void citp_log_fn_drv(const char* msg)
{
if( citp.log_fd < 0 ) {
if( ef_onload_driver_open(&citp.log_fd, OO_STACK_DEV, 1) ) return;
if( citp_fdtable.table )
citp_fdtable.table[citp.log_fd].fdip=fdi_to_fdip(&citp_the_reserved_fd);
/* just to be sure: */
ci_sys_fcntl(citp.log_fd, F_SETFD, FD_CLOEXEC);
}
my_syscall3(ioctl, citp.log_fd, OO_IOC_PRINTK, (long) msg);
}
static int ci_tcp_shutdown_listen(citp_socket* ep, int how, ci_fd_t fd)
{
ci_tcp_socket_listen* tls = SOCK_TO_TCP_LISTEN(ep->s);
if( how == SHUT_WR )
return 0;
ci_sock_lock(ep->netif, &tls->s.b);
ci_netif_lock(ep->netif);
LOG_TC(ci_log(SK_FMT" shutdown(SHUT_RD)", SK_PRI_ARGS(ep)));
__ci_tcp_listen_shutdown(ep->netif, tls, fd);
__ci_tcp_listen_to_normal(ep->netif, tls);
{
ci_fd_t os_sock = ci_get_os_sock_fd(ep, fd);
int flags = ci_sys_fcntl(os_sock, F_GETFL);
flags &= (~O_NONBLOCK);
CI_TRY(ci_sys_fcntl(os_sock, F_SETFL, flags));
ci_rel_os_sock_fd(os_sock);
}
ci_netif_unlock(ep->netif);
ci_sock_unlock(ep->netif, &tls->s.b);
return 0;
}
int ef_onload_handle_move_and_do_cloexec(ef_driver_handle* pfd, int do_cloexec)
{
int fd;
if( do_cloexec )
fd = oo_fcntl_dupfd_cloexec(*pfd, CITP_OPTS.fd_base);
else
fd = ci_sys_fcntl(*pfd, F_DUPFD, CITP_OPTS.fd_base);
/* If we've successfully done the dup then we've also set CLOEXEC if
* needed on the new fd, so we're done.
*/
if( fd >= 0 ) {
ci_tcp_helper_close_no_trampoline(*pfd);
*pfd = fd;
return 0;
}
else {
LOG_NV(ci_log("%s: Failed to move fd from %d, rc %d",
__func__, *pfd, fd));
}
return fd;
}
int citp_ep_dup_fcntl_dup_cloexec(int oldfd, long arg)
{
return ci_sys_fcntl(oldfd, F_DUPFD_CLOEXEC, arg);
}
int citp_ep_dup_fcntl_dup(int oldfd, long arg)
{
return ci_sys_fcntl(oldfd, F_DUPFD, arg);
}
int citp_epoll_create(int size, int flags)
{
citp_fdinfo *fdi;
citp_epoll_fdi *epi;
struct citp_epoll_fd* ep;
int fd;
if( (epi = CI_ALLOC_OBJ(citp_epoll_fdi)) == NULL )
goto fail0;
if( (ep = CI_ALLOC_OBJ(struct citp_epoll_fd)) == NULL )
goto fail1;
fdi = &epi->fdinfo;
citp_fdinfo_init(fdi, &citp_epoll_protocol_impl);
/* Create the epoll fd. */
CITP_FDTABLE_LOCK();
if( (fd = ci_sys_epoll_create_compat(size, flags, 0)) < 0 )
goto fail2;
citp_fdtable_new_fd_set(fd, fdip_busy, TRUE);
/* Init epfd_os */
#ifdef O_CLOEXEC
ep->epfd_os = ci_sys_open(OO_EPOLL_DEV, O_RDWR | O_CLOEXEC);
#else
ep->epfd_os = ci_sys_open(OO_EPOLL_DEV, O_RDWR);
