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 * two in case of loopback
428 conn_cb->ksnr_connected |= BIT(type);
430 case SOCKLND_CONN_BULK_IN:
431 conn_cb->ksnr_blki_conn_count++;
432 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
433 conn_cb->ksnr_connected |= BIT(type);
435 case SOCKLND_CONN_BULK_OUT:
436 conn_cb->ksnr_blko_conn_count++;
437 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
438 conn_cb->ksnr_connected |= BIT(type);
440 case SOCKLND_CONN_ANY:
441 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
442 conn_cb->ksnr_connected |= BIT(type);
449 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
450 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
455 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
458 conn_cb->ksnr_conn_count--;
460 /* check if all connections of the given type got created */
462 case SOCKLND_CONN_CONTROL:
463 conn_cb->ksnr_ctrl_conn_count--;
464 /* there's a single control connection per peer,
465 * two in case of loopback
467 if (conn_cb->ksnr_ctrl_conn_count == 0)
468 conn_cb->ksnr_connected &= ~BIT(type);
470 case SOCKLND_CONN_BULK_IN:
471 conn_cb->ksnr_blki_conn_count--;
472 if (conn_cb->ksnr_blki_conn_count < conn_cb->ksnr_max_conns)
473 conn_cb->ksnr_connected &= ~BIT(type);
475 case SOCKLND_CONN_BULK_OUT:
476 conn_cb->ksnr_blko_conn_count--;
477 if (conn_cb->ksnr_blko_conn_count < conn_cb->ksnr_max_conns)
478 conn_cb->ksnr_connected &= ~BIT(type);
480 case SOCKLND_CONN_ANY:
481 if (conn_cb->ksnr_conn_count < conn_cb->ksnr_max_conns)
482 conn_cb->ksnr_connected &= ~BIT(type);
489 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
490 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
494 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
495 struct ksock_conn *conn)
497 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
498 int type = conn->ksnc_type;
499 struct ksock_interface *iface;
502 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
504 conn->ksnc_conn_cb = conn_cb;
505 ksocknal_conn_cb_addref(conn_cb);
507 if (conn_cb->ksnr_myiface != conn_iface) {
508 if (conn_cb->ksnr_myiface < 0) {
509 /* route wasn't bound locally yet (the initial route) */
510 CDEBUG(D_NET, "Binding %s %pIS to interface %d\n",
511 libcfs_idstr(&peer_ni->ksnp_id),
516 "Rebinding %s %pIS from interface %d to %d\n",
517 libcfs_idstr(&peer_ni->ksnp_id),
519 conn_cb->ksnr_myiface,
522 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
523 conn_cb->ksnr_myiface);
525 iface->ksni_nroutes--;
527 conn_cb->ksnr_myiface = conn_iface;
528 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
529 conn_cb->ksnr_myiface);
531 iface->ksni_nroutes++;
534 ksocknal_incr_conn_count(conn_cb, type);
536 /* Successful connection => further attempts can
537 * proceed immediately
539 conn_cb->ksnr_retry_interval = 0;
543 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
544 struct ksock_conn_cb *conn_cb)
546 struct ksock_conn *conn;
547 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
549 LASSERT(!peer_ni->ksnp_closing);
550 LASSERT(!conn_cb->ksnr_peer);
551 LASSERT(!conn_cb->ksnr_scheduled);
552 LASSERT(!conn_cb->ksnr_connecting);
553 LASSERT(conn_cb->ksnr_connected == 0);
555 conn_cb->ksnr_peer = peer_ni;
556 ksocknal_peer_addref(peer_ni);
558 /* set the conn_cb's interface to the current net's interface */
559 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
560 net->ksnn_interface.ksni_nroutes++;
562 /* peer_ni's route list takes over my ref on 'route' */
563 peer_ni->ksnp_conn_cb = conn_cb;
565 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
566 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
567 (struct sockaddr *)&conn_cb->ksnr_addr))
570 ksocknal_associate_cb_conn_locked(conn_cb, conn);
571 /* keep going (typed conns) */
576 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
578 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
579 struct ksock_interface *iface;
580 struct ksock_conn *conn;
581 struct ksock_conn *cnxt;
583 LASSERT(!conn_cb->ksnr_deleted);
585 /* Close associated conns */
586 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
587 if (conn->ksnc_conn_cb != conn_cb)
590 ksocknal_close_conn_locked(conn, 0);
593 if (conn_cb->ksnr_myiface >= 0) {
594 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
595 conn_cb->ksnr_myiface);
597 iface->ksni_nroutes--;
600 conn_cb->ksnr_deleted = 1;
601 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
602 peer_ni->ksnp_conn_cb = NULL;
604 if (list_empty(&peer_ni->ksnp_conns)) {
605 /* I've just removed the last route to a peer_ni with no active
608 ksocknal_unlink_peer_locked(peer_ni);
613 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id4,
614 struct sockaddr *addr)
616 struct ksock_peer_ni *peer_ni;
617 struct ksock_peer_ni *peer2;
618 struct ksock_conn_cb *conn_cb;
619 struct lnet_processid id;
621 if (id4.nid == LNET_NID_ANY ||
622 id4.pid == LNET_PID_ANY)
626 lnet_nid4_to_nid(id4.nid, &id.nid);
628 /* Have a brand new peer_ni ready... */
629 peer_ni = ksocknal_create_peer(ni, &id);
631 return PTR_ERR(peer_ni);
633 conn_cb = ksocknal_create_conn_cb(addr);
635 ksocknal_peer_decref(peer_ni);
639 write_lock_bh(&ksocknal_data.ksnd_global_lock);
641 /* always called with a ref on ni, so shutdown can't have started */
642 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
645 peer2 = ksocknal_find_peer_locked(ni, &id);
647 ksocknal_peer_decref(peer_ni);
650 /* peer_ni table takes my ref on peer_ni */
651 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
655 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
657 /* Remember conns_per_peer setting at the time
658 * of connection initiation. It will define the
659 * max number of conns per type for this conn_cb
662 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
664 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
670 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
672 struct ksock_conn *conn;
673 struct ksock_conn *cnxt;
674 struct ksock_conn_cb *conn_cb;
676 LASSERT(!peer_ni->ksnp_closing);
678 /* Extra ref prevents peer_ni disappearing until I'm done with it */
679 ksocknal_peer_addref(peer_ni);
680 conn_cb = peer_ni->ksnp_conn_cb;
682 ksocknal_del_conn_cb_locked(conn_cb);
684 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
686 ksocknal_close_conn_locked(conn, 0);
