ssize_t linux_tcp_helper_fop_sendpage(struct file* filp, struct page* page,
int offset, size_t size,
loff_t* ppos, int flags)
{
ci_private_t* priv = filp->private_data;
tcp_helper_resource_t* trs = efab_priv_to_thr(priv);
ci_sock_cmn* s;
OO_DEBUG_VERB(ci_log("%s: %d:%d offset=%d size=%d flags=%x", __FUNCTION__,
NI_ID(&trs->netif), OO_SP_FMT(priv->sock_id), offset,
(int) size, flags));
ci_assert(page);
ci_assert_ge(offset, 0);
ci_assert_gt(size, 0);
ci_assert_le(offset + size, CI_PAGE_SIZE);
#ifndef MSG_SENDPAGE_NOTLAST
/* "flags" is really "more". Convert it. */
if( flags )
flags = MSG_MORE;
/* [more] is sometimes true even for the last page. We get a little
** closer to the truth by spotting that we're not reading to the end of
** the page. - seen on 2.6.18, but not on 2.6.26 or later
*/
if( offset + size < CI_PAGE_SIZE && flags )
flags = 0;
#endif
s = SP_TO_SOCK(&trs->netif, priv->sock_id);
if(CI_LIKELY( s->b.state & CI_TCP_STATE_TCP_CONN ))
return sendpage_copy(&trs->netif,SOCK_TO_TCP(s),page,offset,size,flags);
else
/* Closed or listening. Return epipe. Do not send SIGPIPE, because
** Linux will do it for us. */
return -s->tx_errno;
}
/* 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;
}
/* 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;
}
