored side only a literal
0.0.0.0/0 or ::/0 default route keeps nobody out, and a /0 with host
bits claims that host. Editing a legacy peer that echoes its stored
default route still works.
The groups page lists clients.group_name as stored, while every group
action trims the name it is given. A name with surrounding white space,
saved by a node snapshot or by a client edit, was listed as a separate
group, and deleting, renaming or resetting that row acted on the real
group of the trimmed name instead.
Group names are trimmed wherever a client record is built or updated,
and a one-time seeder folds stored names onto their trimmed form.
DelDepleted turned the emails of depleted usage rows into client
identities through BulkDelete, which trims each email for request
bodies. A depleted row keyed by a padded email, which a node's stats
could store and which belongs to no client, therefore resolved to the
healthy client it resembled and deleted it. Rows whose key carries
surrounding white space name no client and are skipped.
A node snapshot was merged into the master's client records by email, and
the only ownership check filtered one derived copy of it: the node's
settings were still stored and re-synced later, its stats still merged,
and a record linked to the master as well as the node took every value
the node reported, including credentials rendered into the master's own
Xray config.
Each snapshot now passes one gate before anything is written. A record
also linked to the master or another node keeps the master's values
apart from per-link fields and the lifecycle the merge already settles;
a record linked only elsewhere, a detached record, a padded or
case-variant email and a new client holding another client's subId are
dropped from the stored settings as well. A soft-orphaned record can be
taken back only by the node whose merge orphaned it, which the new
clients.sync_orphaned_by column records. A node tunnel no longer owns
the peer keys of a client a master-local tunnel also serves.
A node-sync token, or a master enrolled with the mTLS certificate, may
push inbounds, and the node stored them as already validated. An
inbound is effectively config for a core running as root, so a push
could serve the host's files (hysteria file masquerade), relay to its
own services or the core API (masquerade proxy, fallback dest, REALITY
target), write files (masterKeyLog), use local or templated XDRIVE
storage, fetch URLs (ECH, realm masks), or run tun/http/mixed inbounds.
A push now refuses those fields unless the node already stores the same
value, which keeps rows configured by the node's own admin syncing:
relay targets must be public or another of this node's inbounds, and
only node-eligible protocols are accepted. Every caller is refused env:
references, which the core expands from the panel's environment, and
the inbound listen address is JSON-encoded into the config instead of
quoted by hand.
Every UDP destination a proxy user reached through the AmneziaWG egress
opened a session with a goroutine, a port and a buffer, per association,
and a Mux frame opens a new association, so one user could hold sessions
without limit. All egress UDP sessions now share one budget of 4096;
when it is spent the longest-idle session is closed to make room.
A forwarded port accepts unauthenticated traffic, and every TCP
connection and every new UDP source got its own relay with a goroutine
and a buffer, with no limit, so anyone reaching the port could exhaust
the panel's memory. Live TCP relays and UDP sessions are now capped at
2048 each across all forwards; past that a connection is closed and a
datagram from a new source is dropped, while open flows keep going.
Removing or re-keying a client reconfigured the device in place but left
its TCP and UDP flows standing, and withdrawing a port forward stopped
only new connections. The next client given the same tunnel address
inherited those flows: it could read what outside hosts sent to the
former client and write back into those connections.
After an in-place reconfigure, every stack endpoint and UDP relay
session pinned to an address whose peer was removed or changed is
aborted, and closing a port forward cuts the connections it is relaying.
A TCP port forward's accept loop returned on any accept error, leaving
the socket bound and the forward registered as open, so nothing served
the port again until a panel restart; briefly pushing the panel to its
open-file limit was enough. Forward and egress accept loops now log,
wait 100 ms (or until closed) and keep accepting.
The TCP forwarder handler read the tunnel client's address from the
accepted connection's RemoteAddr, which asks the live endpoint and is
nil once the peer's RST has been processed. A client sending SYN, ACK
and RST through its own tunnel could hit that window and crash the panel
on a goroutine with no recover. The forwarder now hands the handler the
source from the request's endpoint ID, as it already did for the
destination.
Anyone reaching a TUIC port could start QUIC handshakes without ever
reading a reply, each holding state and goroutines, and an unauthenticated
connection got a handler goroutine per stream it opened. Both grew until
the panel ran out of memory.
The listener now counts connections from their first packet until they
authenticate or close: past 64 they must answer a stateless Retry, and
past 1024 in total or 32 per verified source (IPv4 address or IPv6 /64)
they are refused. Before authenticating, a connection may start at most
16 handlers per stream kind; further streams are reset.
