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_processid *id,
614 struct sockaddr *addr)
616 struct ksock_peer_ni *peer_ni;
617 struct ksock_peer_ni *peer2;
618 struct ksock_conn_cb *conn_cb;
620 if (LNET_NID_IS_ANY(&id->nid) ||
621 id->pid == LNET_PID_ANY)
624 /* Have a brand new peer_ni ready... */
625 peer_ni = ksocknal_create_peer(ni, id);
627 return PTR_ERR(peer_ni);
629 conn_cb = ksocknal_create_conn_cb(addr);
631 ksocknal_peer_decref(peer_ni);
635 write_lock_bh(&ksocknal_data.ksnd_global_lock);
637 /* always called with a ref on ni, so shutdown can't have started */
638 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
641 peer2 = ksocknal_find_peer_locked(ni, id);
643 ksocknal_peer_decref(peer_ni);
646 /* peer_ni table takes my ref on peer_ni */
647 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
651 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
653 /* Remember conns_per_peer setting at the time
654 * of connection initiation. It will define the
655 * max number of conns per type for this conn_cb
658 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
660 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
666 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
668 struct ksock_conn *conn;
669 struct ksock_conn *cnxt;
670 struct ksock_conn_cb *conn_cb;
672 LASSERT(!peer_ni->ksnp_closing);
674 /* Extra ref prevents peer_ni disappearing until I'm done with it */
675 ksocknal_peer_addref(peer_ni);
676 conn_cb = peer_ni->ksnp_conn_cb;
678 ksocknal_del_conn_cb_locked(conn_cb);
680 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
682 ksocknal_close_conn_locked(conn, 0);
684 ksocknal_peer_decref(peer_ni);
685 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
689 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id4, __u32 ip)
692 struct hlist_node *pnxt;
693 struct ksock_peer_ni *peer_ni;
698 struct lnet_processid id;
701 lnet_nid4_to_nid(id4.nid, &id.nid);
703 write_lock_bh(&ksocknal_data.ksnd_global_lock);
705 if (!LNET_NID_IS_ANY(&id.nid)) {
706 lo = hash_min(nidhash(&id.nid),
707 HASH_BITS(ksocknal_data.ksnd_peers));
711 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
714 for (i = lo; i <= hi; i++) {
715 hlist_for_each_entry_safe(peer_ni, pnxt,
716 &ksocknal_data.ksnd_peers[i],
718 if (peer_ni->ksnp_ni != ni)
721 if (!((LNET_NID_IS_ANY(&id.nid) ||
722 nid_same(&peer_ni->ksnp_id.nid, &id.nid)) &&
723 (id.pid == LNET_PID_ANY ||
724 peer_ni->ksnp_id.pid == id.pid)))
727 ksocknal_peer_addref(peer_ni); /* a ref for me... */
729 ksocknal_del_peer_locked(peer_ni, ip);
731 if (peer_ni->ksnp_closing &&
732 !list_empty(&peer_ni->ksnp_tx_queue)) {
733 LASSERT(list_empty(&peer_ni->ksnp_conns));
734 LASSERT(peer_ni->ksnp_conn_cb == NULL);
736 list_splice_init(&peer_ni->ksnp_tx_queue,
740 ksocknal_peer_decref(peer_ni); /* ...till here */
742 rc = 0; /* matched! */
746 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
748 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
753 static struct ksock_conn *
754 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
756 struct ksock_peer_ni *peer_ni;
757 struct ksock_conn *conn;
760 read_lock(&ksocknal_data.ksnd_global_lock);
762 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
763 LASSERT(!peer_ni->ksnp_closing);
765 if (peer_ni->ksnp_ni != ni)
768 list_for_each_entry(conn, &peer_ni->ksnp_conns,
773 ksocknal_conn_addref(conn);
774 read_unlock(&ksocknal_data.ksnd_global_lock);
779 read_unlock(&ksocknal_data.ksnd_global_lock);
783 static struct ksock_sched *
784 ksocknal_choose_scheduler_locked(unsigned int cpt)
786 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
