static void start_next_write(pj_ioqueue_key_t *key)
{
if (key_has_pending_write(key)) {
PjUwpSocket *s = (PjUwpSocket*)key->fd;
struct write_operation *op;
op = (struct write_operation*)key->write_list.next;
if (op->op == PJ_IOQUEUE_OP_SEND)
s->Send(op->buf, (pj_ssize_t*)&op->size);
else
s->SendTo(op->buf, (pj_ssize_t*)&op->size, &op->rmt_addr);
}
}
/*
* pj_ioqueue_poll()
*
*/
PJ_DEF(int) pj_ioqueue_poll( pj_ioqueue_t *ioqueue, const pj_time_val *timeout)
{
int i, count, processed;
int msec;
//struct epoll_event *events = ioqueue->events;
//struct queue *queue = ioqueue->queue;
struct epoll_event events[PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL];
struct queue queue[PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL];
pj_timestamp t1, t2;
PJ_CHECK_STACK();
msec = timeout ? PJ_TIME_VAL_MSEC(*timeout) : 9000;
TRACE_((THIS_FILE, "start os_epoll_wait, msec=%d", msec));
pj_get_timestamp(&t1);
//count = os_epoll_wait( ioqueue->epfd, events, ioqueue->max, msec);
count = os_epoll_wait( ioqueue->epfd, events, PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL, msec);
if (count == 0) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
/* Check the closing keys only when there's no activity and when there are
* pending closing keys.
*/
if (count == 0 && !pj_list_empty(&ioqueue->closing_list)) {
pj_lock_acquire(ioqueue->lock);
scan_closing_keys(ioqueue);
pj_lock_release(ioqueue->lock);
}
#endif
TRACE_((THIS_FILE, "os_epoll_wait timed out"));
return count;
}
else if (count < 0) {
TRACE_((THIS_FILE, "os_epoll_wait error"));
return -pj_get_netos_error();
}
pj_get_timestamp(&t2);
TRACE_((THIS_FILE, "os_epoll_wait returns %d, time=%d usec",
count, pj_elapsed_usec(&t1, &t2)));
/* Lock ioqueue. */
pj_lock_acquire(ioqueue->lock);
for (processed=0, i=0; i<count; ++i) {
pj_ioqueue_key_t *h = (pj_ioqueue_key_t*)(epoll_data_type)
events[i].epoll_data;
TRACE_((THIS_FILE, "event %d: events=%d", i, events[i].events));
/*
* Check readability.
*/
if ((events[i].events & EPOLLIN) &&
(key_has_pending_read(h) || key_has_pending_accept(h)) && !IS_CLOSING(h) ) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
queue[processed].key = h;
queue[processed].event_type = READABLE_EVENT;
++processed;
continue;
}
/*
* Check for writeability.
*/
if ((events[i].events & EPOLLOUT) && key_has_pending_write(h) && !IS_CLOSING(h)) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
queue[processed].key = h;
queue[processed].event_type = WRITEABLE_EVENT;
++processed;
continue;
}
#if PJ_HAS_TCP
/*
* Check for completion of connect() operation.
*/
if ((events[i].events & EPOLLOUT) && (h->connecting) && !IS_CLOSING(h)) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
queue[processed].key = h;
queue[processed].event_type = WRITEABLE_EVENT;
++processed;
continue;
}
#endif /* PJ_HAS_TCP */
/*
* Check for error condition.
*/
if ((events[i].events & EPOLLERR) && !IS_CLOSING(h)) {
/*
* We need to handle this exception event. If it's related to us
* connecting, report it as such. If not, just report it as a
* read event and the higher layers will handle it.
