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_processid *id)
157 int cpt = lnet_nid2cpt(&id->nid, ni);
158 struct ksock_net *net = ni->ni_data;
159 struct ksock_peer_ni *peer_ni;
161 LASSERT(!LNET_NID_IS_ANY(&id->nid));
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_idstr(&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_processid *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 (!nid_same(&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_idstr(id),
239 refcount_read(&peer_ni->ksnp_refcount));
245 struct ksock_peer_ni *
246 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_processid *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->pid = peer_ni->ksnp_id.pid;
316 id->nid = lnet_nid_to_nid4(&peer_ni->ksnp_id.nid);
326 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
330 id->pid = peer_ni->ksnp_id.pid;
331 id->nid = lnet_nid_to_nid4(&peer_ni->ksnp_id.nid);
332 *myip = peer_ni->ksnp_passive_ips[j];
341 if (peer_ni->ksnp_conn_cb) {
345 conn_cb = peer_ni->ksnp_conn_cb;
347 id->pid = peer_ni->ksnp_id.pid;
348 id->nid = lnet_nid_to_nid4(&peer_ni->ksnp_id.nid);
349 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
350 struct sockaddr_in *sa =
351 (void *)&conn_cb->ksnr_addr;
353 rc = choose_ipv4_src(myip,
354 conn_cb->ksnr_myiface,
355 ntohl(sa->sin_addr.s_addr),
357 *peer_ip = ntohl(sa->sin_addr.s_addr);
358 *port = ntohs(sa->sin_port);
361 *peer_ip = 0xFFFFFFFF;
365 *conn_count = conn_cb->ksnr_conn_count;
371 read_unlock(&ksocknal_data.ksnd_global_lock);
376 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
379 unsigned int count = 0;
382 case SOCKLND_CONN_CONTROL:
383 count = conn_cb->ksnr_ctrl_conn_count;
385 case SOCKLND_CONN_BULK_IN:
386 count = conn_cb->ksnr_blki_conn_count;
388 case SOCKLND_CONN_BULK_OUT:
389 count = conn_cb->ksnr_blko_conn_count;
391 case SOCKLND_CONN_ANY:
392 count = conn_cb->ksnr_conn_count;
403 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
405 struct lnet_ni *ni = peer_ni->ksnp_ni;
406 struct lnet_ioctl_config_socklnd_tunables *tunables;
410 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
412 return tunables->lnd_conns_per_peer;
416 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
419 conn_cb->ksnr_conn_count++;
421 /* check if all connections of the given type got created */
423 case SOCKLND_CONN_CONTROL:
424 conn_cb->ksnr_ctrl_conn_count++;
425 /* there's a single control connection per peer */
426 conn_cb->ksnr_connected |= BIT(type);
428 case SOCKLND_CONN_BULK_IN:
429 conn_cb->ksnr_blki_conn_count++;
430 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
431 conn_cb->ksnr_connected |= BIT(type);
433 case SOCKLND_CONN_BULK_OUT:
434 conn_cb->ksnr_blko_conn_count++;
435 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
436 conn_cb->ksnr_connected |= BIT(type);
438 case SOCKLND_CONN_ANY:
439 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
440 conn_cb->ksnr_connected |= BIT(type);
447 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
448 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
452 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
453 struct ksock_conn *conn)
455 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
456 int type = conn->ksnc_type;
457 struct ksock_interface *iface;
460 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
462 conn->ksnc_conn_cb = conn_cb;
463 ksocknal_conn_cb_addref(conn_cb);
465 if (conn_cb->ksnr_myiface != conn_iface) {
466 if (conn_cb->ksnr_myiface < 0) {
467 /* route wasn't bound locally yet (the initial route) */
468 CDEBUG(D_NET, "Binding %s %pIS to interface %d\n",
469 libcfs_idstr(&peer_ni->ksnp_id),
474 "Rebinding %s %pIS from interface %d to %d\n",
475 libcfs_idstr(&peer_ni->ksnp_id),
477 conn_cb->ksnr_myiface,
480 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
481 conn_cb->ksnr_myiface);
483 iface->ksni_nroutes--;
485 conn_cb->ksnr_myiface = conn_iface;
486 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
487 conn_cb->ksnr_myiface);
489 iface->ksni_nroutes++;
492 ksocknal_incr_conn_count(conn_cb, type);
494 /* Successful connection => further attempts can
495 * proceed immediately
497 conn_cb->ksnr_retry_interval = 0;
501 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
502 struct ksock_conn_cb *conn_cb)
504 struct ksock_conn *conn;
505 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
507 LASSERT(!peer_ni->ksnp_closing);
508 LASSERT(!conn_cb->ksnr_peer);
509 LASSERT(!conn_cb->ksnr_scheduled);
510 LASSERT(!conn_cb->ksnr_connecting);
511 LASSERT(conn_cb->ksnr_connected == 0);
513 conn_cb->ksnr_peer = peer_ni;
514 ksocknal_peer_addref(peer_ni);
516 /* set the conn_cb's interface to the current net's interface */
517 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
518 net->ksnn_interface.ksni_nroutes++;
520 /* peer_ni's route list takes over my ref on 'route' */
521 peer_ni->ksnp_conn_cb = conn_cb;
523 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
524 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
525 (struct sockaddr *)&conn_cb->ksnr_addr))
528 ksocknal_associate_cb_conn_locked(conn_cb, conn);
529 /* keep going (typed conns) */
534 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
536 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
537 struct ksock_interface *iface;
538 struct ksock_conn *conn;
539 struct ksock_conn *cnxt;
541 LASSERT(!conn_cb->ksnr_deleted);
543 /* Close associated conns */
544 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
545 if (conn->ksnc_conn_cb != conn_cb)
548 ksocknal_close_conn_locked(conn, 0);
551 if (conn_cb->ksnr_myiface >= 0) {
552 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
553 conn_cb->ksnr_myiface);
555 iface->ksni_nroutes--;
558 conn_cb->ksnr_deleted = 1;
559 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
560 peer_ni->ksnp_conn_cb = NULL;
562 if (list_empty(&peer_ni->ksnp_conns)) {
563 /* I've just removed the last route to a peer_ni with no active
566 ksocknal_unlink_peer_locked(peer_ni);
571 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id4,
572 struct sockaddr *addr)
