static int
efab_tcp_drop_from_acceptq(ci_private_t *priv, void *arg)
{
struct oo_op_tcp_drop_from_acceptq *carg = arg;
tcp_helper_resource_t *thr;
tcp_helper_endpoint_t *ep;
citp_waitable *w;
ci_tcp_state *ts;
int rc = -EINVAL;
/* find stack */
rc = efab_thr_table_lookup(NULL, carg->stack_id,
EFAB_THR_TABLE_LOOKUP_CHECK_USER |
EFAB_THR_TABLE_LOOKUP_NO_UL,
&thr);
if( rc < 0 )
return rc;
ci_assert( thr->k_ref_count & TCP_HELPER_K_RC_NO_USERLAND );
/* find endpoint and drop OS socket */
ep = ci_trs_get_valid_ep(thr, carg->sock_id);
if( ep == NULL )
goto fail1;
w = SP_TO_WAITABLE(&thr->netif, carg->sock_id);
if( !(w->state & CI_TCP_STATE_TCP) || w->state == CI_TCP_LISTEN )
goto fail2;
ts = SP_TO_TCP(&thr->netif, carg->sock_id);
ci_assert(ep->os_port_keeper);
ci_assert_equal(ep->os_socket, NULL);
LOG_TV(ci_log("%s: send reset to non-accepted connection", __FUNCTION__));
/* copy from ci_tcp_listen_shutdown_queues() */
ci_assert(ts->s.b.sb_aflags & CI_SB_AFLAG_TCP_IN_ACCEPTQ);
rc = ci_netif_lock(&thr->netif);
if( rc != 0 ) {
ci_assert_equal(rc, -EINTR);
rc = -ERESTARTSYS;
goto fail2;
}
ci_bit_clear(&ts->s.b.sb_aflags, CI_SB_AFLAG_TCP_IN_ACCEPTQ_BIT);
/* We have no way to close this connection from the other side:
* there was no RST from peer. */
ci_assert_nequal(ts->s.b.state, CI_TCP_CLOSED);
ci_assert_nequal(ts->s.b.state, CI_TCP_TIME_WAIT);
ci_tcp_send_rst(&thr->netif, ts);
ci_tcp_drop(&thr->netif, ts, ECONNRESET);
ci_assert_equal(ep->os_port_keeper, NULL);
ci_netif_unlock(&thr->netif);
efab_tcp_helper_k_ref_count_dec(thr, 1);
return 0;
fail1:
efab_thr_release(thr);
fail2:
ci_log("%s: inconsistent ep %d:%d", __func__, carg->stack_id, carg->sock_id);
return rc;
}
int ci_udp_ioctl(citp_socket *ep, ci_fd_t fd, int request, void* arg)
{
ci_netif* ni = ep->netif;
ci_udp_state* us = SOCK_TO_UDP(ep->s);
int rc;
ci_netif_lock(ni);
rc = ci_udp_ioctl_locked(ni, us, fd, request, arg);
ci_netif_unlock(ni);
return rc;
}
ci_fd_t ci_udp_ep_ctor(citp_socket* ep, ci_netif* netif, int domain, int type)
{
ci_udp_state* us;
ci_fd_t fd;
VERB( log(LPFIN "ctor( )" ) );
ci_assert(ep);
ci_assert(netif);
ci_netif_lock(netif);
us = ci_udp_get_state_buf(netif);
if (!us) {
ci_netif_unlock(netif);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__,NI_ID(netif)));
return -ENOMEM;
}
/* It's required to set protocol before ci_tcp_helper_sock_attach()
* since it's used to determine if TCP or UDP file operations should be
* attached to the file descriptor in kernel. */
sock_protocol(&us->s) = IPPROTO_UDP;
/* NB: this attach will close the os_sock_fd */
fd = ci_tcp_helper_sock_attach(ci_netif_get_driver_handle(netif),
SC_SP(&us->s), domain, type);
if( fd < 0 ) {
if( fd == -EAFNOSUPPORT )
LOG_U(ci_log("%s: ci_tcp_helper_sock_attach (domain=%d, type=%d) "
"failed %d", __FUNCTION__, domain, type, fd));
else
LOG_E(ci_log("%s: ci_tcp_helper_sock_attach (domain=%d, type=%d) "
"failed %d", __FUNCTION__, domain, type, fd));
ci_netif_unlock(netif);
return fd;
}
ci_assert(~us->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN);
us->s.rx_errno = 0;
us->s.tx_errno = 0;
us->s.so_error = 0;
us->s.cp.sock_cp_flags |= OO_SCP_UDP_WILD;
ep->s = &us->s;
ep->netif = netif;
CHECK_UEP(ep);
ci_netif_unlock(netif);
return fd;
}
ci_fd_t ci_tcp_ep_ctor(citp_socket* ep, ci_netif* netif, int domain, int type)
{
ci_tcp_state* ts;
ci_fd_t fd;
ci_assert(ep);
ci_assert(netif);
ci_netif_lock(netif);
ts = ci_tcp_get_state_buf(netif);
if( ts == NULL ) {
ci_netif_unlock(netif);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__,NI_ID(netif)));
return -ENOMEM;
}
fd = ci_tcp_helper_sock_attach(ci_netif_get_driver_handle(netif), S_SP(ts),
domain, type);
if( fd < 0 ) {
if( fd == -EAFNOSUPPORT )
LOG_U(ci_log("%s: ci_tcp_helper_sock_attach" \
"(domain=%d, type=%d) failed %d",
__FUNCTION__, domain, type, fd));
else
LOG_E(ci_log("%s: ci_tcp_helper_sock_attach" \
"(domain=%d, type=%d) failed %d",
__FUNCTION__, domain, type, fd));
}
else {
ci_assert(~ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN);
/* Apply default sockbuf sizes now we've updated them from the kernel
** defaults. */
ts->s.so.sndbuf = NI_OPTS(netif).tcp_sndbuf_def;
ts->s.so.rcvbuf = NI_OPTS(netif).tcp_rcvbuf_def;
ep->netif = netif;
ep->s = &ts->s;
CHECK_TEP(ep);
}
ci_netif_unlock(netif);
return fd;
}
void ci_tcp_linger(ci_netif* ni, ci_tcp_state* ts)
