Conversation
… be used Outgoing connections went through five separate socket() calls, each hard-coding protocol 0. One factory, hts_socket_client(), now owns that choice and takes IPPROTO_MPTCP where the run asked for it. Two FTP data sockets pick up socket_set_nosigpipe() on the way, which they never had. --mptcp turns it on, --mptcp=0 off. Unset means on where the kernel exposes net.mptcp.blackhole_timeout and net.mptcp.syn_retrans_before_tcp_fallback: without those a middlebox that drops the SYN carrying the MPTCP option costs the whole connect timeout on every host behind it, so the user has to ask. Reporting goes through MPTCP_INFO, which is refused on a socket the kernel fell back to plain TCP. SO_PROTOCOL cannot answer that question: it reads IPPROTO_MPTCP either way, measured on both a negotiated and a fallen-back loopback connection. The accounting first counted an HTTPS connection twice, because a finished TLS handshake sends the slot back through the connect path. One single-file crawl reported four connections where it had made one. macOS has MPTCP behind connectx() on an AF_MULTIPATH socket, which is a different connect call rather than a different protocol argument, so it is not wired here. Signed-off-by: Xavier Roche <xroche@gmail.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Xavier Roche <roche@httrack.com>
…t could see The default read only whether two sysctls exist. blackhole_timeout is a number, and zero means the kernel's blackhole detection is off, so the file being there never answered the question. Read both values instead. The capability macro had two sources: configure's AC_CHECK_DECLS and a defined(IPPROTO_MPTCP) test in the installed htsnet.h. configure now owns it and puts HTS_INET_MPTCP in config.h, which keeps it out of htsfeatures.h, where only switches that move an installed struct belong. --enable-mptcp joins the other optional features, so --disable-auto-features reaches it, and it gates on host_os first the way --enable-backtrace does, or an explicit ask would fail the build on GNU/Hurd. Tests 397 and debian/rules want every feature named. Two mutants survived every test. Accounting once per mirror rather than once per connection read the same, because each crawl leg only required a non-zero count. Nothing exercised the unset option at all. Both now have a leg, and the crawls that count run with keep-alive off so more than one connection is on the line. The self-test skipped by returning 0, which reads as a pass, and it asserted MPTCP was negotiated on a build whose headers cannot report that. It returns 77 like its siblings, and gates on the new mptcpinfo feature. Dropped: an AF_INET6 retry, a probed flag, and two NULL checks no caller can reach. The counters cover the HTTP backend only, which the field docs now say. Signed-off-by: Xavier Roche <xroche@gmail.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Xavier Roche <roche@httrack.com>
The last leg crawls over HTTPS, so "build (no openssl, --disable-https)" red on it. Its neighbours skip such a leg on HTTPS_SUPPORT, and that loses one leg rather than the whole test, so the other five still run there. Caught by CI and not locally, because the default build has OpenSSL. Rechecked against a --disable-https build, where the test passes, and against --disable-auto-features, where HTS_INET_MPTCP is 0, the test skips, and 397 and 398 both pass. Signed-off-by: Xavier Roche <xroche@gmail.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Xavier Roche <roche@httrack.com>
xroche
changed the title
A front end polls hts_get_stats(), so a counter that only reaches the log is a counter no GUI can show. stat_transport_failures is the shape to copy: live on opt, published into hts_stat_struct at both the mirror-end and the on-demand site. The two Multipath TCP counts now go the same way. Both fields are appended at the tail of hts_stat_struct, so the soname holds. The direction that breaks is a front end built against the new header running against an older engine, which is the usual rebuild-after-merge. -#test=mptcpstats covers it, and runs everywhere because none of it needs a kernel with Multipath TCP: the counters reach the stats, copy_htsopt carries the option but never the counts, and an unset option leaves a set one alone. Signed-off-by: Xavier Roche <xroche@gmail.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Xavier Roche <roche@httrack.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
--mptcproutes every outgoing connection through one socket factory, which asks the kernel for Multipath TCP and falls back to plain TCP. It is on by default only where twonet.mptcpsysctls say the kernel can escape a blackholed SYN by itself.