if( ep->epfd_os >= 0 )
ci_sys_fcntl(ep->epfd_os, F_SETFD, FD_CLOEXEC);
#endif
if( ep->epfd_os < 0 ) {
Log_E(ci_log("%s: ERROR: failed to open(%s) errno=%d",
__FUNCTION__, OO_EPOLL_DEV, errno));
goto fail3;
}
__citp_fdtable_reserve(ep->epfd_os, 1);
ep->shared = mmap(NULL, sizeof(*ep->shared), PROT_READ, MAP_SHARED,
ep->epfd_os, 0);
if( ep->shared == MAP_FAILED ) {
Log_E(ci_log("%s: ERROR: failed to mmap shared segment errno=%d",
__FUNCTION__, errno));
goto fail4;
}
__citp_fdtable_reserve(ep->shared->epfd, 1);
CITP_FDTABLE_UNLOCK();
epi->epoll = ep;
ep->size = size;
oo_wqlock_init(&ep->lock);
ep->not_mt_safe = ! CITP_OPTS.ul_epoll_mt_safe;
ci_dllist_init(&ep->oo_sockets);
ep->oo_sockets_n = 0;
ci_dllist_init(&ep->dead_sockets);
oo_atomic_set(&ep->refcount, 1);
ep->epfd_syncs_needed = 0;
ep->blocking = 0;
citp_fdtable_insert(fdi, fd, 0);
Log_POLL(ci_log("%s: fd=%d driver_fd=%d epfd=%d", __FUNCTION__,
fd, ep->epfd_os, (int) ep->shared->epfd));
return fd;
fail4:
__citp_fdtable_reserve(ep->epfd_os, 0);
ci_sys_close(ep->epfd_os);
fail3:
ci_sys_close(fd);
citp_fdtable_busy_clear(fd, fdip_unknown, 1);
fail2:
CITP_FDTABLE_UNLOCK();
CI_FREE_OBJ(ep);
fail1:
CI_FREE_OBJ(epi);
fail0:
return -1;
}
int ef_onload_driver_open(ef_driver_handle* pfd,
enum oo_device_type dev_type,
int do_cloexec)
{
int rc;
int flags = 0;
int saved_errno = errno;
#ifdef O_CLOEXEC
if( do_cloexec )
flags = O_CLOEXEC;
#endif
ci_assert(pfd);
rc = oo_open(pfd, dev_type, flags);
if( rc != 0 && errno != EMFILE && fd_is_saved[dev_type] >= 0 ) {
ci_clone_fd_t op;
op.do_cloexec = do_cloexec;
LOG_NV(ci_log("%s: open failed, but cloning from saved fd", __func__));
rc = ci_sys_ioctl((ci_fd_t) saved_fd[dev_type],
clone_ioctl[dev_type], &op);
if( rc < 0 )
return rc;
errno = saved_errno;
*pfd = op.fd;
}
if( rc != 0 )
return rc;
/* Our internal driver handles are not visible to the application. It may
* make assumptions about the fd space available to it, and try to dup2/3
* onto one of our driver fds. To try and minimise this we allow the user
* to specify a minimum value for us to use, to try and keep out of their
* way.
*
* We have to be able to cope with them coming along and trying to dup onto
* one of these fds anyway, as they may not have set the option up. As such
* we treat failure to shift the fd as acceptable, and just retain the old
* one.