{
/* This is called at user-level when a socket is closed if linger is
** enabled and has a timeout, and there is TX data outstanding.
**
** Our job is to block until all data is successfully sent and acked, or
** until timeout.
*/
ci_uint64 sleep_seq;
int rc = 0;
ci_uint32 timeout = ts->s.so.linger * 1000;
LOG_TC(log("%s: "NTS_FMT, __FUNCTION__, NTS_PRI_ARGS(ni, ts)));
ci_assert(ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN);
ci_assert(ts->s.b.sb_aflags & CI_SB_AFLAG_IN_SO_LINGER);
ci_assert(ts->s.s_flags & CI_SOCK_FLAG_LINGER);
ci_assert(ts->s.b.state != CI_TCP_LISTEN);
while( 1 ) {
sleep_seq = ts->s.b.sleep_seq.all;
ci_rmb();
if( SEQ_EQ(tcp_enq_nxt(ts), tcp_snd_una(ts)) )
return;
rc = ci_sock_sleep(ni, &ts->s.b, CI_SB_FLAG_WAKE_TX, 0, sleep_seq,
&timeout);
if( rc )
break;
}
if( ! SEQ_EQ(tcp_enq_nxt(ts), tcp_snd_una(ts)) ) {
ci_netif_lock(ni);
/* check we are working with the same socket, and it was not closed and
* dropped under our feet. */
if( ! SEQ_EQ(tcp_enq_nxt(ts), tcp_snd_una(ts)) &&
(ts->s.b.sb_aflags & CI_SB_AFLAG_IN_SO_LINGER) )
ci_tcp_drop(ni, ts, 0);
ci_netif_unlock(ni);
}
}
static int oo_pipe_ctor(ci_netif* netif, struct oo_pipe** out_pipe,
int fds[2], int flags)
{
struct oo_pipe* p;
int rc;
ci_assert(netif);
ci_netif_lock(netif);
p = oo_pipe_buf_get(netif);
if( !p ) {
rc = -1;
errno = ENOMEM;
goto out;
}
if( flags & O_NONBLOCK ) {
p->aflags = (CI_PFD_AFLAG_NONBLOCK << CI_PFD_AFLAG_READER_SHIFT) |
(CI_PFD_AFLAG_NONBLOCK << CI_PFD_AFLAG_WRITER_SHIFT);
}
/* attach */
rc = ci_tcp_helper_pipe_attach(ci_netif_get_driver_handle(netif),
W_SP(&p->b), flags, fds);
if( rc < 0 ) {
LOG_E(ci_log("%s: ci_tcp_helper_pipe_attach %d", __FUNCTION__, rc));
errno = -rc;
rc = -1;
goto out;
}
*out_pipe = p;
out:
ci_netif_unlock(netif);
return rc;
}
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 onload_zc_alloc_buffers(int fd, struct onload_zc_iovec* iovecs,
int iovecs_len,
enum onload_zc_buffer_type_flags flags)
{
int rc = 0, i;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
ci_ip_pkt_fmt *pkt;
unsigned max_len;
Log_CALL(ci_log("%s(%d, %p, %d, %x)", __FUNCTION__, fd, iovecs,
iovecs_len, flags));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < iovecs_len; ++i ) {
max_len = CI_CFG_PKT_BUF_SIZE;
pkt = ci_netif_pkt_tx_tcp_alloc(ni);
if( pkt == NULL ) {
while( --i >= 0 )
ci_netif_pkt_release(ni, (ci_ip_pkt_fmt*)iovecs[i].buf);
rc = -ENOMEM;
ci_netif_unlock(ni);
goto out;
}
/* Make sure this is clear as it affects behaviour when freeing */
pkt->pf.udp.rx_flags = 0;
iovecs[i].buf = (struct oo_zc_buf *)pkt;
if( flags & ONLOAD_ZC_BUFFER_HDR_TCP ) {
if( (citp_fdinfo_get_type(fdi) == CITP_TCP_SOCKET) &&
(epi->sock.s->b.state & CI_TCP_STATE_TCP_CONN) ) {
ci_tcp_state* ts = SOCK_TO_TCP(epi->sock.s);
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base = ((char *)oo_tx_ip_hdr(pkt)) +
ts->outgoing_hdrs_len;
max_len = tcp_eff_mss(ts);
}
else {
/* Best guess. We can fix it up later. Magic 12 leaves
* space for time stamp option (common case)
*/
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base =
(uint8_t*) oo_tx_ip_data(pkt) + sizeof(ci_tcp_hdr) + 12;
}
}
else if( flags & ONLOAD_ZC_BUFFER_HDR_UDP ) {
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base =
(uint8_t*) oo_tx_ip_data(pkt) + sizeof(ci_udp_hdr);
}
else
iovecs[i].iov_base = PKT_START(pkt);
