Fragments carried 1280 bytes of value data, plus the msgpack framing and,
for encrypted peers, the envelope overhead: packets exceeded the 1280-byte
IPv6 minimum MTU. Derive both fragment sizes from the 1232-byte UDP payload
that fits in it, less an upper bound of the fragment framing.
Value data fragments carry no node ID and were merged into a partial
message by source address and transaction ID alone, first come first
served. Transaction IDs increment from a random start and show in any
plaintext packet of a pair, such as the first-contact ping. An attacker
spoofing the sender's address could thus pre-seed forged plaintext
fragments, or a plaintext header, for an encrypted reply: the real parts
were then dropped as overlapping, forging unsigned values or losing the
message.
Parts of a message must now all be plaintext, or all come in envelopes
from the same sender key as its first part. An encrypted part replaces
plaintext state, which anyone can forge from a spoofed address; plaintext
or foreign-key parts never join an encrypted message.
A session was cached for an address whenever we answered a peer there,
including replies to Initial and Hello envelopes, whose source address is
not verified. A spoofed Initial or Hello therefore replaced the session of
the real owner of the address, whose Data envelopes were then dropped until
a retry, and spoofed Initials from many addresses filled the cache and
evicted legitimate sessions, which were picked by address order.
Only cache a session for a peer whose ephemeral key arrived in a Session
or Data envelope, i.e. opened with our key for its address. Replies to
Initial and Hello envelopes derive their keys without caching them, so a
node that sent an Initial only confirms the session, and switches to Data
envelopes, once the peer answered a Session envelope. Evict the least
recently used session when the cache is full.
Sessions are keyed by peer address, so the receiver finds the keys from the source address: Data envelopes drop both keys from the header (30 bytes of overhead instead of 94) and need no key derivation to reject. A session is confirmed when the peer answers our envelope request with the session keys, or sends Data itself. Credentials only travel in key-carrying envelopes, which replaces Peer::acked. A received Initial envelope unconfirms the session, and full session caches evict instead of refusing, so a peer that derived keys keeps them.
Ephemeral keys derive from a per-epoch secret and the peer address, so only that address learns them. A Session envelope then proves return routability: encrypted put, refresh and listen are accepted without token, and messages to encrypted peers no longer carry one (40 bytes saved). Initial envelopes cannot write. Tokens for plaintext peers move from Dht to NetworkEngine and derive from the same epoch secret. Session caches are keyed by address, and cached keys of both epochs are tried before any derivation.
Each node rotates an ephemeral X25519 key every 30 minutes and keeps the
previous one for another period. Envelopes carry the sender static and
ephemeral keys. Session envelopes use keys derived from all four DH
combinations of static and ephemeral keys, erased with the epoch.
Initial envelopes, sent before the peer's ephemeral key is known, mix
the sender ephemeral and static keys with the receiver static key.
A hello is unauthenticated: it is only answered, without adding or
moving its claimed node. A peer is acknowledged only by a reply to a
request sent in an envelope, and a request retry restarts with an
Initial envelope and our credential.
With persist_path, a transport key generated by the node is saved in the
state (owner-only file permissions) and restored on start, keeping the
node ID across restarts. Configured keys are never written to the state.
Nodes using transport encryption derive their node ID from their X25519
transport key, advertise it in plaintext messages ("k") and in node
lists ("k4"/"k6"), and encrypt all traffic with nodes whose key is
known. Keys are bound to node IDs, so they can't be forged, and
plaintext messages never move a node known to support encryption.
Legacy nodes keep talking in plaintext.
Private networks (TransportConfig::authorities) only accept nodes with
a credential issued by one of the network authorities, never send or
accept plaintext, and introduce themselves to addresses with a Hello.
X25519 transport keys, node ID derivation from the public key,
HKDF-SHA256 key agreement, ChaCha20-Poly1305 envelopes that legacy
msgpack parsers reject, and Ed25519 network authority credentials.
Route push notifications to Huawei devices, which have no Google Play
Services and therefore cannot receive FCM. gorush, already used as the
push relay, supports HMS natively as platform 3.
HMS takes its payload as a serialized JSON string in "huawei_data" and
carries its own ttl/priority fields, so it needs its own branch rather
than the generic "data" object used for FCM and APNs.
The enum value is appended last: MSGPACK_ADD_ENUM serializes it as an
integer, so reordering would break compatibility with deployed proxies.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Query::isSatisfiedBy() considers an unfiltered answer sufficient for a
filtered query, which is true for a vanilla get: onGetValuesDone() applies
the operation filter to the values before calling back.
A projected query is different. It receives the fields another query selected
and can no longer evaluate its own filters on them, all the more so as those
filters may target fields that were not selected at all. Two concurrent
queries such as "SELECT id" and "SELECT id WHERE value_type=1" therefore made
the filtered callback report identifiers whose type does not match, and
searchNodeGetDone() additionally cancelled its own request so it never got a
correctly filtered answer.
Add Query::isProjectionSatisfiedBy(), which additionally requires both queries
to apply the same filters, and use it for operations with a query callback.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
SecureDht only forwarded the push token to the DHT it wraps, so the
push platform and topic given to DhtRunner after the proxy client was
created never reached it. A device started without push platform
(e.g. a Jami account created by a build without push support) then
subscribed with an empty platform once given a token, and the proxy
server pushed it as an iOS device.
