4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lnet/klnds/socklnd/socklnd.c
33 * Author: Zach Brown <zab@zabbo.net>
34 * Author: Peter J. Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
36 * Author: Eric Barton <eric@bartonsoftware.com>
39 #include <linux/ethtool.h>
40 #include <linux/inetdevice.h>
42 #include <linux/sunrpc/addr.h>
44 static const struct lnet_lnd the_ksocklnd;
45 struct ksock_nal_data ksocknal_data;
47 static struct ksock_interface *
48 ksocknal_ip2iface(struct lnet_ni *ni, struct sockaddr *addr)
50 struct ksock_net *net = ni->ni_data;
51 struct ksock_interface *iface;
53 iface = &net->ksnn_interface;
55 if (rpc_cmp_addr((struct sockaddr *)&iface->ksni_addr, addr))
61 static struct ksock_interface *
62 ksocknal_index2iface(struct lnet_ni *ni, int index)
64 struct ksock_net *net = ni->ni_data;
65 struct ksock_interface *iface;
67 iface = &net->ksnn_interface;
69 if (iface->ksni_index == index)
75 static int ksocknal_ip2index(struct sockaddr *addr, struct lnet_ni *ni)
77 struct net_device *dev;
79 DECLARE_CONST_IN_IFADDR(ifa);
81 if (addr->sa_family != AF_INET)
82 /* No IPv6 support yet */
86 for_each_netdev(ni->ni_net_ns, dev) {
87 int flags = dev_get_flags(dev);
88 struct in_device *in_dev;
90 if (flags & IFF_LOOPBACK) /* skip the loopback IF */
93 if (!(flags & IFF_UP))
96 in_dev = __in_dev_get_rcu(dev);
100 in_dev_for_each_ifa_rcu(ifa, in_dev) {
101 if (ifa->ifa_local ==
102 ((struct sockaddr_in *)addr)->sin_addr.s_addr)
114 static struct ksock_conn_cb *
115 ksocknal_create_conn_cb(struct sockaddr *addr)
117 struct ksock_conn_cb *conn_cb;
119 LIBCFS_ALLOC(conn_cb, sizeof(*conn_cb));
123 refcount_set(&conn_cb->ksnr_refcount, 1);
124 conn_cb->ksnr_peer = NULL;
125 conn_cb->ksnr_retry_interval = 0; /* OK to connect at any time */
126 rpc_copy_addr((struct sockaddr *)&conn_cb->ksnr_addr, addr);
127 rpc_set_port((struct sockaddr *)&conn_cb->ksnr_addr,
129 conn_cb->ksnr_myiface = -1;
130 conn_cb->ksnr_scheduled = 0;
131 conn_cb->ksnr_connecting = 0;
132 conn_cb->ksnr_connected = 0;
133 conn_cb->ksnr_deleted = 0;
134 conn_cb->ksnr_conn_count = 0;
135 conn_cb->ksnr_ctrl_conn_count = 0;
136 conn_cb->ksnr_blki_conn_count = 0;
137 conn_cb->ksnr_blko_conn_count = 0;
138 conn_cb->ksnr_max_conns = 0;
144 ksocknal_destroy_conn_cb(struct ksock_conn_cb *conn_cb)
146 LASSERT(refcount_read(&conn_cb->ksnr_refcount) == 0);
148 if (conn_cb->ksnr_peer)
149 ksocknal_peer_decref(conn_cb->ksnr_peer);
151 LIBCFS_FREE(conn_cb, sizeof(*conn_cb));
154 static struct ksock_peer_ni *
155 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_process_id id)
157 int cpt = lnet_cpt_of_nid(id.nid, ni);
158 struct ksock_net *net = ni->ni_data;
159 struct ksock_peer_ni *peer_ni;
161 LASSERT(id.nid != LNET_NID_ANY);
162 LASSERT(id.pid != LNET_PID_ANY);
163 LASSERT(!in_interrupt());
165 if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
166 CERROR("Can't create peer_ni: network shutdown\n");
167 return ERR_PTR(-ESHUTDOWN);
170 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
172 atomic_dec(&net->ksnn_npeers);
173 return ERR_PTR(-ENOMEM);
176 peer_ni->ksnp_ni = ni;
177 peer_ni->ksnp_id = id;
178 refcount_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
179 peer_ni->ksnp_closing = 0;
180 peer_ni->ksnp_accepting = 0;
181 peer_ni->ksnp_proto = NULL;
182 peer_ni->ksnp_last_alive = 0;
183 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
184 peer_ni->ksnp_conn_cb = NULL;
186 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
187 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
188 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
189 spin_lock_init(&peer_ni->ksnp_lock);
195 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
197 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
199 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
200 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
202 LASSERT(refcount_read(&peer_ni->ksnp_refcount) == 0);
203 LASSERT(peer_ni->ksnp_accepting == 0);
204 LASSERT(list_empty(&peer_ni->ksnp_conns));
205 LASSERT(peer_ni->ksnp_conn_cb == NULL);
206 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
207 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
209 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
211 /* NB a peer_ni's connections and conn_cb keep a reference on their
212 * peer_ni until they are destroyed, so we can be assured that _all_
213 * state to do with this peer_ni has been cleaned up when its refcount
216 if (atomic_dec_and_test(&net->ksnn_npeers))
217 wake_up_var(&net->ksnn_npeers);
220 struct ksock_peer_ni *
221 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
223 struct ksock_peer_ni *peer_ni;
224 unsigned long hash = nidhash(id.nid);
226 hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
228 LASSERT(!peer_ni->ksnp_closing);
230 if (peer_ni->ksnp_ni != ni)
233 if (peer_ni->ksnp_id.nid != id.nid ||
234 peer_ni->ksnp_id.pid != id.pid)
237 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
238 peer_ni, libcfs_id2str(id),
239 refcount_read(&peer_ni->ksnp_refcount));
245 struct ksock_peer_ni *
246 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
248 struct ksock_peer_ni *peer_ni;
250 read_lock(&ksocknal_data.ksnd_global_lock);
251 peer_ni = ksocknal_find_peer_locked(ni, id);
252 if (peer_ni != NULL) /* +1 ref for caller? */
253 ksocknal_peer_addref(peer_ni);
254 read_unlock(&ksocknal_data.ksnd_global_lock);
260 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
263 struct ksock_interface *iface;
265 for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
266 struct sockaddr_in sa = { .sin_family = AF_INET };
267 LASSERT(i < LNET_INTERFACES_NUM);
268 sa.sin_addr.s_addr = htonl(peer_ni->ksnp_passive_ips[i]);
270 iface = ksocknal_ip2iface(peer_ni->ksnp_ni,
271 (struct sockaddr *)&sa);
273 * All IPs in peer_ni->ksnp_passive_ips[] come from the
274 * interface list, therefore the call must succeed.
