Hi, HAProxy 2.0-dev5 was released on 2019/06/02. It added 92 new commits after version 2.0-dev4.
This version addresses a small number of no-to-trivial issues met while working on fixing other bugs and during benchmarks. One such issue concerns the excess of wakeup/sleep in the H2 mux since 1.9, due to the fact that the connection's mux buffer is the same size as the streams' and that when forwarding large data, each stream is able to fill the buffer and leave no more room to other ones. There are ways to address this at the connection management level but this requires to enforce new internal rules that were not necessarily strict enough yet, and that would require to carefully recheck the code. This will be for 2.1. In the mean time, using a list of buffers (buffer ring) for the connection was way easier and significantly improved the performance, while taking advance on what was expected for 2.1 without taking much risk. We also noticed that some locks and some atomic operations in the scheduler were taking a significant amount of time by causing lines to bounce between all cores. While it's often irritating to see this, I'm reminding myself that two years ago we were crossing fingers hoping that very basic multi- threading would be completed in time for 1.8 and that now we've already reached the point where bouncing CPU cache lines are the factor affecting the scalability, so I must admit it's quite pleasant to revisit such parts which were once consdered very fast, because now they're the bottlenecks! Thus some locks were reviewed again and the ones which were really not used anymore were released. Others have been identified to be reworked in 2.1. Another indicator that we're getting closer to the release is that we're mostly facing bugs affecting stable releases as well. For example there was another bug in the LB code where it's possible to loop forever on a dead server, which also affects 1.9 and 1.8, the queue dispatching code was bogus since 1.9, and we discovered that "http-request reject" rules have never worked well... Aside this, some improvements to the HTX code were made to better deal with 1xx messages. Strictly speaking they were not required for this version but I predicted that we'd face difficult to fix issues if we were to postpone them for 2.1, and in practice we found corner cases which are better addressed since these changes. A user-visible change was made (I hate to do this but it has become a real pain and we'll have more trouble if we don't address it now). It's the presence of similarly-named servers in the config. I figured that 5 years later, we continue to warn instead of reject this. Given all the dynamic stuff people are using, it creates a real havoc to still have to deal with this, so better reject this now than letting people down this breach for another 5 years. The HTX performance was further increased by allowing zero-copy to be performed when receiving an H2 message. This will be visible in H2 to H2 scenarios mostly. In addition, the H2 performance appeared to be horrible when configured with large buffers, due to a few remaining bytes available after each message causing some realigns. This was now fixed. The SOCKS4 upstream server code was merged. At the moment while SOCKS4 happens to work for me, it fails for its author (Alec) and his fix really bothers me. While merging it we noticed that the current connection layer is still affected by a few very dirty corner cases that I really don't like. In short, the handshake code was never used to switch between reading and writing till now. It's not a big deal since this code is going to disappear soon, but when working on the replacement part, Olivier found other deeper issues that I really want to see addressed. Just for this, I'm expecting at least another -dev next week before a final release as we can't afford to have dirty hacks in this area in an LTS version or we risk to pay them the high price. I just found while testing this release that "tcp-request inspect-delay" doesn't work in HTTP+HTX mode (it does not wait). I don't know why yet, but 1.9 has the same issue. Overall most of the performance issues addressed in this version also affect 1.9 and sometimes even 1.8, so we could expect that a few of them will be backported later if they are of particular concern. Now for the next steps, I'd like to emit another dev version this week (Wednesday or Thursday depending how things go with the pending issues), with the hope that we can do 2.0-final the week after, before the 15th. Please find the usual URLs below : Site index : http://www.haproxy.org/ Discourse : http://discourse.haproxy.org/ Slack channel : https://slack.haproxy.org/ Issue tracker : https://github.com/haproxy/haproxy/issues Sources : http://www.haproxy.org/download/2.0/src/ Git repository : http://git.haproxy.org/git/haproxy.git/ Git Web browsing : http://git.haproxy.org/?p=haproxy.git Changelog : http://www.haproxy.org/download/2.0/src/CHANGELOG Cyril's HTML doc : http://cbonte.github.io/haproxy-dconv/ Willy --- Complete changelog : Alexander Liu (1): MEDIUM: connection: Upstream SOCKS4 proxy support Christopher Faulet (43): BUG/MINOR: proto-htx: Try to keep connections alive on redirect BUG/MEDIUM: spoe: Don't use the SPOE applet after releasing it BUG/MINOR: lua: Set right direction and flags on new HTTP objects BUG/MINOR: mux-h2: Count EOM in bytes sent when a HEADERS frame is formatted BUG/MINOR: mux-h1: Report EOI instead EOS on parsing error or H2 upgrade BUG/MEDIUM: proto-htx: Not forward too much data when 1xx reponses are handled BUG/MINOR: htx: Remove a