iovecs[i].iov_len = CI_CFG_PKT_BUF_SIZE -
((char *)iovecs[i].iov_base - (char *)pkt);
if( iovecs[i].iov_len > max_len )
iovecs[i].iov_len = max_len;
}
ni->state->n_async_pkts += iovecs_len;
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
case CITP_PASSTHROUGH_FD:
rc = -ESOCKTNOSUPPORT;
break;
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
out:
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
return rc;
}
int onload_zc_release_buffers(int fd, onload_zc_handle* bufs, int bufs_len)
{
int rc = 0, i;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
Log_CALL(ci_log("%s(%d, %p, %d)", __FUNCTION__, fd, bufs, bufs_len));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < bufs_len; ++i ) {
ci_ip_pkt_fmt* pkt = (ci_ip_pkt_fmt*)bufs[i];
if( pkt->stack_id != ni->state->stack_id ) {
LOG_U(log("%s: attempt to free buffer from stack %d to stack %d",
__FUNCTION__, pkt->stack_id, ni->state->stack_id));
rc = -EINVAL;
break;
}
}
if( rc == 0 ) {
for( i = 0; i < bufs_len; ++i )
ci_netif_pkt_release_check_keep(ni, (ci_ip_pkt_fmt*)bufs[i]);
}
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
return rc;
}
int efab_file_move_to_alien_stack_rsop(ci_private_t *stack_priv, void *arg)
{
ci_fixed_descriptor_t sock_fd = *(ci_fixed_descriptor_t *)arg;
struct file *sock_file = fget(sock_fd);
ci_private_t *sock_priv;
tcp_helper_resource_t *old_thr;
tcp_helper_resource_t *new_thr;
citp_waitable *w;
int rc;
if( sock_file == NULL )
return -EINVAL;
if( !FILE_IS_ENDPOINT_SOCK(sock_file) ||
stack_priv->fd_type != CI_PRIV_TYPE_NETIF ) {
fput(sock_file);
return -EINVAL;
}
sock_priv = sock_file->private_data;
ci_assert(sock_priv->fd_type == CI_PRIV_TYPE_TCP_EP ||
sock_priv->fd_type == CI_PRIV_TYPE_UDP_EP);
old_thr = sock_priv->thr;
new_thr = stack_priv->thr;
ci_assert(old_thr);
ci_assert(new_thr);
if( old_thr == new_thr ) {
fput(sock_file);
return 0;
}
if( tcp_helper_cluster_from_cluster(old_thr) != 0 ) {
LOG_S(ci_log("%s: move_fd() not permitted on clustered stacks", __func__));
fput(sock_file);
return -EINVAL;
}
w = SP_TO_WAITABLE(&old_thr->netif, sock_priv->sock_id);
rc = ci_sock_lock(&old_thr->netif, w);
if( rc != 0 ) {
fput(sock_file);
return rc;
}
rc = ci_netif_lock(&old_thr->netif);
if( rc != 0 ) {
ci_sock_unlock(&old_thr->netif, w);
fput(sock_file);
return rc;
}
efab_thr_ref(new_thr);
rc = efab_file_move_to_alien_stack(sock_priv, &stack_priv->thr->netif);
fput(sock_file);
if( rc != 0 )
efab_thr_release(new_thr);
else
ci_netif_unlock(&new_thr->netif);
return rc;
}
/* 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 onload_zc_release_buffers(int fd, onload_zc_handle* bufs, int bufs_len)
{
int rc = 0, i, rx_pkt, released;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
ci_ip_pkt_fmt* pkt;
Log_CALL(ci_log("%s(%d, %p, %d)", __FUNCTION__, fd, bufs, bufs_len));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < bufs_len; ++i ) {
pkt = (ci_ip_pkt_fmt*)bufs[i];
if( pkt->stack_id != ni->state->stack_id ) {
LOG_U(log("%s: attempt to free buffer from stack %d to stack %d",
__FUNCTION__, pkt->stack_id, ni->state->stack_id));
rc = -EINVAL;
break;
}
}
if( rc == 0 ) {
for( i = 0; i < bufs_len; ++i ) {
pkt = (ci_ip_pkt_fmt*)bufs[i];
/* If we are releasing a packet without the RX_FLAG then the user
* allocated and then freed the packet (without using it).
* We detect this to decrement n_asyn_pkts.
