/**
* give control back to the scheduler after main() exits. This allows
* remaining threads to continue running.
* FIXME: we don't know whether user explicit calls exit() or main() normally returns
* in the previous case, we should exit immediately, while in the later, we should
* join other threads.
* Overriding exit() does not work because normal returning from
* main() also calls exit().
**/
static void exit_func(void)
{
// don't do anything if we're in a forked child process
if( getpid() != capriccio_main_pid )
return;
exit_func_done = 1;
main_exited = 1;
if( !exit_whole_program )
// this will block until all other threads finish
thread_exit(NULL);
// dump the blocking graph before we exit
if( conf_dump_blocking_graph ) {
tdebug("dumping blocking graph from exit_func()\n");
dump_blocking_graph();
}
// FIXME: make sure to kill cloned children
if( conf_dump_timing_info ) {
if( main_timer.running ) stop_timer(&main_timer);
if( scheduler_timer.running ) stop_timer(&scheduler_timer);
if( app_timer.running ) stop_timer(&app_timer);
print_timers();
}
}
int thread_join(thread_t *t, void **ret)
{
if (t == NULL)
return_errno(FALSE, EINVAL);
if ( !( t->joinable ) )
return_errno(FALSE, EINVAL);
assert(t->state != GHOST);
// A thread can be joined only once
if (t->join_thread)
return_errno(FALSE, EACCES);
t->join_thread = current_thread;
// Wait for the thread to complete
tdebug( "**** thread state: %d\n" ,t->state);
if (t->state != ZOMBIE) {
CAP_SET_SYSCALL();
thread_suspend_self(0);
CAP_CLEAR_SYSCALL();
}
// clean up the dead thread
if (ret != NULL)
*ret = t->ret;
free_thread( t );
return TRUE;
}
/**
* Wrapper function for new threads. This allows us to clean up
* correctly if a thread exits without calling thread_exit().
**/
static void* new_thread_wrapper(void *arg)
{
void *ret;
(void) arg;
// set up initial stats
current_thread->curr_stats.files = 0;
current_thread->curr_stats.sockets = 0;
current_thread->curr_stats.heap = 0;
bg_set_current_stats( ¤t_thread->curr_stats );
current_thread->prev_stats = current_thread->curr_stats;
// set up stack limit for new thread
stack_bottom = current_thread->stack_bottom;
stack_fingerprint = current_thread->stack_fingerprint;
// start the thread
tdebug("Initial arg = %p\n", current_thread->initial_arg);
ret = current_thread->initial_func(current_thread->initial_arg);
// call thread_exit() to do the cleanup
thread_exit(ret);
return NULL;
}
void thread_exit(void *ret)
{
thread_t *t = current_thread;
sanity_check_threadcounts();
tdebug("current=%s\n", current_thread?current_thread->name : "NULL");
if (current_thread == main_thread && main_exited == 0) {
// the case when the user calls thread_exit() in main thread is complicated
// we cannot simply terminate the main thread, because we need that stack to terminate the
// whole program normally. so we call exit() to make the c runtime help us get the stack
// context where we can just return to terminate the whole program
// this will call exit_func() and in turn call thread_exit() again
main_exited = 1;
exit (0);
}
// note the thread exit in the blocking graph
t->curr_stats.node = bg_exit_node;
current_thread->prev_stats.node->num_here--;
current_thread->curr_stats.node->num_here++;
if( bg_save_stats ) {
bg_update_stats();
}
// update thread counts
num_runnable_threads--;
if( t->daemon ) num_daemon_threads--;
t->state = ZOMBIE;
num_zombie_threads++;
// deallocate the TCB
// keep the thread, if the thread is Joinable, and we want the return value for something
if ( !( t->joinable ) ) {
// tell the scheduler thread to delete the current one
current_thread_exited = 1;
} else {
t->ret = ret;
if (t->join_thread)
thread_resume(t->join_thread);
}
sanity_check_threadcounts();
// squirrel away the stack limit--not that we'll need it again
current_thread->stack_bottom = stack_bottom;
current_thread->stack_fingerprint = stack_fingerprint;
// give control back to the scheduler
#ifdef NO_SCHEDULER_THREAD
do_scheduler(NULL);
#else
co_call(scheduler_thread->coro, NULL);
#endif
}
// only resume the thread internally
// don't touch the timeout flag and the sleep queue
static void _thread_resume(thread_t *t)
{
tdebug("t=%p\n",t);
if (t->state != SUSPENDED)
return;
num_suspended_threads--;
num_runnable_threads++;
sanity_check_threadcounts();
assert(t->state == SUSPENDED);
t->state = RUNNABLE;
assert( t->sleep == -1 );
sched_add_thread(t);
}
/**
* Entry point for the vcore. Basic job is to either resume the thread that
* was interrupted in the case of a notification coming in, or to find a new
* thread from the user level threading library and launch it.
