static void citp_epoll_dtor(citp_fdinfo* fdi, int fdt_locked)
{
struct citp_epoll_fd* ep = fdi_to_epoll(fdi);
if (!oo_atomic_dec_and_test(&ep->refcount))
return;
CITP_FDTABLE_LOCK();
ci_tcp_helper_close_no_trampoline(ep->shared->epfd);
__citp_fdtable_reserve(ep->shared->epfd, 0);
munmap(ep->shared, sizeof(*ep->shared));
ci_tcp_helper_close_no_trampoline(ep->epfd_os);
__citp_fdtable_reserve(ep->epfd_os, 0);
CITP_FDTABLE_UNLOCK();
CI_FREE_OBJ(ep);
}
/* Move priv file to the alien_ni stack.
* Should be called with the locked priv stack and socket;
* the function returns with this stack being unlocked.
* If rc=0, it returns with alien_ni stack locked;
* otherwise, both stacks are unlocked.
* Socket is always unlocked on return. */
int efab_file_move_to_alien_stack(ci_private_t *priv, ci_netif *alien_ni)
{
tcp_helper_resource_t *old_thr = priv->thr;
tcp_helper_resource_t *new_thr = netif2tcp_helper_resource(alien_ni);
ci_sock_cmn *old_s = SP_TO_SOCK(&old_thr->netif, priv->sock_id);
ci_sock_cmn *new_s;
ci_sock_cmn *mid_s;
tcp_helper_endpoint_t *old_ep, *new_ep;
int rc, i;
int pollwait_register = 0;
#if CI_CFG_FD_CACHING
oo_p sp;
#endif
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d", __func__,
old_thr->id, priv->sock_id, new_thr->id));
/* Poll the old stack - deliver all data to our socket */
ci_netif_poll(&old_thr->netif);
/* Endpoints in epoll list should not be moved, because waitq is already
* in the epoll internal structures (bug 41152). */
if( !list_empty(&priv->_filp->f_ep_links) ) {
rc = -EBUSY;
goto fail1;
}
if( !efab_file_move_supported(&old_thr->netif, old_s) ) {
rc = -EINVAL;
goto fail1;
}
/* Lock the second stack */
i = 0;
while( ! ci_netif_trylock(alien_ni) ) {
ci_netif_unlock(&old_thr->netif);
if( i++ >= 1000 ) {
rc = -EBUSY;
goto fail1_ni_unlocked;
}
rc = ci_netif_lock(&old_thr->netif);
if( rc != 0 )
goto fail1_ni_unlocked;
}
/* Allocate a new socket in the alien_ni stack */
rc = -ENOMEM;
if( old_s->b.state == CI_TCP_STATE_UDP ) {
ci_udp_state *new_us = ci_udp_get_state_buf(alien_ni);
if( new_us == NULL )
goto fail2;
new_s = &new_us->s;
}
else {
ci_tcp_state *new_ts = ci_tcp_get_state_buf(alien_ni);
if( new_ts == NULL )
goto fail2;
new_s = &new_ts->s;
}
/* Allocate an intermediate "socket" outside of everything */
mid_s = ci_alloc(CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
if( mid_s == NULL )
goto fail3;
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d:%d", __func__,
old_thr->id, priv->sock_id,
new_thr->id, new_s->b.bufid));
/* Copy TCP/UDP state */
memcpy(mid_s, old_s, CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
/* do not copy old_s->b.bufid
* and other fields in stack adress space */
mid_s->b.sb_aflags |= CI_SB_AFLAG_ORPHAN;
mid_s->b.bufid = new_s->b.bufid;
mid_s->b.post_poll_link = new_s->b.post_poll_link;
mid_s->b.ready_link = new_s->b.ready_link;
mid_s->reap_link = new_s->reap_link;