Co-authored-by: Copilot (Claude Opus 5.5) <223556219+Copilot@users.noreply.github.com>
The DHT listen callback of a push subscription kept the push type and
topic of the request that created it. Refreshing the subscription
with another platform updated the listener, but values kept being
pushed to the first platform: an Android device that once subscribed
without platform was pushed as an iOS device forever (BadDeviceToken)
and never received values such as connection requests.
Keep the push type and topic in the session context shared with the
callback, and update them on each subscription refresh and when the
state is restored.
Co-authored-by: Copilot (Claude Opus 5.5) <223556219+Copilot@users.noreply.github.com>
Prefix::size_ counts bits, but commonBits walked content_ a byte at a
time up to it. Two prefixes that agree on their whole common length are
therefore read well past their end, which AddressSanitizer reports as a
heap-buffer-overflow, and the count that comes back is eight times too
large. A lookup with a common prefix length hits this as soon as an
index entry matches the key it is looking for, and entry.prefix comes
straight off the network. The flags were indexed the same way, so a
partially known prefix was read at the wrong position too.
swapBit had the other half of the confusion: it flipped bit
(8 - bit) % 8 of the byte, while isActiveBit reads bit 7 - (bit % 8).
getSibling, which flips the last bit of a prefix to name the sibling
node in the trie, therefore returned a prefix differing in some other
bit and sharing the last one, so lookup and insertion looked up the
wrong node. Splitting a leaf and finding a canary both rely on it.
Add tests covering both, which fail before the fix: commonBits returns
193 for two identical 160 bit prefixes, and a prefix and its sibling
share 7 bits rather than all but the last.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Resolve idle-runner push futures immediately under the operation queue lock. A stopped runner has no loop left to execute the queued promise, otherwise consumers can wait indefinitely.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
asio sockets are not thread-safe. http::Connection::close() used to close
the socket synchronously on the calling thread. Request::cancel() and
Request::~Request() (via terminate()) are called from arbitrary threads
(e.g. DhtProxyClient::stop(), cancelAllListeners(), cancelListen()),
so the close raced with in-flight operations on the io_context thread.
In particular, asio::async_connect's range_connect_op closes and
re-opens the socket when trying the next endpoint (common on mobile
where the IPv6 endpoint fails first). A concurrent close() from another
thread corrupts the epoll reactor's per-descriptor state, leading to
crashes in epoll_reactor::deregister_descriptor / register_descriptor
called from range_connect_op::process.
Dispatch the close to the io_context thread when not already on it.
Also keep the connection alive during the TLS handshake, like every
other asynchronous operation, so that the destructor can never run
while an operation is pending.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Track subscribe/permanent-put expiration with std::chrono::system_clock
(wall clock) instead of steady_clock in both the proxy client and server,
so persisted expirations survive restarts and map directly to epoch time.
- DhtProxyServer::clock is now system_clock; expireTimer/expireNotifyTimer
use asio::system_timer.
- DhtProxyClient::refreshSubscriberTimer uses asio::system_timer and
system_clock::now() for subscriber refresh.
Key partial messages by source and transaction ID, buffer bounded fragments received before their header, and avoid charging reordered requests twice against rate limits. Cover source collisions, early fragments, empty data and large values.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
RevocationList::sign() always overwrote the CRL Number extension: random
on the first signature, incremented afterwards. This leaves no way for a
certificate authority replicated over several devices to keep numbers
unique, since every device branching from the same list mints the same
number for a different content.
Add an optional `number` parameter, used verbatim when non-zero, so
callers can partition the number space (RFC 5280 5.2.3 requires
monotonicity for a given CRL scope and issuer, not globally).
Also fix the default increment path: the extension holds a DER integer
with leading zeros stripped, so reading it straight into an uint64_t only
worked for numbers wide enough to occupy all 8 bytes. Signing a list
numbered 5 produced 0x0500000000000001 instead of 6. Read and write the
number as a big-endian byte string instead, which makes the `endian()`
helper unnecessary.
Add tests covering the default numbering, explicit numbers (including
values with leading zero bytes and the largest positive 8-byte integer),
pack/unpack round-trips, and a device-partitioned numbering scheme.
This prevents ThreadPool::stop from hanging if it is called with `wait`
set to true (which is notably the case in ThreadPool::join) and the
pool's task queue is non-empty.
Make Hash default constructor, hex constructor, data(), const
operator[] and fromString constexpr. Change fromString to be
non-throwing: it now uses a validity flag and fills the hash with zeros
on parse error instead of throwing. Add a constexpr toHexArray helper
that returns a std::array<char, N*2> for compile-time hex conversion.
Update tests to include <string_view>, use std::string_view, add
constexpr InfoHash usage and a static_assert for the hex array.
Adrien Beraud
cyrilleandCopilot
vx32
François-Simon Fauteux-Chapleau
copilot-swe-agent[bot]andaberaud
Pavan Koushik Nellore
Adrien Béraud