574 struct ksock_peer_ni *peer_ni;
575 struct ksock_peer_ni *peer2;
576 struct ksock_conn_cb *conn_cb;
577 struct lnet_processid id;
579 if (id4.nid == LNET_NID_ANY ||
580 id4.pid == LNET_PID_ANY)
584 lnet_nid4_to_nid(id4.nid, &id.nid);
586 /* Have a brand new peer_ni ready... */
587 peer_ni = ksocknal_create_peer(ni, &id);
589 return PTR_ERR(peer_ni);
591 conn_cb = ksocknal_create_conn_cb(addr);
593 ksocknal_peer_decref(peer_ni);
597 write_lock_bh(&ksocknal_data.ksnd_global_lock);
599 /* always called with a ref on ni, so shutdown can't have started */
600 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
603 peer2 = ksocknal_find_peer_locked(ni, &id);
605 ksocknal_peer_decref(peer_ni);
608 /* peer_ni table takes my ref on peer_ni */
609 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
613 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
615 /* Remember conns_per_peer setting at the time
616 * of connection initiation. It will define the
617 * max number of conns per type for this conn_cb
620 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
622 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
628 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
630 struct ksock_conn *conn;
631 struct ksock_conn *cnxt;
632 struct ksock_conn_cb *conn_cb;
634 LASSERT(!peer_ni->ksnp_closing);
636 /* Extra ref prevents peer_ni disappearing until I'm done with it */
637 ksocknal_peer_addref(peer_ni);
638 conn_cb = peer_ni->ksnp_conn_cb;
640 ksocknal_del_conn_cb_locked(conn_cb);
642 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
644 ksocknal_close_conn_locked(conn, 0);
646 ksocknal_peer_decref(peer_ni);
647 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
651 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id4, __u32 ip)
654 struct hlist_node *pnxt;
655 struct ksock_peer_ni *peer_ni;
660 struct lnet_processid id;
663 lnet_nid4_to_nid(id4.nid, &id.nid);
665 write_lock_bh(&ksocknal_data.ksnd_global_lock);
667 if (!LNET_NID_IS_ANY(&id.nid)) {
668 lo = hash_min(nidhash(&id.nid),
669 HASH_BITS(ksocknal_data.ksnd_peers));
673 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
676 for (i = lo; i <= hi; i++) {
677 hlist_for_each_entry_safe(peer_ni, pnxt,
678 &ksocknal_data.ksnd_peers[i],
680 if (peer_ni->ksnp_ni != ni)
683 if (!((LNET_NID_IS_ANY(&id.nid) ||
684 nid_same(&peer_ni->ksnp_id.nid, &id.nid)) &&
685 (id.pid == LNET_PID_ANY ||
686 peer_ni->ksnp_id.pid == id.pid)))
689 ksocknal_peer_addref(peer_ni); /* a ref for me... */
691 ksocknal_del_peer_locked(peer_ni, ip);
693 if (peer_ni->ksnp_closing &&
694 !list_empty(&peer_ni->ksnp_tx_queue)) {
695 LASSERT(list_empty(&peer_ni->ksnp_conns));
696 LASSERT(peer_ni->ksnp_conn_cb == NULL);
698 list_splice_init(&peer_ni->ksnp_tx_queue,
702 ksocknal_peer_decref(peer_ni); /* ...till here */
704 rc = 0; /* matched! */
708 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
710 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
715 static struct ksock_conn *
716 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
718 struct ksock_peer_ni *peer_ni;
719 struct ksock_conn *conn;
722 read_lock(&ksocknal_data.ksnd_global_lock);
724 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
725 LASSERT(!peer_ni->ksnp_closing);
727 if (peer_ni->ksnp_ni != ni)
730 list_for_each_entry(conn, &peer_ni->ksnp_conns,
735 ksocknal_conn_addref(conn);
736 read_unlock(&ksocknal_data.ksnd_global_lock);
741 read_unlock(&ksocknal_data.ksnd_global_lock);
745 static struct ksock_sched *
746 ksocknal_choose_scheduler_locked(unsigned int cpt)
748 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
751 if (sched->kss_nthreads == 0) {
752 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
753 if (sched->kss_nthreads > 0) {
754 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
755 cpt, sched->kss_cpt);
766 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
768 struct ksock_connreq *cr;
770 struct sockaddr_storage peer;
772 rc = lnet_sock_getaddr(sock, true, &peer);
774 CERROR("Can't determine new connection's address\n");
778 LIBCFS_ALLOC(cr, sizeof(*cr));
780 LCONSOLE_ERROR_MSG(0x12f,
781 "Dropping connection request from %pIS: memory exhausted\n",
788 cr->ksncr_sock = sock;
790 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
792 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
793 wake_up(&ksocknal_data.ksnd_connd_waitq);
795 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
800 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
803 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
804 return conn_cb->ksnr_connecting;
809 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
810 struct socket *sock, int type)
812 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
814 struct lnet_process_id peerid4;
816 struct ksock_conn *conn;
817 struct ksock_conn *conn2;
818 struct ksock_peer_ni *peer_ni = NULL;
819 struct ksock_peer_ni *peer2;
820 struct ksock_sched *sched;
821 struct ksock_hello_msg *hello;
824 struct ksock_tx *txtmp;
831 active = (conn_cb != NULL);
833 LASSERT(active == (type != SOCKLND_CONN_NONE));
835 LIBCFS_ALLOC(conn, sizeof(*conn));
841 conn->ksnc_peer = NULL;
842 conn->ksnc_conn_cb = NULL;
843 conn->ksnc_sock = sock;
844 /* 2 ref, 1 for conn, another extra ref prevents socket
845 * being closed before establishment of connection */
846 refcount_set(&conn->ksnc_sock_refcount, 2);
847 conn->ksnc_type = type;
848 ksocknal_lib_save_callback(sock, conn);
849 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
851 conn->ksnc_rx_ready = 0;
852 conn->ksnc_rx_scheduled = 0;
854 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
855 conn->ksnc_tx_ready = 0;
856 conn->ksnc_tx_scheduled = 0;
857 conn->ksnc_tx_carrier = NULL;
858 atomic_set (&conn->ksnc_tx_nob, 0);
860 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
861 kshm_ips[LNET_INTERFACES_NUM]));
867 /* stash conn's local and remote addrs */
868 rc = ksocknal_lib_get_conn_addrs(conn);
872 /* Find out/confirm peer_ni's NID and connection type and get the
873 * vector of interfaces she's willing to let me connect to.