276 LASSERT(iface != NULL);
278 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
279 peer_ni, iface, iface->ksni_nroutes);
280 iface->ksni_npeers--;
283 LASSERT(list_empty(&peer_ni->ksnp_conns));
284 LASSERT(peer_ni->ksnp_conn_cb == NULL);
285 LASSERT(!peer_ni->ksnp_closing);
286 peer_ni->ksnp_closing = 1;
287 hlist_del(&peer_ni->ksnp_list);
288 /* lose peerlist's ref */
289 ksocknal_peer_decref(peer_ni);
293 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
294 struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
295 int *port, int *conn_count, int *share_count)
297 struct ksock_peer_ni *peer_ni;
298 struct ksock_conn_cb *conn_cb;
303 read_lock(&ksocknal_data.ksnd_global_lock);
305 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
307 if (peer_ni->ksnp_ni != ni)
310 if (peer_ni->ksnp_n_passive_ips == 0 &&
311 peer_ni->ksnp_conn_cb == NULL) {
315 *id = peer_ni->ksnp_id;
325 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
329 *id = peer_ni->ksnp_id;
330 *myip = peer_ni->ksnp_passive_ips[j];
339 if (peer_ni->ksnp_conn_cb) {
343 conn_cb = peer_ni->ksnp_conn_cb;
345 *id = peer_ni->ksnp_id;
346 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
347 struct sockaddr_in *sa =
348 (void *)&conn_cb->ksnr_addr;
350 rc = choose_ipv4_src(myip,
351 conn_cb->ksnr_myiface,
352 ntohl(sa->sin_addr.s_addr),
354 *peer_ip = ntohl(sa->sin_addr.s_addr);
355 *port = ntohs(sa->sin_port);
358 *peer_ip = 0xFFFFFFFF;
362 *conn_count = conn_cb->ksnr_conn_count;
368 read_unlock(&ksocknal_data.ksnd_global_lock);
373 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
376 unsigned int count = 0;
379 case SOCKLND_CONN_CONTROL:
380 count = conn_cb->ksnr_ctrl_conn_count;
382 case SOCKLND_CONN_BULK_IN:
383 count = conn_cb->ksnr_blki_conn_count;
385 case SOCKLND_CONN_BULK_OUT:
386 count = conn_cb->ksnr_blko_conn_count;
388 case SOCKLND_CONN_ANY:
389 count = conn_cb->ksnr_conn_count;
400 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
402 struct lnet_ni *ni = peer_ni->ksnp_ni;
403 struct lnet_ioctl_config_socklnd_tunables *tunables;
407 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
409 return tunables->lnd_conns_per_peer;
413 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
416 conn_cb->ksnr_conn_count++;
418 /* check if all connections of the given type got created */
420 case SOCKLND_CONN_CONTROL:
421 conn_cb->ksnr_ctrl_conn_count++;
422 /* there's a single control connection per peer */
423 conn_cb->ksnr_connected |= BIT(type);
425 case SOCKLND_CONN_BULK_IN:
426 conn_cb->ksnr_blki_conn_count++;
427 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
428 conn_cb->ksnr_connected |= BIT(type);
430 case SOCKLND_CONN_BULK_OUT:
431 conn_cb->ksnr_blko_conn_count++;
432 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
433 conn_cb->ksnr_connected |= BIT(type);
435 case SOCKLND_CONN_ANY:
436 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
437 conn_cb->ksnr_connected |= BIT(type);
444 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
445 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
449 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
450 struct ksock_conn *conn)
452 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
453 int type = conn->ksnc_type;
454 struct ksock_interface *iface;
457 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
459 conn->ksnc_conn_cb = conn_cb;
460 ksocknal_conn_cb_addref(conn_cb);
462 if (conn_cb->ksnr_myiface != conn_iface) {
463 if (conn_cb->ksnr_myiface < 0) {
464 /* route wasn't bound locally yet (the initial route) */
465 CDEBUG(D_NET, "Binding %s %pIS to interface %d\n",
466 libcfs_id2str(peer_ni->ksnp_id),
471 "Rebinding %s %pIS from interface %d to %d\n",
472 libcfs_id2str(peer_ni->ksnp_id),
474 conn_cb->ksnr_myiface,
477 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
478 conn_cb->ksnr_myiface);
480 iface->ksni_nroutes--;
482 conn_cb->ksnr_myiface = conn_iface;
483 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
484 conn_cb->ksnr_myiface);
486 iface->ksni_nroutes++;
489 ksocknal_incr_conn_count(conn_cb, type);
491 /* Successful connection => further attempts can
492 * proceed immediately
494 conn_cb->ksnr_retry_interval = 0;
498 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
499 struct ksock_conn_cb *conn_cb)
501 struct ksock_conn *conn;
502 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
504 LASSERT(!peer_ni->ksnp_closing);
505 LASSERT(!conn_cb->ksnr_peer);
506 LASSERT(!conn_cb->ksnr_scheduled);
507 LASSERT(!conn_cb->ksnr_connecting);
508 LASSERT(conn_cb->ksnr_connected == 0);
510 conn_cb->ksnr_peer = peer_ni;
511 ksocknal_peer_addref(peer_ni);
513 /* set the conn_cb's interface to the current net's interface */
514 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
515 net->ksnn_interface.ksni_nroutes++;
517 /* peer_ni's route list takes over my ref on 'route' */
518 peer_ni->ksnp_conn_cb = conn_cb;
520 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
521 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
522 (struct sockaddr *)&conn_cb->ksnr_addr))
525 ksocknal_associate_cb_conn_locked(conn_cb, conn);
526 /* keep going (typed conns) */
531 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
533 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
534 struct ksock_interface *iface;
535 struct ksock_conn *conn;
536 struct ksock_conn *cnxt;
538 LASSERT(!conn_cb->ksnr_deleted);
540 /* Close associated conns */
541 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
542 if (conn->ksnc_conn_cb != conn_cb)
545 ksocknal_close_conn_locked(conn, 0);
548 if (conn_cb->ksnr_myiface >= 0) {
549 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
550 conn_cb->ksnr_myiface);
552 iface->ksni_nroutes--;
555 conn_cb->ksnr_deleted = 1;
556 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
557 peer_ni->ksnp_conn_cb = NULL;
559 if (list_empty(&peer_ni->ksnp_conns)) {
560 /* I've just removed the last route to a peer_ni with no active
563 ksocknal_unlink_peer_locked(peer_ni);
568 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id,
569 struct sockaddr *addr)
571 struct ksock_peer_ni *peer_ni;
572 struct ksock_peer_ni *peer2;
573 struct ksock_conn_cb *conn_cb;
575 if (id.nid == LNET_NID_ANY ||
576 id.pid == LNET_PID_ANY)
579 /* Have a brand new peer_ni ready... */
580 peer_ni = ksocknal_create_peer(ni, id);
582 return PTR_ERR(peer_ni);
584 conn_cb = ksocknal_create_conn_cb(addr);
586 ksocknal_peer_decref(peer_ni);
590 write_lock_bh(&ksocknal_data.ksnd_global_lock);
592 /* always called with a ref on ni, so shutdown can't have started */
593 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
596 peer2 = ksocknal_find_peer_locked(ni, id);
598 ksocknal_peer_decref(peer_ni);
601 /* peer_ni table takes my ref on peer_ni */
602 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
606 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
