When the stem process terminates and SIGCHLD is sent to the supervisor,
the signal might be handled by the main thread or any other threads that
aren't blocking SIGCHLD explicitly. Convert last_signal to a std::atomic<int>
such that non-main threads can safely set last_signal without triggering
data race as reported by TSAN. This doesn't make it less racy to work
last_signal, but it appears we only use it for debug printing anyhow and
another option might have been to just remove last_signal altogether.
Follow-up for #4849
* origin/topic/timw/merge-3rdparty-repo-into-main: (30 commits)
Ignore src/3rdparty for pre-commit
src/3rdparty: Port doctest fix for including <ciso646> from upstream
src/3rdparty: Update doctest to v2.4.12
src/3rdparty: Move jthread/stop_token out of std namespace to prevent collisions
src/3rdparty: Fix clang-tidy bugprone-casting-through-void warning
src/3rdparty: Add jthread and stop_token headers
src/3rdparty: Update SQLite to 3.47.1
src/3rdparty: Upgrade sqlite3 to 3.45.0
src/3rdparty: modp: Disable deprecation warning for sprintf
src/3rdparty: Update SQLite to 3.41.2
src/3rdparty: Handle zeek-inet-ntop snprintf() return value correctly
src/3rdparty: Rework inet-ntop change to build on Windows
src/3rdparty: Use snprintf in zeek_inet_ntop.c to silence macOS compiler warnings
src/3rdparty: Use snprintf in patricia.cc to silence macOS compiler warnings
src/3rdparty: Update ConvertUTF Unicode license as per request from LLVM upstream
src/3rdparty: Update doctest to 2.4.9 and sqlite to 3.39.4
src/3rdparty: Fix rampant off-by-one error in last change
src/3rdparty: numeric conversion functions now return the number of characters added
src/3rdparty: Fix GCC 12.2 warning in access to patricia_t members
src/3rdparty: Rebase patricia.{h,c} on upstream version
...
This has been failing on our weekly builds for a bit. It's fixed upstream
but there's no idea when they'll release a new version that includes it.
This cherry-picks 5eb20178ba
Access to members of the larger prefix_t, while safe, throws warnings
when the compiler knows it previously allocated the smaller prefix4_t.
Adapted from 11ae12079ca716f8cfac42cccfd905dad45da18f in the upstream repo.
If Cluster::init() hasn't been invoked yet, Cluster::subscribe() with the
ZeroMQ backend would block because the main_inproc socket didn't
yet have a connection from the child thread. Prevent this by connecting
the main and child socket pair at construction time.
This will queue the subscriptions and start processing them once the
child thread has started.
An URI containing a bracketed or non-bracketed IPv6 address of the form
http://[::1]:42 was previously split on the first colon for port extraction,
causing a subsequent to_count() call to fail. Harden this to check for a
digits in the last :[0-9]+ component.
Fixes#4842