**/
void __attribute__((noreturn)) vcore_entry()
{
/* Grab references to the current vcoreid vcore preemption data, and the
* vcoremap */
assert(in_vcore_context());
uint32_t vcoreid = vcore_id();
struct preempt_data *vcpd = &__procdata.vcore_preempt_data[vcoreid];
struct vcore *vc = &__procinfo.vcoremap[vcoreid];
tdebug("current=%s, vcore=%d\n",
current_thread?current_thread->name : "NULL", vcoreid);
/* Assert that notifications are disabled. Should always have notifications
* disabled when coming in here. */
assert(vcpd->notif_enabled == FALSE);
/* Put this in the loop that deals with notifications. It will return if
* there is no preempt pending. */
if (vc->preempt_pending)
sys_yield(TRUE);
/* When running vcore_entry(), we are using the TLS of the vcore, not any
* particular thread. If current_thread is set in the vcore's TLS, then
* that means the thread did not yield voluntarily, and was, instead,
* interrupted by a notification. We therefore need to restore the thread
* context from the notification trapframe, not the one stored in the
* thread struct itself. */
if (unlikely(current_thread)) {
vcpd->notif_pending = 0;
/* Do one last check for notifs after clearing pending */
// TODO: call the handle_notif() here (first)
/* Copy the notification trapframe into the current
* threads trapframe */
memcpy(¤t_thread->context->utf, &vcpd->notif_tf,
sizeof(struct user_trapframe));
/* Restore the context from the current_thread's trapframe */
restore_context(current_thread->context);
assert(0);
}
/* Otherwise either a vcore is coming up for the first time, or a thread
* has just yielded and vcore_entry() was called directly. In this case we
* need to figure out which thread to schedule next on the vcore */
run_next_thread();
assert(0);
}
static int tc_x11_demultiplex(TCModuleInstance *self,
vframe_list_t *vframe, aframe_list_t *aframe)
{
TCX11PrivateData *priv = NULL;
uint64_t now = 0;
int ret = 0;
TC_MODULE_SELF_CHECK(self, "demultiplex");
priv = self->userdata;
priv->reftime = tc_gettime();
tdebug(priv, "begin demultiplex");
if (aframe != NULL) {
aframe->audio_len = 0; /* no audio from here */
}
if (vframe != NULL) {
tdebug(priv, " begin acquire");
ret = tc_x11source_acquire(&priv->src, vframe->video_buf,
vframe->video_size);
tdebug(priv, " end acquire");
if (ret > 0) {
int64_t naptime = 0;
uint64_t now = 0;
vframe->attributes |= TC_FRAME_IS_KEYFRAME;
vframe->video_len = ret;
now = tc_gettime();
naptime = (priv->frame_delay - (now - priv->reftime));
if (priv->skew >= priv->skew_limit) {
tc_log_info(MOD_NAME, " skew correction (naptime was %lu)",
(unsigned long)naptime);
int64_t t = naptime;
naptime -= priv->skew;
priv->skew = TC_MAX(0, priv->skew - t);
}
if (naptime <= 0) {
/* don't sleep at all if delay is already excessive */
tc_log_info(MOD_NAME, "%-18s", " NO SLEEP!");
priv->expired++;
} else {
tc_log_info(MOD_NAME, "%-18s %lu", " sleep time",
(unsigned long)(naptime));
tc_timer_sleep(&priv->timer, (uint64_t)naptime);
}
}
}
now = tc_gettime();
now -= priv->reftime;
priv->skew += now - priv->frame_delay;
tdebug(priv, "end multiplex");
tc_log_info(MOD_NAME, "%-18s %li", "detected skew", (long)(priv->skew));
return (ret > 0) ?ret :-1;
}
void sockio_poll_poll(long long usecs)
{
int ret, i;
fdstruct_t *fds;
iorequest_t *req;
int timeout = (usecs == -1 ? -1 : usecs / 1000); // translate from microseconds to milliseconds
// get the lock
thread_latch( sockio_poll_latch );
assert(num_outstanding >= 0);
if( num_outstanding == 0 ) {
thread_unlatch( sockio_poll_latch );
return;
}
// NOTE: why a 100us timeout here? this is preventing all runnable threads from proceeding - zf
// removing it
//while( (ret=syscall(SYS_poll, ufds, num_outstanding, timeout)) < 0 && errno == EINTR ) ;
(void) timeout;
while( (ret=syscall(SYS_poll, ufds, num_outstanding, 0)) < 0 && errno == EINTR ) ;
if(ret < 0) {
perror("FATAL ERROR. poll() failed. ");
exit(1);
}
if(ret == 0) {
// release the lock
thread_unlatch( sockio_poll_latch );
return;
}
tdebug("%d fds are ready for IO\n", ret);
// process the returned events
for(i=num_outstanding-1; i>=0; i--) {
// skip past fds w/ no events
if( ufds[i].revents == 0 )
continue;
// find the fdstruct & get the first request
fds = get_fdstruct(ufds[i].fd);
req = view_first_waiter(fds);
// do the IO for as many waiters as possible
//
// FIXME: do this with just one syscall by using readv/writev!!!