688 ksocknal_peer_decref(peer_ni);
689 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
693 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id4, __u32 ip)
696 struct hlist_node *pnxt;
697 struct ksock_peer_ni *peer_ni;
702 struct lnet_processid id;
705 lnet_nid4_to_nid(id4.nid, &id.nid);
707 write_lock_bh(&ksocknal_data.ksnd_global_lock);
709 if (!LNET_NID_IS_ANY(&id.nid)) {
710 lo = hash_min(nidhash(&id.nid),
711 HASH_BITS(ksocknal_data.ksnd_peers));
715 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
718 for (i = lo; i <= hi; i++) {
719 hlist_for_each_entry_safe(peer_ni, pnxt,
720 &ksocknal_data.ksnd_peers[i],
722 if (peer_ni->ksnp_ni != ni)
725 if (!((LNET_NID_IS_ANY(&id.nid) ||
726 nid_same(&peer_ni->ksnp_id.nid, &id.nid)) &&
727 (id.pid == LNET_PID_ANY ||
728 peer_ni->ksnp_id.pid == id.pid)))
731 ksocknal_peer_addref(peer_ni); /* a ref for me... */
733 ksocknal_del_peer_locked(peer_ni, ip);
735 if (peer_ni->ksnp_closing &&
736 !list_empty(&peer_ni->ksnp_tx_queue)) {
737 LASSERT(list_empty(&peer_ni->ksnp_conns));
738 LASSERT(peer_ni->ksnp_conn_cb == NULL);
740 list_splice_init(&peer_ni->ksnp_tx_queue,
744 ksocknal_peer_decref(peer_ni); /* ...till here */
746 rc = 0; /* matched! */
750 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
752 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
757 static struct ksock_conn *
758 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
760 struct ksock_peer_ni *peer_ni;
761 struct ksock_conn *conn;
764 read_lock(&ksocknal_data.ksnd_global_lock);
766 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
767 LASSERT(!peer_ni->ksnp_closing);
769 if (peer_ni->ksnp_ni != ni)
772 list_for_each_entry(conn, &peer_ni->ksnp_conns,
777 ksocknal_conn_addref(conn);
778 read_unlock(&ksocknal_data.ksnd_global_lock);
783 read_unlock(&ksocknal_data.ksnd_global_lock);
787 static struct ksock_sched *
788 ksocknal_choose_scheduler_locked(unsigned int cpt)
790 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
793 if (sched->kss_nthreads == 0) {
794 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
795 if (sched->kss_nthreads > 0) {
796 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
797 cpt, sched->kss_cpt);
808 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
810 struct ksock_connreq *cr;
812 struct sockaddr_storage peer;
814 rc = lnet_sock_getaddr(sock, true, &peer);
816 CERROR("Can't determine new connection's address\n");
820 LIBCFS_ALLOC(cr, sizeof(*cr));
822 LCONSOLE_ERROR_MSG(0x12f,
823 "Dropping connection request from %pIS: memory exhausted\n",
830 cr->ksncr_sock = sock;
832 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
834 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
835 wake_up(&ksocknal_data.ksnd_connd_waitq);
837 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
842 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
845 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
846 return conn_cb->ksnr_connecting;
851 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
852 struct socket *sock, int type)
854 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
856 struct lnet_processid peerid;
858 struct ksock_conn *conn;
859 struct ksock_conn *conn2;
860 struct ksock_peer_ni *peer_ni = NULL;
861 struct ksock_peer_ni *peer2;
862 struct ksock_sched *sched;
863 struct ksock_hello_msg *hello;
866 struct ksock_tx *txtmp;
873 active = (conn_cb != NULL);
875 LASSERT(active == (type != SOCKLND_CONN_NONE));
877 LIBCFS_ALLOC(conn, sizeof(*conn));
883 conn->ksnc_peer = NULL;
884 conn->ksnc_conn_cb = NULL;
885 conn->ksnc_sock = sock;
886 /* 2 ref, 1 for conn, another extra ref prevents socket
887 * being closed before establishment of connection */
888 refcount_set(&conn->ksnc_sock_refcount, 2);
889 conn->ksnc_type = type;
890 ksocknal_lib_save_callback(sock, conn);
891 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
893 conn->ksnc_rx_ready = 0;
894 conn->ksnc_rx_scheduled = 0;
896 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
897 conn->ksnc_tx_ready = 0;
898 conn->ksnc_tx_scheduled = 0;
899 conn->ksnc_tx_carrier = NULL;
900 atomic_set (&conn->ksnc_tx_nob, 0);
902 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
903 kshm_ips[LNET_INTERFACES_NUM]));
909 /* stash conn's local and remote addrs */
910 rc = ksocknal_lib_get_conn_addrs(conn);
914 /* Find out/confirm peer_ni's NID and connection type and get the
915 * vector of interfaces she's willing to let me connect to.
916 * Passive connections use the listener timeout since the peer_ni sends
921 peer_ni = conn_cb->ksnr_peer;
922 LASSERT(ni == peer_ni->ksnp_ni);
924 /* Active connection sends HELLO eagerly */
925 hello->kshm_nips = 0;
926 peerid = peer_ni->ksnp_id;
928 write_lock_bh(global_lock);
929 conn->ksnc_proto = peer_ni->ksnp_proto;
930 write_unlock_bh(global_lock);
932 if (conn->ksnc_proto == NULL) {
933 conn->ksnc_proto = &ksocknal_protocol_v3x;
934 #if SOCKNAL_VERSION_DEBUG
935 if (*ksocknal_tunables.ksnd_protocol == 2)
936 conn->ksnc_proto = &ksocknal_protocol_v2x;
937 else if (*ksocknal_tunables.ksnd_protocol == 1)
938 conn->ksnc_proto = &ksocknal_protocol_v1x;
942 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
946 peerid.nid = LNET_ANY_NID;
947 peerid.pid = LNET_PID_ANY;
949 /* Passive, get protocol from peer_ni */
950 conn->ksnc_proto = NULL;
953 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
957 LASSERT(rc == 0 || active);
958 LASSERT(conn->ksnc_proto != NULL);
959 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
961 cpt = lnet_nid2cpt(&peerid.nid, ni);
964 ksocknal_peer_addref(peer_ni);
965 write_lock_bh(global_lock);
967 peer_ni = ksocknal_create_peer(ni, &peerid);
968 if (IS_ERR(peer_ni)) {
969 rc = PTR_ERR(peer_ni);
973 write_lock_bh(global_lock);
975 /* called with a ref on ni, so shutdown can't have started */
976 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
978 peer2 = ksocknal_find_peer_locked(ni, &peerid);
980 /* NB this puts an "empty" peer_ni in the peer_ni
981 * table (which takes my ref) */
982 hash_add(ksocknal_data.ksnd_peers,
983 &peer_ni->ksnp_list, nidhash(&peerid.nid));
985 ksocknal_peer_decref(peer_ni);
990 ksocknal_peer_addref(peer_ni);
991 peer_ni->ksnp_accepting++;
993 /* Am I already connecting to this guy? Resolve in
994 * favour of higher NID...