The UDP association cap applied per QUIC connection, and nothing limited
how many connections one credential could authenticate, so a single
client could hold enough UDP relays (each a host port) to exhaust the
host's ephemeral range. A client now gets at most 32 connections and 512
open UDP relays across all of them; past that a connection is refused at
authentication and a relay is not opened.
Dissociating a UDP association closed its SOCKS session, but the
response relay could be blocked opening a stream on credit the client
never grants, and only the connection's end released it. A client could
leave one such goroutine per Packet/Dissociate round, past the cap on
live associations. Each relay now runs under a context the association's
removal cancels, whichever path removes it.
max_idle_time and authentication_timeout are second counts multiplied
into a Duration with no bound, so a stored value past the int64 range
wrapped negative and the QUIC listener panicked on the first
ClientHello, taking the panel down at every start until the row was
fixed. Saves now refuse more than a day, and a stored row past it fails
its own listener start instead.
Every handshake initiation builds Jc junk packets of Jmin..Jmax bytes
and the I1-I5 signature packets, and the limits allowed the full uint32
range and lengths of any size, so one connection made the embedded
device allocate without bound. The MTU reaches the device as an int32
and a wider value made it panic. Both ended the whole panel at the next
connection, after every restart.
Jc is now at most 128, Jmin/Jmax and each signature packet at most one
65535-byte datagram, and the MTU at most 65535. Saves refuse more, the
device setup refuses rows stored before, and the stored MTU is capped
where it is read. The form schemas carry the same limits.
Per-client counters are stored up to TrafficMax, which a node can reach
by reporting huge usage, and the group totals, group baselines and the
per-client reset lookup added two such counters inside SQL, overflowing
it: every group listing and reset failed, and no client on that reset
cycle was reset. The counters are now read separately and added in Go,
saturating at TrafficMax.
allocateWireguardAddress walked the scope one address at a time and
tested each against every wide claim, so a claim covering an IPv6 /64
(a node can report such allowedIPs) turned the next address-less client
add into a 2^64-step loop that also held every later client change on
that inbound. The scan now jumps past a wide claim in one step, so it is
bounded by the number of claims.
An interval renewal turned the client's reset days into milliseconds
with no bound, so a value that wraps int64 (a node reports it unchecked)
sent the catch-up loop stepping through time for seconds per client
while the traffic writer held the database. Intervals above 36500 days
no longer renew, and every client write refuses them.
gregorianToJalali walked the calendar one month at a time from a 2024
reference, so an expiry near the int64 limit (a node reports expiries
unchecked) cost seconds of CPU on every subscription fetch that renders
{{JALALI_EXPIRE_DATE}}. Jalali leap years repeat every 33 years, so
whole 33-year blocks are skipped before the walk.
pushClientTraffics decoded the pushed list whole, and each "{}" became a
full traffic row, so a body under the size cap could cost the panel
close to a gigabyte. The list is now decoded entry by entry and refused
past 262144 entries. The capped decoder the node runtime already used
moves to json_util so both share it.
Deleting a client by an email no record holds fell back to a settings
LIKE '%email%' over every inbound with the raw route parameter, which
costs SQLite the settings length times the pattern length while it holds
the database; a long near-miss email froze the panel for minutes, and
wildcards were not escaped. The pre-filter now uses the first 64 bytes
of the email, escaped; the exact client match after it still decides.
A node's online tree and both client-IP reports were stored as sent:
they could name any client of the master or another node, under any
panel's GUID, dated into the future, and in any number. That marked
master clients online, fed source addresses for clients the node does
not serve into the IP limit and its fail2ban log, and planted rows the
retention sweep never removed.
Each report is now cut to the clients linked to the node's inbounds and
to the node's own key and its sub-nodes' origins (never this panel's or
another node's GUID); timestamps are capped at now, and one client's
list from a node keeps at most its 256 newest entries.
After a sync tick disabled clients, the master pushed every affected
node inbound with its own 4 s timeout, a few at a time, and the job
served no other node until all were done; a hanging node could stall it
for as long as it had inbounds to report. Pushes are now grouped by node
and each node shares one budget, still a few inbounds at a time; what
does not make it stays for the reconcile, the node being dirty already.
A per-client reset the node never accepted stopped the delivery loop:
resets queued before it were re-sent on every sync, each time wiping the
usage made since, and those after it were never sent. Each reset now
settles by itself, a refused one moves to the back of the queue, and the
loop goes on until the context is spent.