608 /* Remember conns_per_peer setting at the time
609 * of connection initiation. It will define the
610 * max number of conns per type for this conn_cb
613 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
615 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
621 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
623 struct ksock_conn *conn;
624 struct ksock_conn *cnxt;
625 struct ksock_conn_cb *conn_cb;
627 LASSERT(!peer_ni->ksnp_closing);
629 /* Extra ref prevents peer_ni disappearing until I'm done with it */
630 ksocknal_peer_addref(peer_ni);
631 conn_cb = peer_ni->ksnp_conn_cb;
633 ksocknal_del_conn_cb_locked(conn_cb);
635 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
637 ksocknal_close_conn_locked(conn, 0);
639 ksocknal_peer_decref(peer_ni);
640 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
644 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
647 struct hlist_node *pnxt;
648 struct ksock_peer_ni *peer_ni;
654 write_lock_bh(&ksocknal_data.ksnd_global_lock);
656 if (id.nid != LNET_NID_ANY) {
657 lo = hash_min(nidhash(id.nid),
658 HASH_BITS(ksocknal_data.ksnd_peers));
662 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
665 for (i = lo; i <= hi; i++) {
666 hlist_for_each_entry_safe(peer_ni, pnxt,
667 &ksocknal_data.ksnd_peers[i],
669 if (peer_ni->ksnp_ni != ni)
672 if (!((id.nid == LNET_NID_ANY ||
673 peer_ni->ksnp_id.nid == id.nid) &&
674 (id.pid == LNET_PID_ANY ||
675 peer_ni->ksnp_id.pid == id.pid)))
678 ksocknal_peer_addref(peer_ni); /* a ref for me... */
680 ksocknal_del_peer_locked(peer_ni, ip);
682 if (peer_ni->ksnp_closing &&
683 !list_empty(&peer_ni->ksnp_tx_queue)) {
684 LASSERT(list_empty(&peer_ni->ksnp_conns));
685 LASSERT(peer_ni->ksnp_conn_cb == NULL);
687 list_splice_init(&peer_ni->ksnp_tx_queue,
691 ksocknal_peer_decref(peer_ni); /* ...till here */
693 rc = 0; /* matched! */
697 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
699 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
704 static struct ksock_conn *
705 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
707 struct ksock_peer_ni *peer_ni;
708 struct ksock_conn *conn;
711 read_lock(&ksocknal_data.ksnd_global_lock);
713 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
714 LASSERT(!peer_ni->ksnp_closing);
716 if (peer_ni->ksnp_ni != ni)
719 list_for_each_entry(conn, &peer_ni->ksnp_conns,
724 ksocknal_conn_addref(conn);
725 read_unlock(&ksocknal_data.ksnd_global_lock);
730 read_unlock(&ksocknal_data.ksnd_global_lock);
734 static struct ksock_sched *
735 ksocknal_choose_scheduler_locked(unsigned int cpt)
737 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
740 if (sched->kss_nthreads == 0) {
741 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
742 if (sched->kss_nthreads > 0) {
743 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
744 cpt, sched->kss_cpt);
755 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
757 struct ksock_connreq *cr;
759 struct sockaddr_storage peer;
761 rc = lnet_sock_getaddr(sock, true, &peer);
763 CERROR("Can't determine new connection's address\n");
767 LIBCFS_ALLOC(cr, sizeof(*cr));
769 LCONSOLE_ERROR_MSG(0x12f,
770 "Dropping connection request from %pIS: memory exhausted\n",
777 cr->ksncr_sock = sock;
779 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
781 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
782 wake_up(&ksocknal_data.ksnd_connd_waitq);
784 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
789 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
792 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
793 return conn_cb->ksnr_connecting;
798 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
799 struct socket *sock, int type)
801 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
803 struct lnet_process_id peerid;
805 struct ksock_conn *conn;
806 struct ksock_conn *conn2;
807 struct ksock_peer_ni *peer_ni = NULL;
808 struct ksock_peer_ni *peer2;
809 struct ksock_sched *sched;
810 struct ksock_hello_msg *hello;
813 struct ksock_tx *txtmp;
820 active = (conn_cb != NULL);
822 LASSERT(active == (type != SOCKLND_CONN_NONE));
824 LIBCFS_ALLOC(conn, sizeof(*conn));
830 conn->ksnc_peer = NULL;
831 conn->ksnc_conn_cb = NULL;
832 conn->ksnc_sock = sock;
833 /* 2 ref, 1 for conn, another extra ref prevents socket
834 * being closed before establishment of connection */
835 refcount_set(&conn->ksnc_sock_refcount, 2);
836 conn->ksnc_type = type;
837 ksocknal_lib_save_callback(sock, conn);
838 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
840 conn->ksnc_rx_ready = 0;
841 conn->ksnc_rx_scheduled = 0;
843 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
844 conn->ksnc_tx_ready = 0;
845 conn->ksnc_tx_scheduled = 0;
846 conn->ksnc_tx_carrier = NULL;
847 atomic_set (&conn->ksnc_tx_nob, 0);
849 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
850 kshm_ips[LNET_INTERFACES_NUM]));
856 /* stash conn's local and remote addrs */
857 rc = ksocknal_lib_get_conn_addrs(conn);
861 /* Find out/confirm peer_ni's NID and connection type and get the
862 * vector of interfaces she's willing to let me connect to.
863 * Passive connections use the listener timeout since the peer_ni sends
868 peer_ni = conn_cb->ksnr_peer;
869 LASSERT(ni == peer_ni->ksnp_ni);
871 /* Active connection sends HELLO eagerly */
872 hello->kshm_nips = 0;
873 peerid = peer_ni->ksnp_id;
875 write_lock_bh(global_lock);
876 conn->ksnc_proto = peer_ni->ksnp_proto;
877 write_unlock_bh(global_lock);
879 if (conn->ksnc_proto == NULL) {
880 conn->ksnc_proto = &ksocknal_protocol_v3x;
881 #if SOCKNAL_VERSION_DEBUG
882 if (*ksocknal_tunables.ksnd_protocol == 2)
883 conn->ksnc_proto = &ksocknal_protocol_v2x;
884 else if (*ksocknal_tunables.ksnd_protocol == 1)
885 conn->ksnc_proto = &ksocknal_protocol_v1x;
889 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
893 peerid.nid = LNET_NID_ANY;
894 peerid.pid = LNET_PID_ANY;
896 /* Passive, get protocol from peer_ni */
897 conn->ksnc_proto = NULL;
900 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
904 LASSERT(rc == 0 || active);
905 LASSERT(conn->ksnc_proto != NULL);
906 LASSERT(peerid.nid != LNET_NID_ANY);
908 cpt = lnet_cpt_of_nid(peerid.nid, ni);
911 ksocknal_peer_addref(peer_ni);
912 write_lock_bh(global_lock);
914 peer_ni = ksocknal_create_peer(ni, peerid);
915 if (IS_ERR(peer_ni)) {
916 rc = PTR_ERR(peer_ni);
920 write_lock_bh(global_lock);
922 /* called with a ref on ni, so shutdown can't have started */
923 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
925 peer2 = ksocknal_find_peer_locked(ni, peerid);
927 /* NB this puts an "empty" peer_ni in the peer_ni
928 * table (which takes my ref) */
929 hash_add(ksocknal_data.ksnd_peers,
930 &peer_ni->ksnp_list, nidhash(peerid.nid));
932 ksocknal_peer_decref(peer_ni);
937 ksocknal_peer_addref(peer_ni);
938 peer_ni->ksnp_accepting++;
940 /* Am I already connecting to this guy? Resolve in
941 * favour of higher NID...