*/
if (h->connecting) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
queue[processed].key = h;
queue[processed].event_type = EXCEPTION_EVENT;
++processed;
} else if (key_has_pending_read(h) || key_has_pending_accept(h)) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
queue[processed].key = h;
queue[processed].event_type = READABLE_EVENT;
++processed;
}
continue;
}
}
for (i=0; i<processed; ++i) {
if (queue[i].key->grp_lock)
pj_grp_lock_add_ref_dbg(queue[i].key->grp_lock, "ioqueue", 0);
}
PJ_RACE_ME(5);
pj_lock_release(ioqueue->lock);
PJ_RACE_ME(5);
/* Now process the events. */
for (i=0; i<processed; ++i) {
switch (queue[i].event_type) {
case READABLE_EVENT:
ioqueue_dispatch_read_event(ioqueue, queue[i].key);
break;
case WRITEABLE_EVENT:
ioqueue_dispatch_write_event(ioqueue, queue[i].key);
break;
case EXCEPTION_EVENT:
ioqueue_dispatch_exception_event(ioqueue, queue[i].key);
break;
case NO_EVENT:
pj_assert(!"Invalid event!");
break;
}
#if PJ_IOQUEUE_HAS_SAFE_UNREG
decrement_counter(queue[i].key);
#endif
if (queue[i].key->grp_lock)
pj_grp_lock_dec_ref_dbg(queue[i].key->grp_lock,
"ioqueue", 0);
}
/* Special case:
* When epoll returns > 0 but no descriptors are actually set!
*/
if (count > 0 && !processed && msec > 0) {
pj_thread_sleep(msec);
}
pj_get_timestamp(&t1);
TRACE_((THIS_FILE, "ioqueue_poll() returns %d, time=%d usec",
processed, pj_elapsed_usec(&t2, &t1)));
return processed;
}
/*
* pj_ioqueue_poll()
*
* Few things worth written:
*
* - we used to do only one callback called per poll, but it didn't go
* very well. The reason is because on some situation, the write
* callback gets called all the time, thus doesn't give the read
* callback to get called. This happens, for example, when user
* submit write operation inside the write callback.
* As the result, we changed the behaviour so that now multiple
* callbacks are called in a single poll. It should be fast too,
* just that we need to be carefull with the ioqueue data structs.
*
* - to guarantee preemptiveness etc, the poll function must strictly
* work on fd_set copy of the ioqueue (not the original one).
*/
PJ_DEF(int) pj_ioqueue_poll( pj_ioqueue_t *ioqueue, const pj_time_val *timeout)
{
pj_fd_set_t rfdset, wfdset, xfdset;
int count, counter;
pj_ioqueue_key_t *h;
struct event
{
pj_ioqueue_key_t *key;
enum ioqueue_event_type event_type;
} event[PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL];
PJ_ASSERT_RETURN(ioqueue, -PJ_EINVAL);
/* Lock ioqueue before making fd_set copies */
pj_lock_acquire(ioqueue->lock);
/* We will only do select() when there are sockets to be polled.
* Otherwise select() will return error.
*/
if (PJ_FD_COUNT(&ioqueue->rfdset)==0 &&
PJ_FD_COUNT(&ioqueue->wfdset)==0
#if defined(PJ_HAS_TCP) && PJ_HAS_TCP!=0
&& PJ_FD_COUNT(&ioqueue->xfdset)==0
#endif
)
{
#if PJ_IOQUEUE_HAS_SAFE_UNREG
scan_closing_keys(ioqueue);
#endif
pj_lock_release(ioqueue->lock);
TRACE__((THIS_FILE, " poll: no fd is set"));
if (timeout)
pj_thread_sleep(PJ_TIME_VAL_MSEC(*timeout));
return 0;
}
/* Copy ioqueue's pj_fd_set_t to local variables. */
pj_memcpy(&rfdset, &ioqueue->rfdset, sizeof(pj_fd_set_t));
pj_memcpy(&wfdset, &ioqueue->wfdset, sizeof(pj_fd_set_t));
#if PJ_HAS_TCP
pj_memcpy(&xfdset, &ioqueue->xfdset, sizeof(pj_fd_set_t));
#else
PJ_FD_ZERO(&xfdset);
#endif
#if VALIDATE_FD_SET
validate_sets(ioqueue, &rfdset, &wfdset, &xfdset);
#endif
/* Unlock ioqueue before select(). */
pj_lock_release(ioqueue->lock);
count = pj_sock_select(ioqueue->nfds+1, &rfdset, &wfdset, &xfdset,
timeout);
if (count == 0)
return 0;
else if (count < 0)
return -pj_get_netos_error();
else if (count > PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL)
count = PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL;
/* Scan descriptor sets for event and add the events in the event
* array to be processed later in this function. We do this so that
* events can be processed in parallel without holding ioqueue lock.
*/
pj_lock_acquire(ioqueue->lock);
counter = 0;
/* Scan for writable sockets first to handle piggy-back data
* coming with accept().