996 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
997 ksocknal_connecting(peer_ni->ksnp_conn_cb,
998 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
1000 warn = "connection race resolution";
1005 if (peer_ni->ksnp_closing ||
1006 (active && conn_cb->ksnr_deleted)) {
1007 /* peer_ni/conn_cb got closed under me */
1009 warn = "peer_ni/conn_cb removed";
1013 if (peer_ni->ksnp_proto == NULL) {
1014 /* Never connected before.
1015 * NB recv_hello may have returned EPROTO to signal my peer_ni
1016 * wants a different protocol than the one I asked for.
1018 LASSERT(list_empty(&peer_ni->ksnp_conns));
1020 peer_ni->ksnp_proto = conn->ksnc_proto;
1021 peer_ni->ksnp_incarnation = incarnation;
1024 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1025 peer_ni->ksnp_incarnation != incarnation) {
1026 /* peer_ni rebooted or I've got the wrong protocol version */
1027 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1029 peer_ni->ksnp_proto = NULL;
1031 warn = peer_ni->ksnp_incarnation != incarnation ?
1032 "peer_ni rebooted" :
1033 "wrong proto version";
1043 warn = "lost conn race";
1046 warn = "retry with different protocol version";
1050 /* Refuse to duplicate an existing connection, unless this is a
1051 * loopback connection */
1052 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1053 (struct sockaddr *)&conn->ksnc_myaddr)) {
1054 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1056 (struct sockaddr *)&conn2->ksnc_peeraddr,
1057 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1059 (struct sockaddr *)&conn2->ksnc_myaddr,
1060 (struct sockaddr *)&conn->ksnc_myaddr) ||
1061 conn2->ksnc_type != conn->ksnc_type)
1065 /* If max conns per type is not registered in conn_cb
1066 * as ksnr_max_conns, use ni's conns_per_peer
1068 if ((peer_ni->ksnp_conn_cb &&
1069 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1070 (!peer_ni->ksnp_conn_cb &&
1071 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1074 /* Reply on a passive connection attempt so the peer_ni
1075 * realises we're connected.
1085 /* If the connection created by this route didn't bind to the IP
1086 * address the route connected to, the connection/route matching
1087 * code below probably isn't going to work.
1090 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1091 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1092 CERROR("Route %s %pIS connected to %pIS\n",
1093 libcfs_idstr(&peer_ni->ksnp_id),
1094 &conn_cb->ksnr_addr,
1095 &conn->ksnc_peeraddr);
1098 /* Search for a conn_cb corresponding to the new connection and
1099 * create an association. This allows incoming connections created
1100 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1101 * continually create duplicate conn_cbs.
1103 conn_cb = peer_ni->ksnp_conn_cb;
1105 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1106 (struct sockaddr *)&conn_cb->ksnr_addr))
1107 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1109 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1110 peer_ni->ksnp_last_alive = ktime_get_seconds();
1111 peer_ni->ksnp_send_keepalive = 0;
1112 peer_ni->ksnp_error = 0;
1114 sched = ksocknal_choose_scheduler_locked(cpt);
1116 CERROR("no schedulers available. node is unhealthy\n");
1120 * The cpt might have changed if we ended up selecting a non cpt
1121 * native scheduler. So use the scheduler's cpt instead.
1123 cpt = sched->kss_cpt;
1124 sched->kss_nconns++;
1125 conn->ksnc_scheduler = sched;
1127 conn->ksnc_tx_last_post = ktime_get_seconds();
1128 /* Set the deadline for the outgoing HELLO to drain */
1129 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1130 conn->ksnc_tx_deadline = ktime_get_seconds() +
1132 smp_mb(); /* order with adding to peer_ni's conn list */
1134 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1135 ksocknal_conn_addref(conn);
1137 ksocknal_new_packet(conn, 0);
1139 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1141 /* Take packets blocking for this connection. */
1142 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1143 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1147 list_del(&tx->tx_list);
1148 ksocknal_queue_tx_locked(tx, conn);
1151 write_unlock_bh(global_lock);
1152 /* We've now got a new connection. Any errors from here on are just
1153 * like "normal" comms errors and we close the connection normally.
1154 * NB (a) we still have to send the reply HELLO for passive
1156 * (b) normal I/O on the conn is blocked until I setup and call the
1160 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1161 " incarnation:%lld sched[%d]\n",
1162 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1163 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1167 hello->kshm_nips = 0;
1168 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1171 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1172 kshm_ips[LNET_INTERFACES_NUM]));
1174 /* setup the socket AFTER I've received hello (it disables
1175 * SO_LINGER). I might call back to the acceptor who may want
1176 * to send a protocol version response and then close the
1177 * socket; this ensures the socket only tears down after the
1178 * response has been sent.