* RX packets (kept via ONLOAD_ZC_KEEP) are counted differently
* so don't decrement here. (But may release)
*/
rx_pkt = pkt->flags & CI_PKT_FLAG_RX;
released = ci_netif_pkt_release_check_keep(ni, pkt);
if ( ! rx_pkt ) {
ci_assert(released == 1);
(void) released;
--ni->state->n_async_pkts;
}
}
}
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
return rc;
}
static int ci_udp_ioctl_locked(ci_netif* ni, ci_udp_state* us,
ci_fd_t fd, int request, void* arg)
{
int rc;
switch( request ) {
case FIONREAD: /* synonym of SIOCINQ */
if( ! CI_IOCTL_ARG_OK(int, arg) )
return -EFAULT;
rc = 1;
#if CI_CFG_ZC_RECV_FILTER
# ifndef __KERNEL__
if( us->recv_q_filter ) {
ci_netif_unlock(ni);
ci_sock_lock(ni, &us->s.b);
rc = ci_udp_recv_q_readable(ni, us);
ci_sock_unlock(ni, &us->s.b);
ci_netif_lock(ni);
}
# endif
#endif
if( rc ) {
/* Return the size of the datagram at the head of the receive queue.
*
* Careful: extract side of receive queue is owned by sock lock,
* which we don't have. However, freeing of bufs is owned by netif
* lock, which we do have. So we're safe so long as we only read
* [extract] once.
*/
oo_pkt_p extract = us->recv_q.extract;
if( OO_PP_NOT_NULL(extract) ) {
ci_ip_pkt_fmt* pkt = PKT_CHK(ni, extract);
if( (pkt->pf.udp.rx_flags & CI_IP_PKT_FMT_PREFIX_UDP_RX_CONSUMED) &&
OO_PP_NOT_NULL(pkt->next) )
pkt = PKT_CHK(ni, pkt->next);
if( !(pkt->pf.udp.rx_flags & CI_IP_PKT_FMT_PREFIX_UDP_RX_CONSUMED) ) {
*(int*) arg = pkt->pf.udp.pay_len;
return 0;
}
}
}
/* Nothing in userlevel receive queue: So take the value returned by
* the O/S socket.
*/
if( !(us->s.os_sock_status & OO_OS_STATUS_RX) ) {
*(int*)arg = 0;
return 0;
}
goto sys_ioctl;
case TIOCOUTQ: /* synonym of SIOCOUTQ */
if( ! CI_IOCTL_ARG_OK(int, arg) )
return -EFAULT;
*(int*)arg = us->tx_count + oo_atomic_read(&us->tx_async_q_level);
return 0;
case SIOCGSTAMP:
#if defined( __linux__) && defined(__KERNEL__)
/* The following code assumes the width of the timespec and timeval fields */
# error "Need to consider 32-on-64 bit setting of timeval arg"
#endif
if( ! (us->udpflags & CI_UDPF_LAST_RECV_ON) )
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
return ci_udp_ioctl_siocgstamp(ni, us, arg, 1);
case SIOCGSTAMPNS:
if( ! (us->udpflags & CI_UDPF_LAST_RECV_ON) )
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
return ci_udp_ioctl_siocgstamp(ni, us, arg, 0);
}
return ci_udp_ioctl_slow(ni, us, fd, request, arg);
sys_ioctl:
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
}
/* c_ni is assumed to be locked on enterance and is always unlocked on
* exit. */
int ci_tcp_connect_lo_toconn(ci_netif *c_ni, oo_sp c_id, ci_uint32 dst,
ci_netif *l_ni, oo_sp l_id)
{
ci_tcp_state *ts;
ci_tcp_socket_listen *tls, *alien_tls;
citp_waitable_obj *wo;
citp_waitable *w;
int rc;
ci_assert(ci_netif_is_locked(c_ni));
ci_assert(OO_SP_NOT_NULL(c_id));
ci_assert(OO_SP_NOT_NULL(l_id));
LOG_TC(log("%s: connect %d:%d to %d:%d", __FUNCTION__,
c_ni->state->stack_id, OO_SP_TO_INT(c_id),
l_ni->state->stack_id, OO_SP_TO_INT(l_id)));
alien_tls = SP_TO_TCP_LISTEN(l_ni, l_id);
if( (int)ci_tcp_acceptq_n(alien_tls) >= alien_tls->acceptq_max ) {
ci_netif_unlock(c_ni);
return -EBUSY;
}
/* In c_ni, create shadow listening socket tls (copy l_id) */
ts = ci_tcp_get_state_buf(c_ni);
if( ts == NULL ) {
ci_netif_unlock(c_ni);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__, NI_ID(c_ni)));
return -ENOMEM;
}
/* init common tcp fields */
ts->s.so = alien_tls->s.so;
ts->s.cp.ip_ttl = alien_tls->s.cp.ip_ttl;
S_TCP_HDR(&ts->s)->tcp_source_be16 =
S_TCP_HDR(&alien_tls->s)->tcp_source_be16;
ts->s.domain = alien_tls->s.domain;
ts->c = alien_tls->c;
ts->c.tcp_defer_accept = OO_TCP_DEFER_ACCEPT_OFF;
/* make sure nobody will ever connect to our "shadow" socket
* except us */
ci_bit_set(&ts->s.b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
ci_tcp_set_slow_state(c_ni, ts, CI_TCP_LISTEN);
tls = SOCK_TO_TCP_LISTEN(&ts->s);
/* no timer: */
tls->s.s_flags = alien_tls->s.s_flags | CI_SOCK_FLAG_BOUND_ALIEN;
tls->acceptq_max = 1;
rc = ci_tcp_listen_init(c_ni, tls);
if( rc != 0 ) {
citp_waitable_obj_free(c_ni, &tls->s.b);
return rc;
}
