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_process_id peerid4;
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 peerid4 = lnet_pid_to_pid4(&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, peerid4.nid, hello);
946 peerid4.nid = LNET_NID_ANY;
947 peerid4.pid = LNET_PID_ANY;
949 /* Passive, get protocol from peer_ni */
950 conn->ksnc_proto = NULL;
953 rc = ksocknal_recv_hello(ni, conn, hello, &peerid4, &incarnation);
957 LASSERT(rc == 0 || active);
958 LASSERT(conn->ksnc_proto != NULL);
959 LASSERT(peerid4.nid != LNET_NID_ANY);
961 cpt = lnet_cpt_of_nid(peerid4.nid, ni);
964 ksocknal_peer_addref(peer_ni);
965 write_lock_bh(global_lock);
967 struct lnet_processid peerid;
969 lnet_pid4_to_pid(peerid4, &peerid);
970 peer_ni = ksocknal_create_peer(ni, &peerid);
971 if (IS_ERR(peer_ni)) {
972 rc = PTR_ERR(peer_ni);
976 write_lock_bh(global_lock);
978 /* called with a ref on ni, so shutdown can't have started */
979 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
981 peer2 = ksocknal_find_peer_locked(ni, &peerid);
983 /* NB this puts an "empty" peer_ni in the peer_ni
984 * table (which takes my ref) */
985 hash_add(ksocknal_data.ksnd_peers,
986 &peer_ni->ksnp_list, nidhash(&peerid.nid));
988 ksocknal_peer_decref(peer_ni);
993 ksocknal_peer_addref(peer_ni);
994 peer_ni->ksnp_accepting++;
996 /* Am I already connecting to this guy? Resolve in
997 * favour of higher NID...
999 if (peerid4.nid < lnet_nid_to_nid4(&ni->ni_nid) &&
1000 ksocknal_connecting(peer_ni->ksnp_conn_cb,
1001 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
1003 warn = "connection race resolution";
1008 if (peer_ni->ksnp_closing ||
1009 (active && conn_cb->ksnr_deleted)) {
1010 /* peer_ni/conn_cb got closed under me */
1012 warn = "peer_ni/conn_cb removed";
1016 if (peer_ni->ksnp_proto == NULL) {
1017 /* Never connected before.
1018 * NB recv_hello may have returned EPROTO to signal my peer_ni
1019 * wants a different protocol than the one I asked for.
1021 LASSERT(list_empty(&peer_ni->ksnp_conns));
1023 peer_ni->ksnp_proto = conn->ksnc_proto;
1024 peer_ni->ksnp_incarnation = incarnation;
1027 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1028 peer_ni->ksnp_incarnation != incarnation) {
1029 /* peer_ni rebooted or I've got the wrong protocol version */
1030 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1032 peer_ni->ksnp_proto = NULL;
1034 warn = peer_ni->ksnp_incarnation != incarnation ?
1035 "peer_ni rebooted" :
1036 "wrong proto version";
1046 warn = "lost conn race";
1049 warn = "retry with different protocol version";
1053 /* Refuse to duplicate an existing connection, unless this is a
1054 * loopback connection */
1055 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1056 (struct sockaddr *)&conn->ksnc_myaddr)) {
1057 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1059 (struct sockaddr *)&conn2->ksnc_peeraddr,
1060 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1062 (struct sockaddr *)&conn2->ksnc_myaddr,
1063 (struct sockaddr *)&conn->ksnc_myaddr) ||
1064 conn2->ksnc_type != conn->ksnc_type)
1068 /* If max conns per type is not registered in conn_cb
1069 * as ksnr_max_conns, use ni's conns_per_peer
1071 if ((peer_ni->ksnp_conn_cb &&
1072 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1073 (!peer_ni->ksnp_conn_cb &&
1074 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1077 /* Reply on a passive connection attempt so the peer_ni
1078 * realises we're connected.
1088 /* If the connection created by this route didn't bind to the IP
1089 * address the route connected to, the connection/route matching
1090 * code below probably isn't going to work.
1093 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1094 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1095 CERROR("Route %s %pIS connected to %pIS\n",
1096 libcfs_idstr(&peer_ni->ksnp_id),
1097 &conn_cb->ksnr_addr,
1098 &conn->ksnc_peeraddr);
1101 /* Search for a conn_cb corresponding to the new connection and
1102 * create an association. This allows incoming connections created
1103 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1104 * continually create duplicate conn_cbs.
1106 conn_cb = peer_ni->ksnp_conn_cb;
1108 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1109 (struct sockaddr *)&conn_cb->ksnr_addr))
1110 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1112 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1113 peer_ni->ksnp_last_alive = ktime_get_seconds();
1114 peer_ni->ksnp_send_keepalive = 0;
1115 peer_ni->ksnp_error = 0;
1117 sched = ksocknal_choose_scheduler_locked(cpt);
1119 CERROR("no schedulers available. node is unhealthy\n");
1123 * The cpt might have changed if we ended up selecting a non cpt
1124 * native scheduler. So use the scheduler's cpt instead.