1181 rc = ksocknal_lib_setup_sock(sock);
1183 write_lock_bh(global_lock);
1185 /* NB my callbacks block while I hold ksnd_global_lock */
1186 ksocknal_lib_set_callback(sock, conn);
1189 peer_ni->ksnp_accepting--;
1191 write_unlock_bh(global_lock);
1194 write_lock_bh(global_lock);
1195 if (!conn->ksnc_closing) {
1196 /* could be closed by another thread */
1197 ksocknal_close_conn_locked(conn, rc);
1199 write_unlock_bh(global_lock);
1200 } else if (ksocknal_connsock_addref(conn) == 0) {
1201 /* Allow I/O to proceed. */
1202 ksocknal_read_callback(conn);
1203 ksocknal_write_callback(conn);
1204 ksocknal_connsock_decref(conn);
1207 ksocknal_connsock_decref(conn);
1208 ksocknal_conn_decref(conn);
1213 if (!peer_ni->ksnp_closing &&
1214 list_empty(&peer_ni->ksnp_conns) &&
1215 peer_ni->ksnp_conn_cb == NULL) {
1216 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1217 ksocknal_unlink_peer_locked(peer_ni);
1220 write_unlock_bh(global_lock);
1224 CERROR("Not creating conn %s type %d: %s\n",
1225 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1227 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1228 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1233 /* Request retry by replying with CONN_NONE
1234 * ksnc_proto has been set already
1236 conn->ksnc_type = SOCKLND_CONN_NONE;
1237 hello->kshm_nips = 0;
1238 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1241 write_lock_bh(global_lock);
1242 peer_ni->ksnp_accepting--;
1243 write_unlock_bh(global_lock);
1247 * If we get here without an error code, just use -EALREADY.
1248 * Depending on how we got here, the error may be positive
1249 * or negative. Normalize the value for ksocknal_txlist_done().
1251 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1252 ksocknal_txlist_done(ni, &zombies, rc2);
1253 ksocknal_peer_decref(peer_ni);
1257 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1258 kshm_ips[LNET_INTERFACES_NUM]));
1260 LIBCFS_FREE(conn, sizeof(*conn));
1269 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1271 /* This just does the immmediate housekeeping, and queues the
1272 * connection for the reaper to terminate.
1273 * Caller holds ksnd_global_lock exclusively in irq context */
1274 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1275 struct ksock_conn_cb *conn_cb;
1276 struct ksock_conn *conn2;
1278 int duplicate_count = 0;
1280 LASSERT(peer_ni->ksnp_error == 0);
1281 LASSERT(!conn->ksnc_closing);
1282 conn->ksnc_closing = 1;
1284 /* ksnd_deathrow_conns takes over peer_ni's ref */
1285 list_del(&conn->ksnc_list);
1287 conn_cb = conn->ksnc_conn_cb;
1288 if (conn_cb != NULL) {
1289 /* dissociate conn from cb... */
1290 LASSERT(!conn_cb->ksnr_deleted);
1292 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1294 /* connected bit is set only if all connections
1295 * of the given type got created
1297 if (conn_count == conn_cb->ksnr_max_conns)
1298 LASSERT((conn_cb->ksnr_connected &
1299 BIT(conn->ksnc_type)) != 0);
1301 if (conn_count == 1) {
1302 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1304 if (conn2->ksnc_conn_cb == conn_cb &&
1305 conn2->ksnc_type == conn->ksnc_type)
1306 duplicate_count += 1;
1308 if (duplicate_count > 0)
1309 CERROR("Found %d duplicate conns type %d\n",
1313 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1315 conn->ksnc_conn_cb = NULL;
1317 /* drop conn's ref on conn_cb */
1318 ksocknal_conn_cb_decref(conn_cb);
1321 if (list_empty(&peer_ni->ksnp_conns)) {
1322 /* No more connections to this peer_ni */
1324 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1325 struct ksock_tx *tx;
1327 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1329 /* throw them to the last connection...,
1330 * these TXs will be send to /dev/null by scheduler */
1331 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1333 ksocknal_tx_prep(conn, tx);
1335 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1336 list_splice_init(&peer_ni->ksnp_tx_queue,
1337 &conn->ksnc_tx_queue);
1338 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1341 /* renegotiate protocol version */
1342 peer_ni->ksnp_proto = NULL;
1343 /* stash last conn close reason */
1344 peer_ni->ksnp_error = error;
1346 if (peer_ni->ksnp_conn_cb == NULL) {
1347 /* I've just closed last conn belonging to a
1348 * peer_ni with no connections to it
1350 ksocknal_unlink_peer_locked(peer_ni);
1354 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1356 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1357 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1359 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1363 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1365 bool notify = false;
1366 time64_t last_alive = 0;
1368 /* There has been a connection failure or comms error; but I'll only
1369 * tell LNET I think the peer_ni is dead if it's to another kernel and
1370 * there are no connections or connection attempts in existence. */
1372 read_lock(&ksocknal_data.ksnd_global_lock);
1374 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1375 list_empty(&peer_ni->ksnp_conns) &&
1376 peer_ni->ksnp_accepting == 0 &&
1377 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1379 last_alive = peer_ni->ksnp_last_alive;
1382 read_unlock(&ksocknal_data.ksnd_global_lock);
1385 lnet_notify(peer_ni->ksnp_ni,
1386 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1387 false, false, last_alive);
1391 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1393 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1394 struct ksock_tx *tx;
1395 struct ksock_tx *tmp;
1398 /* NB safe to finalize TXs because closing of socket will
1399 * abort all buffered data */
1400 LASSERT(conn->ksnc_sock == NULL);
1402 spin_lock(&peer_ni->ksnp_lock);
1404 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1406 if (tx->tx_conn != conn)
1409 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1411 tx->tx_msg.ksm_zc_cookies[0] = 0;
1412 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1413 list_move(&tx->tx_zc_list, &zlist);
1416 spin_unlock(&peer_ni->ksnp_lock);
1418 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1419 tx_zc_list)) != NULL) {
1420 list_del(&tx->tx_zc_list);
1421 ksocknal_tx_decref(tx);
1426 ksocknal_terminate_conn(struct ksock_conn *conn)
1428 /* This gets called by the reaper (guaranteed thread context) to
1429 * disengage the socket from its callbacks and close it.