943 if (peerid.nid < lnet_nid_to_nid4(&ni->ni_nid) &&
944 ksocknal_connecting(peer_ni->ksnp_conn_cb,
945 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
947 warn = "connection race resolution";
952 if (peer_ni->ksnp_closing ||
953 (active && conn_cb->ksnr_deleted)) {
954 /* peer_ni/conn_cb got closed under me */
956 warn = "peer_ni/conn_cb removed";
960 if (peer_ni->ksnp_proto == NULL) {
961 /* Never connected before.
962 * NB recv_hello may have returned EPROTO to signal my peer_ni
963 * wants a different protocol than the one I asked for.
965 LASSERT(list_empty(&peer_ni->ksnp_conns));
967 peer_ni->ksnp_proto = conn->ksnc_proto;
968 peer_ni->ksnp_incarnation = incarnation;
971 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
972 peer_ni->ksnp_incarnation != incarnation) {
973 /* peer_ni rebooted or I've got the wrong protocol version */
974 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
976 peer_ni->ksnp_proto = NULL;
978 warn = peer_ni->ksnp_incarnation != incarnation ?
980 "wrong proto version";
990 warn = "lost conn race";
993 warn = "retry with different protocol version";
997 /* Refuse to duplicate an existing connection, unless this is a
998 * loopback connection */
999 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1000 (struct sockaddr *)&conn->ksnc_myaddr)) {
1001 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1003 (struct sockaddr *)&conn2->ksnc_peeraddr,
1004 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1006 (struct sockaddr *)&conn2->ksnc_myaddr,
1007 (struct sockaddr *)&conn->ksnc_myaddr) ||
1008 conn2->ksnc_type != conn->ksnc_type)
1012 /* If max conns per type is not registered in conn_cb
1013 * as ksnr_max_conns, use ni's conns_per_peer
1015 if ((peer_ni->ksnp_conn_cb &&
1016 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1017 (!peer_ni->ksnp_conn_cb &&
1018 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1021 /* Reply on a passive connection attempt so the peer_ni
1022 * realises we're connected.
1032 /* If the connection created by this route didn't bind to the IP
1033 * address the route connected to, the connection/route matching
1034 * code below probably isn't going to work.
1037 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1038 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1039 CERROR("Route %s %pIS connected to %pIS\n",
1040 libcfs_id2str(peer_ni->ksnp_id),
1041 &conn_cb->ksnr_addr,
1042 &conn->ksnc_peeraddr);
1045 /* Search for a conn_cb corresponding to the new connection and
1046 * create an association. This allows incoming connections created
1047 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1048 * continually create duplicate conn_cbs.
1050 conn_cb = peer_ni->ksnp_conn_cb;
1052 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1053 (struct sockaddr *)&conn_cb->ksnr_addr))
1054 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1056 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1057 peer_ni->ksnp_last_alive = ktime_get_seconds();
1058 peer_ni->ksnp_send_keepalive = 0;
1059 peer_ni->ksnp_error = 0;
1061 sched = ksocknal_choose_scheduler_locked(cpt);
1063 CERROR("no schedulers available. node is unhealthy\n");
1067 * The cpt might have changed if we ended up selecting a non cpt
1068 * native scheduler. So use the scheduler's cpt instead.
1070 cpt = sched->kss_cpt;
1071 sched->kss_nconns++;
1072 conn->ksnc_scheduler = sched;
1074 conn->ksnc_tx_last_post = ktime_get_seconds();
1075 /* Set the deadline for the outgoing HELLO to drain */
1076 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1077 conn->ksnc_tx_deadline = ktime_get_seconds() +
1079 smp_mb(); /* order with adding to peer_ni's conn list */
1081 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1082 ksocknal_conn_addref(conn);
1084 ksocknal_new_packet(conn, 0);
1086 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1088 /* Take packets blocking for this connection. */
1089 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1090 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1094 list_del(&tx->tx_list);
1095 ksocknal_queue_tx_locked(tx, conn);
1098 write_unlock_bh(global_lock);
1100 /* We've now got a new connection. Any errors from here on are just
1101 * like "normal" comms errors and we close the connection normally.
1102 * NB (a) we still have to send the reply HELLO for passive
1104 * (b) normal I/O on the conn is blocked until I setup and call the
1108 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1109 " incarnation:%lld sched[%d]\n",
1110 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1111 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1115 hello->kshm_nips = 0;
1116 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1119 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1120 kshm_ips[LNET_INTERFACES_NUM]));
1122 /* setup the socket AFTER I've received hello (it disables
1123 * SO_LINGER). I might call back to the acceptor who may want
1124 * to send a protocol version response and then close the
1125 * socket; this ensures the socket only tears down after the
1126 * response has been sent.
1129 rc = ksocknal_lib_setup_sock(sock);
1131 write_lock_bh(global_lock);
1133 /* NB my callbacks block while I hold ksnd_global_lock */
1134 ksocknal_lib_set_callback(sock, conn);
1137 peer_ni->ksnp_accepting--;
1139 write_unlock_bh(global_lock);
1142 write_lock_bh(global_lock);
1143 if (!conn->ksnc_closing) {
1144 /* could be closed by another thread */
1145 ksocknal_close_conn_locked(conn, rc);
1147 write_unlock_bh(global_lock);
1148 } else if (ksocknal_connsock_addref(conn) == 0) {
1149 /* Allow I/O to proceed. */
1150 ksocknal_read_callback(conn);
1151 ksocknal_write_callback(conn);
1152 ksocknal_connsock_decref(conn);
1155 ksocknal_connsock_decref(conn);
1156 ksocknal_conn_decref(conn);
1161 if (!peer_ni->ksnp_closing &&
1162 list_empty(&peer_ni->ksnp_conns) &&
1163 peer_ni->ksnp_conn_cb == NULL) {
1164 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1165 ksocknal_unlink_peer_locked(peer_ni);
1168 write_unlock_bh(global_lock);
1172 CERROR("Not creating conn %s type %d: %s\n",
1173 libcfs_id2str(peerid), conn->ksnc_type, warn);
1175 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1176 libcfs_id2str(peerid), conn->ksnc_type, warn);
1181 /* Request retry by replying with CONN_NONE
1182 * ksnc_proto has been set already
1184 conn->ksnc_type = SOCKLND_CONN_NONE;
1185 hello->kshm_nips = 0;
1186 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1189 write_lock_bh(global_lock);
1190 peer_ni->ksnp_accepting--;
1191 write_unlock_bh(global_lock);
1195 * If we get here without an error code, just use -EALREADY.