/* Connect c_id to tls */
ts = SP_TO_TCP(c_ni, c_id);
rc = ci_tcp_connect_lo_samestack(c_ni, ts, tls->s.b.bufid);
/* Accept as from tls */
if( !ci_tcp_acceptq_not_empty(tls) ) {
/* it is possible, for example, if ci_tcp_listenq_try_promote() failed
* because there are no endpoints */
ci_tcp_listenq_drop_all(c_ni, tls);
citp_waitable_obj_free(c_ni, &tls->s.b);
ci_netif_unlock(c_ni);
return -EBUSY;
}
w = ci_tcp_acceptq_get(c_ni, tls);
ci_assert(w);
LOG_TV(ci_log("%s: %d:%d to %d:%d shadow %d:%d accepted %d:%d",
__FUNCTION__,
c_ni->state->stack_id, OO_SP_TO_INT(c_id),
l_ni->state->stack_id, OO_SP_TO_INT(l_id),
c_ni->state->stack_id, tls->s.b.bufid,
c_ni->state->stack_id, w->bufid));
ci_assert(w->state & CI_TCP_STATE_TCP);
ci_assert(w->state != CI_TCP_LISTEN);
/* Destroy tls.
* NB: nobody could possibly connect to it, so no need to do proper
* shutdown.
*/
ci_assert_equal(ci_tcp_acceptq_n(tls), 0);
ci_tcp_listenq_drop_all(c_ni, tls);
citp_waitable_obj_free(c_ni, &tls->s.b);
ci_netif_unlock(c_ni);
/* Keep a port reference */
{
tcp_helper_endpoint_t *l_ep, *a_ep;
struct oo_file_ref* os_sock_ref;
ci_irqlock_state_t lock_flags;
l_ep = ci_trs_ep_get(netif2tcp_helper_resource(l_ni), l_id);
a_ep = ci_trs_ep_get(netif2tcp_helper_resource(c_ni), W_SP(w));
ci_irqlock_lock(&l_ep->thr->lock, &lock_flags);
os_sock_ref = l_ep->os_socket;
ci_assert_equal(a_ep->os_port_keeper, NULL);
if( os_sock_ref != NULL ) {
os_sock_ref = oo_file_ref_add(os_sock_ref);
os_sock_ref = oo_file_ref_xchg(&a_ep->os_port_keeper, os_sock_ref);
ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
if( os_sock_ref != NULL )
oo_file_ref_drop(os_sock_ref);
}
else {
ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
goto cleanup;
}
}
/* lock l_ni: Check that l_id is the same socket it used to be */
/* create ref-sock in l_ni, put it into acc q */
if( ci_netif_lock(l_ni) != 0 )
goto cleanup;
if( alien_tls->s.b.state != CI_TCP_LISTEN ||
(alien_tls->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN) ||
S_TCP_HDR(&alien_tls->s)->tcp_source_be16 != TS_TCP(ts)->tcp_dest_be16 ||
(alien_tls->s.pkt.ip.ip_saddr_be32 != INADDR_ANY &&
alien_tls->s.pkt.ip.ip_saddr_be32 != ts->s.pkt.ip.ip_daddr_be32) ) {
ci_netif_unlock(l_ni);
goto cleanup;
}
ci_bit_mask_set(&w->sb_aflags,
CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);
wo = citp_waitable_obj_alloc(l_ni);
if( wo == NULL ) {
ci_netif_unlock(l_ni);
goto cleanup;
}
wo->waitable.state = CI_TCP_CLOSED;
wo->waitable.sb_aflags |= CI_SB_AFLAG_MOVED_AWAY;
wo->waitable.moved_to_stack_id = c_ni->state->stack_id;
wo->waitable.moved_to_sock_id = W_SP(w);
LOG_TC(log("%s: put to acceptq %d:%d referencing %d:%d", __func__,
l_ni->state->stack_id, OO_SP_TO_INT(W_SP(&wo->waitable)),
c_ni->state->stack_id, OO_SP_TO_INT(W_SP(w))));
ci_tcp_acceptq_put(l_ni, alien_tls, &wo->waitable);
citp_waitable_wake_not_in_poll(l_ni, &alien_tls->s.b, CI_SB_FLAG_WAKE_RX);
ci_netif_unlock(l_ni);
return rc;
cleanup:
ci_assert(w->sb_aflags & CI_SB_AFLAG_ORPHAN);
ci_bit_mask_clear(&w->sb_aflags,
CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);
efab_tcp_helper_close_endpoint(netif2tcp_helper_resource(c_ni), w->bufid);
/* we can not guarantee c_ni lock, so we can' call
* ci_tcp_drop(c_ni, ts). So, we return error; UL will handover
* and close ts endpoint. */
return -EBUSY;
}
/* Returns:
* 0 on success
*
* CI_SOCKET_ERROR (and errno set)
* this is a normal error that is returned to the
* the application
*
* CI_SOCKET_HANDOVER we tell the upper layers to handover, no need
* to set errno since it isn't a real error
*/
int ci_tcp_connect(citp_socket* ep, const struct sockaddr* serv_addr,
socklen_t addrlen, ci_fd_t fd, int *p_moved)
{
/* Address family is validated earlier. */
struct sockaddr_in* inaddr = (struct sockaddr_in*) serv_addr;
ci_sock_cmn* s = ep->s;
ci_tcp_state* ts = &SOCK_TO_WAITABLE_OBJ(s)->tcp;
int rc = 0, crc;
ci_uint32 dst_be32;
if( NI_OPTS(ep->netif).tcp_connect_handover )
return CI_SOCKET_HANDOVER;
/* Make sure we're up-to-date. */
ci_netif_lock(ep->netif);
CHECK_TEP(ep);
ci_netif_poll(ep->netif);
/*
* 1. Check if state of the socket is OK for connect operation.