1430 * ksnc_refcount will eventually hit zero, and then the reaper will
1433 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1434 struct ksock_sched *sched = conn->ksnc_scheduler;
1435 bool failed = false;
1437 LASSERT(conn->ksnc_closing);
1439 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1440 spin_lock_bh(&sched->kss_lock);
1442 /* a closing conn is always ready to tx */
1443 conn->ksnc_tx_ready = 1;
1445 if (!conn->ksnc_tx_scheduled &&
1446 !list_empty(&conn->ksnc_tx_queue)) {
1447 list_add_tail(&conn->ksnc_tx_list,
1448 &sched->kss_tx_conns);
1449 conn->ksnc_tx_scheduled = 1;
1450 /* extra ref for scheduler */
1451 ksocknal_conn_addref(conn);
1453 wake_up(&sched->kss_waitq);
1456 spin_unlock_bh(&sched->kss_lock);
1458 /* serialise with callbacks */
1459 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1461 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1463 /* OK, so this conn may not be completely disengaged from its
1464 * scheduler yet, but it _has_ committed to terminate...
1466 conn->ksnc_scheduler->kss_nconns--;
1468 if (peer_ni->ksnp_error != 0) {
1469 /* peer_ni's last conn closed in error */
1470 LASSERT(list_empty(&peer_ni->ksnp_conns));
1472 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1475 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1478 ksocknal_peer_failed(peer_ni);
1480 /* The socket is closed on the final put; either here, or in
1481 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1482 * when the connection was established, this will close the socket
1483 * immediately, aborting anything buffered in it. Any hung
1484 * zero-copy transmits will therefore complete in finite time.
1486 ksocknal_connsock_decref(conn);
1490 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1492 /* Queue the conn for the reaper to destroy */
1493 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1494 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1496 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1497 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1499 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1503 ksocknal_destroy_conn(struct ksock_conn *conn)
1507 /* Final coup-de-grace of the reaper */
1508 CDEBUG(D_NET, "connection %p\n", conn);
1510 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1511 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1512 LASSERT(conn->ksnc_sock == NULL);
1513 LASSERT(conn->ksnc_conn_cb == NULL);
1514 LASSERT(!conn->ksnc_tx_scheduled);
1515 LASSERT(!conn->ksnc_rx_scheduled);
1516 LASSERT(list_empty(&conn->ksnc_tx_queue));
1518 /* complete current receive if any */
1519 switch (conn->ksnc_rx_state) {
1520 case SOCKNAL_RX_LNET_PAYLOAD:
1521 last_rcv = conn->ksnc_rx_deadline -
1523 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1524 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1526 &conn->ksnc_peeraddr,
1527 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1528 ktime_get_seconds() - last_rcv);
1529 if (conn->ksnc_lnet_msg)
1530 conn->ksnc_lnet_msg->msg_health_status =
1531 LNET_MSG_STATUS_REMOTE_ERROR;
1532 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1534 case SOCKNAL_RX_LNET_HEADER:
1535 if (conn->ksnc_rx_started)
1536 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1537 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1538 &conn->ksnc_peeraddr,
1539 conn->ksnc_proto->pro_version);
1541 case SOCKNAL_RX_KSM_HEADER:
1542 if (conn->ksnc_rx_started)
1543 CERROR("Incomplete receive of ksock message from %s, ip %pISp, with error, protocol: %d.x.\n",
1544 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1545 &conn->ksnc_peeraddr,
1546 conn->ksnc_proto->pro_version);
1548 case SOCKNAL_RX_SLOP:
1549 if (conn->ksnc_rx_started)
1550 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1551 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1552 &conn->ksnc_peeraddr);
1559 ksocknal_peer_decref(conn->ksnc_peer);
1561 LIBCFS_FREE(conn, sizeof(*conn));
1565 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1566 struct sockaddr *addr, int why)
1568 struct ksock_conn *conn;
1569 struct ksock_conn *cnxt;
1572 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1575 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1577 ksocknal_close_conn_locked(conn, why);
1585 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1587 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1590 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1592 count = ksocknal_close_peer_conns_locked(
1593 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1595 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1601 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1603 struct ksock_peer_ni *peer_ni;
1604 struct hlist_node *pnxt;
1609 struct sockaddr_in sa = {.sin_family = AF_INET};
1611 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1613 if (!LNET_NID_IS_ANY(&id->nid)) {
1614 lo = hash_min(nidhash(&id->nid),
1615 HASH_BITS(ksocknal_data.ksnd_peers));
1619 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1622 sa.sin_addr.s_addr = htonl(ipaddr);
1623 for (i = lo; i <= hi; i++) {
1624 hlist_for_each_entry_safe(peer_ni, pnxt,
1625 &ksocknal_data.ksnd_peers[i],
1628 if (!((LNET_NID_IS_ANY(&id->nid) ||
1629 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1630 (id->pid == LNET_PID_ANY ||
1631 id->pid == peer_ni->ksnp_id.pid)))
1634 count += ksocknal_close_peer_conns_locked(
1636 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1640 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1642 /* wildcards always succeed */
1643 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1647 return (count == 0 ? -ENOENT : 0);
1651 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1653 /* The router is telling me she's been notified of a change in
1656 struct lnet_processid id = {
1657 .pid = LNET_PID_ANY,
1661 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1663 /* If the gateway crashed, close all open connections... */
1664 ksocknal_close_matching_conns(&id, 0);
1667 /* We can only establish new connections
1668 * if we have autroutes, and these connect on demand.