874 * Passive connections use the listener timeout since the peer_ni sends
879 peer_ni = conn_cb->ksnr_peer;
880 LASSERT(ni == peer_ni->ksnp_ni);
882 /* Active connection sends HELLO eagerly */
883 hello->kshm_nips = 0;
884 peerid4 = lnet_pid_to_pid4(&peer_ni->ksnp_id);
886 write_lock_bh(global_lock);
887 conn->ksnc_proto = peer_ni->ksnp_proto;
888 write_unlock_bh(global_lock);
890 if (conn->ksnc_proto == NULL) {
891 conn->ksnc_proto = &ksocknal_protocol_v3x;
892 #if SOCKNAL_VERSION_DEBUG
893 if (*ksocknal_tunables.ksnd_protocol == 2)
894 conn->ksnc_proto = &ksocknal_protocol_v2x;
895 else if (*ksocknal_tunables.ksnd_protocol == 1)
896 conn->ksnc_proto = &ksocknal_protocol_v1x;
900 rc = ksocknal_send_hello(ni, conn, peerid4.nid, hello);
904 peerid4.nid = LNET_NID_ANY;
905 peerid4.pid = LNET_PID_ANY;
907 /* Passive, get protocol from peer_ni */
908 conn->ksnc_proto = NULL;
911 rc = ksocknal_recv_hello(ni, conn, hello, &peerid4, &incarnation);
915 LASSERT(rc == 0 || active);
916 LASSERT(conn->ksnc_proto != NULL);
917 LASSERT(peerid4.nid != LNET_NID_ANY);
919 cpt = lnet_cpt_of_nid(peerid4.nid, ni);
922 ksocknal_peer_addref(peer_ni);
923 write_lock_bh(global_lock);
925 struct lnet_processid peerid;
927 lnet_pid4_to_pid(peerid4, &peerid);
928 peer_ni = ksocknal_create_peer(ni, &peerid);
929 if (IS_ERR(peer_ni)) {
930 rc = PTR_ERR(peer_ni);
934 write_lock_bh(global_lock);
936 /* called with a ref on ni, so shutdown can't have started */
937 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
939 peer2 = ksocknal_find_peer_locked(ni, &peerid);
941 /* NB this puts an "empty" peer_ni in the peer_ni
942 * table (which takes my ref) */
943 hash_add(ksocknal_data.ksnd_peers,
944 &peer_ni->ksnp_list, nidhash(&peerid.nid));
946 ksocknal_peer_decref(peer_ni);
951 ksocknal_peer_addref(peer_ni);
952 peer_ni->ksnp_accepting++;
954 /* Am I already connecting to this guy? Resolve in
955 * favour of higher NID...
957 if (peerid4.nid < lnet_nid_to_nid4(&ni->ni_nid) &&
958 ksocknal_connecting(peer_ni->ksnp_conn_cb,
959 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
961 warn = "connection race resolution";
966 if (peer_ni->ksnp_closing ||
967 (active && conn_cb->ksnr_deleted)) {
968 /* peer_ni/conn_cb got closed under me */
970 warn = "peer_ni/conn_cb removed";
974 if (peer_ni->ksnp_proto == NULL) {
975 /* Never connected before.
976 * NB recv_hello may have returned EPROTO to signal my peer_ni
977 * wants a different protocol than the one I asked for.
979 LASSERT(list_empty(&peer_ni->ksnp_conns));
981 peer_ni->ksnp_proto = conn->ksnc_proto;
982 peer_ni->ksnp_incarnation = incarnation;
985 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
986 peer_ni->ksnp_incarnation != incarnation) {
987 /* peer_ni rebooted or I've got the wrong protocol version */
988 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
990 peer_ni->ksnp_proto = NULL;
992 warn = peer_ni->ksnp_incarnation != incarnation ?
994 "wrong proto version";
1004 warn = "lost conn race";
1007 warn = "retry with different protocol version";
1011 /* Refuse to duplicate an existing connection, unless this is a
1012 * loopback connection */
1013 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1014 (struct sockaddr *)&conn->ksnc_myaddr)) {
1015 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1017 (struct sockaddr *)&conn2->ksnc_peeraddr,
1018 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1020 (struct sockaddr *)&conn2->ksnc_myaddr,
1021 (struct sockaddr *)&conn->ksnc_myaddr) ||
1022 conn2->ksnc_type != conn->ksnc_type)
1026 /* If max conns per type is not registered in conn_cb
1027 * as ksnr_max_conns, use ni's conns_per_peer
1029 if ((peer_ni->ksnp_conn_cb &&
1030 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1031 (!peer_ni->ksnp_conn_cb &&
1032 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1035 /* Reply on a passive connection attempt so the peer_ni
1036 * realises we're connected.
1046 /* If the connection created by this route didn't bind to the IP
1047 * address the route connected to, the connection/route matching
1048 * code below probably isn't going to work.
1051 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1052 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1053 CERROR("Route %s %pIS connected to %pIS\n",
1054 libcfs_idstr(&peer_ni->ksnp_id),
1055 &conn_cb->ksnr_addr,
1056 &conn->ksnc_peeraddr);
1059 /* Search for a conn_cb corresponding to the new connection and
1060 * create an association. This allows incoming connections created
1061 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1062 * continually create duplicate conn_cbs.
1064 conn_cb = peer_ni->ksnp_conn_cb;
1066 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1067 (struct sockaddr *)&conn_cb->ksnr_addr))
1068 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1070 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1071 peer_ni->ksnp_last_alive = ktime_get_seconds();
1072 peer_ni->ksnp_send_keepalive = 0;
1073 peer_ni->ksnp_error = 0;
1075 sched = ksocknal_choose_scheduler_locked(cpt);
1077 CERROR("no schedulers available. node is unhealthy\n");
1081 * The cpt might have changed if we ended up selecting a non cpt
1082 * native scheduler. So use the scheduler's cpt instead.
1084 cpt = sched->kss_cpt;
1085 sched->kss_nconns++;
1086 conn->ksnc_scheduler = sched;
1088 conn->ksnc_tx_last_post = ktime_get_seconds();
1089 /* Set the deadline for the outgoing HELLO to drain */
1090 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1091 conn->ksnc_tx_deadline = ktime_get_seconds() +
1093 smp_mb(); /* order with adding to peer_ni's conn list */
1095 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1096 ksocknal_conn_addref(conn);
1098 ksocknal_new_packet(conn, 0);
1100 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1102 /* Take packets blocking for this connection. */
1103 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1104 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1108 list_del(&tx->tx_list);
1109 ksocknal_queue_tx_locked(tx, conn);
1112 write_unlock_bh(global_lock);
1114 /* We've now got a new connection. Any errors from here on are just
1115 * like "normal" comms errors and we close the connection normally.