if( old_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *mid_ts = SOCK_TO_TCP(mid_s);
mid_ts->timeout_q_link = new_ts->timeout_q_link;
mid_ts->tx_ready_link = new_ts->tx_ready_link;
mid_ts->rto_tid = new_ts->rto_tid;
mid_ts->delack_tid = new_ts->delack_tid;
mid_ts->zwin_tid = new_ts->zwin_tid;
mid_ts->kalive_tid = new_ts->kalive_tid;
mid_ts->cork_tid = new_ts->cork_tid;
ci_ip_queue_init(&mid_ts->recv1);
ci_ip_queue_init(&mid_ts->recv2);
ci_ip_queue_init(&mid_ts->send);
ci_ip_queue_init(&mid_ts->retrans);
mid_ts->send_prequeue = OO_PP_ID_NULL;
new_ts->retrans_ptr = OO_PP_NULL;
mid_ts->tmpl_head = OO_PP_NULL;
oo_atomic_set(&mid_ts->send_prequeue_in, 0);
*new_ts = *mid_ts;
ci_pmtu_state_init(alien_ni, &new_ts->s, &new_ts->pmtus,
CI_IP_TIMER_PMTU_DISCOVER);
#if CI_CFG_FD_CACHING
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_link, sp, "epch");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_link);
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_fd_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_fd_link, sp, "ecfd");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_fd_link);
#endif
/* free temporary mid_ts storage */
CI_FREE_OBJ(mid_ts);
}
else {
ci_udp_state *mid_us = SOCK_TO_UDP(mid_s);
*SOCK_TO_UDP(new_s) = *mid_us;
CI_FREE_OBJ(mid_us);
}
/* Move the filter */
old_ep = ci_trs_ep_get(old_thr, priv->sock_id);
new_ep = ci_trs_ep_get(new_thr, new_s->b.bufid);
rc = tcp_helper_endpoint_move_filters_pre(old_ep, new_ep);
if( rc != 0 ) {
rc = -EINVAL;
goto fail3;
}
/* Allocate a new file for the new endpoint */
rc = onload_alloc_file(new_thr, new_s->b.bufid, priv->_filp->f_flags,
priv->fd_type, &old_ep->alien_ref);
if( rc != 0 )
goto fail4;
ci_assert(old_ep->alien_ref);
/* Copy F_SETOWN_EX, F_SETSIG to the new file */
#ifdef F_SETOWN_EX
rcu_read_lock();
__f_setown(old_ep->alien_ref->_filp, priv->_filp->f_owner.pid,
priv->_filp->f_owner.pid_type, 1);
rcu_read_unlock();
#endif
old_ep->alien_ref->_filp->f_owner.signum = priv->_filp->f_owner.signum;
old_ep->alien_ref->_filp->f_flags |= priv->_filp->f_flags & O_NONBLOCK;
/* Move os_socket from one ep to another */
if( tcp_helper_endpoint_set_aflags(new_ep, OO_THR_EP_AFLAG_ATTACHED) &
OO_THR_EP_AFLAG_ATTACHED ) {
fput(old_ep->alien_ref->_filp);
rc = -EBUSY;
goto fail2; /* state & filters are cleared by fput() */
}
/********* Point of no return **********/
ci_wmb();
priv->fd_type = CI_PRIV_TYPE_ALIEN_EP;
priv->_filp->f_op = &linux_tcp_helper_fops_alien;
ci_wmb();
oo_file_moved(priv);
/* Read all already-arrived packets after the filters move but before
* copying of the receive queue. */
ci_netif_poll(&old_thr->netif);
tcp_helper_endpoint_move_filters_post(old_ep, new_ep);
ci_assert( efab_file_move_supported(&old_thr->netif, old_s));
/* There's a gap between un-registering the old ep, and registering the
* the new. However, the notifications shouldn't be in use for sockets
* that are in a state that can be moved, so this shouldn't be a problem.