// FIXME: it's VERY VERY BAD to hold sockio_poll_latch across these
// syscalls. To fix, sockio_poll_add_request needs to drop waiting
// requests into a list somewhere, so only the list needs to be locked
while( req != NULL ) {
do {
switch (req->type) {
case READ: ret = syscall(SYS_read, fds->fd, req->args.rw.buf, req->args.rw.count); break;
case WRITE: ret = syscall(SYS_write, fds->fd, req->args.rw.buf, req->args.rw.count); break;
case POLL1: ret = 1; req->args.poll1.ufds[0].revents = ufds[i].revents; break;
case CONNECT: ret=0; break; // system call already done in blocking_io.c
case ACCEPT: case SEND: case RECV:
ret = syscall(SYS_socketcall, req->args.scall.which, req->args.scall.argv); break;
default: assert(0);
}
} while(ret==-1 && errno==EINTR);
// the request would have blocked. Keep the fd in the poll set, and try again later
if( ret == -1 && (errno==EAGAIN || errno==EWOULDBLOCK))
req = NULL;
// there was some other error - return this error to all waiters
else if( ret == -1 ) {
while( (req=remove_first_waiter(fds)) != NULL ) {
req->ret = -1;
req->err = errno;
thread_resume( req->thread );
}
}
// the call succeeded
else {
remove_first_waiter(fds);
req->ret = ret;
req->err = 0;
thread_resume( req->thread );
// a read or write succeeded, but we didn't get the full count.
if( (req->type == READ || req->type == WRITE) && (size_t) ret < req->args.rw.count )
req = NULL;
// for everything else, we get the next request
else
req = view_first_waiter(fds);
}
}
// update the poll flags for the fd
req = view_first_waiter(fds);
if( req != NULL ) {
// add flags for the next poll() call
debug("more waiters for %d - will poll again", fds->fd);
switch( req->type ) {
case READ: case RECV: case ACCEPT: ufds[i].events = POLLIN|POLLPRI; break;
case WRITE: case SEND: case CONNECT: ufds[i].events = POLLOUT; break;
case POLL1: ufds[i].events = req->args.poll1.ufds[0].events; break;
default: assert(0);
}
} else {
// plug the hole in the request list
num_outstanding--;
ufds[i] = ufds[num_outstanding];
}
}
// release the lock
thread_unlatch( sockio_poll_latch );
}
/**
* perform necessary management to yield the current thread
* if suspended == TRUE && timeout != 0 -> the thread is added
* to the sleep queue and later waken up when the clock times out
* returns FALSE if time-out actually happens, TRUE if waken up
* by other threads, INTERRUPTED if interrupted by a signal
**/
static int thread_yield_internal(int suspended, unsigned long long timeout)
{
// now we use a per-thread errno stored in thread_t
int savederrno;
int rv = OK;
tdebug("current_thread=%p\n",current_thread);
savederrno = errno;
// decide what to do with the thread
if( !suspended ) // just add it to the runlist
sched_add_thread( current_thread );
else if( timeout ) // add to the sleep list
sleepq_add_thread( current_thread, timeout);
{
#ifdef SHOW_EDGE_TIMES
cpu_tick_t start, end, rstart, rend;
GET_CPU_TICKS(start);
GET_REAL_CPU_TICKS(rstart);
#endif
// figure out the current node in the graph
if( !conf_no_stacktrace )
bg_backtrace_set_node();
// FIXME: fake out what cil would do... current_thread->curr_stats.node = bg_dummy_node;
// we should already have been told the node by CIL or directly by the programmer
assert( current_thread->curr_stats.node != NULL );
// update node counts
current_thread->prev_stats.node->num_here--;
current_thread->curr_stats.node->num_here++;
// update the blocking graph info
if( bg_save_stats )