1126 cpt = sched->kss_cpt;
1127 sched->kss_nconns++;
1128 conn->ksnc_scheduler = sched;
1130 conn->ksnc_tx_last_post = ktime_get_seconds();
1131 /* Set the deadline for the outgoing HELLO to drain */
1132 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1133 conn->ksnc_tx_deadline = ktime_get_seconds() +
1135 smp_mb(); /* order with adding to peer_ni's conn list */
1137 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1138 ksocknal_conn_addref(conn);
1140 ksocknal_new_packet(conn, 0);
1142 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1144 /* Take packets blocking for this connection. */
1145 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1146 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1150 list_del(&tx->tx_list);
1151 ksocknal_queue_tx_locked(tx, conn);
1154 write_unlock_bh(global_lock);
1156 /* We've now got a new connection. Any errors from here on are just
1157 * like "normal" comms errors and we close the connection normally.
1158 * NB (a) we still have to send the reply HELLO for passive
1160 * (b) normal I/O on the conn is blocked until I setup and call the
1164 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1165 " incarnation:%lld sched[%d]\n",
1166 libcfs_id2str(peerid4), conn->ksnc_proto->pro_version,
1167 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1171 hello->kshm_nips = 0;
1172 rc = ksocknal_send_hello(ni, conn, peerid4.nid, hello);
1175 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1176 kshm_ips[LNET_INTERFACES_NUM]));
1178 /* setup the socket AFTER I've received hello (it disables
1179 * SO_LINGER). I might call back to the acceptor who may want
1180 * to send a protocol version response and then close the
1181 * socket; this ensures the socket only tears down after the
1182 * response has been sent.
1185 rc = ksocknal_lib_setup_sock(sock);
1187 write_lock_bh(global_lock);
1189 /* NB my callbacks block while I hold ksnd_global_lock */
1190 ksocknal_lib_set_callback(sock, conn);
1193 peer_ni->ksnp_accepting--;
1195 write_unlock_bh(global_lock);
1198 write_lock_bh(global_lock);
1199 if (!conn->ksnc_closing) {
1200 /* could be closed by another thread */
1201 ksocknal_close_conn_locked(conn, rc);
1203 write_unlock_bh(global_lock);
1204 } else if (ksocknal_connsock_addref(conn) == 0) {
1205 /* Allow I/O to proceed. */
1206 ksocknal_read_callback(conn);
1207 ksocknal_write_callback(conn);
1208 ksocknal_connsock_decref(conn);
1211 ksocknal_connsock_decref(conn);
1212 ksocknal_conn_decref(conn);
1217 if (!peer_ni->ksnp_closing &&
1218 list_empty(&peer_ni->ksnp_conns) &&
1219 peer_ni->ksnp_conn_cb == NULL) {
1220 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1221 ksocknal_unlink_peer_locked(peer_ni);
1224 write_unlock_bh(global_lock);
1228 CERROR("Not creating conn %s type %d: %s\n",
1229 libcfs_id2str(peerid4), conn->ksnc_type, warn);
1231 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1232 libcfs_id2str(peerid4), conn->ksnc_type, warn);
1237 /* Request retry by replying with CONN_NONE
1238 * ksnc_proto has been set already
1240 conn->ksnc_type = SOCKLND_CONN_NONE;
1241 hello->kshm_nips = 0;
1242 ksocknal_send_hello(ni, conn, peerid4.nid, hello);
1245 write_lock_bh(global_lock);
1246 peer_ni->ksnp_accepting--;
1247 write_unlock_bh(global_lock);
1251 * If we get here without an error code, just use -EALREADY.
1252 * Depending on how we got here, the error may be positive
1253 * or negative. Normalize the value for ksocknal_txlist_done().
1255 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1256 ksocknal_txlist_done(ni, &zombies, rc2);
1257 ksocknal_peer_decref(peer_ni);
1261 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1262 kshm_ips[LNET_INTERFACES_NUM]));
1264 LIBCFS_FREE(conn, sizeof(*conn));
1273 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1275 /* This just does the immmediate housekeeping, and queues the
1276 * connection for the reaper to terminate.