*/
h = ioqueue->active_list.next;
for ( ; h!=&ioqueue->active_list && counter<count; h = h->next) {
if ( (key_has_pending_write(h) || key_has_pending_connect(h))
&& PJ_FD_ISSET(h->fd, &wfdset) && !IS_CLOSING(h))
{
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
event[counter].key = h;
event[counter].event_type = WRITEABLE_EVENT;
++counter;
}
/* Scan for readable socket. */
if ((key_has_pending_read(h) || key_has_pending_accept(h))
&& PJ_FD_ISSET(h->fd, &rfdset) && !IS_CLOSING(h) &&
counter<count)
{
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
event[counter].key = h;
event[counter].event_type = READABLE_EVENT;
++counter;
}
#if PJ_HAS_TCP
if (key_has_pending_connect(h) && PJ_FD_ISSET(h->fd, &xfdset) &&
!IS_CLOSING(h) && counter<count)
{
#if PJ_IOQUEUE_HAS_SAFE_UNREG
increment_counter(h);
#endif
event[counter].key = h;
event[counter].event_type = EXCEPTION_EVENT;
++counter;
}
#endif
}
pj_lock_release(ioqueue->lock);
count = counter;
/* Now process all events. The dispatch functions will take care
* of locking in each of the key
*/
for (counter=0; counter<count; ++counter) {
switch (event[counter].event_type) {
case READABLE_EVENT:
ioqueue_dispatch_read_event(ioqueue, event[counter].key);
break;
case WRITEABLE_EVENT:
ioqueue_dispatch_write_event(ioqueue, event[counter].key);
break;
case EXCEPTION_EVENT:
ioqueue_dispatch_exception_event(ioqueue, event[counter].key);
break;
case NO_EVENT:
pj_assert(!"Invalid event!");
break;
}
#if PJ_IOQUEUE_HAS_SAFE_UNREG
decrement_counter(event[counter].key);
#endif
}
return count;
}
/*
* ioqueue_dispatch_event()
*
* Report occurence of an event in the key to be processed by the
* framework.
*/
void ioqueue_dispatch_write_event(pj_ioqueue_t *ioqueue, pj_ioqueue_key_t *h)
{
/* Lock the key. */
pj_mutex_lock(h->mutex);
if (IS_CLOSING(h)) {
pj_mutex_unlock(h->mutex);
return;
}
#if defined(PJ_HAS_TCP) && PJ_HAS_TCP!=0
if (h->connecting) {
/* Completion of connect() operation */
pj_ssize_t bytes_transfered;
pj_bool_t has_lock;
/* Clear operation. */
h->connecting = 0;
ioqueue_remove_from_set(ioqueue, h, WRITEABLE_EVENT);
ioqueue_remove_from_set(ioqueue, h, EXCEPTION_EVENT);
#if (defined(PJ_HAS_SO_ERROR) && PJ_HAS_SO_ERROR!=0)
/* from connect(2):
* On Linux, use getsockopt to read the SO_ERROR option at
* level SOL_SOCKET to determine whether connect() completed
* successfully (if SO_ERROR is zero).
*/
{
int value;
int vallen = sizeof(value);
int gs_rc = pj_sock_getsockopt(h->fd, SOL_SOCKET, SO_ERROR,
&value, &vallen);
if (gs_rc != 0) {
/* Argh!! What to do now???
* Just indicate that the socket is connected. The
* application will get error as soon as it tries to use
* the socket to send/receive.
*/
bytes_transfered = 0;
} else {
bytes_transfered = value;
}
}
#elif defined(PJ_WIN32) && PJ_WIN32!=0
bytes_transfered = 0; /* success */
#else
/* Excellent information in D.J. Bernstein page:
* http://cr.yp.to/docs/connect.html
*
* Seems like the most portable way of detecting connect()
* failure is to call getpeername(). If socket is connected,
* getpeername() will return 0. If the socket is not connected,
* it will return ENOTCONN, and read(fd, &ch, 1) will produce
* the right errno through error slippage. This is a combination
* of suggestions from Douglas C. Schmidt and Ken Keys.