1673 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1677 struct ksock_conn *conn;
1679 for (index = 0; ; index++) {
1680 read_lock(&ksocknal_data.ksnd_global_lock);
1685 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1687 ksocknal_conn_addref(conn);
1692 read_unlock(&ksocknal_data.ksnd_global_lock);
1697 ksocknal_lib_push_conn (conn);
1698 ksocknal_conn_decref(conn);
1703 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1710 if (!LNET_NID_IS_ANY(&id->nid)) {
1711 lo = hash_min(nidhash(&id->nid),
1712 HASH_BITS(ksocknal_data.ksnd_peers));
1716 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1719 for (bkt = lo; bkt <= hi; bkt++) {
1720 int peer_off; /* searching offset in peer_ni hash table */
1722 for (peer_off = 0; ; peer_off++) {
1723 struct ksock_peer_ni *peer_ni;
1726 read_lock(&ksocknal_data.ksnd_global_lock);
1727 hlist_for_each_entry(peer_ni,
1728 &ksocknal_data.ksnd_peers[bkt],
1730 if (!((LNET_NID_IS_ANY(&id->nid) ||
1732 &peer_ni->ksnp_id.nid)) &&
1733 (id->pid == LNET_PID_ANY ||
1734 id->pid == peer_ni->ksnp_id.pid)))
1737 if (i++ == peer_off) {
1738 ksocknal_peer_addref(peer_ni);
1742 read_unlock(&ksocknal_data.ksnd_global_lock);
1744 if (i <= peer_off) /* no match */
1748 ksocknal_push_peer(peer_ni);
1749 ksocknal_peer_decref(peer_ni);
1756 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1758 struct lnet_process_id id4 = {};
1759 struct lnet_processid id = {};
1760 struct libcfs_ioctl_data *data = arg;
1764 case IOC_LIBCFS_GET_INTERFACE: {
1765 struct ksock_net *net = ni->ni_data;
1766 struct ksock_interface *iface;
1767 struct sockaddr_in *sa;
1769 read_lock(&ksocknal_data.ksnd_global_lock);
1771 if (data->ioc_count >= 1) {
1775 iface = &net->ksnn_interface;
1777 sa = (void *)&iface->ksni_addr;
1778 if (sa->sin_family == AF_INET)
1779 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1781 data->ioc_u32[0] = 0xFFFFFFFF;
1782 data->ioc_u32[1] = iface->ksni_netmask;
1783 data->ioc_u32[2] = iface->ksni_npeers;
1784 data->ioc_u32[3] = iface->ksni_nroutes;
1787 read_unlock(&ksocknal_data.ksnd_global_lock);
1791 case IOC_LIBCFS_GET_PEER: {
1796 int share_count = 0;
1798 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1799 &id4, &myip, &ip, &port,
1800 &conn_count, &share_count);
1804 data->ioc_nid = id4.nid;
1805 data->ioc_count = share_count;
1806 data->ioc_u32[0] = ip;
1807 data->ioc_u32[1] = port;
1808 data->ioc_u32[2] = myip;
1809 data->ioc_u32[3] = conn_count;
1810 data->ioc_u32[4] = id4.pid;
1814 case IOC_LIBCFS_ADD_PEER: {
1815 struct sockaddr_in sa = {.sin_family = AF_INET};
1817 id4.nid = data->ioc_nid;
1818 id4.pid = LNET_PID_LUSTRE;
1819 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1820 sa.sin_port = htons(data->ioc_u32[1]);
1821 return ksocknal_add_peer(ni, id4, (struct sockaddr *)&sa);
1823 case IOC_LIBCFS_DEL_PEER:
1824 id4.nid = data->ioc_nid;
1825 id4.pid = LNET_PID_ANY;
1826 return ksocknal_del_peer(ni, id4,
1827 data->ioc_u32[0]); /* IP */
1829 case IOC_LIBCFS_GET_CONN: {
1833 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1834 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1835 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1840 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1842 data->ioc_count = txmem;
1843 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1844 data->ioc_flags = nagle;
1845 if (psa->sin_family == AF_INET)
1846 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1848 data->ioc_u32[0] = 0xFFFFFFFF;
1849 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1850 &conn->ksnc_peeraddr);
1851 if (mysa->sin_family == AF_INET)
1852 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1854 data->ioc_u32[2] = 0xFFFFFFFF;
1855 data->ioc_u32[3] = conn->ksnc_type;
1856 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1857 data->ioc_u32[5] = rxmem;
1858 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1859 ksocknal_conn_decref(conn);
1863 case IOC_LIBCFS_CLOSE_CONNECTION:
1864 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1865 id.pid = LNET_PID_ANY;
1866 return ksocknal_close_matching_conns(&id,
1869 case IOC_LIBCFS_REGISTER_MYNID:
1870 /* Ignore if this is a noop */
1871 if (nid_is_nid4(&ni->ni_nid) &&
1872 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1875 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1876 libcfs_nid2str(data->ioc_nid),
1877 libcfs_nidstr(&ni->ni_nid));
1880 case IOC_LIBCFS_PUSH_CONNECTION:
1881 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1882 id.pid = LNET_PID_ANY;
1883 return ksocknal_push(ni, &id);
1892 ksocknal_free_buffers (void)
1894 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1896 if (ksocknal_data.ksnd_schedulers != NULL)
1897 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1899 spin_lock(&ksocknal_data.ksnd_tx_lock);
1901 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1903 struct ksock_tx *tx;
1905 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1906 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1908 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1909 tx_list)) != NULL) {
1910 list_del(&tx->tx_list);
1911 LIBCFS_FREE(tx, tx->tx_desc_size);
1914 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1918 static int ksocknal_get_link_status(struct net_device *dev)
1924 if (!netif_running(dev)) {
1926 CDEBUG(D_NET, "device not running\n");
1928 /* Some devices may not be providing link settings */
1929 else if (dev->ethtool_ops->get_link) {
1930 ret = dev->ethtool_ops->get_link(dev);
1931 CDEBUG(D_NET, "get_link returns %u\n", ret);
1938 ksocknal_handle_link_state_change(struct net_device *dev,
1939 unsigned char operstate)
1941 struct lnet_ni *ni = NULL;
1942 struct ksock_net *net;
1943 struct ksock_net *cnxt;
1945 unsigned char link_down = !(operstate == IF_OPER_UP);
1946 struct in_device *in_dev;
1947 bool found_ip = false;
1948 struct ksock_interface *ksi = NULL;
1949 struct sockaddr_in *sa;
1950 DECLARE_CONST_IN_IFADDR(ifa);
1952 ifindex = dev->ifindex;
1954 if (!ksocknal_data.ksnd_nnets)
1957 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1960 ksi = &net->ksnn_interface;
1961 sa = (void *)&ksi->ksni_addr;
1964 if (ksi->ksni_index != ifindex ||
1965 strcmp(ksi->ksni_name, dev->name))
1970 in_dev = __in_dev_get_rtnl(dev);
1972 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1974 CDEBUG(D_NET, "set link fatal state to 1\n");
1975 atomic_set(&ni->ni_fatal_error_on, 1);
1978 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1979 if (sa->sin_addr.s_addr == ifa->ifa_local)
1985 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1987 CDEBUG(D_NET, "set link fatal state to 1\n");
1988 atomic_set(&ni->ni_fatal_error_on, 1);
1993 CDEBUG(D_NET, "set link fatal state to 1\n");
1994 atomic_set(&ni->ni_fatal_error_on, link_down);
1996 CDEBUG(D_NET, "set link fatal state to %u\n",
1997 (ksocknal_get_link_status(dev) == 0));
1998 atomic_set(&ni->ni_fatal_error_on,
1999 (ksocknal_get_link_status(dev) == 0));
2008 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2011 struct ksock_net *net;