1277 * Caller holds ksnd_global_lock exclusively in irq context */
1278 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1279 struct ksock_conn_cb *conn_cb;
1280 struct ksock_conn *conn2;
1282 int duplicate_count = 0;
1284 LASSERT(peer_ni->ksnp_error == 0);
1285 LASSERT(!conn->ksnc_closing);
1286 conn->ksnc_closing = 1;
1288 /* ksnd_deathrow_conns takes over peer_ni's ref */
1289 list_del(&conn->ksnc_list);
1291 conn_cb = conn->ksnc_conn_cb;
1292 if (conn_cb != NULL) {
1293 /* dissociate conn from cb... */
1294 LASSERT(!conn_cb->ksnr_deleted);
1296 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1298 /* connected bit is set only if all connections
1299 * of the given type got created
1301 if (conn_count == conn_cb->ksnr_max_conns)
1302 LASSERT((conn_cb->ksnr_connected &
1303 BIT(conn->ksnc_type)) != 0);
1305 if (conn_count == 1) {
1306 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1308 if (conn2->ksnc_conn_cb == conn_cb &&
1309 conn2->ksnc_type == conn->ksnc_type)
1310 duplicate_count += 1;
1312 if (duplicate_count > 0)
1313 CERROR("Found %d duplicate conns type %d\n",
1317 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1319 conn->ksnc_conn_cb = NULL;
1321 /* drop conn's ref on conn_cb */
1322 ksocknal_conn_cb_decref(conn_cb);
1325 if (list_empty(&peer_ni->ksnp_conns)) {
1326 /* No more connections to this peer_ni */
1328 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1329 struct ksock_tx *tx;
1331 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1333 /* throw them to the last connection...,
1334 * these TXs will be send to /dev/null by scheduler */
1335 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1337 ksocknal_tx_prep(conn, tx);
1339 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1340 list_splice_init(&peer_ni->ksnp_tx_queue,
1341 &conn->ksnc_tx_queue);
1342 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1345 /* renegotiate protocol version */
1346 peer_ni->ksnp_proto = NULL;
1347 /* stash last conn close reason */
1348 peer_ni->ksnp_error = error;
1350 if (peer_ni->ksnp_conn_cb == NULL) {
1351 /* I've just closed last conn belonging to a
1352 * peer_ni with no connections to it
1354 ksocknal_unlink_peer_locked(peer_ni);
1358 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1360 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1361 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1363 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1367 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1369 bool notify = false;
1370 time64_t last_alive = 0;
1372 /* There has been a connection failure or comms error; but I'll only
1373 * tell LNET I think the peer_ni is dead if it's to another kernel and
1374 * there are no connections or connection attempts in existence. */
1376 read_lock(&ksocknal_data.ksnd_global_lock);
1378 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1379 list_empty(&peer_ni->ksnp_conns) &&
1380 peer_ni->ksnp_accepting == 0 &&
1381 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1383 last_alive = peer_ni->ksnp_last_alive;
1386 read_unlock(&ksocknal_data.ksnd_global_lock);
1389 lnet_notify(peer_ni->ksnp_ni,
1390 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1391 false, false, last_alive);
1395 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1397 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1398 struct ksock_tx *tx;
1399 struct ksock_tx *tmp;
1402 /* NB safe to finalize TXs because closing of socket will
1403 * abort all buffered data */
1404 LASSERT(conn->ksnc_sock == NULL);
1406 spin_lock(&peer_ni->ksnp_lock);
1408 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1410 if (tx->tx_conn != conn)
1413 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1415 tx->tx_msg.ksm_zc_cookies[0] = 0;
1416 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1417 list_move(&tx->tx_zc_list, &zlist);
1420 spin_unlock(&peer_ni->ksnp_lock);
1422 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1423 tx_zc_list)) != NULL) {
1424 list_del(&tx->tx_zc_list);
1425 ksocknal_tx_decref(tx);
1430 ksocknal_terminate_conn(struct ksock_conn *conn)
1432 /* This gets called by the reaper (guaranteed thread context) to
1433 * disengage the socket from its callbacks and close it.
1434 * ksnc_refcount will eventually hit zero, and then the reaper will
1437 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1438 struct ksock_sched *sched = conn->ksnc_scheduler;
1439 bool failed = false;
1441 LASSERT(conn->ksnc_closing);
1443 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1444 spin_lock_bh(&sched->kss_lock);
1446 /* a closing conn is always ready to tx */
1447 conn->ksnc_tx_ready = 1;
1449 if (!conn->ksnc_tx_scheduled &&
1450 !list_empty(&conn->ksnc_tx_queue)) {
1451 list_add_tail(&conn->ksnc_tx_list,
1452 &sched->kss_tx_conns);
1453 conn->ksnc_tx_scheduled = 1;
1454 /* extra ref for scheduler */
1455 ksocknal_conn_addref(conn);
1457 wake_up(&sched->kss_waitq);
1460 spin_unlock_bh(&sched->kss_lock);
1462 /* serialise with callbacks */
1463 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1465 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1467 /* OK, so this conn may not be completely disengaged from its
1468 * scheduler yet, but it _has_ committed to terminate...