*/
if( *pfd < CITP_OPTS.fd_base )
if( ef_onload_handle_move_and_do_cloexec(pfd, do_cloexec) == 0 )
return 0;
if( do_cloexec ) {
#if defined(O_CLOEXEC)
static int o_cloexec_fails = -1;
if( o_cloexec_fails < 0 ) {
int arg;
rc = ci_sys_fcntl(*(int *)pfd, F_GETFD, &arg);
if( rc == 0 && (arg & FD_CLOEXEC) )
o_cloexec_fails = 0;
else
o_cloexec_fails = 1;
}
#else
static const int o_cloexec_fails = 1;
#endif
if( o_cloexec_fails )
CI_DEBUG_TRY(ci_sys_fcntl(*(int *)pfd, F_SETFD, FD_CLOEXEC));
}
return 0;
}
/* fixme kostik: this is partially copy-paste from citp_sock_fcntl */
static int citp_pipe_fcntl(citp_fdinfo* fdinfo, int cmd, long arg)
{
int rc = 0;
citp_pipe_fdi* epi = fdi_to_pipe_fdi(fdinfo);
struct oo_pipe* p = epi->pipe;
switch ( cmd ) {
case F_GETFL: {
ci_uint32 flag_nonb = CI_PFD_AFLAG_NONBLOCK;
if( ! fdi_is_reader(fdinfo) ) {
rc = O_WRONLY;
flag_nonb <<= CI_PFD_AFLAG_WRITER_SHIFT;
}
else
flag_nonb <<= CI_PFD_AFLAG_READER_SHIFT;
if ( p->aflags & flag_nonb ) rc |= O_NONBLOCK;
break;
}
case F_SETFL: {
ci_uint32 bit;
rc = ci_sys_fcntl(fdinfo->fd, cmd, arg);
if( rc < 0 )
break;
bit = CI_PFD_AFLAG_NONBLOCK <<
(fdi_is_reader(fdinfo) ? CI_PFD_AFLAG_READER_SHIFT :
CI_PFD_AFLAG_WRITER_SHIFT);
if( arg & (O_NONBLOCK | O_NDELAY) )
ci_bit_mask_set(&p->aflags, bit);
else
ci_bit_mask_clear(&p->aflags, bit);
break;
}
case F_DUPFD:
rc = citp_ep_dup(fdinfo->fd, citp_ep_dup_fcntl_dup, arg);
break;
#ifdef F_DUPFD_CLOEXEC
case F_DUPFD_CLOEXEC:
rc = citp_ep_dup(fdinfo->fd, citp_ep_dup_fcntl_dup_cloexec, arg);
break;
#endif
case F_GETFD:
case F_SETFD:
rc = ci_sys_fcntl(fdinfo->fd, cmd, arg);
break;
case F_GETLK:
case F_SETLK:
case F_SETLKW:
/* File locks not supported on sockets */
Log_U(ci_log("%s: cmd %d not supported on sockets!",__FUNCTION__,
cmd));
errno = ENOTSUP;
rc = CI_SOCKET_ERROR;
break;
case F_GETOWN:
case F_SETOWN:
#ifdef F_GETOWN_EX
case F_GETOWN_EX:
#endif
#ifdef F_SETOWN_EX
case F_SETOWN_EX:
#endif
rc = ci_sys_fcntl(fdinfo->fd, cmd, arg);
if( rc != 0 )
break;
p->b.sigown = arg;
if( p->b.sigown && (p->b.sb_aflags & CI_SB_AFLAG_O_ASYNC) )
ci_bit_set(&p->b.wake_request, CI_SB_FLAG_WAKE_RX_B);
break;
#ifdef F_SETPIPE_SZ
case F_SETPIPE_SZ:
/* System pipe buf size is rounded up to power of two. We
* cannot replicate this.
*/
rc = ci_pipe_set_size(epi->ni, p, arg);
if( rc < 0 ) {
errno = EINVAL;
rc = CI_SOCKET_ERROR;
break;
}
rc = 0;
break;
#endif
#ifdef F_GETPIPE_SZ
case F_GETPIPE_SZ:
rc = (p->bufs_max - 1) * OO_PIPE_BUF_MAX_SIZE;
break;
#endif
default:
/* fixme kostik: logging should include some pipe identification */
errno = ENOTSUP;
rc = CI_SOCKET_ERROR;
}
Log_VSC(log("%s(%d, %d, %ld) = %d (errno=%d)",
__FUNCTION__, fdinfo->fd, cmd, arg, rc, errno));
return rc;
}