*/
start_again:
if( (rc = ci_tcp_connect_handle_so_error(s)) != 0) {
CI_SET_ERROR(rc, rc);
goto unlock_out;
}
if( s->b.state != CI_TCP_CLOSED ) {
/* see if progress can be made on this socket before
** determining status (e.g. non-blocking connect and connect poll)*/
if( s->b.state & CI_TCP_STATE_SYNCHRONISED ) {
if( ts->tcpflags & CI_TCPT_FLAG_NONBLOCK_CONNECT ) {
ts->tcpflags &= ~CI_TCPT_FLAG_NONBLOCK_CONNECT;
rc = 0;
goto unlock_out;
}
if( serv_addr->sa_family == AF_UNSPEC )
LOG_E(ci_log("Onload does not support TCP disconnect via "
"connect(addr->sa_family==AF_UNSPEC)"));
CI_SET_ERROR(rc, EISCONN);
}
else if( s->b.state == CI_TCP_LISTEN ) {
#if CI_CFG_POSIX_CONNECT_AFTER_LISTEN
CI_SET_ERROR(rc, EOPNOTSUPP);
#else
if( ci_tcp_validate_sa(s->domain, serv_addr, addrlen) ) {
/* Request should be forwarded to OS */
rc = CI_SOCKET_HANDOVER;
goto unlock_out;
}
if( serv_addr->sa_family == AF_UNSPEC ) {
/* Linux does listen shutdown on disconnect (AF_UNSPEC) */
ci_netif_unlock(ep->netif);
rc = ci_tcp_shutdown(ep, SHUT_RD, fd);
goto out;
} else {
/* Linux has curious error reporting in this case */
CI_SET_ERROR(rc, EISCONN);
}
#endif
}
else {
/* Socket is in SYN-SENT state. Let's block for receiving SYN-ACK */
ci_assert_equal(s->b.state, CI_TCP_SYN_SENT);
if( s->b.sb_aflags & (CI_SB_AFLAG_O_NONBLOCK | CI_SB_AFLAG_O_NDELAY) )
CI_SET_ERROR(rc, EALREADY);
else
goto syn_sent;
}
goto unlock_out;
}
/* Check if we've ever been connected. */
if( ts->tcpflags & CI_TCPT_FLAG_WAS_ESTAB ) {
CI_SET_ERROR(rc, EISCONN);
goto unlock_out;
}
/*
* 2. Check address parameter, if it's inappropriate for handover
* decision or handover should be done, try to to call OS and
* do handover on success.
*/
if (
/* Af first, check that address family and length is OK. */
ci_tcp_validate_sa(s->domain, serv_addr, addrlen)
/* rfc793 p54 if the foreign socket is unspecified return */
/* "error: foreign socket unspecified" (EINVAL), but keep it to OS */
|| (dst_be32 = ci_get_ip4_addr(inaddr->sin_family, serv_addr)) == 0
/* Zero destination port is tricky as well, keep it to OS */
|| inaddr->sin_port == 0 )
{
rc = CI_SOCKET_HANDOVER;
goto unlock_out;
}
/* is this a socket that we can handle? */
rc = ci_tcp_connect_check_dest(ep, dst_be32, inaddr->sin_port);
if( rc ) goto unlock_out;
if( (ts->s.pkt.flags & CI_IP_CACHE_IS_LOCALROUTE) &&
OO_SP_IS_NULL(ts->local_peer) ) {
/* Try to connect to another stack; handover if can't */
struct oo_op_loopback_connect op;
op.dst_port = inaddr->sin_port;
op.dst_addr = dst_be32;
/* this operation unlocks netif */
rc = oo_resource_op(fd, OO_IOC_TCP_LOOPBACK_CONNECT, &op);
if( rc < 0)
return CI_SOCKET_HANDOVER;
if( op.out_moved )
*p_moved = 1;
if( op.out_rc == -EINPROGRESS )
RET_WITH_ERRNO( EINPROGRESS );
else if( op.out_rc == -EAGAIN )
return -EAGAIN;
else if( op.out_rc != 0 )
return CI_SOCKET_HANDOVER;
return 0;
}
/* filters can't handle alien source address */
if( (s->s_flags & CI_SOCK_FLAG_BOUND_ALIEN) &&
! (ts->s.pkt.flags & CI_IP_CACHE_IS_LOCALROUTE) ) {
rc = CI_SOCKET_HANDOVER;
goto unlock_out;
}
crc = ci_tcp_connect_ul_start(ep->netif, ts, dst_be32, inaddr->sin_port, &rc);
if( crc != CI_CONNECT_UL_OK ) {
switch( crc ) {
case CI_CONNECT_UL_FAIL:
goto unlock_out;
case CI_CONNECT_UL_LOCK_DROPPED:
goto out;
case CI_CONNECT_UL_START_AGAIN:
goto start_again;
}
}
CI_TCP_STATS_INC_ACTIVE_OPENS( ep->netif );
syn_sent:
rc = ci_tcp_connect_ul_syn_sent(ep->netif, ts);
unlock_out:
ci_netif_unlock(ep->netif);
out:
return rc;
}
/* This function must be called with netif lock not held and it always
* returns with the netif lock not held.