1470 conn->ksnc_scheduler->kss_nconns--;
1472 if (peer_ni->ksnp_error != 0) {
1473 /* peer_ni's last conn closed in error */
1474 LASSERT(list_empty(&peer_ni->ksnp_conns));
1476 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1479 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1482 ksocknal_peer_failed(peer_ni);
1484 /* The socket is closed on the final put; either here, or in
1485 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1486 * when the connection was established, this will close the socket
1487 * immediately, aborting anything buffered in it. Any hung
1488 * zero-copy transmits will therefore complete in finite time.
1490 ksocknal_connsock_decref(conn);
1494 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1496 /* Queue the conn for the reaper to destroy */
1497 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1498 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1500 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1501 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1503 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1507 ksocknal_destroy_conn(struct ksock_conn *conn)
1511 /* Final coup-de-grace of the reaper */
1512 CDEBUG(D_NET, "connection %p\n", conn);
1514 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1515 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1516 LASSERT(conn->ksnc_sock == NULL);
1517 LASSERT(conn->ksnc_conn_cb == NULL);
1518 LASSERT(!conn->ksnc_tx_scheduled);
1519 LASSERT(!conn->ksnc_rx_scheduled);
1520 LASSERT(list_empty(&conn->ksnc_tx_queue));
1522 /* complete current receive if any */
1523 switch (conn->ksnc_rx_state) {
1524 case SOCKNAL_RX_LNET_PAYLOAD:
1525 last_rcv = conn->ksnc_rx_deadline -
1527 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1528 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1530 &conn->ksnc_peeraddr,
1531 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1532 ktime_get_seconds() - last_rcv);
1533 if (conn->ksnc_lnet_msg)
1534 conn->ksnc_lnet_msg->msg_health_status =
1535 LNET_MSG_STATUS_REMOTE_ERROR;
1536 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1538 case SOCKNAL_RX_LNET_HEADER:
1539 if (conn->ksnc_rx_started)
1540 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1541 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1542 &conn->ksnc_peeraddr,
1543 conn->ksnc_proto->pro_version);
1545 case SOCKNAL_RX_KSM_HEADER:
1546 if (conn->ksnc_rx_started)
1547 CERROR("Incomplete receive of ksock message from %s, ip %pISp, with error, protocol: %d.x.\n",
1548 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1549 &conn->ksnc_peeraddr,
1550 conn->ksnc_proto->pro_version);
1552 case SOCKNAL_RX_SLOP:
1553 if (conn->ksnc_rx_started)
1554 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1555 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1556 &conn->ksnc_peeraddr);
1563 ksocknal_peer_decref(conn->ksnc_peer);
1565 LIBCFS_FREE(conn, sizeof(*conn));
1569 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1570 struct sockaddr *addr, int why)
1572 struct ksock_conn *conn;
1573 struct ksock_conn *cnxt;
1576 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1579 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1581 ksocknal_close_conn_locked(conn, why);
1589 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1591 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1594 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1596 count = ksocknal_close_peer_conns_locked(
1597 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1599 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1605 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1607 struct ksock_peer_ni *peer_ni;
1608 struct hlist_node *pnxt;
1613 struct sockaddr_in sa = {.sin_family = AF_INET};
1615 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1617 if (!LNET_NID_IS_ANY(&id->nid)) {
1618 lo = hash_min(nidhash(&id->nid),
1619 HASH_BITS(ksocknal_data.ksnd_peers));
1623 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1626 sa.sin_addr.s_addr = htonl(ipaddr);
1627 for (i = lo; i <= hi; i++) {
1628 hlist_for_each_entry_safe(peer_ni, pnxt,
1629 &ksocknal_data.ksnd_peers[i],
1632 if (!((LNET_NID_IS_ANY(&id->nid) ||
1633 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1634 (id->pid == LNET_PID_ANY ||
1635 id->pid == peer_ni->ksnp_id.pid)))
1638 count += ksocknal_close_peer_conns_locked(
1640 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1644 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1646 /* wildcards always succeed */
1647 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1651 return (count == 0 ? -ENOENT : 0);
1655 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1657 /* The router is telling me she's been notified of a change in
1660 struct lnet_processid id = {
1661 .pid = LNET_PID_ANY,
1664 CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1666 lnet_nid4_to_nid(gw_nid, &id.nid);
1667 /* If the gateway crashed, close all open connections... */
1668 ksocknal_close_matching_conns(&id, 0);
1671 /* We can only establish new connections
1672 * if we have autroutes, and these connect on demand. */
1676 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1680 struct ksock_conn *conn;
