thread_pthread: remove timer-thread by restructuring GVL
To reduce resource use and reduce CI failure; remove
timer-thread. Single-threaded Ruby processes (including forked
children) will never see extra thread overhead. This prevents
glibc and jemalloc from going into multi-threaded mode and
initializing locks or causing fragmentation via arena explosion.
The GVL is implements its own wait-queue as a ccan/list to
permit controlling wakeup order. Timeslice under contention is
handled by a designated timer thread (similar to choosing a
"patrol_thread" for current deadlock checking).
There is only one self-pipe, now, as wakeups for timeslice are
done independently using condition variables. This reduces FD
pressure slightly.
Signal handling is handled directly by a Ruby Thread (instead
of timer-thread) by exposing signal self-pipe to callers of
rb_thread_fd_select, native_sleep, rb_wait_for_single_fd, etc...
Acquiring, using, and releasing the self-pipe is exposed via 4
new internal functions:
rb_sigwait_fd_sleep - sleep and wait for signal (and no other FDs)
rb_sigwait_fd_put - release acquired result from rb_sigwait_fd_get
rb_sigwait_fd_migrate - migrate signal handling to another thread
after calling rb_sigwait_fd_put.
rb_sigwait_fd_migrate is necessary for waitpid callers because
only one thread can wait on self-pipe at a time, otherwise a
deadlock will occur if threads fight over the self-pipe.
TRAP_INTERRUPT_MASK is now set for the main thread directly in
signal handler via rb_thread_wakeup_timer_thread.
Originally, I wanted to use POSIX timers
(timer_create/timer_settime) for this. Unfortunately, this
proved unfeasible as Mutex#sleep resumes on spurious wakeups and
test/thread/test_cv.rb::test_condvar_timed_wait failed. Using
pthread_sigmask to mask out SIGVTALRM fixed that test, but
test/fiddle/test_function.rb::test_nogvl_poll proved there'd be
some unavoidable (and frequent) incompatibilities from that
approach.
Finally, this allows us to drop thread_destruct_lock and
interrupt current ec directly.
We don't need to rely on vm->thread_destruct_lock or a coherent
vm->running_thread on any platform. Separate timer-thread for
time slice and signal handling is relegated to thread_win32.c,
now.
thread_pthread: remove timer-thread by restructuring GVL
To reduce resource use and reduce CI failure; remove
timer-thread. Single-threaded Ruby processes (including forked
children) will never see extra thread overhead. This prevents
glibc and jemalloc from going into multi-threaded mode and
initializing locks or causing fragmentation via arena explosion.
The GVL is implements its own wait-queue as a ccan/list to
permit controlling wakeup order. Timeslice under contention is
handled by a designated timer thread (similar to choosing a
"patrol_thread" for current deadlock checking).
There is only one self-pipe, now, as wakeups for timeslice are
done independently using condition variables. This reduces FD
pressure slightly.
Signal handling is handled directly by a Ruby Thread (instead
of timer-thread) by exposing signal self-pipe to callers of
rb_thread_fd_select, native_sleep, rb_wait_for_single_fd, etc...
Acquiring, using, and releasing the self-pipe is exposed via 4
new internal functions:
rb_sigwait_fd_get - exclusively acquire timer_thread_pipe.normal[0]
rb_sigwait_fd_sleep - sleep and wait for signal (and no other FDs)
rb_sigwait_fd_put - release acquired result from rb_sigwait_fd_get
rb_sigwait_fd_migrate - migrate signal handling to another thread
after calling rb_sigwait_fd_put.
rb_sigwait_fd_migrate is necessary for waitpid callers because
only one thread can wait on self-pipe at a time, otherwise a
deadlock will occur if threads fight over the self-pipe.
TRAP_INTERRUPT_MASK is now set for the main thread directly in
signal handler via rb_thread_wakeup_timer_thread.
Originally, I wanted to use POSIX timers
(timer_create/timer_settime) for this. Unfortunately, this
proved unfeasible as Mutex#sleep resumes on spurious wakeups and
test/thread/test_cv.rb::test_condvar_timed_wait failed. Using
pthread_sigmask to mask out SIGVTALRM fixed that test, but
test/fiddle/test_function.rb::test_nogvl_poll proved there'd be
some unavoidable (and frequent) incompatibilities from that
approach.
Finally, this allows us to drop thread_destruct_lock and
interrupt current ec directly.
We don't need to rely on vm->thread_destruct_lock or a coherent
vm->running_thread on any platform. Separate timer-thread for
time slice and signal handling is relegated to thread_win32.c,
now.
[ruby-core:88088] [Misc #14937]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64107 b2dd03c8-39d4-4d8f-98ff-823fe69b080e