static void
clock_sleep_ns_abs (guint64 ns_abs)
{
#ifdef HAVE_CLOCK_NANOSLEEP
int ret;
struct timespec then;
then.tv_sec = ns_abs / 1000000000;
then.tv_nsec = ns_abs % 1000000000;
do {
ret = clock_nanosleep (sampling_posix_clock, TIMER_ABSTIME, &then, NULL);
if (ret != 0 && ret != EINTR)
g_error ("%s: clock_nanosleep () returned %d", __func__, ret);
} while (ret == EINTR && mono_atomic_load_i32 (&sampling_thread_running));
#else
int ret;
gint64 diff;
struct timespec req;
/*
* What follows is a crude attempt at emulating clock_nanosleep () on OSs
* which don't provide it (e.g. FreeBSD).
*
* The problem with nanosleep () is that if it is interrupted by a signal,
* time will drift as a result of having to restart the call after the
* signal handler has finished. For this reason, we avoid using the rem
* argument of nanosleep (). Instead, before every nanosleep () call, we
* check if enough time has passed to satisfy the sleep request. If yes, we
* simply return. If not, we calculate the difference and do another sleep.
*
* This should reduce the amount of drift that happens because we account
* for the time spent executing the signal handler, which nanosleep () is
* not guaranteed to do for the rem argument.
*
* The downside to this approach is that it is slightly expensive: We have
* to make an extra system call to retrieve the current time whenever we're
* going to restart a nanosleep () call. This is unlikely to be a problem
* in practice since the sampling thread won't be receiving many signals in
* the first place (it's a tools thread, so no STW), and because typical
* sleep periods for the thread are many orders of magnitude bigger than
* the time it takes to actually perform that system call (just a few
* nanoseconds).
*/
do {
diff = (gint64) ns_abs - (gint64) clock_get_time_ns ();
if (diff <= 0)
break;
req.tv_sec = diff / 1000000000;
req.tv_nsec = diff % 1000000000;
if ((ret = nanosleep (&req, NULL)) == -1 && errno != EINTR)
g_error ("%s: nanosleep () returned -1, errno = %d", __func__, errno);
} while (ret == -1 && mono_atomic_load_i32 (&sampling_thread_running));
#endif
}
static inline void
request_interrupt (gpointer thread_info, HANDLE native_thread_handle, gint32 pending_apc_slot, PAPCFUNC apc_callback, DWORD tid)
{
/*
* On Windows platforms, an async interrupt/abort request queues an APC
* that needs to be processed by target thread before it can return from an
* alertable OS wait call and complete the mono interrupt/abort request.
* Uncontrolled queuing of APC's could flood the APC queue preventing the target thread
* to return from its alertable OS wait call, blocking the interrupt/abort requests to complete
* This check makes sure that only one APC per type gets queued, preventing potential flooding
* of the APC queue. NOTE, this code will execute regardless if targeted thread is currently in
* an alertable wait or not. This is done to prevent races between interrupt/abort requests and
* alertable wait calls. Threads already in an alertable wait should handle WAIT_IO_COMPLETION
* return scenarios and restart the alertable wait operation if needed or take other actions
* (like service the interrupt/abort request).
*/
MonoThreadInfo *info = (MonoThreadInfo *)thread_info;
gint32 old_wait_info, new_wait_info;
do {
old_wait_info = mono_atomic_load_i32 (&info->thread_wait_info);
if (old_wait_info & pending_apc_slot)
return;
new_wait_info = old_wait_info | pending_apc_slot;
} while (mono_atomic_cas_i32 (&info->thread_wait_info, new_wait_info, old_wait_info) != old_wait_info);
THREADS_INTERRUPT_DEBUG ("%06d - Interrupting/Aborting syscall in thread %06d", GetCurrentThreadId (), tid);
QueueUserAPC (apc_callback, native_thread_handle, (ULONG_PTR)NULL);
}
static void
clock_sleep_ns_abs (guint64 ns_abs)
{
kern_return_t ret;
mach_timespec_t then, remain_unused;
then.tv_sec = ns_abs / 1000000000;
then.tv_nsec = ns_abs % 1000000000;
do {
ret = clock_sleep (sampling_clock_service, TIME_ABSOLUTE, then, &remain_unused);
if (ret != KERN_SUCCESS && ret != KERN_ABORTED)
g_error ("%s: clock_sleep () returned %d", __func__, ret);
} while (ret == KERN_ABORTED && mono_atomic_load_i32 (&sampling_thread_running));
}
void
mono_runtime_shutdown_stat_profiler (void)
{
mono_atomic_store_i32 (&sampling_thread_running, 0);
mono_profiler_sampling_thread_post ();
#ifndef HOST_DARWIN
/*
* There is a slight problem when we're using CLOCK_PROCESS_CPUTIME_ID: If
* we're shutting down and there's largely no activity in the process other
* than waiting for the sampler thread to shut down, it can take upwards of
* 20 seconds (depending on a lot of factors) for us to shut down because
* the sleep progresses very slowly as a result of the low CPU activity.