1682 for (index = 0; ; index++) {
1683 read_lock(&ksocknal_data.ksnd_global_lock);
1688 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1690 ksocknal_conn_addref(conn);
1695 read_unlock(&ksocknal_data.ksnd_global_lock);
1700 ksocknal_lib_push_conn (conn);
1701 ksocknal_conn_decref(conn);
1706 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1713 if (!LNET_NID_IS_ANY(&id->nid)) {
1714 lo = hash_min(nidhash(&id->nid),
1715 HASH_BITS(ksocknal_data.ksnd_peers));
1719 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1722 for (bkt = lo; bkt <= hi; bkt++) {
1723 int peer_off; /* searching offset in peer_ni hash table */
1725 for (peer_off = 0; ; peer_off++) {
1726 struct ksock_peer_ni *peer_ni;
1729 read_lock(&ksocknal_data.ksnd_global_lock);
1730 hlist_for_each_entry(peer_ni,
1731 &ksocknal_data.ksnd_peers[bkt],
1733 if (!((LNET_NID_IS_ANY(&id->nid) ||
1735 &peer_ni->ksnp_id.nid)) &&
1736 (id->pid == LNET_PID_ANY ||
1737 id->pid == peer_ni->ksnp_id.pid)))
1740 if (i++ == peer_off) {
1741 ksocknal_peer_addref(peer_ni);
1745 read_unlock(&ksocknal_data.ksnd_global_lock);
1747 if (i <= peer_off) /* no match */
1751 ksocknal_push_peer(peer_ni);
1752 ksocknal_peer_decref(peer_ni);
1759 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1761 struct lnet_process_id id4 = {};
1762 struct lnet_processid id = {};
1763 struct libcfs_ioctl_data *data = arg;
1767 case IOC_LIBCFS_GET_INTERFACE: {
1768 struct ksock_net *net = ni->ni_data;
1769 struct ksock_interface *iface;
1770 struct sockaddr_in *sa;
1772 read_lock(&ksocknal_data.ksnd_global_lock);
1774 if (data->ioc_count >= 1) {
1778 iface = &net->ksnn_interface;
1780 sa = (void *)&iface->ksni_addr;
1781 if (sa->sin_family == AF_INET)
1782 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1784 data->ioc_u32[0] = 0xFFFFFFFF;
1785 data->ioc_u32[1] = iface->ksni_netmask;
1786 data->ioc_u32[2] = iface->ksni_npeers;
1787 data->ioc_u32[3] = iface->ksni_nroutes;
1790 read_unlock(&ksocknal_data.ksnd_global_lock);
1794 case IOC_LIBCFS_GET_PEER: {
1799 int share_count = 0;
1801 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1802 &id4, &myip, &ip, &port,
1803 &conn_count, &share_count);
1807 data->ioc_nid = id4.nid;
1808 data->ioc_count = share_count;
1809 data->ioc_u32[0] = ip;
1810 data->ioc_u32[1] = port;
1811 data->ioc_u32[2] = myip;
1812 data->ioc_u32[3] = conn_count;
1813 data->ioc_u32[4] = id4.pid;
1817 case IOC_LIBCFS_ADD_PEER: {
1818 struct sockaddr_in sa = {.sin_family = AF_INET};
1820 id4.nid = data->ioc_nid;
1821 id4.pid = LNET_PID_LUSTRE;
1822 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1823 sa.sin_port = htons(data->ioc_u32[1]);
1824 return ksocknal_add_peer(ni, id4, (struct sockaddr *)&sa);
1826 case IOC_LIBCFS_DEL_PEER:
1827 id4.nid = data->ioc_nid;
1828 id4.pid = LNET_PID_ANY;
1829 return ksocknal_del_peer(ni, id4,
1830 data->ioc_u32[0]); /* IP */
1832 case IOC_LIBCFS_GET_CONN: {
1836 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1837 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1838 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1843 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1845 data->ioc_count = txmem;
1846 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1847 data->ioc_flags = nagle;
1848 if (psa->sin_family == AF_INET)
1849 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1851 data->ioc_u32[0] = 0xFFFFFFFF;
1852 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1853 &conn->ksnc_peeraddr);
1854 if (mysa->sin_family == AF_INET)
1855 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1857 data->ioc_u32[2] = 0xFFFFFFFF;
1858 data->ioc_u32[3] = conn->ksnc_type;
1859 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1860 data->ioc_u32[5] = rxmem;
1861 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1862 ksocknal_conn_decref(conn);
1866 case IOC_LIBCFS_CLOSE_CONNECTION:
1867 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1868 id.pid = LNET_PID_ANY;
1869 return ksocknal_close_matching_conns(&id,
1872 case IOC_LIBCFS_REGISTER_MYNID:
1873 /* Ignore if this is a noop */
1874 if (nid_is_nid4(&ni->ni_nid) &&
1875 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1878 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1879 libcfs_nid2str(data->ioc_nid),
1880 libcfs_nidstr(&ni->ni_nid));
1883 case IOC_LIBCFS_PUSH_CONNECTION:
1884 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1885 id.pid = LNET_PID_ANY;
1886 return ksocknal_push(ni, &id);
1895 ksocknal_free_buffers (void)
1897 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1899 if (ksocknal_data.ksnd_schedulers != NULL)
1900 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1902 spin_lock(&ksocknal_data.ksnd_tx_lock);
1904 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1906 struct ksock_tx *tx;
1908 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1909 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1911 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1912 tx_list)) != NULL) {
1913 list_del(&tx->tx_list);
1914 LIBCFS_FREE(tx, tx->tx_desc_size);
1917 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1921 static int ksocknal_get_link_status(struct net_device *dev)
1927 if (!netif_running(dev)) {