2012 struct ksock_net *cnxt;
2013 struct net_device *event_netdev = ifa->ifa_dev->dev;
2015 struct ksock_interface *ksi = NULL;
2016 struct sockaddr_in *sa;
2018 if (!ksocknal_data.ksnd_nnets)
2021 ifindex = event_netdev->ifindex;
2023 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2026 ksi = &net->ksnn_interface;
2027 sa = (void *)&ksi->ksni_addr;
2029 if (ksi->ksni_index != ifindex ||
2030 strcmp(ksi->ksni_name, event_netdev->name))
2033 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2034 CDEBUG(D_NET, "set link fatal state to %u\n",
2035 (event == NETDEV_DOWN));
2037 atomic_set(&ni->ni_fatal_error_on,
2038 (event == NETDEV_DOWN));
2045 /************************************
2046 * Net device notifier event handler
2047 ************************************/
2048 static int ksocknal_device_event(struct notifier_block *unused,
2049 unsigned long event, void *ptr)
2051 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2052 unsigned char operstate;
2054 operstate = dev->operstate;
2056 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2057 event, dev->name, dev->ifindex, operstate);
2063 ksocknal_handle_link_state_change(dev, operstate);
2070 /************************************
2071 * Inetaddr notifier event handler
2072 ************************************/
2073 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2074 unsigned long event, void *ptr)
2076 struct in_ifaddr *ifa = ptr;
2078 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2079 event, &ifa->ifa_address, &ifa->ifa_mask);
2085 ksocknal_handle_inetaddr_change(ifa, event);
2092 static struct notifier_block ksocknal_dev_notifier_block = {
2093 .notifier_call = ksocknal_device_event,
2096 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2097 .notifier_call = ksocknal_inetaddr_event,
2101 ksocknal_base_shutdown(void)
2103 struct ksock_sched *sched;
2104 struct ksock_peer_ni *peer_ni;
2107 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2108 libcfs_kmem_read());
2109 LASSERT (ksocknal_data.ksnd_nnets == 0);
2111 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2112 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2113 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2116 switch (ksocknal_data.ksnd_init) {
2121 case SOCKNAL_INIT_ALL:
2122 case SOCKNAL_INIT_DATA:
2123 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2126 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2127 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2128 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2129 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2130 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2132 if (ksocknal_data.ksnd_schedulers != NULL) {
2133 cfs_percpt_for_each(sched, i,
2134 ksocknal_data.ksnd_schedulers) {
2136 LASSERT(list_empty(&sched->kss_tx_conns));
2137 LASSERT(list_empty(&sched->kss_rx_conns));
2138 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2139 LASSERT(sched->kss_nconns == 0);
2143 /* flag threads to terminate; wake and wait for them to die */
2144 ksocknal_data.ksnd_shuttingdown = 1;
2145 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2146 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2148 if (ksocknal_data.ksnd_schedulers != NULL) {
2149 cfs_percpt_for_each(sched, i,
2150 ksocknal_data.ksnd_schedulers)
2151 wake_up_all(&sched->kss_waitq);
2154 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2155 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2156 "waiting for %d threads to terminate\n",
2157 atomic_read(&ksocknal_data.ksnd_nthreads));
2159 ksocknal_free_buffers();
2161 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2165 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2166 libcfs_kmem_read());
2168 module_put(THIS_MODULE);
2172 ksocknal_base_startup(void)
2174 struct ksock_sched *sched;
2178 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2179 LASSERT(ksocknal_data.ksnd_nnets == 0);
2181 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2183 hash_init(ksocknal_data.ksnd_peers);
2185 rwlock_init(&ksocknal_data.ksnd_global_lock);
2186 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2188 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2189 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2190 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2191 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2192 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2194 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2195 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2196 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2197 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2199 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2200 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2202 /* NB memset above zeros whole of ksocknal_data */
2204 /* flag lists/ptrs/locks initialised */
2205 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2206 if (!try_module_get(THIS_MODULE))
2209 /* Create a scheduler block per available CPT */
2210 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2212 if (ksocknal_data.ksnd_schedulers == NULL)
2215 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2219 * make sure not to allocate more threads than there are
2220 * cores/CPUs in teh CPT
2222 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2223 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2224 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2227 * max to half of CPUs, assume another half should be
2228 * reserved for upper layer modules
2230 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2233 sched->kss_nthreads_max = nthrs;
2236 spin_lock_init(&sched->kss_lock);
2237 INIT_LIST_HEAD(&sched->kss_rx_conns);
2238 INIT_LIST_HEAD(&sched->kss_tx_conns);
2239 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2240 init_waitqueue_head(&sched->kss_waitq);
2243 ksocknal_data.ksnd_connd_starting = 0;
2244 ksocknal_data.ksnd_connd_failed_stamp = 0;
2245 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2246 /* must have at least 2 connds to remain responsive to accepts while
2248 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2249 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2251 if (*ksocknal_tunables.ksnd_nconnds_max <
2252 *ksocknal_tunables.ksnd_nconnds) {
2253 ksocknal_tunables.ksnd_nconnds_max =
2254 ksocknal_tunables.ksnd_nconnds;
2257 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2258 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2259 ksocknal_data.ksnd_connd_starting++;
2260 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2262 rc = ksocknal_thread_start(ksocknal_connd,
2263 (void *)((uintptr_t)i),
2264 "socknal_cd%02d", i);
2266 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2267 ksocknal_data.ksnd_connd_starting--;
2268 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2269 CERROR("Can't spawn socknal connd: %d\n", rc);