A node also delivered its master's reset to its own sub-nodes inline, so
one silent sub-node held the master's request until it timed out and was
replayed; the inline attempt now waits at most a second and the rest
finishes, or stays queued, in the background.
The node traffic sync restarted the master's own Xray with force after
any tick that disabled clients, so a node reporting a fresh over-quota
client every few seconds dropped every local user's connections each
time, though nodes serve none of them. Use the non-forced restart the
local traffic job uses: an unchanged local config keeps the core, and a
change still restarts when restartXrayOnClientDisable asks for it.
Reaching RestartXray needs a running core, so no unit test covers this.
Disabling depleted clients rewrote each affected inbound's settings and
aborted the whole lifecycle tick when one blob did not parse, so a node
storing a malformed blob on the master switched expiry and quota checks
off for every inbound. Unparseable node settings are now treated like a
snapshot that drops the hub's clients (the hub keeps its copy), and an
unparseable stored blob is left as it is while the client's rows are
still disabled; the rewrite also works one level deep like the others.
Dedupe, tombstone stripping and the lifecycle lift decoded a node's
whole settings blob into generic maps and wrote it back indented, so a
deeply nested value anywhere in it cost the square of its depth on every
merge (10 KB became 50 MB). They now split the blob one level deep, read
only email, expiryTime and enable of each client entry, keep everything
else as raw text and write the result compactly.
Three statement shapes bound one SQL variable per row and failed past
SQLite's limit, which a node can reach by itself with the clients and
inbounds it reports: the depleted-client lookup and disable in
disableInvalidClients (stopping every expiry and quota check), the
whole-slice Saves of renewed inbounds and traffics (stopping renewals),
and Preload("ClientStats") over all inbounds (stopping config builds).
The IN lists are chunked, the Saves go 100 rows at a time, and client
stats are attached in chunks instead of preloaded.
The node Remote keeps a fingerprint, a remote id and possibly an alias
per inbound tag. When the master swept an inbound the node stopped
reporting, those entries stayed, so a node rotating tags grew the maps
for as long as the master ran. The sweep now hands every removed tag to
Remote.ForgetInbound, which drops all of it in both tag forms, and a
master-driven delete uses it too. The sweep's call is wiring;
TestRemoteForgetInboundDropsEveryTrace covers what it drops.
Two shapes of a node's inbound list cost the master without bound:
- an inbound listed many times was merged once per listing, every pass
walking its whole client set again; the first listing now wins and the
repeats are dropped;
- disabled inbounds were never swept, an exemption meant for inbounds the
master created but has not delivered yet, so a node could report any
number of them and the master kept them all; the exemption now covers
only rows without an origin, i.e. never reported by the node.
A peer's panelGuid is copied into nodes.guid and into every inbound row
it reports (an indexed column), a forwarded origin GUID likewise, and a
GUID-keyed IP tree or a master's global-traffic push keys one row per
client. One oversized value therefore multiplied into gigabytes stored.
GUIDs over 64 bytes (panels mint 36-character UUIDs) now fail the
heartbeat, fall back to the reporting node's key, or are dropped.
SSRFGuardedDialContext tried each address of a name in turn with a fresh
10 s, so a name resolving to thousands of unanswering addresses held a
dial (and the caller's goroutine) for hours; a node can choose such a
name for itself. The whole call now shares one 10 s budget, divided over
the remaining addresses with net.Dialer's 2 s floor. Multi-address DNS
answers cannot be produced in a unit test here, so no test reaches it.
Node responses are untrusted input on the master, and a few shapes still
cost it without bound under the 64 MiB body cap:
- a success=false message became an error (logged, stored as lastError)
in full; it is now cut to the 8 KiB the non-200 diagnostic already uses;
- each "{}" of a reported list decodes into a whole struct, so lists are
now decoded entry by entry with per-endpoint caps (1024 descendants,
65536 inbounds and host groups, 1M client-IP rows), and the GUID-keyed
trees and maps are capped at 1M values;
- a TLS handshake the node never answers kept a socket and goroutines
alive forever, since the transport keeps dialing for its pool after the
request gives up; node transports now give up after 10 s.
The master's node clients followed redirects to any host, presenting the
same panel-wide credentials (the mTLS client certificate) and replaying
the request there, so one node could have the master push to another.
Node clients and the certificate-pin probe now stop at a redirect off
the original host:port and report the 3xx as the node's answer. The
probe change is wiring; TestNodeClientsDoNotFollowARedirectToAnotherHost
covers the guard.