forgotten while loop in htx_defrag() MINOR: htx: Remove the macro IS_HTX_SMP() and always use IS_HTX_STRM() instead MINOR: htx: Move the macro IS_HTX_STRM() in proto/stream.h MINOR: htx: Store the head position instead of the wrap one MINOR: htx: Store start-line block's position instead of address of its payload MINOR: htx: Add functions to get the first block of an HTX message MINOR: mux-h2/htx: Get the start-line from the head when HEADERS frame is built MINOR: htx: Replace the function http_find_stline() by http_get_stline() CLEANUP: htx: Remove unused function htx_get_stline() MINOR: http/htx: Use sl_pos directly to replace the start-line MEDIUM: http/htx: Perform analysis relatively to the first block MINOR: channel/htx: Call channel_htx_recv_max() from channel_recv_max() MINOR: htx: Add function htx_get_max_blksz() BUG/MINOR: htx: Change htx_xfer_blk() to also count metadata MEDIUM: mux-h1: Use the count value received from the SI in h1_rcv_buf() MINOR: mux-h2: Use the count value received from the SI in h2_rcv_buf() MINOR: stream-int: Don't use the flag CO_RFL_KEEP_RSV anymore in si_cs_recv() MINOR: connection: Remove the unused flag CO_RFL_KEEP_RSV MINOR: mux-h2/htx: Support zero-copy when possible in h2_rcv_buf() MINOR: htx: Add a field to set the memory used by headers in the HTX start-line MINOR: h2/htx: Set hdrs_bytes on the SL when an HTX message is produced MINOR: mux-h1: Set hdrs_bytes on the SL when an HTX message is produced MINOR: htx: Be sure to xfer all headers in one time in htx_xfer_blks() MEDIUM: htx: 1xx messages are now part of the final reponses MINOR: channel/htx: Add function to forward headers of an HTX message MINOR: filters/htx: Use channel_htx_fwd_headers() after headers filtering MINOR: proto-htx: Use channel_htx_fwd_headers() to forward 1xx responses MEDIUM: htx: Store the first block position instead of the start-line one MINOR: stats/htx: don't use the first block position but the head one MINOR: channel/htx: Add functions to forward a part or all HTX payload MINOR: proto-htx: Use channel_htx_fwd_all() when unfiltered body are forwarded MEDIUM: filters/htx: Filter body relatively to the first block MINOR: htx: Optimize htx_drain() when all data are drained MINOR: htx: don't rely on htx_find_blk() anymore in the function htx_truncate() MINOR: htx: remove the unused function htx_find_blk() MINOR: htx: Remove support of pseudo headers because it is unused BUG/MEDIUM: mux-h1: Don't skip the TCP splicing when there is no more data to read Ilya Shipitsin (2): BUG/MINOR: ssl_sock: Fix memory leak when disabling compression BUILD: ssl: fix latest LibreSSL reg-test error Michael Prokop (1): DOC: fix typos Olivier Houchard (6): MINOR: ssl: Make sure the underlying xprt's init method doesn't fail. MINOR: ssl: Don't forget to call the close method of the underlying xprt. MEDIUM: sessions: Introduce session flags. BUG/MEDIUM: h2: Don't forget to set h2s->cs to NULL after having free'd cs. BUG/MEDIUM: connection: Use the session to get the origin address if needed. MEDIUM: tasks: Get rid of active_tasks_mask. Willy Tarreau (39): BUILD: watchdog: use si_value.sival_int, not si_int for the timer's value BUILD: signals: FreeBSD has SI_LWP instead of SI_TKILL BUILD: watchdog: condition it to USE_RT MINOR: raw_sock: report global traffic statistics MINOR: stats: report the global output bit rate in human readable form OPTIM: freq-ctr: don't take the date lock for most updates MEDIUM: mux-h2: avoid doing expensive buffer realigns when not absolutely needed CLEANUP: debug: remove the TRACE() macro MINOR: buffer: introduce b_make() to make a buffer from its parameters MINOR: buffer: add a new buffer ring API to manipulate rings of buffers MEDIUM: mux-h2: replace all occurrences of mbuf with a buffer ring MEDIUM: mux-h2: make the conditions to send based on mbuf, not just its tail MINOR: mux-h2: introduce h2_release_mbuf() to release all buffers in the mbuf ring MEDIUM: mux-h2: make the send() function iterate over all mux buffers CLEANUP: mux-h2: consistently use a local variable for the mbuf MINOR: mux-h2: report the mbuf's head and tail in "show fd" MAJOR: mux-h2: switch to next mux buffer on buffer full condition. BUILD: connections: shut up gcc about impossible out-of-bounds warning MINOR: cli/activity: remove "fd_del" and "fd_skip" from show activity MINOR: cli/activity: add 3 general purpose counters in development mode BUG/MAJOR: lb/threads: make sure the avoided server is not full on second pass BUG/MEDIUM: queue: fix the tree walk in pendconn_redistribute. BUG/MEDIUM: threads: fix double-word CAS on non-optimized 32-bit platforms MEDIUM: config: now alert when two servers have the same name BUG/MEDIUM: http: fix "http-request reject" when not final MINOR: htx: rename htx_append_blk_value() to htx_add_data_atonce() MINOR: htx: make htx_add_data() return the transmitted byte count MEDIUM: htx: make htx_add_data() never defragment the buffer MINOR: activity: write totals on the "show activity" output MINOR: activity: report totals and average separately MEDIUM: poller: separate the wait time from the wake events MINOR: activity: report the number of failed pool/buffer allocations MEDIUM: buffers: relax the buffer lock a little bit MINOR: task: turn the WQ lock to an RW_LOCK MEDIUM: task: don't grab the WR lock just to check the WQ BUG/MEDIUM: mux-h2: fix the conditions to end the h2_send() loop BUG/MEDIUM: mux-h2: don't refrain from offering oneself a used buffer BUILD: contrib/prometheus: fix build breakage caused by move of idle_pct BUG/MINOR: deinit/threads: make hard-stop-after perform a clean exit ---