1116 * NB (a) we still have to send the reply HELLO for passive
1118 * (b) normal I/O on the conn is blocked until I setup and call the
1122 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1123 " incarnation:%lld sched[%d]\n",
1124 libcfs_id2str(peerid4), conn->ksnc_proto->pro_version,
1125 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1129 hello->kshm_nips = 0;
1130 rc = ksocknal_send_hello(ni, conn, peerid4.nid, hello);
1133 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1134 kshm_ips[LNET_INTERFACES_NUM]));
1136 /* setup the socket AFTER I've received hello (it disables
1137 * SO_LINGER). I might call back to the acceptor who may want
1138 * to send a protocol version response and then close the
1139 * socket; this ensures the socket only tears down after the
1140 * response has been sent.
1143 rc = ksocknal_lib_setup_sock(sock);
1145 write_lock_bh(global_lock);
1147 /* NB my callbacks block while I hold ksnd_global_lock */
1148 ksocknal_lib_set_callback(sock, conn);
1151 peer_ni->ksnp_accepting--;
1153 write_unlock_bh(global_lock);
1156 write_lock_bh(global_lock);
1157 if (!conn->ksnc_closing) {
1158 /* could be closed by another thread */
1159 ksocknal_close_conn_locked(conn, rc);
1161 write_unlock_bh(global_lock);
1162 } else if (ksocknal_connsock_addref(conn) == 0) {
1163 /* Allow I/O to proceed. */
1164 ksocknal_read_callback(conn);
1165 ksocknal_write_callback(conn);
1166 ksocknal_connsock_decref(conn);
1169 ksocknal_connsock_decref(conn);
1170 ksocknal_conn_decref(conn);
1175 if (!peer_ni->ksnp_closing &&
1176 list_empty(&peer_ni->ksnp_conns) &&
1177 peer_ni->ksnp_conn_cb == NULL) {
1178 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1179 ksocknal_unlink_peer_locked(peer_ni);
1182 write_unlock_bh(global_lock);
1186 CERROR("Not creating conn %s type %d: %s\n",
1187 libcfs_id2str(peerid4), conn->ksnc_type, warn);
1189 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1190 libcfs_id2str(peerid4), conn->ksnc_type, warn);
1195 /* Request retry by replying with CONN_NONE
1196 * ksnc_proto has been set already
1198 conn->ksnc_type = SOCKLND_CONN_NONE;
1199 hello->kshm_nips = 0;
1200 ksocknal_send_hello(ni, conn, peerid4.nid, hello);
1203 write_lock_bh(global_lock);
1204 peer_ni->ksnp_accepting--;
1205 write_unlock_bh(global_lock);
1209 * If we get here without an error code, just use -EALREADY.
1210 * Depending on how we got here, the error may be positive
1211 * or negative. Normalize the value for ksocknal_txlist_done().
1213 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1214 ksocknal_txlist_done(ni, &zombies, rc2);
1215 ksocknal_peer_decref(peer_ni);
1219 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1220 kshm_ips[LNET_INTERFACES_NUM]));
1222 LIBCFS_FREE(conn, sizeof(*conn));
1231 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1233 /* This just does the immmediate housekeeping, and queues the
1234 * connection for the reaper to terminate.
1235 * Caller holds ksnd_global_lock exclusively in irq context */
1236 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1237 struct ksock_conn_cb *conn_cb;
1238 struct ksock_conn *conn2;
1240 LASSERT(peer_ni->ksnp_error == 0);
1241 LASSERT(!conn->ksnc_closing);
1242 conn->ksnc_closing = 1;
1244 /* ksnd_deathrow_conns takes over peer_ni's ref */
1245 list_del(&conn->ksnc_list);
1247 conn_cb = conn->ksnc_conn_cb;
1248 if (conn_cb != NULL) {
1249 /* dissociate conn from cb... */
1250 LASSERT(!conn_cb->ksnr_deleted);
1252 /* connected bit is set only if all connections
1253 * of the given type got created
1255 if (ksocknal_get_conn_count_by_type(conn_cb, conn->ksnc_type) ==
1256 conn_cb->ksnr_max_conns)
1257 LASSERT((conn_cb->ksnr_connected &
1258 BIT(conn->ksnc_type)) != 0);
1260 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1261 if (conn2->ksnc_conn_cb == conn_cb &&
1262 conn2->ksnc_type == conn->ksnc_type)
1265 conn_cb->ksnr_connected &= ~BIT(conn->ksnc_type);
1267 conn->ksnc_conn_cb = NULL;
1269 /* drop conn's ref on conn_cb */
1270 ksocknal_conn_cb_decref(conn_cb);
1273 if (list_empty(&peer_ni->ksnp_conns)) {
1274 /* No more connections to this peer_ni */
1276 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1277 struct ksock_tx *tx;
1279 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1281 /* throw them to the last connection...,
1282 * these TXs will be send to /dev/null by scheduler */
1283 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1285 ksocknal_tx_prep(conn, tx);
1287 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1288 list_splice_init(&peer_ni->ksnp_tx_queue,
1289 &conn->ksnc_tx_queue);
1290 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1293 /* renegotiate protocol version */
1294 peer_ni->ksnp_proto = NULL;
1295 /* stash last conn close reason */
1296 peer_ni->ksnp_error = error;
1298 if (peer_ni->ksnp_conn_cb == NULL) {
1299 /* I've just closed last conn belonging to a
1300 * peer_ni with no connections to it
1302 ksocknal_unlink_peer_locked(peer_ni);
1306 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1308 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1309 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1311 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1315 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1317 bool notify = false;
1318 time64_t last_alive = 0;
1320 /* There has been a connection failure or comms error; but I'll only
1321 * tell LNET I think the peer_ni is dead if it's to another kernel and
1322 * there are no connections or connection attempts in existence. */
1324 read_lock(&ksocknal_data.ksnd_global_lock);
1326 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1327 list_empty(&peer_ni->ksnp_conns) &&
1328 peer_ni->ksnp_accepting == 0 &&
1329 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1331 last_alive = peer_ni->ksnp_last_alive;
1334 read_unlock(&ksocknal_data.ksnd_global_lock);
1337 lnet_notify(peer_ni->ksnp_ni,
1338 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1339 false, false, last_alive);
1343 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1345 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1346 struct ksock_tx *tx;
1347 struct ksock_tx *tmp;
1350 /* NB safe to finalize TXs because closing of socket will
1351 * abort all buffered data */
1352 LASSERT(conn->ksnc_sock == NULL);
1354 spin_lock(&peer_ni->ksnp_lock);
1356 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1358 if (tx->tx_conn != conn)
1361 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1363 tx->tx_msg.ksm_zc_cookies[0] = 0;
1364 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1365 list_move(&tx->tx_zc_list, &zlist);
1368 spin_unlock(&peer_ni->ksnp_lock);
1370 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1371 tx_zc_list)) != NULL) {
1372 list_del(&tx->tx_zc_list);
1373 ksocknal_tx_decref(tx);
1378 ksocknal_terminate_conn(struct ksock_conn *conn)
1380 /* This gets called by the reaper (guaranteed thread context) to
1381 * disengage the socket from its callbacks and close it.