*/
int efab_tcp_helper_reuseport_bind(ci_private_t *priv, void *arg)
{
oo_tcp_reuseport_bind_t* trb = arg;
ci_netif* ni = &priv->thr->netif;
tcp_helper_cluster_t* thc;
tcp_helper_resource_t* thr = NULL;
citp_waitable* waitable;
ci_sock_cmn* sock = SP_TO_SOCK(ni, priv->sock_id);
struct oof_manager* fm = efab_tcp_driver.filter_manager;
struct oof_socket* oofilter;
struct oof_socket dummy_oofilter;
int protocol = thc_get_sock_protocol(sock);
char name[CI_CFG_CLUSTER_NAME_LEN + 1];
int rc, rc1;
int flags = 0;
tcp_helper_cluster_t* named_thc,* ported_thc;
int alloced = 0;
/* No clustering on sockets bound to alien addresses */
if( sock->s_flags & CI_SOCK_FLAG_BOUND_ALIEN )
return 0;
if( NI_OPTS(ni).cluster_ignore == 1 ) {
LOG_NV(ci_log("%s: Ignored attempt to use clusters due to "
"EF_CLUSTER_IGNORE option.", __FUNCTION__));
return 0;
}
if( trb->port_be16 == 0 ) {
ci_log("%s: Reuseport on port=0 is not supported", __FUNCTION__);
return -EINVAL;
}
if( trb->cluster_size < 2 ) {
ci_log("%s: Cluster sizes < 2 are not supported", __FUNCTION__);
return -EINVAL;
}
if( sock->s_flags & (CI_SOCK_FLAG_TPROXY | CI_SOCK_FLAG_MAC_FILTER) ) {
ci_log("%s: Scalable filter sockets cannot be clustered",
__FUNCTION__);
return -EINVAL;
}
oofilter = &ci_trs_ep_get(priv->thr, priv->sock_id)->oofilter;
if( oofilter->sf_local_port != NULL ) {
ci_log("%s: Socket that already have filter cannot be clustered",
__FUNCTION__);
return -EINVAL;
}
if( priv->thr->thc ) {
/* Reserve proto:port[:ip] until bind (or close)*/
rc = oof_socket_add(fm, oofilter,
OOF_SOCKET_ADD_FLAG_CLUSTERED |
OOF_SOCKET_ADD_FLAG_DUMMY,
protocol, trb->addr_be32, trb->port_be16, 0, 0,
&ported_thc);
if( rc > 0 )
rc = 0;
if( rc == 0 )
sock->s_flags |= CI_SOCK_FLAG_FILTER;
return rc;
}
mutex_lock(&thc_init_mutex);
/* We are going to be iterating over clusters, make sure they don't
* change.
*/
mutex_lock(&thc_mutex);
/* Lookup a suitable cluster to use */
/* We try to add dummy filter to oof to reserve proto:port[:ip] tuple,
* if there is already a cluster at the tuple we will get reference to it,
*/
oof_socket_ctor(&dummy_oofilter);
rc = oof_socket_add(fm, &dummy_oofilter,
OOF_SOCKET_ADD_FLAG_CLUSTERED |
OOF_SOCKET_ADD_FLAG_DUMMY |
OOF_SOCKET_ADD_FLAG_NO_STACK,
protocol, trb->addr_be32, trb->port_be16, 0, 0,
&ported_thc);
if( rc < 0 ) /* non-clustered socket on the tuple */
goto alloc_fail0;
if( ! gen_cluster_name(trb->cluster_name, name) ) {
/* user requested a cluster by name. But we need to make sure
* that the oof_local_port that the user is interested in is not
* being used by another cluster. We search for cluster by name
* and use results of prior protp:port[:ip] search oof_local_port
* to then do some sanity checking.