InitLocalizer filled the shared bundle in place, and the Telegram bot
runs it again while the panel is serving, so a request could read the
bundle's maps while they were being written and the process died on a
concurrent map access. Build the bundle privately and publish it, and
the bot and web localizers, through atomic pointers; the lazy load in
the middleware publishes only if nothing else did first.
The shared zstd decoder grew its output per frame, copying everything
decoded so far for each extra frame, so 30 KB of tiny frames cost about
8 GB of allocation and seconds of CPU per request. The output buffer is
now sized once from the first frame's declared size (Compress writes a
single frame) with decoding capped at that capacity, and callers waiting
for a decoder worker no longer each hold a buffer.
The envelope middleware buffered and decoded any request carrying its
headers, GETs included, which the router-wide body cap skips and which
read-only tokens may call; its raw read had no cap of its own. Refuse an
envelope on safe methods and read at most the decoded-size limit.
The live-update socket accepted any Origin whose host name matched, but
cookies do not separate ports or schemes, so a page served by another
service on the same host name could open it with the admin's session.
Only http(s) origins are accepted now (https when the panel terminates
TLS itself), and the port must match whenever the Host header has one.
SetLoginUser added the user to whatever session the login request
carried, so its CSRF token outlived the login. Anyone can fetch an
anonymous session with its token, and a host on a sibling domain can
plant that cookie in the admin's browser. Clear the session before the
user is stored; the next page load or token fetch mints a new token.
A failed login keeps its username in the in-memory log buffer, the
limiter table and the notifier queue, and the route accepted any body
the global cap allowed. The login route now takes at most 16 KiB and
answers a username over 256 bytes with the generic failure before any
of those see it; the user setters refuse such a name too, so no stored
username can be one the login route turns away.
The login limiter counted failures per (address, username) only after
the check finished, so attempts sent together all passed before the
first failure was recorded, and one address could own every record of
the bounded table and push out the one holding its own failures.
An admitted attempt now holds a slot until it reports back, and an
address is blocked as a whole after five times the per-username limit
spread over any names, which also caps the records one peer can own.
Remote routing sources are fetched documents the panel re-renders on
every subscription request. A Clash YAML document stays small through
aliases, which decode to one shared value but are written out in full,
so it now has to fit a 4 MiB written-size budget and 16 levels of
nesting when it is accepted (and when a persisted copy is loaded). A JSON
routing profile could repeat route groups, each copy repeating every
site list; a group now counts once, since a repeat can never match.
The YAML encoder writes multi-line strings as block scalars and gets
CR and the Unicode line breaks (NEL, LS, PS) wrong inside them, so a
remark, path or header name with such characters could continue as YAML
outside its scalar. Strings with a non-printable rune or invalid UTF-8
are now written double-quoted with escapes, map keys included.
A client's external subscription is fetched from a third party on every
subscriber refresh, and the fetcher trusted it more than its siblings:
- it dialled any address, redirects included; it now goes through the
SSRF guard like the outbound subscription and remote routing fetchers;
- it forwarded any line containing "://"; it now keeps only server share
links, so app deeplinks and plain URLs never reach the subscriber's app;
- it bounded the body but not the link count (now 1000) or the headers
(now 64 KiB), and stored the raw error, URL included, on every row; the
stored error is now host-only and at most 1 KiB.
The link parser also refuses links that nest JSON past 16 levels, since
the JSON subscription pretty-prints it at quadratic cost, and external
configs are marshalled compactly until the final, single indent pass.
The traffic poll read every Xray counter in one QueryStats reply under
gRPC's default 4 MiB receive limit; past it the whole poll failed and
no client was counted, so quotas stopped being enforced. Outbound
counters carry tags derived from third-party link remarks, so one long
remark was enough.
Poll the user, inbound and outbound families separately with a 64 MiB
limit, keep counting the families that answer, and cap a subscription
link's slug at 64 runes so the tag stays small to begin with.
The panel runs xray-core's conf.Build() in its own process on inbounds
and on outbounds that come from pushes and subscriptions. Some values
make it allocate or loop without bound (an XHTTP sessionIDLength, udphop
port lists, any number of finalmask entries, each xmc one costing an RSA
key), and certificate paths are read whole whatever they name.
Refuse those before any build: sessionIDLength above 1024, more than one
port range's worth of udphop ports per stream, more than 16 masks per
stream, more than 64 xmc masks per outbound subscription, and certificate
or key files that are not regular files of at most 1 MiB. Inbound saves,
the TUIC readers and the certificate-pin helper apply the file rule too;
the pin helper also stops looking at paths that name files on a node.
MHSanaei