1929 CDEBUG(D_NET, "device not running\n");
1931 /* Some devices may not be providing link settings */
1932 else if (dev->ethtool_ops->get_link) {
1933 ret = dev->ethtool_ops->get_link(dev);
1934 CDEBUG(D_NET, "get_link returns %u\n", ret);
1941 ksocknal_handle_link_state_change(struct net_device *dev,
1942 unsigned char operstate)
1944 struct lnet_ni *ni = NULL;
1945 struct ksock_net *net;
1946 struct ksock_net *cnxt;
1948 unsigned char link_down = !(operstate == IF_OPER_UP);
1949 struct in_device *in_dev;
1950 bool found_ip = false;
1951 struct ksock_interface *ksi = NULL;
1952 struct sockaddr_in *sa;
1953 DECLARE_CONST_IN_IFADDR(ifa);
1955 ifindex = dev->ifindex;
1957 if (!ksocknal_data.ksnd_nnets)
1960 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1963 ksi = &net->ksnn_interface;
1964 sa = (void *)&ksi->ksni_addr;
1967 if (ksi->ksni_index != ifindex ||
1968 strcmp(ksi->ksni_name, dev->name))
1973 in_dev = __in_dev_get_rtnl(dev);
1975 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1977 CDEBUG(D_NET, "set link fatal state to 1\n");
1978 atomic_set(&ni->ni_fatal_error_on, 1);
1981 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1982 if (sa->sin_addr.s_addr == ifa->ifa_local)
1988 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1990 CDEBUG(D_NET, "set link fatal state to 1\n");
1991 atomic_set(&ni->ni_fatal_error_on, 1);
1996 CDEBUG(D_NET, "set link fatal state to 1\n");
1997 atomic_set(&ni->ni_fatal_error_on, link_down);
1999 CDEBUG(D_NET, "set link fatal state to %u\n",
2000 (ksocknal_get_link_status(dev) == 0));
2001 atomic_set(&ni->ni_fatal_error_on,
2002 (ksocknal_get_link_status(dev) == 0));
2011 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2014 struct ksock_net *net;
2015 struct ksock_net *cnxt;
2016 struct net_device *event_netdev = ifa->ifa_dev->dev;
2018 struct ksock_interface *ksi = NULL;
2019 struct sockaddr_in *sa;
2021 if (!ksocknal_data.ksnd_nnets)
2024 ifindex = event_netdev->ifindex;
2026 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2029 ksi = &net->ksnn_interface;
2030 sa = (void *)&ksi->ksni_addr;
2032 if (ksi->ksni_index != ifindex ||
2033 strcmp(ksi->ksni_name, event_netdev->name))
2036 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2037 CDEBUG(D_NET, "set link fatal state to %u\n",
2038 (event == NETDEV_DOWN));
2040 atomic_set(&ni->ni_fatal_error_on,
2041 (event == NETDEV_DOWN));
2048 /************************************
2049 * Net device notifier event handler
2050 ************************************/
2051 static int ksocknal_device_event(struct notifier_block *unused,
2052 unsigned long event, void *ptr)
2054 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2055 unsigned char operstate;
2057 operstate = dev->operstate;
2059 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2060 event, dev->name, dev->ifindex, operstate);
2066 ksocknal_handle_link_state_change(dev, operstate);
2073 /************************************
2074 * Inetaddr notifier event handler
2075 ************************************/
2076 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2077 unsigned long event, void *ptr)
2079 struct in_ifaddr *ifa = ptr;
2081 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2082 event, &ifa->ifa_address, &ifa->ifa_mask);
2088 ksocknal_handle_inetaddr_change(ifa, event);
2095 static struct notifier_block ksocknal_dev_notifier_block = {
2096 .notifier_call = ksocknal_device_event,
2099 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2100 .notifier_call = ksocknal_inetaddr_event,
2104 ksocknal_base_shutdown(void)
2106 struct ksock_sched *sched;
2107 struct ksock_peer_ni *peer_ni;
2110 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2111 libcfs_kmem_read());
2112 LASSERT (ksocknal_data.ksnd_nnets == 0);
2114 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2115 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2116 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2119 switch (ksocknal_data.ksnd_init) {
2124 case SOCKNAL_INIT_ALL:
2125 case SOCKNAL_INIT_DATA:
2126 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2129 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2130 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2131 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2132 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2133 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2135 if (ksocknal_data.ksnd_schedulers != NULL) {
2136 cfs_percpt_for_each(sched, i,
2137 ksocknal_data.ksnd_schedulers) {
2139 LASSERT(list_empty(&sched->kss_tx_conns));
2140 LASSERT(list_empty(&sched->kss_rx_conns));
2141 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2142 LASSERT(sched->kss_nconns == 0);
2146 /* flag threads to terminate; wake and wait for them to die */
2147 ksocknal_data.ksnd_shuttingdown = 1;
2148 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2149 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2151 if (ksocknal_data.ksnd_schedulers != NULL) {
2152 cfs_percpt_for_each(sched, i,
2153 ksocknal_data.ksnd_schedulers)
2154 wake_up_all(&sched->kss_waitq);
2157 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2158 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2159 "waiting for %d threads to terminate\n",