*/
if( old_ep->os_sock_pt.whead ) {
pollwait_register = 1;
efab_tcp_helper_os_pollwait_unregister(old_ep);
}
ci_assert_equal(new_ep->os_socket, NULL);
new_ep->os_socket = oo_file_ref_xchg(&old_ep->os_socket, NULL);
ci_assert_equal(old_ep->os_socket, NULL);
if( pollwait_register )
efab_tcp_helper_os_pollwait_register(new_ep);
ci_bit_clear(&new_s->b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
if( new_s->b.state == CI_TCP_ESTABLISHED )
CI_TCP_STATS_INC_CURR_ESTAB(alien_ni);
/* Copy recv queue */
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *old_ts = SOCK_TO_TCP(old_s);
int i;
/* Stop timers */
ci_ip_timer_clear(&old_thr->netif, &old_ts->kalive_tid);
ci_ip_timer_clear(&old_thr->netif, &old_ts->delack_tid);
efab_ip_queue_copy(alien_ni, &new_ts->recv1,
&old_thr->netif, &old_ts->recv1);
efab_ip_queue_copy(alien_ni, &new_ts->recv2,
&old_thr->netif, &old_ts->recv2);
new_ts->recv1_extract = new_ts->recv1.head;
/* Drop reorder buffer */
ci_ip_queue_init(&new_ts->rob);
new_ts->dsack_block = OO_PP_INVALID;
new_ts->dsack_start = new_ts->dsack_end = 0;
for( i = 0; i <= CI_TCP_SACK_MAX_BLOCKS; i++ )
new_ts->last_sack[i] = OO_PP_NULL;
}
else {
/* There should not be any recv q, but drop it to be sure */
ci_udp_recv_q_init(&SOCK_TO_UDP(new_s)->recv_q);
}
/* Old stack can be unlocked */
old_s->b.sb_flags |= CI_SB_FLAG_MOVED;
ci_netif_unlock(&old_thr->netif);
ci_assert( efab_file_move_supported(alien_ni, new_s) );
/* Move done: poll for any new data. */
ci_netif_poll(alien_ni);
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
/* Timers setup: delack, keepalive */
if( (new_ts->acks_pending & CI_TCP_ACKS_PENDING_MASK) > 0)
ci_tcp_timeout_delack(alien_ni, new_ts);
ci_tcp_kalive_reset(alien_ni, new_ts);
}
/* Old ep: we are done. */
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_BIT);
old_s->b.moved_to_stack_id = alien_ni->state->stack_id;
old_s->b.moved_to_sock_id = new_s->b.bufid;
if( ! list_empty(&priv->_filp->f_ep_links) )
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL_BIT);
ci_sock_unlock(&old_thr->netif, &old_s->b);
ci_sock_unlock(alien_ni, &new_s->b);
ci_assert(ci_netif_is_locked(alien_ni));
OO_DEBUG_TCPH(ci_log("%s: -> [%d:%d] %s", __func__,
new_thr->id, new_s->b.bufid,
ci_tcp_state_str(new_s->b.state)));
return 0;
fail4:
/* We clear the filters from the new ep.
* For now, we do not need to re-insert old filters because hw filters
* are alredy here (in case of accepted socket) or not needed.
* We have not removed old sw filters yet. */
tcp_helper_endpoint_move_filters_undo(old_ep, new_ep);
fail3:
if( new_s->b.state & CI_TCP_STATE_TCP )
ci_tcp_state_free(alien_ni, SOCK_TO_TCP(new_s));
else
ci_udp_state_free(alien_ni, SOCK_TO_UDP(new_s));
fail2:
ci_netif_unlock(alien_ni);
fail1:
ci_netif_unlock(&old_thr->netif);
fail1_ni_unlocked:
ci_sock_unlock(&old_thr->netif, &old_s->b);
OO_DEBUG_TCPH(ci_log("%s: rc=%d", __func__, rc));
return rc;
}
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;
}
/* we don't register protocol impl */
int citp_pipe_create(int fds[2], int flags)
{
citp_pipe_fdi* epi_read;
citp_pipe_fdi* epi_write;
struct oo_pipe* p = NULL; /* make compiler happy */
ci_netif* ni;
int rc = -1;
ef_driver_handle fd = -1;
Log_V(log(LPF "pipe()"));
/* citp_netif_exists() does not need citp_ul_lock here */
if( CITP_OPTS.ul_pipe == CI_UNIX_PIPE_ACCELERATE_IF_NETIF &&
! citp_netif_exists() ) {
return CITP_NOT_HANDLED;
}
rc = citp_netif_alloc_and_init(&fd, &ni);
if( rc != 0 ) {
if( rc == CI_SOCKET_HANDOVER ) {
/* This implies EF_DONT_ACCELERATE is set, so we handover
* regardless of CITP_OPTS.no_fail */
