Hi, HAProxy 2.6.36 was released on 2026/11/18. It added 48 new commits after version 2.6.35.
This release is made of a subset of the fixes shipped in 2.8.31. It starts with four important fixes. A peer could send a message whose payload did not fit in the channel once the buffer reserve was accounted for: the stream was then marked as waiting for room although nothing had been transferred, a state that can never be resolved. The H2, H3 and QUIC multiplexers treated it as an internal error and killed the worker; they now report a regular stream error. In the same conditions a hard assertion in the buffer shift code could fire, which is now relaxed to only reject truly impossible moves. Crashes could be experienced because of these issues, so do not delay the upgrade. This branch also receives QUIC work that was delayed: the handshake no longer depends on the client resending its ClientHello when it arrives split into many CRYPTO frames (which is what Chrome does), a transfer no longer freezes when an empty STREAM frame carrying only the FIN bit has to be retransmitted, a connection whose stream was released before sending anything no longer waits for its idle timeout, and "wait-for-handshake" now really holds the streams until the QUIC handshake is complete. The rest of the release fixes truncated or stalled HTTP/1 responses, a leak of file descriptors, two crashes in the resolvers while a name is resolved, a wrong result in two sample converters, and a few other issues that can exhaust resources over time. One behaviour change deserves attention: "cache-use"/"cache-store" rules declared in a "defaults" section, which never worked and could crash or silently break caching, are now rejected with an explicit alert at startup, so they must be moved to the proxies. * buf/muxes: A stream could be left waiting for room that could never come. When data could not be moved from a receive buffer to the channel because the channel was empty but its reserve left no usable room, the stream was flagged as waiting for room even though nothing had been transferred, and that state could never be resolved. The H2, H3 and QUIC multiplexers detected it as an internal error and aborted the process, so a single request with a payload sized against the buffer could kill the worker. They now report a stream error, so only the affected stream is closed. Under the same conditions, a hard assertion in the buffer shift code could also fire; it is relaxed to only reject truly impossible moves. Reported with a reproducer and a fix proposal by Adam Crawford; the buffer assertion was found by code analysis (Yuqi Qiu and Xiang Li, Nankai University). * cache: A leak and a misplacement of the cache rules were fixed. Two cache bugs were fixed: a configuration with two "cache-use" rules that both hit leaked an applet and a cache row on every request, until memory and cache space were exhausted (any client could trigger it by requesting cached objects), and "cache-use"/"cache-store" rules declared in a "defaults" section never worked and could crash or silently break caching; they are now rejected with an explicit alert at startup, so they must be moved to the proxies. * mux-h2: Host and path based rules could be bypassed on absolute URIs. When an HTTP/2 request carried an absolute URI (e.g. "GET http://host/path"), the authority and the path sent to the server were not extracted the same way as HAProxy's own analyzers do it: the authority stopped at the first '/' instead of the first '/', '?' or '#', and the userinfo was stripped up to the first '@' instead of the last one. A request evaluated by HAProxy for one host or path could therefore be delivered to the server with another one, bypassing host and path based rules. The mux now splits absolute URIs exactly like the HTTP analyzers do. * quic: A handshake stall could be caused by an unset loss timer. An unset loss timer could hide a real one and make the server loop on padded Initial packets without ever retransmitting the Handshake data, stalling the handshake for minutes with clients that do not arm the anti-deadlock probe. The selection now only considers the packet number spaces that have an armed loss time. * quic: The handshake no longer relies on retransmitting the whole packet. A ClientHello may be split across several CRYPTO frames mixed in a single QUIC packet, which is what Chrome does to keep its first Initial packet opaque to middleboxes. Frames were handled strictly in order, so heavily fragmented CRYPTO data could not be buffered and the whole packet was rejected and left unacknowledged, unless the client re-emitted its ClientHello. The parsing loop now stores the frames it cannot buffer yet and parses the packet again, up to three times, so the handshake completes on the first attempt. A crash could occur on the error path of this reparse loop, which is fixed as well. * quic: A transfer could freeze when its FIN was lost. STREAM frames have a dedicated retransmission path that drops the ranges already acknowledged. An empty STREAM frame carrying only the FIN bit covers no range at all, so it was discarded on retransmission and the client waited forever for the end of the response. Such a frame is now kept until its FIN is acknowledged, which also required extending the release state of the stream descriptor. A connection could also freeze until its idle timeout when a stream was released before sending anything (for instance when an error was raised while encoding the response headers), because the stream was not removed from the connection's tree. * quic: "wait-for-handshake" is now effective over QUIC. The "wait-for-handshake" http-request action did nothing on a QUIC frontend: streams were processed as soon as early data arrived. The