*/
rc1 = thc_search_by_name(name, protocol, trb->port_be16, ci_geteuid(),
&named_thc);
if( rc1 < 0 ) {
rc = rc1;
goto alloc_fail;
}
if( rc1 == 0 ) {
if( rc == 1 ) {
/* search by oof_local_port found a cluster which search by
* name didn't find. */
LOG_E(ci_log("Error: Cluster with requested name %s already "
"bound to %s", name, ported_thc->thc_name));
rc = -EEXIST;
goto alloc_fail;
}
else {
/* Neither searches found a cluster. So allocate one below.
*/
}
}
else {
if( rc == 1 ) {
/* Both searches found clusters. Fine if they are the same or
* else error. */
if( named_thc != ported_thc ) {
LOG_E(ci_log("Error: Cluster %s does not handle socket %s:%d. "
"Cluster %s does", name, FMT_PROTOCOL(protocol),
trb->port_be16, named_thc->thc_name));
rc = -EEXIST;
goto alloc_fail;
}
}
/* Search by name found a cluster no conflict with search by tuple
* (the ported cluster is either none or the same as named)*/
thc = named_thc;
goto cont;
}
}
else {
/* No cluster name requested. We have already looked for a cluster handling
* the tuple. If none found, then try to use an existing
* cluster this process created. If none found, then allocate one.
*/
/* If rc == 0, then no cluster found - try to allocate one.
* If rc == 1, we found cluster - make sure that euids match and continue. */
if( rc == 1 ) {
thc = ported_thc;
if( thc->thc_euid != ci_geteuid() ) {
rc = -EADDRINUSE;
goto alloc_fail;
}
goto cont;
}
rc = thc_search_by_name(name, protocol, trb->port_be16, ci_geteuid(),
&thc);
if( rc < 0 )
goto alloc_fail;
if( rc == 1 )
goto cont;
}
/* When an interface is in tproxy mode, all clustered listening socket
* are assumed to be part of tproxy passive side. This requires
* rss context to use altered rss hashing based solely on src ip:port.
*/
flags = tcp_helper_cluster_thc_flags(&NI_OPTS(ni));
if( (rc = thc_alloc(name, protocol, trb->port_be16, ci_geteuid(),
trb->cluster_size, flags, &thc)) != 0 )
goto alloc_fail;
alloced = 1;
cont:
tcp_helper_cluster_ref(thc);
/* At this point we have our cluster with one additional reference */
/* Find a suitable stack within the cluster to use */
rc = thc_get_thr(thc, &dummy_oofilter, &thr);
if( rc != 0 )
rc = thc_alloc_thr(thc, trb->cluster_restart_opt,
&ni->opts, ni->flags, &thr);
/* If get or alloc succeeded thr holds reference to the cluster,
* so the cluster cannot go away. We'll drop our reference and also
* will not be accessing state within the cluster anymore so we can
* drop the lock. */
mutex_unlock(&thc_mutex);
if( alloced && rc == 0 && (flags & THC_FLAG_TPROXY) != 0 ) {
/* Tproxy filter is allocated as late as here,
* the reason is that this needs to be preceded by stack allocation
* (firmware needs initialized vi) */
rc = thc_install_tproxy(thc, NI_OPTS(ni).scalable_filter_ifindex);
if( rc != 0 )
efab_thr_release(thr);
}
tcp_helper_cluster_release(thc, NULL);
if( rc != 0 ) {
oof_socket_del(fm, &dummy_oofilter);
goto alloc_fail_unlocked;
}
/* We have thr and we hold single reference to it. */
/* Move the socket into the new stack */
if( (rc = ci_netif_lock(ni)) != 0 )
goto drop_and_done;
waitable = SP_TO_WAITABLE(ni, priv->sock_id);
rc = ci_sock_lock(ni, waitable);
if( rc != 0 ) {
ci_netif_unlock(ni);
goto drop_and_done;
}
/* thr referencing scheme comes from efab_file_move_to_alien_stack_rsop */
efab_thr_ref(thr);
rc = efab_file_move_to_alien_stack(priv, &thr->netif, 0);
if( rc != 0 )
efab_thr_release(thr);
else {
/* beside us, socket now holds its own reference to thr */
oofilter = &ci_trs_ep_get(thr, sock->b.moved_to_sock_id)->oofilter;
oof_socket_replace(fm, &dummy_oofilter, oofilter);
SP_TO_SOCK(&thr->netif, sock->b.moved_to_sock_id)->s_flags |= CI_SOCK_FLAG_FILTER;
ci_netif_unlock(&thr->netif);
}
drop_and_done:
if( rc != 0 )
oof_socket_del(fm, &dummy_oofilter);
/* Drop the reference we got from thc_get_thr or thc_alloc_thr().
* If things went wrong both stack and cluster might disappear. */
efab_thr_release(thr);
oof_socket_dtor(&dummy_oofilter);
mutex_unlock(&thc_init_mutex);
return rc;
alloc_fail:
oof_socket_del(fm, &dummy_oofilter);
alloc_fail0:
mutex_unlock(&thc_mutex);
alloc_fail_unlocked:
oof_socket_dtor(&dummy_oofilter);
mutex_unlock(&thc_init_mutex);
return rc;
}