1196 * Depending on how we got here, the error may be positive
1197 * or negative. Normalize the value for ksocknal_txlist_done().
1199 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1200 ksocknal_txlist_done(ni, &zombies, rc2);
1201 ksocknal_peer_decref(peer_ni);
1205 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1206 kshm_ips[LNET_INTERFACES_NUM]));
1208 LIBCFS_FREE(conn, sizeof(*conn));
1217 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1219 /* This just does the immmediate housekeeping, and queues the
1220 * connection for the reaper to terminate.
1221 * Caller holds ksnd_global_lock exclusively in irq context */
1222 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1223 struct ksock_conn_cb *conn_cb;
1224 struct ksock_conn *conn2;
1226 LASSERT(peer_ni->ksnp_error == 0);
1227 LASSERT(!conn->ksnc_closing);
1228 conn->ksnc_closing = 1;
1230 /* ksnd_deathrow_conns takes over peer_ni's ref */
1231 list_del(&conn->ksnc_list);
1233 conn_cb = conn->ksnc_conn_cb;
1234 if (conn_cb != NULL) {
1235 /* dissociate conn from cb... */
1236 LASSERT(!conn_cb->ksnr_deleted);
1238 /* connected bit is set only if all connections
1239 * of the given type got created
1241 if (ksocknal_get_conn_count_by_type(conn_cb, conn->ksnc_type) ==
1242 conn_cb->ksnr_max_conns)
1243 LASSERT((conn_cb->ksnr_connected &
1244 BIT(conn->ksnc_type)) != 0);
1246 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1247 if (conn2->ksnc_conn_cb == conn_cb &&
1248 conn2->ksnc_type == conn->ksnc_type)
1251 conn_cb->ksnr_connected &= ~BIT(conn->ksnc_type);
1253 conn->ksnc_conn_cb = NULL;
1255 /* drop conn's ref on conn_cb */
1256 ksocknal_conn_cb_decref(conn_cb);
1259 if (list_empty(&peer_ni->ksnp_conns)) {
1260 /* No more connections to this peer_ni */
1262 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1263 struct ksock_tx *tx;
1265 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1267 /* throw them to the last connection...,
1268 * these TXs will be send to /dev/null by scheduler */
1269 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1271 ksocknal_tx_prep(conn, tx);
1273 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1274 list_splice_init(&peer_ni->ksnp_tx_queue,
1275 &conn->ksnc_tx_queue);
1276 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1279 /* renegotiate protocol version */
1280 peer_ni->ksnp_proto = NULL;
1281 /* stash last conn close reason */
1282 peer_ni->ksnp_error = error;
1284 if (peer_ni->ksnp_conn_cb == NULL) {
1285 /* I've just closed last conn belonging to a
1286 * peer_ni with no connections to it
1288 ksocknal_unlink_peer_locked(peer_ni);
1292 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1294 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1295 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1297 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1301 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1303 bool notify = false;
1304 time64_t last_alive = 0;
1306 /* There has been a connection failure or comms error; but I'll only
1307 * tell LNET I think the peer_ni is dead if it's to another kernel and
1308 * there are no connections or connection attempts in existence. */
1310 read_lock(&ksocknal_data.ksnd_global_lock);
1312 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1313 list_empty(&peer_ni->ksnp_conns) &&
1314 peer_ni->ksnp_accepting == 0 &&
1315 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1317 last_alive = peer_ni->ksnp_last_alive;
1320 read_unlock(&ksocknal_data.ksnd_global_lock);
1323 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid,
1324 false, false, last_alive);
1328 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1330 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1331 struct ksock_tx *tx;
1332 struct ksock_tx *tmp;
1335 /* NB safe to finalize TXs because closing of socket will
1336 * abort all buffered data */
1337 LASSERT(conn->ksnc_sock == NULL);
1339 spin_lock(&peer_ni->ksnp_lock);
1341 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1343 if (tx->tx_conn != conn)
1346 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1348 tx->tx_msg.ksm_zc_cookies[0] = 0;
1349 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1350 list_move(&tx->tx_zc_list, &zlist);
1353 spin_unlock(&peer_ni->ksnp_lock);
1355 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1356 tx_zc_list)) != NULL) {
1357 list_del(&tx->tx_zc_list);
1358 ksocknal_tx_decref(tx);
1363 ksocknal_terminate_conn(struct ksock_conn *conn)
1365 /* This gets called by the reaper (guaranteed thread context) to
1366 * disengage the socket from its callbacks and close it.
1367 * ksnc_refcount will eventually hit zero, and then the reaper will
1370 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1371 struct ksock_sched *sched = conn->ksnc_scheduler;
1372 bool failed = false;
1374 LASSERT(conn->ksnc_closing);
1376 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1377 spin_lock_bh(&sched->kss_lock);
1379 /* a closing conn is always ready to tx */
1380 conn->ksnc_tx_ready = 1;
1382 if (!conn->ksnc_tx_scheduled &&
1383 !list_empty(&conn->ksnc_tx_queue)) {
1384 list_add_tail(&conn->ksnc_tx_list,
1385 &sched->kss_tx_conns);
1386 conn->ksnc_tx_scheduled = 1;
1387 /* extra ref for scheduler */
1388 ksocknal_conn_addref(conn);
1390 wake_up(&sched->kss_waitq);
1393 spin_unlock_bh(&sched->kss_lock);
1395 /* serialise with callbacks */
1396 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1398 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1400 /* OK, so this conn may not be completely disengaged from its
1401 * scheduler yet, but it _has_ committed to terminate...
1403 conn->ksnc_scheduler->kss_nconns--;
1405 if (peer_ni->ksnp_error != 0) {
1406 /* peer_ni's last conn closed in error */
1407 LASSERT(list_empty(&peer_ni->ksnp_conns));
1409 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1412 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1415 ksocknal_peer_failed(peer_ni);
1417 /* The socket is closed on the final put; either here, or in
1418 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1419 * when the connection was established, this will close the socket
1420 * immediately, aborting anything buffered in it. Any hung
1421 * zero-copy transmits will therefore complete in finite time.