*
* We fix this by repeatedly sending the profiler signal to the sampler
* thread in order to interrupt the sleep. clock_sleep_ns_abs () will check
* sampling_thread_running upon an interrupt and return immediately if it's
* zero. profiler_signal_handler () has a special case to ignore the signal
* for the sampler thread.
*/
MonoThreadInfo *info;
// Did it shut down already?
if ((info = mono_thread_info_lookup (sampling_thread))) {
while (!mono_atomic_load_i32 (&sampling_thread_exiting)) {
mono_threads_pthread_kill (info, profiler_signal);
mono_thread_info_usleep (10 * 1000 /* 10ms */);
}
// Make sure info can be freed.
mono_hazard_pointer_clear (mono_hazard_pointer_get (), 1);
}
#endif
mono_os_event_wait_one (&sampling_thread_exited, MONO_INFINITE_WAIT, FALSE);
mono_os_event_destroy (&sampling_thread_exited);
/*
* We can't safely remove the signal handler because we have no guarantee
* that all pending signals have been delivered at this point. This should
* not really be a problem anyway.
*/
//remove_signal_handler (profiler_signal);
}
static gsize
sampling_thread_func (gpointer unused)
{
MonoInternalThread *thread = mono_thread_internal_current ();
thread->flags |= MONO_THREAD_FLAG_DONT_MANAGE;
ERROR_DECL (error);
MonoString *name = mono_string_new_checked (mono_get_root_domain (), "Profiler Sampler", error);
mono_error_assert_ok (error);
mono_thread_set_name_internal (thread, name, FALSE, FALSE, error);
mono_error_assert_ok (error);
mono_thread_info_set_flags (MONO_THREAD_INFO_FLAGS_NO_GC | MONO_THREAD_INFO_FLAGS_NO_SAMPLE);
int old_policy;
struct sched_param old_sched;
pthread_getschedparam (pthread_self (), &old_policy, &old_sched);
/*
* Attempt to switch the thread to real time scheduling. This will not
* necessarily work on all OSs; for example, most Linux systems will give
* us EPERM here unless configured to allow this.
*
* TODO: This does not work on Mac (and maybe some other OSs). On Mac, we
* have to use the Mach thread policy routines to switch to real-time
* scheduling. This is quite tricky as we need to specify how often we'll
* be doing work (easy), the normal processing time needed (also easy),
* and the maximum amount of processing time needed (hard). This is
* further complicated by the fact that if we misbehave and take too long
* to do our work, the kernel may knock us back down to the normal thread
* scheduling policy without telling us.
*/
struct sched_param sched = { .sched_priority = sched_get_priority_max (SCHED_FIFO) };
pthread_setschedparam (pthread_self (), SCHED_FIFO, &sched);
MonoProfilerSampleMode mode;
init:
mono_profiler_get_sample_mode (NULL, &mode, NULL);
if (mode == MONO_PROFILER_SAMPLE_MODE_NONE) {
mono_profiler_sampling_thread_wait ();
if (!mono_atomic_load_i32 (&sampling_thread_running))
goto done;
goto init;
}
clock_init (mode);
for (guint64 sleep = clock_get_time_ns (); mono_atomic_load_i32 (&sampling_thread_running); clock_sleep_ns_abs (sleep)) {
uint32_t freq;
MonoProfilerSampleMode new_mode;
mono_profiler_get_sample_mode (NULL, &new_mode, &freq);
if (new_mode != mode) {
clock_cleanup ();
goto init;
}
sleep += 1000000000 / freq;
FOREACH_THREAD_SAFE_EXCLUDE (info, MONO_THREAD_INFO_FLAGS_NO_SAMPLE) {
g_assert (mono_thread_info_get_tid (info) != sampling_thread);
/*
* Require an ack for the last sampling signal sent to the thread
* so that we don't overflow the signal queue, leading to all sorts
* of problems (e.g. GC STW failing).
*/
if (profiler_signal != SIGPROF && !mono_atomic_cas_i32 (&info->profiler_signal_ack, 0, 1))
continue;
mono_threads_pthread_kill (info, profiler_signal);
mono_atomic_inc_i32 (&profiler_signals_sent);
} FOREACH_THREAD_SAFE_END
}
