Added
-
--client-stream-idle-timeout(seconds, default 30) cancels a
client-streaming RPC whose client has sent nothing for that long, whether
before the first message or between messages. Without it, a client that
opensGetTaddressBalanceStreamand sends no address parks a handler
goroutine indefinitely, so enough open streams exhaust memory. Set it to 0
to disable (GHSA-5xg7-x58j-m5vr, GHSA-4wjx-vhhv-vc8c). -
--max-concurrent-recvbounds how many requests the server receives at
once across every connection; the default is 32. gRPC reads a whole request
into memory before any handler runs, so many large requests arriving at
once could exhaust memory. A request over the limit is refused with
UNAVAILABLEbefore its body is read, and one whose body hasn't arrived
within 20 seconds is canceled. Client-streaming calls, including gRPC
reflection, count against a separate pool of the same size for as long as
they're open. Set it to 0 for no limit (GHSA-xrx5-jc6q-9h8r).When a pool is full, a new request can take the oldest slot of the
connection holding the most, so one connection can't lock out the rest.
Clients spread over many connections still can, so a per-IP connection
limit at a proxy or firewall is a good idea. -
--max-concurrent-sendtx-mibbounds the transaction data, in MiB, that
SendTransactionworks on at once; the default is 16. Each transaction is
copied several times on its way to the backend. Requests over the limit are
refused withUNAVAILABLE. Set it to 0 for no limit
(GHSA-xrx5-jc6q-9h8r).
Changed
-
Repeated
poolTypesentries in aGetBlockRange,GetBlockRangeNullifiers
orGetMempoolTxrequest are now ignored after the first, so the server's
filtering cost no longer grows with the length of the list
(GHSA-8vh8-6jg3-8f5v). -
The largest request lightwalletd accepts is now 2 MiB, down from gRPC's
default of 4 MiB. That still fits the largest transaction
SendTransactionaccepts. This doesn't apply in darkside mode, and neither
does--max-concurrent-recv. -
A
GetMempoolTxmempool refresh that fails no longer holds off other
callers' refreshes for two seconds. A client that cancels, or stops
reading, no longer stops the mempool view from being refreshed. A
transaction that fails to parse is skipped rather than failing every
refresh until it leaves the mempool (GHSA-w64p-hjp5-3mj7). -
GetMempoolStreamno longer holds its shared lock while calling the
backend, so a slow backend no longer stalls every other stream client or
keeps them from noticing a cancel. A client that cancels no longer stops
the refresh it started, and a transaction that fails to parse is skipped
rather than failing every refresh until the next block
(GHSA-hp3q-vfr4-v8m6, GHSA-pj5c-jx29-gmxr).
Fixed
-
A request with a huge number of list elements is now refused before it is
decoded. An empty element takes 2 bytes on the wire but 16 to 24 bytes once
decoded, so a single 4 MiB request could cost almost 90 MiB before the
per-request limits on addresses or excluded txids could reject it, and a
few at once could exhaust the server's memory. The limits checked before
decoding are twice the existing ones, so a request just over a limit still
gets the sameInvalidArgumentas before (GHSA-p4pq-3cf7-45rv). -
GetMempoolTxno longer sorts the shared mempool list in place. Concurrent
callers each read that list without holding the lock, so the in-place sort let
one caller reorder it under another, which could send a transaction twice and
skip another.MempoolFilternow works on copies and leaves its arguments
untouched. -
The Prometheus metrics HTTP listener (
--http-bind-addr) now has request
timeouts: 10 seconds for a request's headers to arrive, 30 seconds for the
whole request and for the response, and 60 seconds idle. Before, a peer
that sent a partial request held its connection, and a file descriptor,
for as long as it liked, and enough of them starved the gRPC listener in
the same process (GHSA-w24w-x7vw-f2w8). A failure to start the listener is
now logged instead of ignored.The listener also uses its own handler mux, so it no longer serves
/debug/vars, which exposed the command line, including an RPC password
given there (GHSA-j79f-98gc-33jw). -
A block with no transactions, or whose first transaction has no
transparent input, is now rejected by the parser instead of crashing the
process when its height is read. Consensus requires a coinbase with one
input, so only darkside mode or a misbehaving backend could supply such a
block (GHSA-56hr-g827-v3rj, GHSA-jcrq-vv43-968c). -
Darkside
AddAddressUtxonow rejects a transaction ID that isn't 32 bytes
withInvalidArgumentinstead of crashing (GHSA-pp3p-5vmm-3cvw). -
A panic in a gRPC handler now fails that one request with
Internaland
logs a stack trace, instead of ending the process. -
Transparent-address validation compiles its regular expression once
instead of on every address, so a request with many addresses no longer
costs thousands of compilations (GHSA-x943-ghgq-v675). -
The startup debug log line that prints the parsed options no longer
includes the backend RPC password (CodeQL alert #4).
Security
-
Built with Go 1.27.4, which fixes an HTTP/1 server crash in
net/http
(GO-2026-6859). Only lightwalletd's HTTP listener uses HTTP/1, and it binds
to localhost by default. Also updatedgolang.org/x/netto v0.62.0, which
fixes two issues in itshtmlpackage (GO-2026-6860, GO-2026-6866).
lightwalletd doesn't use that package. -
Updated
golang.org/x/netto v0.60.0, which fixes five HTTP/2 server
vulnerabilities reachable through gRPC (GO-2026-6603, GO-2026-6610,
GO-2026-6611, GO-2026-6612, GO-2026-6617). Two of them let a client crash
the process or exhaust its memory (GHSA-gq9m-j9cv-6w9p).