bool select_call::wait_os(bool zero_timeout)
{
timeval to, *pto = NULL;
timespec to_pselect, *pto_pselect = NULL;
/* Avner: I put it in comment, because this logic is wrong
// optimization: do not call os select if ALL fds are excluded
// extend check to write/except fds
if (m_rfd_count == m_n_exclude_fds)
return;
*/
if (zero_timeout) {
to.tv_sec = to.tv_usec = 0;
pto = &to;
}
else {
pto = m_timeout;
}
// Restore original sets
if (m_b_run_prepare_to_poll) {
if (m_readfds) FD_COPY(m_readfds, &m_os_rfds, m_nfds);
if (m_writefds) FD_COPY(m_writefds, &m_os_wfds, m_nfds);
if (m_exceptfds)FD_COPY(m_exceptfds, &m_orig_exceptfds, m_nfds);
}
__log_func("calling os select: %d", m_nfds);
if (m_sigmask) {
if (pto) {
to_pselect.tv_sec = pto->tv_sec;
to_pselect.tv_nsec = pto->tv_usec * 1000;
pto_pselect = &to_pselect;
}
m_n_all_ready_fds = orig_os_api.pselect(m_nfds, m_readfds, m_writefds, m_exceptfds, pto_pselect, m_sigmask);
} else {
m_n_all_ready_fds = orig_os_api.select(m_nfds, m_readfds, m_writefds, m_exceptfds, pto);
}
if (m_n_all_ready_fds < 0) {
vma_throw_object(io_mux_call::io_error);
}
if (m_n_all_ready_fds > 0) {
__log_func("wait_os() returned with %d", m_n_all_ready_fds);
}
return false; // No cq_fd in select() event
}
epoll_wait_call::epoll_wait_call(epoll_event *extra_events_buffer, offloaded_mode_t *off_modes_buffer,
int epfd, epoll_event *events, int maxevents,
int timeout, const sigset_t *sigmask /* = NULL */) :
io_mux_call(NULL, off_modes_buffer, 0, sigmask), // TODO: rethink on these arguments
m_epfd(epfd), m_events(events), m_maxevents(maxevents), m_timeout(timeout),
m_p_ready_events(extra_events_buffer)
{
// get epfd_info
m_epfd_info = fd_collection_get_epfd(epfd);
if (!m_epfd_info || maxevents <= 0) {
__log_dbg("error, epfd %d not found or maxevents <= 0 (=%d)", epfd, maxevents);
errno = maxevents <= 0 ? EINVAL : EBADF;
vma_throw_object(io_mux_call::io_error);
}
// create stats
m_p_stats = &m_epfd_info->stats()->stats;
}
select_call::select_call(int *off_fds_buffer, offloaded_mode_t *off_modes_buffer,
int nfds, fd_set *readfds, fd_set *writefds,
fd_set *exceptfds, timeval *timeout, const sigset_t *__sigmask /* = NULL */) :
io_mux_call(off_fds_buffer, off_modes_buffer, nfds, __sigmask),
m_nfds(nfds), m_readfds(readfds), m_writefds(writefds),
m_exceptfds(exceptfds), m_timeout(timeout), m_nfds_with_cq(0), m_b_run_prepare_to_poll(false)
{
int fd;
//socket_fd_api* temp_sock_fd_api = NULL;
if (m_nfds > FD_SETSIZE) {
errno = ENOMEM;
vma_throw_object(io_mux_call::io_error);
}
// create stats
m_p_stats = &g_select_stats;
vma_stats_instance_get_select_block(m_p_stats);
bool offloaded_read = !!m_readfds;
bool offloaded_write = !!m_writefds;
if (offloaded_read || offloaded_write) {
FD_ZERO(&m_os_rfds, m_nfds);
FD_ZERO(&m_os_wfds, m_nfds);
//covers the case of select(readfds = NULL)
if(!m_readfds) {
FD_ZERO(&m_cq_rfds, m_nfds);
m_readfds = &m_cq_rfds;
}
// get offloaded fds in read set
for (fd = 0; fd < m_nfds; ++fd) {
bool check_read = offloaded_read && FD_ISSET(fd, m_readfds);
bool check_write = offloaded_write && FD_ISSET(fd, m_writefds);
socket_fd_api* psock = fd_collection_get_sockfd(fd);
if (psock && psock->get_type() == FD_TYPE_SOCKET) {
offloaded_mode_t off_mode = OFF_NONE;
if (check_read) off_mode = (offloaded_mode_t)(off_mode | OFF_READ);
if (check_write) off_mode = (offloaded_mode_t)(off_mode | OFF_WRITE);
if (off_mode) {
__log_func("---> fd=%d IS SET for read or write!", fd);
m_p_all_offloaded_fds[m_num_all_offloaded_fds] = fd;
m_p_offloaded_modes[m_num_all_offloaded_fds] = off_mode;
m_num_all_offloaded_fds++;
if (! psock->skip_os_select()) {
if (check_read) {
FD_SET(fd, &m_os_rfds);
if (psock->is_readable(NULL)) {
io_mux_call::update_fd_array(&m_fd_ready_array, fd);
m_n_ready_rfds++;
m_n_all_ready_fds++;
} else {