1382 * ksnc_refcount will eventually hit zero, and then the reaper will
1385 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1386 struct ksock_sched *sched = conn->ksnc_scheduler;
1387 bool failed = false;
1389 LASSERT(conn->ksnc_closing);
1391 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1392 spin_lock_bh(&sched->kss_lock);
1394 /* a closing conn is always ready to tx */
1395 conn->ksnc_tx_ready = 1;
1397 if (!conn->ksnc_tx_scheduled &&
1398 !list_empty(&conn->ksnc_tx_queue)) {
1399 list_add_tail(&conn->ksnc_tx_list,
1400 &sched->kss_tx_conns);
1401 conn->ksnc_tx_scheduled = 1;
1402 /* extra ref for scheduler */
1403 ksocknal_conn_addref(conn);
1405 wake_up(&sched->kss_waitq);
1408 spin_unlock_bh(&sched->kss_lock);
1410 /* serialise with callbacks */
1411 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1413 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1415 /* OK, so this conn may not be completely disengaged from its
1416 * scheduler yet, but it _has_ committed to terminate...
1418 conn->ksnc_scheduler->kss_nconns--;
1420 if (peer_ni->ksnp_error != 0) {
1421 /* peer_ni's last conn closed in error */
1422 LASSERT(list_empty(&peer_ni->ksnp_conns));
1424 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1427 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1430 ksocknal_peer_failed(peer_ni);
1432 /* The socket is closed on the final put; either here, or in
1433 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1434 * when the connection was established, this will close the socket
1435 * immediately, aborting anything buffered in it. Any hung
1436 * zero-copy transmits will therefore complete in finite time.
1438 ksocknal_connsock_decref(conn);
1442 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1444 /* Queue the conn for the reaper to destroy */
1445 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1446 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1448 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1449 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1451 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1455 ksocknal_destroy_conn(struct ksock_conn *conn)
1459 /* Final coup-de-grace of the reaper */
1460 CDEBUG(D_NET, "connection %p\n", conn);
1462 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1463 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1464 LASSERT(conn->ksnc_sock == NULL);
1465 LASSERT(conn->ksnc_conn_cb == NULL);
1466 LASSERT(!conn->ksnc_tx_scheduled);
1467 LASSERT(!conn->ksnc_rx_scheduled);
1468 LASSERT(list_empty(&conn->ksnc_tx_queue));
1470 /* complete current receive if any */
1471 switch (conn->ksnc_rx_state) {
1472 case SOCKNAL_RX_LNET_PAYLOAD:
1473 last_rcv = conn->ksnc_rx_deadline -
1475 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1476 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1478 &conn->ksnc_peeraddr,
1479 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1480 ktime_get_seconds() - last_rcv);
1481 if (conn->ksnc_lnet_msg)
1482 conn->ksnc_lnet_msg->msg_health_status =
1483 LNET_MSG_STATUS_REMOTE_ERROR;
1484 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1486 case SOCKNAL_RX_LNET_HEADER:
1487 if (conn->ksnc_rx_started)
1488 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1489 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1490 &conn->ksnc_peeraddr,
1491 conn->ksnc_proto->pro_version);
1493 case SOCKNAL_RX_KSM_HEADER:
1494 if (conn->ksnc_rx_started)
1495 CERROR("Incomplete receive of ksock message from %s, ip %pISp, with error, protocol: %d.x.\n",
1496 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1497 &conn->ksnc_peeraddr,
1498 conn->ksnc_proto->pro_version);
1500 case SOCKNAL_RX_SLOP:
1501 if (conn->ksnc_rx_started)
1502 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1503 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1504 &conn->ksnc_peeraddr);
1511 ksocknal_peer_decref(conn->ksnc_peer);
1513 LIBCFS_FREE(conn, sizeof(*conn));
1517 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1518 struct sockaddr *addr, int why)
1520 struct ksock_conn *conn;
1521 struct ksock_conn *cnxt;
1524 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1527 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1529 ksocknal_close_conn_locked(conn, why);
1537 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1539 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1542 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1544 count = ksocknal_close_peer_conns_locked(
1545 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1547 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1553 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1555 struct ksock_peer_ni *peer_ni;
1556 struct hlist_node *pnxt;
1561 struct sockaddr_in sa = {.sin_family = AF_INET};
1563 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1565 if (!LNET_NID_IS_ANY(&id->nid)) {
1566 lo = hash_min(nidhash(&id->nid),
1567 HASH_BITS(ksocknal_data.ksnd_peers));
1571 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1574 sa.sin_addr.s_addr = htonl(ipaddr);
1575 for (i = lo; i <= hi; i++) {
1576 hlist_for_each_entry_safe(peer_ni, pnxt,
1577 &ksocknal_data.ksnd_peers[i],
1580 if (!((LNET_NID_IS_ANY(&id->nid) ||
1581 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1582 (id->pid == LNET_PID_ANY ||
1583 id->pid == peer_ni->ksnp_id.pid)))
1586 count += ksocknal_close_peer_conns_locked(
1588 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1592 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1594 /* wildcards always succeed */
1595 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1599 return (count == 0 ? -ENOENT : 0);
1603 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1605 /* The router is telling me she's been notified of a change in
1608 struct lnet_processid id = {
1609 .pid = LNET_PID_ANY,
1612 CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1614 lnet_nid4_to_nid(gw_nid, &id.nid);
1615 /* If the gateway crashed, close all open connections... */
1616 ksocknal_close_matching_conns(&id, 0);
1619 /* We can only establish new connections
1620 * if we have autroutes, and these connect on demand. */
1624 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1628 struct ksock_conn *conn;
1630 for (index = 0; ; index++) {
1631 read_lock(&ksocknal_data.ksnd_global_lock);
1636 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1638 ksocknal_conn_addref(conn);
1643 read_unlock(&ksocknal_data.ksnd_global_lock);
1648 ksocknal_lib_push_conn (conn);
1649 ksocknal_conn_decref(conn);
1654 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1661 if (!LNET_NID_IS_ANY(&id->nid)) {
1662 lo = hash_min(nidhash(&id->nid),
1663 HASH_BITS(ksocknal_data.ksnd_peers));
1667 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1670 for (bkt = lo; bkt <= hi; bkt++) {