return CITP_NOT_HANDLED;
}
/* may be lib mismatch - errno will be ELIBACC */
goto fail1;
}
rc = -1;
CI_MAGIC_CHECK(ni, NETIF_MAGIC);
/* add another reference as we have 2 fdis */
citp_netif_add_ref(ni);
epi_read = citp_pipe_epi_alloc(ni, O_RDONLY);
if( epi_read == NULL )
goto fail2;
epi_write = citp_pipe_epi_alloc(ni, O_WRONLY);
if( epi_write == NULL )
goto fail3;
/* oo_pipe init code */
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
rc = oo_pipe_ctor(ni, &p, fds, flags);
if( rc < 0 )
goto fail4;
citp_fdtable_new_fd_set(fds[0], fdip_busy, fdtable_strict());
citp_fdtable_new_fd_set(fds[1], fdip_busy, fdtable_strict());
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
LOG_PIPE("%s: pipe=%p id=%d", __FUNCTION__, p, p->b.bufid);
/* as pipe is created it should be attached to the end-points */
epi_read->pipe = p;
epi_write->pipe = p;
/* We're ready. Unleash us onto the world! */
ci_assert(epi_read->pipe->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_assert(epi_write->pipe->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi_read->pipe->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi_read->pipe->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
citp_fdtable_insert(&epi_read->fdinfo, fds[0], 0);
citp_fdtable_insert(&epi_write->fdinfo, fds[1], 0);
CI_MAGIC_CHECK(ni, NETIF_MAGIC);
return 0;
fail4:
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
fail3:
CI_FREE_OBJ(epi_write);
fail2:
CI_FREE_OBJ(epi_read);
citp_netif_release_ref(ni, 0);
citp_netif_release_ref(ni, 0);
fail1:
if( CITP_OPTS.no_fail && errno != ELIBACC ) {
Log_U(ci_log("%s: failed (errno:%d) - PASSING TO OS", __FUNCTION__, errno));
return CITP_NOT_HANDLED;
}
return rc;
}
static int citp_udp_socket(int domain, int type, int protocol)
{
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ef_driver_handle fd;
int rc;
ci_netif* ni;
Log_V(log(LPF "socket(%d, %d, %d)", domain, type, protocol));
epi = CI_ALLOC_OBJ(citp_sock_fdi);
if( ! epi ) {
Log_U(ci_log(LPF "socket: failed to allocate epi"));
errno = ENOMEM;
goto fail1;
}
fdi = &epi->fdinfo;
citp_fdinfo_init(fdi, &citp_udp_protocol_impl);
rc = citp_netif_alloc_and_init(&fd, &ni);
if( rc != 0 ) {
if( rc == CI_SOCKET_HANDOVER ) {
/* This implies EF_DONT_ACCELERATE is set, so we handover
* regardless of CITP_OPTS.no_fail */
CI_FREE_OBJ(epi);
return rc;
}
goto fail2;
}
/* Protect the fdtable entry until we're done initialising. */
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
if((fd = ci_udp_ep_ctor(&epi->sock, ni, domain, type)) < 0) {
/*! ?? \TODO unpick the ci_udp_ep_ctor according to how failed */
Log_U(ci_log(LPF "socket: udp_ep_ctor failed"));
errno = -fd;
goto fail3;
}
citp_fdtable_new_fd_set(fd, fdip_busy, fdtable_strict());
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
CI_DEBUG(epi->sock.s->pid = getpid());
/* We're ready. Unleash us onto the world! */
ci_assert(epi->sock.s->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi->sock.s->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
citp_fdtable_insert(fdi, fd, 0);
Log_VSS(log(LPF "socket(%d, %d, %d) = "EF_FMT, domain, type, protocol,
EF_PRI_ARGS(epi,fd)));
return fd;
fail3:
if( CITP_OPTS.no_fail && errno != ELIBACC )
CITP_STATS_NETIF(++ni->state->stats.udp_handover_socket);
citp_netif_release_ref(ni, 0);
fail2:
CI_FREE_OBJ(epi);
fail1:
/* BUG1408: Graceful failure. We'll only fail outright if there's a
* driver/library mismatch */
if( CITP_OPTS.no_fail && errno != ELIBACC ) {
Log_U(ci_log("%s: failed (errno:%d) - PASSING TO OS", __FUNCTION__, errno));
return CI_SOCKET_HANDOVER;
}
return -1;
}
