On BSDs, anonymous semaphores are implemented using a file descriptor
which is not marked CLOEXEC, so it gets leaked into child processes.
Use ordinary pipes instead of semaphores everywhere except Linux.
Fixes#7304
With the prior commit, the topic_monitor only writes to the pipe if a
thread is known to be waiting. This is effectively a binary semaphore, and
on systems that support anon semaphores (yes Linux, but not Mac) we can use
them. These are more efficient than self-pipes.
We add a binary_semaphore_t class which uses sem_t if sem_init succeeds,
and a self-pipe if it fails.
On Linux the seq_echo benchmark (run 1024 times) goes from 12.40 seconds to
11.59 seconds, about an 11% improvement.
The topic monitor is what allows a thread to wait for any of a set of
events. Events are identified by a bit in a "pending update" mask. Prior to
this fix, post() would atomically set the bit, and if it was newly set,
announce the change by unconditionally writing to a self-pipe. Threads
could wait for new posts by reading from the pipe.
This is less efficient than it could be; in particular if no thread is
waiting on the pipe, then the write() is unnecessary. This slows down our
signal handler.
Change the design in the following way: if a thread is committed to
waiting, then it atomically sets the "pending update" mask (now just called
status) to a sentinel value STATUS_NEEDS_WAKEUP. Then post() will only
write to the self-pipe if it sees that there is a thread waiting. This
reduces the number of syscalls.
The total effect is hardly noticeable (usually there is a thread waiting)
but it will be important for the next commit.
The topic monitor allows a client to wait for multiple events, e.g. sigchld
or an internal process exit. Prior to this change a client had to specify
the list of generations and the list of topics they are interested in.
Simplify this to just the list of generations, with a max-value generation
meaning the topic is not interesting.
Also remove the use of enum_set and enum_array, it was too complex for what
it offered.
As of GCC 7.4 (at least under macOS 10.10), the previous workaround of
casting a must-use result to `(void)` to avoid warnings about unused
code no longer works.
This workaround is uglier but it quiets these warnings.
This fixes a race condition in the topic monitor. A thread may decide to
enter the wait queue, but before it does the generation list changes, and
so our thread will wait forever, resulting in a hang.
It also simplifies the implementation of the topic monitor considerably;
on reflection the whole "metagen" thing isn't providing any value and we
should just compare generations directly.
In the new design, we have a lock-protected list of current generations,
along with a boolean as to whether someone is reading from the pipe. The
reader (only one at a time) is responsible for broadcasting notifications
via a condition variable.
tsan does funny things to signals, preventing signals from being delivered
in a blocking read. Switch the topic monitor to non-blocking reads under
tsan.
topic_monitor allows for querying changes posted to one or more topics,
initially sigchld. This will eventually replace the waitpid logic in
process_mark_finished_children().
Comment from the new header:
Topic monitoring support. Topics are conceptually "a thing that can
happen." For example, delivery of a SIGINT, a child process exits, etc. It
is possible to post to a topic, which means that that thing happened.
Associated with each topic is a current generation, which is a 64 bit
value. When you query a topic, you get back a generation. If on the next
query the generation has increased, then it indicates someone posted to
the topic.
For example, if you are monitoring a child process, you can query the
sigchld topic. If it has increased since your last query, it is possible
that your child process has exited.
Topic postings may be coalesced. That is there may be two posts to a given
topic, yet the generation only increases by 1. The only guarantee is that
after a topic post, the current generation value is larger than any value
previously queried.
Tying this all together is the topic_monitor_t. This provides the current
topic generations, and also provides the ability to perform a blocking
wait for any topic to change in a particular topic set. This is the real
power of topics: you can wait for a sigchld signal OR a thread exit.