1671 int peer_off; /* searching offset in peer_ni hash table */
1673 for (peer_off = 0; ; peer_off++) {
1674 struct ksock_peer_ni *peer_ni;
1677 read_lock(&ksocknal_data.ksnd_global_lock);
1678 hlist_for_each_entry(peer_ni,
1679 &ksocknal_data.ksnd_peers[bkt],
1681 if (!((LNET_NID_IS_ANY(&id->nid) ||
1683 &peer_ni->ksnp_id.nid)) &&
1684 (id->pid == LNET_PID_ANY ||
1685 id->pid == peer_ni->ksnp_id.pid)))
1688 if (i++ == peer_off) {
1689 ksocknal_peer_addref(peer_ni);
1693 read_unlock(&ksocknal_data.ksnd_global_lock);
1695 if (i <= peer_off) /* no match */
1699 ksocknal_push_peer(peer_ni);
1700 ksocknal_peer_decref(peer_ni);
1707 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1709 struct lnet_process_id id4 = {};
1710 struct lnet_processid id = {};
1711 struct libcfs_ioctl_data *data = arg;
1715 case IOC_LIBCFS_GET_INTERFACE: {
1716 struct ksock_net *net = ni->ni_data;
1717 struct ksock_interface *iface;
1718 struct sockaddr_in *sa;
1720 read_lock(&ksocknal_data.ksnd_global_lock);
1722 if (data->ioc_count >= 1) {
1726 iface = &net->ksnn_interface;
1728 sa = (void *)&iface->ksni_addr;
1729 if (sa->sin_family == AF_INET)
1730 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1732 data->ioc_u32[0] = 0xFFFFFFFF;
1733 data->ioc_u32[1] = iface->ksni_netmask;
1734 data->ioc_u32[2] = iface->ksni_npeers;
1735 data->ioc_u32[3] = iface->ksni_nroutes;
1738 read_unlock(&ksocknal_data.ksnd_global_lock);
1742 case IOC_LIBCFS_GET_PEER: {
1747 int share_count = 0;
1749 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1750 &id4, &myip, &ip, &port,
1751 &conn_count, &share_count);
1755 data->ioc_nid = id4.nid;
1756 data->ioc_count = share_count;
1757 data->ioc_u32[0] = ip;
1758 data->ioc_u32[1] = port;
1759 data->ioc_u32[2] = myip;
1760 data->ioc_u32[3] = conn_count;
1761 data->ioc_u32[4] = id4.pid;
1765 case IOC_LIBCFS_ADD_PEER: {
1766 struct sockaddr_in sa = {.sin_family = AF_INET};
1768 id4.nid = data->ioc_nid;
1769 id4.pid = LNET_PID_LUSTRE;
1770 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1771 sa.sin_port = htons(data->ioc_u32[1]);
1772 return ksocknal_add_peer(ni, id4, (struct sockaddr *)&sa);
1774 case IOC_LIBCFS_DEL_PEER:
1775 id4.nid = data->ioc_nid;
1776 id4.pid = LNET_PID_ANY;
1777 return ksocknal_del_peer(ni, id4,
1778 data->ioc_u32[0]); /* IP */
1780 case IOC_LIBCFS_GET_CONN: {
1784 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1785 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1786 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1791 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1793 data->ioc_count = txmem;
1794 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1795 data->ioc_flags = nagle;
1796 if (psa->sin_family == AF_INET)
1797 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1799 data->ioc_u32[0] = 0xFFFFFFFF;
1800 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1801 &conn->ksnc_peeraddr);
1802 if (mysa->sin_family == AF_INET)
1803 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1805 data->ioc_u32[2] = 0xFFFFFFFF;
1806 data->ioc_u32[3] = conn->ksnc_type;
1807 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1808 data->ioc_u32[5] = rxmem;
1809 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1810 ksocknal_conn_decref(conn);
1814 case IOC_LIBCFS_CLOSE_CONNECTION:
1815 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1816 id.pid = LNET_PID_ANY;
1817 return ksocknal_close_matching_conns(&id,
1820 case IOC_LIBCFS_REGISTER_MYNID:
1821 /* Ignore if this is a noop */
1822 if (nid_is_nid4(&ni->ni_nid) &&
1823 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1826 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1827 libcfs_nid2str(data->ioc_nid),
1828 libcfs_nidstr(&ni->ni_nid));
1831 case IOC_LIBCFS_PUSH_CONNECTION:
1832 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1833 id.pid = LNET_PID_ANY;
1834 return ksocknal_push(ni, &id);
1843 ksocknal_free_buffers (void)
1845 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1847 if (ksocknal_data.ksnd_schedulers != NULL)
1848 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1850 spin_lock(&ksocknal_data.ksnd_tx_lock);
1852 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1854 struct ksock_tx *tx;
1856 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1857 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1859 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1860 tx_list)) != NULL) {
1861 list_del(&tx->tx_list);
1862 LIBCFS_FREE(tx, tx->tx_desc_size);
1865 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1869 static int ksocknal_get_link_status(struct net_device *dev)
1875 if (!netif_running(dev)) {
1877 CDEBUG(D_NET, "device not running\n");
1879 /* Some devices may not be providing link settings */
1880 else if (dev->ethtool_ops->get_link) {
1881 ret = dev->ethtool_ops->get_link(dev);
1882 CDEBUG(D_NET, "get_link returns %u\n", ret);
1889 ksocknal_handle_link_state_change(struct net_device *dev,
1890 unsigned char operstate)
1892 struct lnet_ni *ni = NULL;
1893 struct ksock_net *net;
1894 struct ksock_net *cnxt;
1896 unsigned char link_down = !(operstate == IF_OPER_UP);
1897 struct in_device *in_dev;
1898 bool found_ip = false;
1899 struct ksock_interface *ksi = NULL;
1900 struct sockaddr_in *sa;
1901 DECLARE_CONST_IN_IFADDR(ifa);
1903 ifindex = dev->ifindex;
1905 if (!ksocknal_data.ksnd_nnets)
1908 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1911 ksi = &net->ksnn_interface;
1912 sa = (void *)&ksi->ksni_addr;
1915 if (ksi->ksni_index != ifindex ||
1916 strcmp(ksi->ksni_name, dev->name))
1921 in_dev = __in_dev_get_rtnl(dev);
1923 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1925 CDEBUG(D_NET, "set link fatal state to 1\n");
1926 atomic_set(&ni->ni_fatal_error_on, 1);
1929 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1930 if (sa->sin_addr.s_addr == ifa->ifa_local)
1936 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1938 CDEBUG(D_NET, "set link fatal state to 1\n");
1939 atomic_set(&ni->ni_fatal_error_on, 1);
1944 CDEBUG(D_NET, "set link fatal state to 1\n");
1945 atomic_set(&ni->ni_fatal_error_on, link_down);
1947 CDEBUG(D_NET, "set link fatal state to %u\n",
1948 (ksocknal_get_link_status(dev) == 0));
1949 atomic_set(&ni->ni_fatal_error_on,
1950 (ksocknal_get_link_status(dev) == 0));
1959 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
1962 struct ksock_net *net;
1963 struct ksock_net *cnxt;
1964 struct net_device *event_netdev = ifa->ifa_dev->dev;
1966 struct ksock_interface *ksi = NULL;
1967 struct sockaddr_in *sa;
1969 if (!ksocknal_data.ksnd_nnets)
1972 ifindex = event_netdev->ifindex;
1974 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1977 ksi = &net->ksnn_interface;
1978 sa = (void *)&ksi->ksni_addr;