// Instructing the socket to sample the OS immediately to prevent hitting EAGAIN on recvfrom(),
// after iomux returned a shadow fd as ready (only for non-blocking sockets)
psock->set_immediate_os_sample();
}
}
if (check_write) {
FD_SET(fd, &m_os_wfds);
}
}
else
__log_func("fd=%d must be skipped from os r select()", fd);
}
}
else {
if (check_read) {
FD_SET(fd, &m_os_rfds);
}
if (check_write) {
FD_SET(fd, &m_os_wfds);
}
}
}
}
__log_func("num all offloaded_fds=%d", m_num_all_offloaded_fds);
}
bool select_call::wait(const timeval &elapsed)
{
timeval timeout, *pto = NULL;
timespec to_pselect, *pto_pselect = NULL;
BULLSEYE_EXCLUDE_BLOCK_START
if (m_n_all_ready_fds > 0) {
__log_panic("wait() called when there are ready fd's!!!");
// YossiE TODO make this and some more checks as debug assertions
// In all functions
}
BULLSEYE_EXCLUDE_BLOCK_END
// Restore original sets
if (m_b_run_prepare_to_poll) {
if (m_readfds) FD_COPY(m_readfds, &m_os_rfds, m_nfds);
if (m_writefds) FD_COPY(m_writefds, &m_os_wfds, m_nfds);
if (m_exceptfds)FD_COPY(m_exceptfds, &m_orig_exceptfds, m_nfds);
}
// Call OS select() on original sets + CQ epfd in read set
if (m_readfds)
FD_SET(m_cqepfd, m_readfds);
if (m_timeout) {
tv_sub(m_timeout, &elapsed, &timeout);
if (timeout.tv_sec < 0 || timeout.tv_usec < 0) {
// Already reached timeout
return false;
}
pto = &timeout;
}
__log_func("going to wait on select CQ+OS nfds=%d cqfd=%d pto=%p!!!", m_nfds_with_cq, m_cqepfd, pto);
// ACTUAL CALL TO SELECT
if (m_sigmask) {
if (pto) {
to_pselect.tv_sec = pto->tv_sec;
to_pselect.tv_nsec = pto->tv_usec * 1000;
pto_pselect = &to_pselect;
}
m_n_all_ready_fds = orig_os_api.pselect(m_nfds, m_readfds, m_writefds, m_exceptfds, pto_pselect, m_sigmask);
} else {
m_n_all_ready_fds = orig_os_api.select(m_nfds_with_cq, m_readfds, m_writefds, m_exceptfds, pto);
}
__log_func("done select CQ+OS nfds=%d cqfd=%d pto=%p ready=%d!!!", m_nfds_with_cq, m_cqepfd, pto, m_n_all_ready_fds);
if (m_n_all_ready_fds < 0) {
vma_throw_object(io_mux_call::io_error);
}
// Clear CQ from the set and don't count it
if (m_readfds)
{
if (FD_ISSET(m_cqepfd, m_readfds)) {
FD_CLR(m_cqepfd, m_readfds); // Not needed if m_readfds is NULL
--m_n_all_ready_fds;
return true;
}
}
return false;
}
bool epoll_wait_call::_wait(int timeout)
{
int i, ready_fds, fd;
bool cq_ready = false;
__log_func("calling os epoll: %d", m_epfd);
if (timeout) {
lock();
if (m_epfd_info->m_ready_fds.empty()) {
m_epfd_info->going_to_sleep();
} else {
timeout = 0;
}
unlock();
}
if (m_sigmask) {
ready_fds = orig_os_api.epoll_pwait(m_epfd, m_p_ready_events, m_maxevents, timeout, m_sigmask);
} else {
ready_fds = orig_os_api.epoll_wait(m_epfd, m_p_ready_events, m_maxevents, timeout);
}
if (timeout) {
lock();
m_epfd_info->return_from_sleep();
unlock();
}
if (ready_fds < 0) {
vma_throw_object(io_mux_call::io_error);
}
// convert the returned events to user events and mark offloaded fds
m_n_all_ready_fds = 0;
for (i = 0; i < ready_fds; ++i) {
fd = m_p_ready_events[i].data.fd;
// wakeup event
if(m_epfd_info->is_wakeup_fd(fd))
{
lock();
m_epfd_info->remove_wakeup_fd();
unlock();
continue;
}
// If it's CQ
if (m_epfd_info->is_cq_fd(m_p_ready_events[i].data.u64)) {
cq_ready = true;
continue;
}
if ((m_p_ready_events[i].events & EPOLLIN)) {
socket_fd_api* temp_sock_fd_api = fd_collection_get_sockfd(fd);
if (temp_sock_fd_api) {
// Instructing the socket to sample the OS immediately to prevent hitting EAGAIN on recvfrom(),
// after iomux returned a shadow fd as ready (only for non-blocking sockets)
temp_sock_fd_api->set_immediate_os_sample();
}
}
// Copy event bits and data
m_events[m_n_all_ready_fds].events = m_p_ready_events[i].events;
if (!m_epfd_info->get_data_by_fd(fd, &m_events[m_n_all_ready_fds].data)) {
continue;
}
++m_n_all_ready_fds;
}
return cq_ready;
}