2160 atomic_read(&ksocknal_data.ksnd_nthreads));
2162 ksocknal_free_buffers();
2164 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2168 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2169 libcfs_kmem_read());
2171 module_put(THIS_MODULE);
2175 ksocknal_base_startup(void)
2177 struct ksock_sched *sched;
2181 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2182 LASSERT(ksocknal_data.ksnd_nnets == 0);
2184 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2186 hash_init(ksocknal_data.ksnd_peers);
2188 rwlock_init(&ksocknal_data.ksnd_global_lock);
2189 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2191 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2192 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2193 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2194 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2195 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2197 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2198 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2199 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2200 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2202 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2203 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2205 /* NB memset above zeros whole of ksocknal_data */
2207 /* flag lists/ptrs/locks initialised */
2208 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2209 if (!try_module_get(THIS_MODULE))
2212 /* Create a scheduler block per available CPT */
2213 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2215 if (ksocknal_data.ksnd_schedulers == NULL)
2218 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2222 * make sure not to allocate more threads than there are
2223 * cores/CPUs in teh CPT
2225 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2226 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2227 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2230 * max to half of CPUs, assume another half should be
2231 * reserved for upper layer modules
2233 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2236 sched->kss_nthreads_max = nthrs;
2239 spin_lock_init(&sched->kss_lock);
2240 INIT_LIST_HEAD(&sched->kss_rx_conns);
2241 INIT_LIST_HEAD(&sched->kss_tx_conns);
2242 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2243 init_waitqueue_head(&sched->kss_waitq);
2246 ksocknal_data.ksnd_connd_starting = 0;
2247 ksocknal_data.ksnd_connd_failed_stamp = 0;
2248 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2249 /* must have at least 2 connds to remain responsive to accepts while
2251 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2252 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2254 if (*ksocknal_tunables.ksnd_nconnds_max <
2255 *ksocknal_tunables.ksnd_nconnds) {
2256 ksocknal_tunables.ksnd_nconnds_max =
2257 ksocknal_tunables.ksnd_nconnds;
2260 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2261 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2262 ksocknal_data.ksnd_connd_starting++;
2263 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2265 rc = ksocknal_thread_start(ksocknal_connd,
2266 (void *)((uintptr_t)i),
2267 "socknal_cd%02d", i);
2269 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2270 ksocknal_data.ksnd_connd_starting--;
2271 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2272 CERROR("Can't spawn socknal connd: %d\n", rc);
2277 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2279 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2283 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2284 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2286 /* flag everything initialised */
2287 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2292 ksocknal_base_shutdown();
2297 ksocknal_debug_peerhash(struct lnet_ni *ni)
2299 struct ksock_peer_ni *peer_ni;
2302 read_lock(&ksocknal_data.ksnd_global_lock);
2304 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2305 struct ksock_conn_cb *conn_cb;
2306 struct ksock_conn *conn;
2308 if (peer_ni->ksnp_ni != ni)
2311 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2312 libcfs_idstr(&peer_ni->ksnp_id),
2313 refcount_read(&peer_ni->ksnp_refcount),
2314 peer_ni->ksnp_closing,
2315 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2316 peer_ni->ksnp_zc_next_cookie,
2317 !list_empty(&peer_ni->ksnp_tx_queue),
2318 !list_empty(&peer_ni->ksnp_zc_req_list));
2320 conn_cb = peer_ni->ksnp_conn_cb;
2322 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2323 refcount_read(&conn_cb->ksnr_refcount),
2324 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2325 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2328 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2329 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2330 refcount_read(&conn->ksnc_conn_refcount),
2331 refcount_read(&conn->ksnc_sock_refcount),
2332 conn->ksnc_type, conn->ksnc_closing);
2337 read_unlock(&ksocknal_data.ksnd_global_lock);
2342 ksocknal_shutdown(struct lnet_ni *ni)
2344 struct ksock_net *net = ni->ni_data;
2345 struct lnet_process_id anyid = {
2346 .nid = LNET_NID_ANY,
2347 .pid = LNET_PID_ANY,
2350 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2351 LASSERT(ksocknal_data.ksnd_nnets > 0);