1980 if (ksi->ksni_index != ifindex ||
1981 strcmp(ksi->ksni_name, event_netdev->name))
1984 if (sa->sin_addr.s_addr == ifa->ifa_local) {
1985 CDEBUG(D_NET, "set link fatal state to %u\n",
1986 (event == NETDEV_DOWN));
1988 atomic_set(&ni->ni_fatal_error_on,
1989 (event == NETDEV_DOWN));
1996 /************************************
1997 * Net device notifier event handler
1998 ************************************/
1999 static int ksocknal_device_event(struct notifier_block *unused,
2000 unsigned long event, void *ptr)
2002 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2003 unsigned char operstate;
2005 operstate = dev->operstate;
2007 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2008 event, dev->name, dev->ifindex, operstate);
2014 ksocknal_handle_link_state_change(dev, operstate);
2021 /************************************
2022 * Inetaddr notifier event handler
2023 ************************************/
2024 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2025 unsigned long event, void *ptr)
2027 struct in_ifaddr *ifa = ptr;
2029 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2030 event, &ifa->ifa_address, &ifa->ifa_mask);
2036 ksocknal_handle_inetaddr_change(ifa, event);
2043 static struct notifier_block ksocknal_dev_notifier_block = {
2044 .notifier_call = ksocknal_device_event,
2047 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2048 .notifier_call = ksocknal_inetaddr_event,
2052 ksocknal_base_shutdown(void)
2054 struct ksock_sched *sched;
2055 struct ksock_peer_ni *peer_ni;
2058 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2059 libcfs_kmem_read());
2060 LASSERT (ksocknal_data.ksnd_nnets == 0);
2062 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2063 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2064 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2067 switch (ksocknal_data.ksnd_init) {
2072 case SOCKNAL_INIT_ALL:
2073 case SOCKNAL_INIT_DATA:
2074 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2077 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2078 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2079 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2080 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2081 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2083 if (ksocknal_data.ksnd_schedulers != NULL) {
2084 cfs_percpt_for_each(sched, i,
2085 ksocknal_data.ksnd_schedulers) {
2087 LASSERT(list_empty(&sched->kss_tx_conns));
2088 LASSERT(list_empty(&sched->kss_rx_conns));
2089 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2090 LASSERT(sched->kss_nconns == 0);
2094 /* flag threads to terminate; wake and wait for them to die */
2095 ksocknal_data.ksnd_shuttingdown = 1;
2096 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2097 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2099 if (ksocknal_data.ksnd_schedulers != NULL) {
2100 cfs_percpt_for_each(sched, i,
2101 ksocknal_data.ksnd_schedulers)
2102 wake_up_all(&sched->kss_waitq);
2105 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2106 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2107 "waiting for %d threads to terminate\n",
2108 atomic_read(&ksocknal_data.ksnd_nthreads));
2110 ksocknal_free_buffers();
2112 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2116 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2117 libcfs_kmem_read());
2119 module_put(THIS_MODULE);
2123 ksocknal_base_startup(void)
2125 struct ksock_sched *sched;
2129 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2130 LASSERT(ksocknal_data.ksnd_nnets == 0);
2132 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2134 hash_init(ksocknal_data.ksnd_peers);
2136 rwlock_init(&ksocknal_data.ksnd_global_lock);
2137 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2139 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2140 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2141 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2142 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2143 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2145 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2146 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2147 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2148 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2150 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2151 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2153 /* NB memset above zeros whole of ksocknal_data */
2155 /* flag lists/ptrs/locks initialised */
2156 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2157 if (!try_module_get(THIS_MODULE))
2160 /* Create a scheduler block per available CPT */
2161 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2163 if (ksocknal_data.ksnd_schedulers == NULL)
2166 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2170 * make sure not to allocate more threads than there are
2171 * cores/CPUs in teh CPT
2173 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2174 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2175 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2178 * max to half of CPUs, assume another half should be
2179 * reserved for upper layer modules
2181 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2184 sched->kss_nthreads_max = nthrs;
2187 spin_lock_init(&sched->kss_lock);
2188 INIT_LIST_HEAD(&sched->kss_rx_conns);
2189 INIT_LIST_HEAD(&sched->kss_tx_conns);
2190 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2191 init_waitqueue_head(&sched->kss_waitq);
2194 ksocknal_data.ksnd_connd_starting = 0;
2195 ksocknal_data.ksnd_connd_failed_stamp = 0;
2196 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2197 /* must have at least 2 connds to remain responsive to accepts while
2199 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2200 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2202 if (*ksocknal_tunables.ksnd_nconnds_max <
2203 *ksocknal_tunables.ksnd_nconnds) {
2204 ksocknal_tunables.ksnd_nconnds_max =
2205 ksocknal_tunables.ksnd_nconnds;
2208 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2209 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2210 ksocknal_data.ksnd_connd_starting++;
2211 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2213 rc = ksocknal_thread_start(ksocknal_connd,
2214 (void *)((uintptr_t)i),
2215 "socknal_cd%02d", i);
2217 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2218 ksocknal_data.ksnd_connd_starting--;
2219 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2220 CERROR("Can't spawn socknal connd: %d\n", rc);
2225 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2227 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2231 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2232 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2234 /* flag everything initialised */