1423 ksocknal_connsock_decref(conn);
1427 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1429 /* Queue the conn for the reaper to destroy */
1430 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1431 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1433 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1434 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1436 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1440 ksocknal_destroy_conn(struct ksock_conn *conn)
1444 /* Final coup-de-grace of the reaper */
1445 CDEBUG(D_NET, "connection %p\n", conn);
1447 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1448 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1449 LASSERT(conn->ksnc_sock == NULL);
1450 LASSERT(conn->ksnc_conn_cb == NULL);
1451 LASSERT(!conn->ksnc_tx_scheduled);
1452 LASSERT(!conn->ksnc_rx_scheduled);
1453 LASSERT(list_empty(&conn->ksnc_tx_queue));
1455 /* complete current receive if any */
1456 switch (conn->ksnc_rx_state) {
1457 case SOCKNAL_RX_LNET_PAYLOAD:
1458 last_rcv = conn->ksnc_rx_deadline -
1460 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1461 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1462 &conn->ksnc_peeraddr,
1463 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1464 ktime_get_seconds() - last_rcv);
1465 if (conn->ksnc_lnet_msg)
1466 conn->ksnc_lnet_msg->msg_health_status =
1467 LNET_MSG_STATUS_REMOTE_ERROR;
1468 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1470 case SOCKNAL_RX_LNET_HEADER:
1471 if (conn->ksnc_rx_started)
1472 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1473 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1474 &conn->ksnc_peeraddr,
1475 conn->ksnc_proto->pro_version);
1477 case SOCKNAL_RX_KSM_HEADER:
1478 if (conn->ksnc_rx_started)
1479 CERROR("Incomplete receive of ksock message from %s, ip %pISp, with error, protocol: %d.x.\n",
1480 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1481 &conn->ksnc_peeraddr,
1482 conn->ksnc_proto->pro_version);
1484 case SOCKNAL_RX_SLOP:
1485 if (conn->ksnc_rx_started)
1486 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1487 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1488 &conn->ksnc_peeraddr);
1495 ksocknal_peer_decref(conn->ksnc_peer);
1497 LIBCFS_FREE (conn, sizeof (*conn));
1501 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1502 struct sockaddr *addr, int why)
1504 struct ksock_conn *conn;
1505 struct ksock_conn *cnxt;
1508 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1511 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1513 ksocknal_close_conn_locked(conn, why);
1521 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1523 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1526 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1528 count = ksocknal_close_peer_conns_locked(
1529 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1531 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1537 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1539 struct ksock_peer_ni *peer_ni;
1540 struct hlist_node *pnxt;
1545 struct sockaddr_in sa = {.sin_family = AF_INET};
1547 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1549 if (id.nid != LNET_NID_ANY) {
1550 lo = hash_min(nidhash(id.nid),
1551 HASH_BITS(ksocknal_data.ksnd_peers));
1555 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1558 sa.sin_addr.s_addr = htonl(ipaddr);
1559 for (i = lo; i <= hi; i++) {
1560 hlist_for_each_entry_safe(peer_ni, pnxt,
1561 &ksocknal_data.ksnd_peers[i],
1564 if (!((id.nid == LNET_NID_ANY ||
1565 id.nid == peer_ni->ksnp_id.nid) &&
1566 (id.pid == LNET_PID_ANY ||
1567 id.pid == peer_ni->ksnp_id.pid)))
1570 count += ksocknal_close_peer_conns_locked(
1572 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1576 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1578 /* wildcards always succeed */
1579 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1582 return (count == 0 ? -ENOENT : 0);
1586 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1588 /* The router is telling me she's been notified of a change in
1591 struct lnet_process_id id = {
1593 .pid = LNET_PID_ANY,
1596 CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1598 /* If the gateway crashed, close all open connections... */
1599 ksocknal_close_matching_conns(id, 0);
1602 /* We can only establish new connections
1603 * if we have autroutes, and these connect on demand. */
1607 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1611 struct ksock_conn *conn;
1613 for (index = 0; ; index++) {
1614 read_lock(&ksocknal_data.ksnd_global_lock);
1619 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1621 ksocknal_conn_addref(conn);
1626 read_unlock(&ksocknal_data.ksnd_global_lock);
1631 ksocknal_lib_push_conn (conn);
1632 ksocknal_conn_decref(conn);
1637 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1644 if (id.nid != LNET_NID_ANY) {
1645 lo = hash_min(nidhash(id.nid),
1646 HASH_BITS(ksocknal_data.ksnd_peers));
1650 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1653 for (bkt = lo; bkt <= hi; bkt++) {
1654 int peer_off; /* searching offset in peer_ni hash table */
1656 for (peer_off = 0; ; peer_off++) {
1657 struct ksock_peer_ni *peer_ni;
1660 read_lock(&ksocknal_data.ksnd_global_lock);
1661 hlist_for_each_entry(peer_ni,
1662 &ksocknal_data.ksnd_peers[bkt],
1664 if (!((id.nid == LNET_NID_ANY ||
1665 id.nid == peer_ni->ksnp_id.nid) &&
1666 (id.pid == LNET_PID_ANY ||
1667 id.pid == peer_ni->ksnp_id.pid)))
1670 if (i++ == peer_off) {
1671 ksocknal_peer_addref(peer_ni);
1675 read_unlock(&ksocknal_data.ksnd_global_lock);
1677 if (i <= peer_off) /* no match */
1681 ksocknal_push_peer(peer_ni);
1682 ksocknal_peer_decref(peer_ni);
1689 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1691 struct lnet_process_id id = {0};
1692 struct libcfs_ioctl_data *data = arg;
1696 case IOC_LIBCFS_GET_INTERFACE: {
1697 struct ksock_net *net = ni->ni_data;
1698 struct ksock_interface *iface;
1699 struct sockaddr_in *sa;
1701 read_lock(&ksocknal_data.ksnd_global_lock);
1703 if (data->ioc_count >= 1) {
1707 iface = &net->ksnn_interface;
1709 sa = (void *)&iface->ksni_addr;
1710 if (sa->sin_family == AF_INET)
1711 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1713 data->ioc_u32[0] = 0xFFFFFFFF;
1714 data->ioc_u32[1] = iface->ksni_netmask;
1715 data->ioc_u32[2] = iface->ksni_npeers;
1716 data->ioc_u32[3] = iface->ksni_nroutes;
1719 read_unlock(&ksocknal_data.ksnd_global_lock);
1723 case IOC_LIBCFS_GET_PEER: {
1728 int share_count = 0;
1730 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1731 &id, &myip, &ip, &port,
1732 &conn_count, &share_count);
1736 data->ioc_nid = id.nid;
1737 data->ioc_count = share_count;
1738 data->ioc_u32[0] = ip;
1739 data->ioc_u32[1] = port;
1740 data->ioc_u32[2] = myip;
1741 data->ioc_u32[3] = conn_count;
1742 data->ioc_u32[4] = id.pid;
1746 case IOC_LIBCFS_ADD_PEER: {
1747 struct sockaddr_in sa = {.sin_family = AF_INET};
1749 id.nid = data->ioc_nid;
1750 id.pid = LNET_PID_LUSTRE;
1751 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1752 sa.sin_port = htons(data->ioc_u32[1]);
1753 return ksocknal_add_peer(ni, id, (struct sockaddr *)&sa);
1755 case IOC_LIBCFS_DEL_PEER:
1756 id.nid = data->ioc_nid;
1757 id.pid = LNET_PID_ANY;
1758 return ksocknal_del_peer (ni, id,
1759 data->ioc_u32[0]); /* IP */
1761 case IOC_LIBCFS_GET_CONN: {
1765 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1766 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1767 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1772 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1774 data->ioc_count = txmem;