bg_update_stats();
#ifdef SHOW_EDGE_TIMES
GET_CPU_TICKS(end);
GET_REAL_CPU_TICKS(rend);
{
thread_stats_t *curr = ¤t_thread->curr_stats;
thread_stats_t *prev = ¤t_thread->prev_stats;
output(" %3d -> %-3d %7lld ticks (%lld ms) %7lld rticks (%lld ms) ",
prev->node->node_num, curr->node->node_num,
curr->cpu_ticks - prev->cpu_ticks,
(curr->cpu_ticks - prev->cpu_ticks) / ticks_per_millisecond,
# ifdef USE_PERFCTR
curr->real_ticks - prev->real_ticks,
(curr->real_ticks - prev->real_ticks) / ticks_per_millisecond
# else
curr->cpu_ticks - prev->cpu_ticks,
(curr->cpu_ticks - prev->cpu_ticks) / ticks_per_millisecond
# endif
);
output("update bg node %d: %lld (%lld ms) real: %lld (%lld ms)\n",
current_thread->curr_stats.node->node_num,
(end-start), (end-start)/ticks_per_millisecond,
(rend-rstart), (rend-rstart)/ticks_per_millisecond);
}
#endif
}
// squirrel away the stack limit for next time
current_thread->stack_bottom = stack_bottom;
current_thread->stack_fingerprint = stack_fingerprint;
// switch to the scheduler thread
#ifdef NO_SCHEDULER_THREAD
do_scheduler(NULL);
#else
co_call(scheduler_thread->coro, NULL);
#endif
// set up stack limit for new thread
stack_bottom = current_thread->stack_bottom;
stack_fingerprint = current_thread->stack_fingerprint;
// rotate the stats
if( bg_save_stats ) {
current_thread->prev_stats = current_thread->curr_stats;
// update thread time, to skip time asleep
GET_CPU_TICKS( current_thread->prev_stats.cpu_ticks );
current_thread->prev_stats.cpu_ticks -= ticks_diff; // FIXME: subtract out time to do debug output
#ifdef USE_PERFCTR
GET_REAL_CPU_TICKS( current_thread->prev_stats.real_ticks );
current_thread->prev_stats.real_ticks -= ticks_rdiff; // FIXME: subtract out time to do debug output
#endif
} else {
current_thread->prev_stats.node = current_thread->curr_stats.node;
}
// check whether time-out happens
if (suspended && timeout && current_thread->timeout) {
rv = TIMEDOUT;
current_thread->timeout = 0;
}
// check for and process pending signals
if ( likely(!current_thread->sig_waiting) ) {
if (sig_process_pending())
rv = INTERRUPTED;
} else {
// if sig_waiting is 1, sigwait() itself will handle the remaining
rv = INTERRUPTED;
}
errno = savederrno;
return rv;
}
/**
* Main scheduling loop
**/
static void* do_scheduler(void *arg)
{
static cpu_tick_t next_poll=0, next_overload_check=0, next_info_dump=0, next_graph_stats=0, now=0;
static int pollcount=1000;
static int init_done = 0;
(void) arg; // suppress GCC "unused parameter" warning
in_scheduler = 1;
// make sure we start out by saving edge stats for a while
if( !init_done ) {
init_done = 1;
if (conf_no_statcollect)
bg_save_stats = 0;
else
bg_save_stats = 1;
GET_REAL_CPU_TICKS( now );
next_graph_stats = now + 1 * ticks_per_second;
start_timer(&scheduler_timer);
}
while( 1 ) {
//current_thread = scheduler_thread;
sanity_check_threadcounts();
sanity_check_io_stats();
// wake up threads that have timeouts
sleepq_check(0);
sanity_check_threadcounts();
// break out if there are only daemon threads
if(unlikely (num_suspended_threads == 0 && num_runnable_threads == num_daemon_threads)) {
// dump the blocking graph
if( exit_func_done && conf_dump_blocking_graph ) {
tdebug("dumping blocking graph from do_scheduler()\n");
dump_blocking_graph();
}
// go back to mainthread, which should now be in exit_func()
current_thread = main_thread;
in_scheduler = 0;
co_call(main_thread->coro, NULL);