*/
{
int gp_rc;
struct sockaddr_in addr;
socklen_t addrlen = sizeof(addr);
gp_rc = getpeername(h->fd, (struct sockaddr*)&addr, &addrlen);
bytes_transfered = (gp_rc < 0) ? gp_rc : -gp_rc;
}
#endif
/* Unlock; from this point we don't need to hold key's mutex
* (unless concurrency is disabled, which in this case we should
* hold the mutex while calling the callback) */
if (h->allow_concurrent) {
/* concurrency may be changed while we're in the callback, so
* save it to a flag.
*/
has_lock = PJ_FALSE;
pj_mutex_unlock(h->mutex);
} else {
has_lock = PJ_TRUE;
}
/* Call callback. */
if (h->cb.on_connect_complete && !IS_CLOSING(h))
(*h->cb.on_connect_complete)(h, bytes_transfered);
/* Unlock if we still hold the lock */
if (has_lock) {
pj_mutex_unlock(h->mutex);
}
/* Done. */
} else
#endif /* PJ_HAS_TCP */
if (key_has_pending_write(h)) {
/* Socket is writable. */
struct write_operation *write_op;
pj_ssize_t sent;
pj_status_t send_rc;
/* Get the first in the queue. */
write_op = h->write_list.next;
/* For datagrams, we can remove the write_op from the list
* so that send() can work in parallel.
*/
if (h->fd_type == pj_SOCK_DGRAM()) {
pj_list_erase(write_op);
if (pj_list_empty(&h->write_list))
ioqueue_remove_from_set(ioqueue, h, WRITEABLE_EVENT);
}
/* Send the data.
* Unfortunately we must do this while holding key's mutex, thus
* preventing parallel write on a single key.. :-((
*/
sent = write_op->size - write_op->written;
if (write_op->op == PJ_IOQUEUE_OP_SEND) {
send_rc = pj_sock_send(h->fd, write_op->buf+write_op->written,
&sent, write_op->flags);
/* Can't do this. We only clear "op" after we're finished sending
* the whole buffer.
*/
//write_op->op = 0;
} else if (write_op->op == PJ_IOQUEUE_OP_SEND_TO) {
send_rc = pj_sock_sendto(h->fd,
write_op->buf+write_op->written,
&sent, write_op->flags,
&write_op->rmt_addr,
write_op->rmt_addrlen);
/* Can't do this. We only clear "op" after we're finished sending
* the whole buffer.
*/
//write_op->op = 0;
} else {
pj_assert(!"Invalid operation type!");
write_op->op = PJ_IOQUEUE_OP_NONE;
send_rc = PJ_EBUG;
}
if (send_rc == PJ_SUCCESS) {
write_op->written += sent;
} else {
pj_assert(send_rc > 0);
write_op->written = -send_rc;
}
/* Are we finished with this buffer? */
if (send_rc!=PJ_SUCCESS ||
write_op->written == (pj_ssize_t)write_op->size ||
h->fd_type == pj_SOCK_DGRAM())
{
pj_bool_t has_lock;
write_op->op = PJ_IOQUEUE_OP_NONE;
if (h->fd_type != pj_SOCK_DGRAM()) {
/* Write completion of the whole stream. */
pj_list_erase(write_op);
/* Clear operation if there's no more data to send. */
if (pj_list_empty(&h->write_list))
ioqueue_remove_from_set(ioqueue, h, WRITEABLE_EVENT);
}
/* Unlock; from this point we don't need to hold key's mutex
* (unless concurrency is disabled, which in this case we should
* hold the mutex while calling the callback) */
if (h->allow_concurrent) {
/* concurrency may be changed while we're in the callback, so
* save it to a flag.
*/
has_lock = PJ_FALSE;
pj_mutex_unlock(h->mutex);
} else {
has_lock = PJ_TRUE;
}
/* Call callback. */
if (h->cb.on_write_complete && !IS_CLOSING(h)) {
(*h->cb.on_write_complete)(h,
(pj_ioqueue_op_key_t*)write_op,
write_op->written);
}
if (has_lock) {
pj_mutex_unlock(h->mutex);
}
} else {
pj_mutex_unlock(h->mutex);
}
/* Done. */
} else {
/*
* This is normal; execution may fall here when multiple threads
* are signalled for the same event, but only one thread eventually
* able to process the event.
*/
pj_mutex_unlock(h->mutex);
}
}