1775 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
1776 data->ioc_flags = nagle;
1777 if (psa->sin_family == AF_INET)
1778 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1780 data->ioc_u32[0] = 0xFFFFFFFF;
1781 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1782 &conn->ksnc_peeraddr);
1783 if (mysa->sin_family == AF_INET)
1784 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1786 data->ioc_u32[2] = 0xFFFFFFFF;
1787 data->ioc_u32[3] = conn->ksnc_type;
1788 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1789 data->ioc_u32[5] = rxmem;
1790 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1791 ksocknal_conn_decref(conn);
1795 case IOC_LIBCFS_CLOSE_CONNECTION:
1796 id.nid = data->ioc_nid;
1797 id.pid = LNET_PID_ANY;
1798 return ksocknal_close_matching_conns (id,
1801 case IOC_LIBCFS_REGISTER_MYNID:
1802 /* Ignore if this is a noop */
1803 if (nid_is_nid4(&ni->ni_nid) &&
1804 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1807 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1808 libcfs_nid2str(data->ioc_nid),
1809 libcfs_nidstr(&ni->ni_nid));
1812 case IOC_LIBCFS_PUSH_CONNECTION:
1813 id.nid = data->ioc_nid;
1814 id.pid = LNET_PID_ANY;
1815 return ksocknal_push(ni, id);
1824 ksocknal_free_buffers (void)
1826 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1828 if (ksocknal_data.ksnd_schedulers != NULL)
1829 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1831 spin_lock(&ksocknal_data.ksnd_tx_lock);
1833 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1835 struct ksock_tx *tx;
1837 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1838 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1840 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1841 tx_list)) != NULL) {
1842 list_del(&tx->tx_list);
1843 LIBCFS_FREE(tx, tx->tx_desc_size);
1846 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1850 static int ksocknal_get_link_status(struct net_device *dev)
1856 if (!netif_running(dev))
1858 /* Some devices may not be providing link settings */
1859 else if (dev->ethtool_ops->get_link)
1860 ret = dev->ethtool_ops->get_link(dev);
1866 ksocknal_handle_link_state_change(struct net_device *dev,
1867 unsigned char operstate)
1870 struct ksock_net *net;
1871 struct ksock_net *cnxt;
1873 unsigned char link_down = !(operstate == IF_OPER_UP);
1875 ifindex = dev->ifindex;
1877 if (!ksocknal_data.ksnd_nnets)
1880 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1882 if (net->ksnn_interface.ksni_index != ifindex)
1886 atomic_set(&ni->ni_fatal_error_on, link_down);
1888 atomic_set(&ni->ni_fatal_error_on,
1889 (ksocknal_get_link_status(dev) == 0));
1896 /************************************
1897 * Net device notifier event handler
1898 ************************************/
1899 static int ksocknal_device_event(struct notifier_block *unused,
1900 unsigned long event, void *ptr)
1902 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1903 unsigned char operstate;
1905 operstate = dev->operstate;
1911 ksocknal_handle_link_state_change(dev, operstate);
1918 static struct notifier_block ksocknal_notifier_block = {
1919 .notifier_call = ksocknal_device_event,
1923 ksocknal_base_shutdown(void)
1925 struct ksock_sched *sched;
1926 struct ksock_peer_ni *peer_ni;
1929 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
1930 libcfs_kmem_read());
1931 LASSERT (ksocknal_data.ksnd_nnets == 0);
1933 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL)
1934 unregister_netdevice_notifier(&ksocknal_notifier_block);
1936 switch (ksocknal_data.ksnd_init) {
1941 case SOCKNAL_INIT_ALL:
1942 case SOCKNAL_INIT_DATA:
1943 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
1946 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
1947 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
1948 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
1949 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
1950 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
1952 if (ksocknal_data.ksnd_schedulers != NULL) {
1953 cfs_percpt_for_each(sched, i,
1954 ksocknal_data.ksnd_schedulers) {
1956 LASSERT(list_empty(&sched->kss_tx_conns));
1957 LASSERT(list_empty(&sched->kss_rx_conns));
1958 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
1959 LASSERT(sched->kss_nconns == 0);
1963 /* flag threads to terminate; wake and wait for them to die */
1964 ksocknal_data.ksnd_shuttingdown = 1;
1965 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
1966 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1968 if (ksocknal_data.ksnd_schedulers != NULL) {
1969 cfs_percpt_for_each(sched, i,
1970 ksocknal_data.ksnd_schedulers)
1971 wake_up_all(&sched->kss_waitq);
1974 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
1975 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
1976 "waiting for %d threads to terminate\n",
1977 atomic_read(&ksocknal_data.ksnd_nthreads));
1979 ksocknal_free_buffers();
1981 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
1985 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
1986 libcfs_kmem_read());
1988 module_put(THIS_MODULE);
1992 ksocknal_base_startup(void)
1994 struct ksock_sched *sched;
1998 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
1999 LASSERT(ksocknal_data.ksnd_nnets == 0);
2001 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2003 hash_init(ksocknal_data.ksnd_peers);
2005 rwlock_init(&ksocknal_data.ksnd_global_lock);
2006 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2008 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2009 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2010 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2011 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2012 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2014 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2015 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2016 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2017 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2019 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2020 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2022 /* NB memset above zeros whole of ksocknal_data */
2024 /* flag lists/ptrs/locks initialised */
2025 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2026 if (!try_module_get(THIS_MODULE))
2029 /* Create a scheduler block per available CPT */
2030 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2032 if (ksocknal_data.ksnd_schedulers == NULL)
2035 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2039 * make sure not to allocate more threads than there are
2040 * cores/CPUs in teh CPT
2042 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2043 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2044 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2047 * max to half of CPUs, assume another half should be
2048 * reserved for upper layer modules
2050 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2053 sched->kss_nthreads_max = nthrs;
2056 spin_lock_init(&sched->kss_lock);
2057 INIT_LIST_HEAD(&sched->kss_rx_conns);
2058 INIT_LIST_HEAD(&sched->kss_tx_conns);
2059 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2060 init_waitqueue_head(&sched->kss_waitq);
2063 ksocknal_data.ksnd_connd_starting = 0;
2064 ksocknal_data.ksnd_connd_failed_stamp = 0;
2065 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2066 /* must have at least 2 connds to remain responsive to accepts while
2068 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2069 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2071 if (*ksocknal_tunables.ksnd_nconnds_max <
2072 *ksocknal_tunables.ksnd_nconnds) {