in_scheduler = 1;
if( unlikely(current_thread_exited) ) { // free memory from deleted threads
current_thread_exited=0;
if (current_thread != main_thread) // main_thread is needed for whole program exit
free_thread( current_thread );
}
return NULL;
}
// cheesy way of handling things with timing requirements
{
GET_REAL_CPU_TICKS( now );
// toggle stats collection
if( conf_no_statcollect == 0 && next_graph_stats < now ) {
bg_save_stats = 1 - bg_save_stats;
if( bg_save_stats ) {
// record stats for 100 ms
next_graph_stats = now + 100 * ticks_per_millisecond;
// update the stats epoch, to allow proper handling of the first data items
bg_stats_epoch++;
}
else {
// avoid stats for 2000 ms
next_graph_stats = now + 2000 * ticks_per_millisecond;
}
//output(" *********************** graph stats %s\n", bg_save_stats ? "ON" : "OFF" );
}
// resource utalization
if( unlikely (next_overload_check < now) ) {
check_overload( now );
next_overload_check = now + OVERLOAD_CHECK_INTERVAL;
}
// poll
if( likely( (int)io_polling_func) ) {
if( num_runnable_threads==0 || --pollcount <= 0 || next_poll < now ) {
//if( num_runnable_threads==0 ) {
// poll
long long timeout = 0;
if( num_runnable_threads==0 ) {
if (first_wake_usecs == 0) {
timeout = -1;
} else {
// there are threads in the sleep queue
// so poll for i/o till at most that time
unsigned long long now;
now = current_usecs();
tdebug ("first_wake: %lld, now: %lld\n", first_wake_usecs, now);
if (first_wake_usecs > now)
timeout = first_wake_usecs - now;
}
}
stop_timer(&scheduler_timer);
//if( timeout != -1 ) output("timeout is not zero\n");
io_polling_func( timeout ); // allow blocking
start_timer(&scheduler_timer);
sanity_check_threadcounts();
#ifndef USE_NIO
// sleep for a bit, if there was nothing to do
// FIXME: let the IO functions block instead??
if( num_runnable_threads == 0 ) {
syscall(SYS_sched_yield);
}
#endif
// vary the poll rate depending on the workload
#if 0
if( num_runnable_threads < 5 ) {
next_poll = now + (10*ticks_per_millisecond);
pollcount = 1000;
} else if( num_runnable_threads < 10 ) {
next_poll = now + (50*ticks_per_millisecond);
pollcount = 2000;
} else {
next_poll = now + (100*ticks_per_millisecond);
pollcount = 3000;
}
#else
next_poll = now + (ticks_per_millisecond << 13);
pollcount = 10000;
#endif
}
}
// debug stats
if( 0 && next_info_dump < now ) {
dump_debug_info();
next_info_dump = now + 5 * ticks_per_second;
}
}
// get the head of the run list
current_thread = sched_next_thread();
// scheduler gave an invlid even though there are runnable
// threads. This indicates that every runnable thead is likely to
// require use of an overloaded resource.
if( !valid_thread(current_thread) ) {
pollcount = 0;
continue;
}
// barf, if the returned thread is still on the sleep queue
assert( current_thread->sleep == -1 );
tdebug("running TID %d (%s)\n", current_thread->tid, current_thread->name ? current_thread->name : "no name");
sanity_check_threadcounts();
// call thread
stop_timer(&scheduler_timer);
start_timer(&app_timer);
in_scheduler = 0;
co_call(current_thread->coro, NULL);
in_scheduler = 1;
stop_timer(&app_timer);
start_timer(&scheduler_timer);
if( unlikely(current_thread_exited) ) { // free memory from deleted threads
current_thread_exited=0;
if (current_thread != main_thread) // main_thread is needed for whole program exit
free_thread( current_thread );
}
#ifdef NO_SCHEDULER_THREAD
return NULL;
#endif
}
return NULL;
}