2274 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2276 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2280 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2281 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2283 /* flag everything initialised */
2284 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2289 ksocknal_base_shutdown();
2294 ksocknal_debug_peerhash(struct lnet_ni *ni)
2296 struct ksock_peer_ni *peer_ni;
2299 read_lock(&ksocknal_data.ksnd_global_lock);
2301 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2302 struct ksock_conn_cb *conn_cb;
2303 struct ksock_conn *conn;
2305 if (peer_ni->ksnp_ni != ni)
2308 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2309 libcfs_idstr(&peer_ni->ksnp_id),
2310 refcount_read(&peer_ni->ksnp_refcount),
2311 peer_ni->ksnp_closing,
2312 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2313 peer_ni->ksnp_zc_next_cookie,
2314 !list_empty(&peer_ni->ksnp_tx_queue),
2315 !list_empty(&peer_ni->ksnp_zc_req_list));
2317 conn_cb = peer_ni->ksnp_conn_cb;
2319 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2320 refcount_read(&conn_cb->ksnr_refcount),
2321 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2322 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2325 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2326 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2327 refcount_read(&conn->ksnc_conn_refcount),
2328 refcount_read(&conn->ksnc_sock_refcount),
2329 conn->ksnc_type, conn->ksnc_closing);
2334 read_unlock(&ksocknal_data.ksnd_global_lock);
2339 ksocknal_shutdown(struct lnet_ni *ni)
2341 struct ksock_net *net = ni->ni_data;
2342 struct lnet_process_id anyid = {
2343 .nid = LNET_NID_ANY,
2344 .pid = LNET_PID_ANY,
2347 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2348 LASSERT(ksocknal_data.ksnd_nnets > 0);
2350 /* prevent new peers */
2351 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2353 /* Delete all peers */
2354 ksocknal_del_peer(ni, anyid, 0);
2356 /* Wait for all peer_ni state to clean up */
2357 wait_var_event_warning(&net->ksnn_npeers,
2358 atomic_read(&net->ksnn_npeers) ==
2359 SOCKNAL_SHUTDOWN_BIAS,
2360 "waiting for %d peers to disconnect\n",
2361 ksocknal_debug_peerhash(ni) +
2362 atomic_read(&net->ksnn_npeers) -
2363 SOCKNAL_SHUTDOWN_BIAS);
2365 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2366 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2368 list_del(&net->ksnn_list);
2369 LIBCFS_FREE(net, sizeof(*net));
2371 ksocknal_data.ksnd_nnets--;
2372 if (ksocknal_data.ksnd_nnets == 0)
2373 ksocknal_base_shutdown();
2377 ksocknal_search_new_ipif(struct ksock_net *net)
2380 char *ifnam = &net->ksnn_interface.ksni_name[0];
2381 char *colon = strchr(ifnam, ':');
2383 struct ksock_net *tmp;
2388 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2389 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2390 char *colon2 = strchr(ifnam2, ':');
2395 found = strcmp(ifnam, ifnam2) == 0;
2408 ksocknal_start_schedulers(struct ksock_sched *sched)
2414 if (sched->kss_nthreads == 0) {
2415 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2416 nthrs = sched->kss_nthreads_max;
2418 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2420 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2421 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2423 nthrs = min(nthrs, sched->kss_nthreads_max);
2425 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2426 /* increase two threads if there is new interface */
2427 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2430 for (i = 0; i < nthrs; i++) {
2433 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2434 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2435 "socknal_sd%02d_%02d",
2437 (int)KSOCK_THREAD_SID(id));
2441 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2442 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2446 sched->kss_nthreads += i;
2451 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2453 int newif = ksocknal_search_new_ipif(net);
2457 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2460 for (i = 0; i < ncpts; i++) {
2461 struct ksock_sched *sched;
2462 int cpt = (cpts == NULL) ? i : cpts[i];
2464 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2465 sched = ksocknal_data.ksnd_schedulers[cpt];
2467 if (!newif && sched->kss_nthreads > 0)
2470 rc = ksocknal_start_schedulers(sched);
2478 ksocknal_startup(struct lnet_ni *ni)
2480 struct ksock_net *net;
2481 struct ksock_interface *ksi = NULL;
2482 struct lnet_inetdev *ifaces = NULL;
2483 struct sockaddr_in *sa;
2487 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2488 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2489 rc = ksocknal_base_startup();
2493 LIBCFS_ALLOC(net, sizeof(*net));
2496 net->ksnn_incarnation = ktime_get_real_ns();
2499 ksocknal_tunables_setup(ni);
2501 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2505 ksi = &net->ksnn_interface;
2507 /* Use the first discovered interface or look in the list */
2508 if (ni->ni_interface) {
2509 for (i = 0; i < rc; i++)
2510 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2513 /* ni_interfaces doesn't contain the interface we want */
2515 CERROR("ksocklnd: failed to find interface %s\n",
2521 ni->ni_dev_cpt = ifaces[i].li_cpt;
2522 sa = (void *)&ksi->ksni_addr;
2523 memset(sa, 0, sizeof(*sa));
2524 sa->sin_family = AF_INET;
2525 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2526 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2527 ksi->ksni_netmask = ifaces[i].li_netmask;
2528 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2530 /* call it before add it to ksocknal_data.ksnd_nets */
2531 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2536 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2537 ni->ni_nid.nid_addr[0] =
2538 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2539 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2541 ksocknal_data.ksnd_nnets++;
2546 LIBCFS_FREE(net, sizeof(*net));
2548 if (ksocknal_data.ksnd_nnets == 0)
2549 ksocknal_base_shutdown();
2554 static void __exit ksocklnd_exit(void)
2556 lnet_unregister_lnd(&the_ksocklnd);
2559 static const struct lnet_lnd the_ksocklnd = {
2560 .lnd_type = SOCKLND,
2561 .lnd_startup = ksocknal_startup,
2562 .lnd_shutdown = ksocknal_shutdown,
2563 .lnd_ctl = ksocknal_ctl,
2564 .lnd_send = ksocknal_send,
2565 .lnd_recv = ksocknal_recv,
2566 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2567 .lnd_accept = ksocknal_accept,
2570 static int __init ksocklnd_init(void)
2574 /* check ksnr_connected/connecting field large enough */
2575 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2576 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2578 rc = ksocknal_tunables_init();
2582 lnet_register_lnd(&the_ksocklnd);
2587 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2588 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2589 MODULE_VERSION("2.8.0");
2590 MODULE_LICENSE("GPL");
2592 module_init(ksocklnd_init);
2593 module_exit(ksocklnd_exit);