2235 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2240 ksocknal_base_shutdown();
2245 ksocknal_debug_peerhash(struct lnet_ni *ni)
2247 struct ksock_peer_ni *peer_ni;
2250 read_lock(&ksocknal_data.ksnd_global_lock);
2252 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2253 struct ksock_conn_cb *conn_cb;
2254 struct ksock_conn *conn;
2256 if (peer_ni->ksnp_ni != ni)
2259 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2260 libcfs_idstr(&peer_ni->ksnp_id),
2261 refcount_read(&peer_ni->ksnp_refcount),
2262 peer_ni->ksnp_closing,
2263 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2264 peer_ni->ksnp_zc_next_cookie,
2265 !list_empty(&peer_ni->ksnp_tx_queue),
2266 !list_empty(&peer_ni->ksnp_zc_req_list));
2268 conn_cb = peer_ni->ksnp_conn_cb;
2270 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2271 refcount_read(&conn_cb->ksnr_refcount),
2272 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2273 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2276 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2277 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2278 refcount_read(&conn->ksnc_conn_refcount),
2279 refcount_read(&conn->ksnc_sock_refcount),
2280 conn->ksnc_type, conn->ksnc_closing);
2285 read_unlock(&ksocknal_data.ksnd_global_lock);
2290 ksocknal_shutdown(struct lnet_ni *ni)
2292 struct ksock_net *net = ni->ni_data;
2293 struct lnet_process_id anyid = {
2294 .nid = LNET_NID_ANY,
2295 .pid = LNET_PID_ANY,
2298 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2299 LASSERT(ksocknal_data.ksnd_nnets > 0);
2301 /* prevent new peers */
2302 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2304 /* Delete all peers */
2305 ksocknal_del_peer(ni, anyid, 0);
2307 /* Wait for all peer_ni state to clean up */
2308 wait_var_event_warning(&net->ksnn_npeers,
2309 atomic_read(&net->ksnn_npeers) ==
2310 SOCKNAL_SHUTDOWN_BIAS,
2311 "waiting for %d peers to disconnect\n",
2312 ksocknal_debug_peerhash(ni) +
2313 atomic_read(&net->ksnn_npeers) -
2314 SOCKNAL_SHUTDOWN_BIAS);
2316 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2317 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2319 list_del(&net->ksnn_list);
2320 LIBCFS_FREE(net, sizeof(*net));
2322 ksocknal_data.ksnd_nnets--;
2323 if (ksocknal_data.ksnd_nnets == 0)
2324 ksocknal_base_shutdown();
2328 ksocknal_search_new_ipif(struct ksock_net *net)
2331 char *ifnam = &net->ksnn_interface.ksni_name[0];
2332 char *colon = strchr(ifnam, ':');
2334 struct ksock_net *tmp;
2339 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2340 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2341 char *colon2 = strchr(ifnam2, ':');
2346 found = strcmp(ifnam, ifnam2) == 0;
2359 ksocknal_start_schedulers(struct ksock_sched *sched)
2365 if (sched->kss_nthreads == 0) {
2366 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2367 nthrs = sched->kss_nthreads_max;
2369 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2371 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2372 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2374 nthrs = min(nthrs, sched->kss_nthreads_max);
2376 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2377 /* increase two threads if there is new interface */
2378 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2381 for (i = 0; i < nthrs; i++) {
2384 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2385 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2386 "socknal_sd%02d_%02d",
2388 (int)KSOCK_THREAD_SID(id));
2392 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2393 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2397 sched->kss_nthreads += i;
2402 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2404 int newif = ksocknal_search_new_ipif(net);
2408 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2411 for (i = 0; i < ncpts; i++) {
2412 struct ksock_sched *sched;
2413 int cpt = (cpts == NULL) ? i : cpts[i];
2415 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2416 sched = ksocknal_data.ksnd_schedulers[cpt];
2418 if (!newif && sched->kss_nthreads > 0)
2421 rc = ksocknal_start_schedulers(sched);
2429 ksocknal_startup(struct lnet_ni *ni)
2431 struct ksock_net *net;
2432 struct ksock_interface *ksi = NULL;
2433 struct lnet_inetdev *ifaces = NULL;
2434 struct sockaddr_in *sa;
2438 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2439 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2440 rc = ksocknal_base_startup();
2444 LIBCFS_ALLOC(net, sizeof(*net));
2447 net->ksnn_incarnation = ktime_get_real_ns();
2450 ksocknal_tunables_setup(ni);
2452 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2456 ksi = &net->ksnn_interface;
2458 /* Use the first discovered interface or look in the list */
2459 if (ni->ni_interface) {
2460 for (i = 0; i < rc; i++)
2461 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2464 /* ni_interfaces doesn't contain the interface we want */
2466 CERROR("ksocklnd: failed to find interface %s\n",
2472 ni->ni_dev_cpt = ifaces[i].li_cpt;
2473 sa = (void *)&ksi->ksni_addr;
2474 memset(sa, 0, sizeof(*sa));
2475 sa->sin_family = AF_INET;
2476 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2477 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2478 ksi->ksni_netmask = ifaces[i].li_netmask;
2479 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2481 /* call it before add it to ksocknal_data.ksnd_nets */
2482 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2487 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2488 ni->ni_nid.nid_addr[0] =
2489 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2490 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2492 ksocknal_data.ksnd_nnets++;
2497 LIBCFS_FREE(net, sizeof(*net));
2499 if (ksocknal_data.ksnd_nnets == 0)
2500 ksocknal_base_shutdown();
2505 static void __exit ksocklnd_exit(void)
2507 lnet_unregister_lnd(&the_ksocklnd);
2510 static const struct lnet_lnd the_ksocklnd = {
2511 .lnd_type = SOCKLND,
2512 .lnd_startup = ksocknal_startup,
2513 .lnd_shutdown = ksocknal_shutdown,
2514 .lnd_ctl = ksocknal_ctl,
2515 .lnd_send = ksocknal_send,
2516 .lnd_recv = ksocknal_recv,
2517 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2518 .lnd_accept = ksocknal_accept,
2521 static int __init ksocklnd_init(void)
2525 /* check ksnr_connected/connecting field large enough */
2526 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2527 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2529 rc = ksocknal_tunables_init();
2533 lnet_register_lnd(&the_ksocklnd);
2538 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2539 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2540 MODULE_VERSION("2.8.0");
2541 MODULE_LICENSE("GPL");
2543 module_init(ksocklnd_init);
2544 module_exit(ksocklnd_exit);