2353 /* prevent new peers */
2354 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2356 /* Delete all peers */
2357 ksocknal_del_peer(ni, anyid, 0);
2359 /* Wait for all peer_ni state to clean up */
2360 wait_var_event_warning(&net->ksnn_npeers,
2361 atomic_read(&net->ksnn_npeers) ==
2362 SOCKNAL_SHUTDOWN_BIAS,
2363 "waiting for %d peers to disconnect\n",
2364 ksocknal_debug_peerhash(ni) +
2365 atomic_read(&net->ksnn_npeers) -
2366 SOCKNAL_SHUTDOWN_BIAS);
2368 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2369 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2371 list_del(&net->ksnn_list);
2372 LIBCFS_FREE(net, sizeof(*net));
2374 ksocknal_data.ksnd_nnets--;
2375 if (ksocknal_data.ksnd_nnets == 0)
2376 ksocknal_base_shutdown();
2380 ksocknal_search_new_ipif(struct ksock_net *net)
2383 char *ifnam = &net->ksnn_interface.ksni_name[0];
2384 char *colon = strchr(ifnam, ':');
2386 struct ksock_net *tmp;
2391 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2392 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2393 char *colon2 = strchr(ifnam2, ':');
2398 found = strcmp(ifnam, ifnam2) == 0;
2411 ksocknal_start_schedulers(struct ksock_sched *sched)
2417 if (sched->kss_nthreads == 0) {
2418 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2419 nthrs = sched->kss_nthreads_max;
2421 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2423 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2424 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2426 nthrs = min(nthrs, sched->kss_nthreads_max);
2428 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2429 /* increase two threads if there is new interface */
2430 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2433 for (i = 0; i < nthrs; i++) {
2436 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2437 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2438 "socknal_sd%02d_%02d",
2440 (int)KSOCK_THREAD_SID(id));
2444 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2445 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2449 sched->kss_nthreads += i;
2454 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2456 int newif = ksocknal_search_new_ipif(net);
2460 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2463 for (i = 0; i < ncpts; i++) {
2464 struct ksock_sched *sched;
2465 int cpt = (cpts == NULL) ? i : cpts[i];
2467 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2468 sched = ksocknal_data.ksnd_schedulers[cpt];
2470 if (!newif && sched->kss_nthreads > 0)
2473 rc = ksocknal_start_schedulers(sched);
2481 ksocknal_startup(struct lnet_ni *ni)
2483 struct ksock_net *net;
2484 struct ksock_interface *ksi = NULL;
2485 struct lnet_inetdev *ifaces = NULL;
2486 struct sockaddr_in *sa;
2490 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2491 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2492 rc = ksocknal_base_startup();
2496 LIBCFS_ALLOC(net, sizeof(*net));
2499 net->ksnn_incarnation = ktime_get_real_ns();
2502 ksocknal_tunables_setup(ni);
2504 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2508 ksi = &net->ksnn_interface;
2510 /* Use the first discovered interface or look in the list */
2511 if (ni->ni_interface) {
2512 for (i = 0; i < rc; i++)
2513 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2516 /* ni_interfaces doesn't contain the interface we want */
2518 CERROR("ksocklnd: failed to find interface %s\n",
2524 ni->ni_dev_cpt = ifaces[i].li_cpt;
2525 sa = (void *)&ksi->ksni_addr;
2526 memset(sa, 0, sizeof(*sa));
2527 sa->sin_family = AF_INET;
2528 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2529 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2530 ksi->ksni_netmask = ifaces[i].li_netmask;
2531 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2533 /* call it before add it to ksocknal_data.ksnd_nets */
2534 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2539 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2540 ni->ni_nid.nid_addr[0] =
2541 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2542 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2544 ksocknal_data.ksnd_nnets++;
2549 LIBCFS_FREE(net, sizeof(*net));
2551 if (ksocknal_data.ksnd_nnets == 0)
2552 ksocknal_base_shutdown();
2557 static void __exit ksocklnd_exit(void)
2559 lnet_unregister_lnd(&the_ksocklnd);
2562 static const struct lnet_lnd the_ksocklnd = {
2563 .lnd_type = SOCKLND,
2564 .lnd_startup = ksocknal_startup,
2565 .lnd_shutdown = ksocknal_shutdown,
2566 .lnd_ctl = ksocknal_ctl,
2567 .lnd_send = ksocknal_send,
2568 .lnd_recv = ksocknal_recv,
2569 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2570 .lnd_accept = ksocknal_accept,
2573 static int __init ksocklnd_init(void)
2577 /* check ksnr_connected/connecting field large enough */
2578 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2579 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2581 rc = ksocknal_tunables_init();
2585 lnet_register_lnd(&the_ksocklnd);
2590 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2591 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2592 MODULE_VERSION("2.8.0");
2593 MODULE_LICENSE("GPL");
2595 module_init(ksocklnd_init);
2596 module_exit(ksocklnd_exit);