QUIC multiplexer now holds the streams until the handshake is complete and releases them afterwards, so rules placed after this action really run against a negotiated connection. * quic: Packet parsing was reworked before the fixes above. These patches do not change the behaviour on their own; they prepare the CRYPTO reparse loop by returning a dedicated status from CRYPTO frame handling, by allocating the frame descriptor instead of keeping it on the stack, and by simplifying the return path of the packet parser. * h3: A malformed stream could stall a transfer instead of closing it. A request stream closed with a standalone FIN before any valid request or response was received is now rejected instead of being treated as an empty complete message, so it no longer stays accounted and its memory is released promptly. The "maxrewrite" limit is now enforced on request headers, and the :path pseudo-header is checked for all scheme values (it was only checked for http and https, so an invalid path was accepted with any other scheme). * mux-h1: Responses were truncated, and pipelined requests stalled. A soft-stop while the last response was still being downloaded released the connection and dropped the remaining data, and a pipelined request ending exactly at the end of the buffer was not parsed until "timeout http-request" (or was rejected with a 414 on a valid request). Both are fixed. A 101 response following a "100 Continue" in the same segment was also rejected with a 502 because the websocket key was searched in the wrong message, and "option idle-close-on-response" is now honoured during a soft-stop combined with "close-spread-time". * mux-h1/h1: Other robustness fixes. - "h1-case-adjust" rules now match header names regardless of case, as the documentation states. - "option accept-unsafe-violations-in-http-{request,response}" remains effective on the final response after a 1xx, and after parsing restarted on a full output buffer. - The Connection/Upgrade header cleanup is now idempotent when the headers are received in several parts. - In TCP mode, the Upgrade header is no longer modified in flight while HTTP sample fetches are evaluated. - An FCGI response whose headers arrive in several parts no longer loses its content-length. - The "h1_open_streams" counter exposed in the stats no longer drifts (and can no longer become negative) when a stream cannot be created. - "%Ti" and "%Tt" are now correct for streams on an H2/H3 frontend with an H1 server. * ssl: Certificate management and wolfSSL builds were fixed. When "commit ssl cert" was interrupted (e.g. the CLI client closed the connection) after building its new instances in several passes, the freed transaction was still referenced: the next certificate command on the same certificate could crash the process or corrupt memory. A certificate used by more than 10 instances was enough to make the command multi-pass. With wolfSSL builds, certificate loading could fail even without any .sctl file present: this library provides the custom-extension registration call as a stub that always fails, while HAProxy relies on it to load the SCT list of a certificate. The capability is no longer advertised for wolfSSL, so such builds load their certificates normally again. * peers: Stick-table updates could be lost or misapplied between peers. When an update carrying a "server_key" did not fit in the peer's output buffer, the new dictionary entry was kept in the transmission cache while the value was never sent. The remote peer then resolved its identifier to an unknown or stale entry, losing the server name for the whole session. In addition, when a peer reconnected, the sync task deadline was overwritten by each peer, so heartbeats to idle peers could be delayed until the remote peer closed the connection for inactivity. On the stick-table side, the dictionary reference held by "server_key" was only released on one destruction path, so memory grew with the number of distinct values that went through the table. * resolvers: Two crashes and an out-of-bounds read were fixed. A server of a proxy using a SRV name but without "resolvers" keyword was kept attached to the SRV request of the other servers: the first SRV response could then crash the process, and a reload with a server state file crashed too. Separately, "do-resolve" with a client-controlled value such as hdr(host) could build a DNS name with a 256-byte label or an embedded NUL, leading to out-of-bounds reads and disclosure of heap data in DNS queries. * mux-fcgi: Responses were truncated for slow clients. The FCGI connection timeout was enforced even when streams were still attached. With a slow client, HAProxy stops reading from the FastCGI server once buffers are full, nothing happens on the connection for a while, and it was killed in the middle of the response: the client received a truncated response, logged as "SD--", and the server got a reset while blocked in write(). The timeout is now only enforced when no stream remains attached, as already done for H2; a FastCGI server that does not answer at all still gets a 504 after "timeout server". * fd/connection: A descriptor leak could exhaust "maxsock", and a descriptor could be closed twice. When an accepted connection could not be prepared, for instance because "maxsslconn" was reached, its file descriptor was left open. Incoming connections on that listener kept accumulating leaked descriptors even after SSL connections were released, until "maxsock" was exhausted, which also prevented non-SSL frontends from accepting. Conversely, when a connection failed during a health check and the socket had already been closed, the descriptor could be closed a second time, closing an unrelated file descriptor. * task: A task could be destroyed twice in master-worker mode. A task already flagged as killed but still linked in a queue could be destroyed a second time when a forked child cleans the tasks before executing (master-worker mode with a "peers" section and a local peer). This aborted the child with SIGILL on older branches; the leftover window is now closed by ignoring already-killed tasks, which the scheduler frees when dequeuing them. * samples: Two wrong results were fixed in converters. "jwt_member_query" only checked one of the two error codes of the base64 decoder, and the "ungrpc" converter could return a field taken from the wrong message when a body carried several gRPC messages. * cfgparse: Unsupported password hashes are rejected. "crypt()" may return NULL or an error token for unsupported hash types (on Debian/Ubuntu, strings starting with '*'), and this was not checked, so a user or group password using such a hash was accepted at startup and could never authenticate. It is now detected during configuration validation. It is strongly recommended to upgrade to this version. Crashes could be experienced because of the four fixes listed first. No configuration change is needed to benefit from them, and none of them alters the protocol or the configuration syntax: any 2.6.x build can be replaced in place. Deployments exposing HTTP/2 or HTTP/3 to the Internet should treat this release as mandatory. Thanks everyone for your help on this release. Please find the usual URLs below : Site index : https://www.haproxy.org/ Documentation : https://docs.haproxy.org/ Wiki : https://github.com/haproxy/wiki/wiki Discourse : https://discourse.haproxy.org/ Slack channel : https://slack.haproxy.org/ Issue tracker : https://github.com/haproxy/haproxy/issues Sources : https://www.haproxy.org/download/2.6/src/ Git repository : https://git.haproxy.org/git/haproxy-2.6.git/ Git Web browsing : https://git.haproxy.org/?p=haproxy-2.6.git Changelog : https://www.haproxy.org/download/2.6/src/CHANGELOG Dataplane API : https://github.com/haproxytech/dataplaneapi/releases/latest Pending bugs : https://www.haproxy.org/l/pending-bugs Reviewed bugs : https://www.haproxy.org/l/reviewed-bugs Code reports : https://www.haproxy.org/l/code-reports Latest builds : https://www.haproxy.org/l/dev-packages --- Complete changelog : Aleksandar Lazic (1): BUG/MEDIUM: mux-fcgi: do not enforce the timeout on connections with streams Amaury Denoyelle (12): MINOR: quic: convert qc_stream_desc release field to flags BUG/MEDIUM: quic: handle retransmit for standalone FIN STREAM BUG/MINOR: quic: prevent freeze after early QCS closure BUG/MEDIUM: quic: support wait-for-handshake MINOR: quic: simplify qc_parse_pkt_frms() return path MINOR: quic: use dynamically allocated frame on parsing MINOR: quic: extend return value of CRYPTO parsing BUG/MINOR: quic: repeat packet parsing to deal with fragmented CRYPTO BUG/MEDIUM: quic: prevent crash due to CRYPTO parsing error BUG/MINOR: h3: enforce maxrewrite on request headers BUG/MAJOR: h3: always check :path for all scheme values BUG/MEDIUM: h3: reject standalone FIN on invalid request/response Chris Staite (1): BUG/MEDIUM: ssl/cli: reset the ongoing transaction when "commit ssl cert" is interrupted Christopher Faulet (15): BUG/MEDIUM: h1: make the Connection/Upgrade headers cleanup idempotent BUG/MINOR: h1: keep accepting unsafe violations after a parser reset BUG/MEDIUM: h1: don't rewrite the Upgrade header when not cleaning headers BUG/MEDIUM: h1-htx: don't stall on messages ending exactly at the buffer's end BUG/MEDIUM: mux-h1: don't truncate responses on soft-stop BUG/MINOR: mux-h1: only reset the session's timing info on the frontend side BUG/MINOR: mux-h1: fix the open streams counter on stream creation failure BUG/MINOR: h1: forget a previous pass' results when parsing headers only BUG/MINOR: mux-h1: convert the <from> argument of h1-case-adjust to lower case BUG/MINOR: mux-h1: search the websocket key in the right message BUG/MEDIUM: mux-h1: don't spread-close idle conns with idle-close-on-response BUG/MEDIUM: connection: don't leak the fd of an accepted connection BUG/MEDIUM: resolvers: detach servers without resolvers from the SRV request BUG/MEDIUM: cache: do not create a second applet when two cache-use rules hit BUG/MEDIUM: cache: forbid cache rules in defaults sections Frederic Lecaille (5): BUG/MEDIUM: peers: keep the earliest sync task deadline BUG/MEDIUM: peers: do not keep a dict entry cached when its update was not sent BUG/MINOR: sample: realign the ungrpc converter on message boundaries BUG/MEDIUM: quic: fix loss time selection in quic_loss_pktns() BUG/MINOR: stick-table: release server_key on any stksess destruction Olivier Houchard (3): BUG/MEDIUM: resolvers: reject too long labels when encoding a name BUG/MAJOR: mux-h2: split absolute URIs like the HTTP analysers do on output BUG/MEDIUM: fd: Mark a fd as dead on early close Remi Tricot-Le Breton (1): BUG/MINOR: sample: Fix wrong 'base64urldec' return value check in 'jwt_member_query' William Lallemand (6): BUG/MEDIUM: task: do not destroy twice the tasks already killed in mworker_cleantasks() BUG/MINOR: cfgparse: detect unsupported crypt() hashes on Debian/Ubuntu BUG/MINOR: ssl: exclude wolfSSL from custom extension support SCRIPTS: vtest: allow overriding the Git repository URL CI: github: use haproxytech/VTest2 for regression tests CI: github: use haproxytech/VTest2 for musl regression tests Willy Tarreau (4): BUG/MAJOR: buf: make the shift distance check less strict in b_move() BUG/MAJOR: h3: do not panic on HTX_FL_PARSING_ERROR after htx_xfer() BUG/MAJOR: mux-quic: fail on impossible data transfers in rcv_buf() BUG/MAJOR: mux-h2: fail on impossible data transfers in rcv_buf() -- Christopher Faulet