2073 ksocknal_tunables.ksnd_nconnds_max =
2074 ksocknal_tunables.ksnd_nconnds;
2077 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2078 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2079 ksocknal_data.ksnd_connd_starting++;
2080 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2082 rc = ksocknal_thread_start(ksocknal_connd,
2083 (void *)((uintptr_t)i),
2084 "socknal_cd%02d", i);
2086 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2087 ksocknal_data.ksnd_connd_starting--;
2088 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2089 CERROR("Can't spawn socknal connd: %d\n", rc);
2094 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2096 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2100 register_netdevice_notifier(&ksocknal_notifier_block);
2102 /* flag everything initialised */
2103 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2108 ksocknal_base_shutdown();
2113 ksocknal_debug_peerhash(struct lnet_ni *ni)
2115 struct ksock_peer_ni *peer_ni;
2118 read_lock(&ksocknal_data.ksnd_global_lock);
2120 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2121 struct ksock_conn_cb *conn_cb;
2122 struct ksock_conn *conn;
2124 if (peer_ni->ksnp_ni != ni)
2127 CWARN("Active peer_ni on shutdown: %s, ref %d, "
2128 "closing %d, accepting %d, err %d, zcookie %llu, "
2129 "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2130 refcount_read(&peer_ni->ksnp_refcount),
2131 peer_ni->ksnp_closing,
2132 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2133 peer_ni->ksnp_zc_next_cookie,
2134 !list_empty(&peer_ni->ksnp_tx_queue),
2135 !list_empty(&peer_ni->ksnp_zc_req_list));
2137 conn_cb = peer_ni->ksnp_conn_cb;
2139 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2140 refcount_read(&conn_cb->ksnr_refcount),
2141 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2142 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2145 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2146 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2147 refcount_read(&conn->ksnc_conn_refcount),
2148 refcount_read(&conn->ksnc_sock_refcount),
2149 conn->ksnc_type, conn->ksnc_closing);
2154 read_unlock(&ksocknal_data.ksnd_global_lock);
2159 ksocknal_shutdown(struct lnet_ni *ni)
2161 struct ksock_net *net = ni->ni_data;
2162 struct lnet_process_id anyid = {
2163 .nid = LNET_NID_ANY,
2164 .pid = LNET_PID_ANY,
2167 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2168 LASSERT(ksocknal_data.ksnd_nnets > 0);
2170 /* prevent new peers */
2171 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2173 /* Delete all peers */
2174 ksocknal_del_peer(ni, anyid, 0);
2176 /* Wait for all peer_ni state to clean up */
2177 wait_var_event_warning(&net->ksnn_npeers,
2178 atomic_read(&net->ksnn_npeers) ==
2179 SOCKNAL_SHUTDOWN_BIAS,
2180 "waiting for %d peers to disconnect\n",
2181 ksocknal_debug_peerhash(ni) +
2182 atomic_read(&net->ksnn_npeers) -
2183 SOCKNAL_SHUTDOWN_BIAS);
2185 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2186 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2188 list_del(&net->ksnn_list);
2189 LIBCFS_FREE(net, sizeof(*net));
2191 ksocknal_data.ksnd_nnets--;
2192 if (ksocknal_data.ksnd_nnets == 0)
2193 ksocknal_base_shutdown();
2197 ksocknal_search_new_ipif(struct ksock_net *net)
2200 char *ifnam = &net->ksnn_interface.ksni_name[0];
2201 char *colon = strchr(ifnam, ':');
2203 struct ksock_net *tmp;
2208 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2209 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2210 char *colon2 = strchr(ifnam2, ':');
2215 found = strcmp(ifnam, ifnam2) == 0;
2228 ksocknal_start_schedulers(struct ksock_sched *sched)
2234 if (sched->kss_nthreads == 0) {
2235 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2236 nthrs = sched->kss_nthreads_max;
2238 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2240 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2241 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2243 nthrs = min(nthrs, sched->kss_nthreads_max);
2245 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2246 /* increase two threads if there is new interface */
2247 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2250 for (i = 0; i < nthrs; i++) {
2253 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2254 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2255 "socknal_sd%02d_%02d",
2257 (int)KSOCK_THREAD_SID(id));
2261 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2262 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2266 sched->kss_nthreads += i;
2271 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2273 int newif = ksocknal_search_new_ipif(net);
2277 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2280 for (i = 0; i < ncpts; i++) {
2281 struct ksock_sched *sched;
2282 int cpt = (cpts == NULL) ? i : cpts[i];
2284 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2285 sched = ksocknal_data.ksnd_schedulers[cpt];
2287 if (!newif && sched->kss_nthreads > 0)
2290 rc = ksocknal_start_schedulers(sched);
2298 ksocknal_startup(struct lnet_ni *ni)
2300 struct ksock_net *net;
2301 struct ksock_interface *ksi = NULL;
2302 struct lnet_inetdev *ifaces = NULL;
2303 struct sockaddr_in *sa;
2307 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2308 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2309 rc = ksocknal_base_startup();
2313 LIBCFS_ALLOC(net, sizeof(*net));
2316 net->ksnn_incarnation = ktime_get_real_ns();
2319 ksocknal_tunables_setup(ni);
2321 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2325 ksi = &net->ksnn_interface;
2327 /* Use the first discovered interface or look in the list */
2328 if (ni->ni_interface) {
2329 for (i = 0; i < rc; i++)
2330 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2333 /* ni_interfaces doesn't contain the interface we want */
2335 CERROR("ksocklnd: failed to find interface %s\n",
2341 ni->ni_dev_cpt = ifaces[i].li_cpt;
2342 sa = (void *)&ksi->ksni_addr;
2343 memset(sa, 0, sizeof(*sa));
2344 sa->sin_family = AF_INET;
2345 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2346 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2347 ksi->ksni_netmask = ifaces[i].li_netmask;
2348 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2350 /* call it before add it to ksocknal_data.ksnd_nets */
2351 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2356 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2357 ni->ni_nid.nid_addr[0] =
2358 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2359 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2361 ksocknal_data.ksnd_nnets++;
2366 LIBCFS_FREE(net, sizeof(*net));
2368 if (ksocknal_data.ksnd_nnets == 0)
2369 ksocknal_base_shutdown();
2374 static void __exit ksocklnd_exit(void)
2376 lnet_unregister_lnd(&the_ksocklnd);
2379 static const struct lnet_lnd the_ksocklnd = {
2380 .lnd_type = SOCKLND,
2381 .lnd_startup = ksocknal_startup,
2382 .lnd_shutdown = ksocknal_shutdown,
2383 .lnd_ctl = ksocknal_ctl,
2384 .lnd_send = ksocknal_send,
2385 .lnd_recv = ksocknal_recv,
2386 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2387 .lnd_accept = ksocknal_accept,
2390 static int __init ksocklnd_init(void)
2394 /* check ksnr_connected/connecting field large enough */
2395 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2396 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2398 rc = ksocknal_tunables_init();
2402 lnet_register_lnd(&the_ksocklnd);
2407 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2408 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2409 MODULE_VERSION("2.8.0");
2410 MODULE_LICENSE("GPL");
2412 module_init(ksocklnd_init);
2413 module_exit(ksocklnd_exit);