789 if (sched->kss_nthreads == 0) {
790 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
791 if (sched->kss_nthreads > 0) {
792 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
793 cpt, sched->kss_cpt);
804 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
806 struct ksock_connreq *cr;
808 struct sockaddr_storage peer;
810 rc = lnet_sock_getaddr(sock, true, &peer);
812 CERROR("Can't determine new connection's address\n");
816 LIBCFS_ALLOC(cr, sizeof(*cr));
818 LCONSOLE_ERROR_MSG(0x12f,
819 "Dropping connection request from %pIS: memory exhausted\n",
826 cr->ksncr_sock = sock;
828 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
830 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
831 wake_up(&ksocknal_data.ksnd_connd_waitq);
833 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
838 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
841 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
842 return conn_cb->ksnr_connecting;
847 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
848 struct socket *sock, int type)
850 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
852 struct lnet_processid peerid;
854 struct ksock_conn *conn;
855 struct ksock_conn *conn2;
856 struct ksock_peer_ni *peer_ni = NULL;
857 struct ksock_peer_ni *peer2;
858 struct ksock_sched *sched;
859 struct ksock_hello_msg *hello;
862 struct ksock_tx *txtmp;
869 active = (conn_cb != NULL);
871 LASSERT(active == (type != SOCKLND_CONN_NONE));
873 LIBCFS_ALLOC(conn, sizeof(*conn));
879 conn->ksnc_peer = NULL;
880 conn->ksnc_conn_cb = NULL;
881 conn->ksnc_sock = sock;
882 /* 2 ref, 1 for conn, another extra ref prevents socket
883 * being closed before establishment of connection */
884 refcount_set(&conn->ksnc_sock_refcount, 2);
885 conn->ksnc_type = type;
886 ksocknal_lib_save_callback(sock, conn);
887 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
889 conn->ksnc_rx_ready = 0;
890 conn->ksnc_rx_scheduled = 0;
892 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
893 conn->ksnc_tx_ready = 0;
894 conn->ksnc_tx_scheduled = 0;
895 conn->ksnc_tx_carrier = NULL;
896 atomic_set (&conn->ksnc_tx_nob, 0);
898 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
899 kshm_ips[LNET_INTERFACES_NUM]));
905 /* stash conn's local and remote addrs */
906 rc = ksocknal_lib_get_conn_addrs(conn);
910 /* Find out/confirm peer_ni's NID and connection type and get the
911 * vector of interfaces she's willing to let me connect to.
912 * Passive connections use the listener timeout since the peer_ni sends
917 peer_ni = conn_cb->ksnr_peer;
918 LASSERT(ni == peer_ni->ksnp_ni);
920 /* Active connection sends HELLO eagerly */
921 hello->kshm_nips = 0;
922 peerid = peer_ni->ksnp_id;
924 write_lock_bh(global_lock);
925 conn->ksnc_proto = peer_ni->ksnp_proto;
926 write_unlock_bh(global_lock);
928 if (conn->ksnc_proto == NULL) {
929 conn->ksnc_proto = &ksocknal_protocol_v3x;
930 #if SOCKNAL_VERSION_DEBUG
931 if (*ksocknal_tunables.ksnd_protocol == 2)
932 conn->ksnc_proto = &ksocknal_protocol_v2x;
933 else if (*ksocknal_tunables.ksnd_protocol == 1)
934 conn->ksnc_proto = &ksocknal_protocol_v1x;
938 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
942 peerid.nid = LNET_ANY_NID;
943 peerid.pid = LNET_PID_ANY;
945 /* Passive, get protocol from peer_ni */
946 conn->ksnc_proto = NULL;
949 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
953 LASSERT(rc == 0 || active);
954 LASSERT(conn->ksnc_proto != NULL);
955 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
957 cpt = lnet_nid2cpt(&peerid.nid, ni);
960 ksocknal_peer_addref(peer_ni);
961 write_lock_bh(global_lock);
963 peer_ni = ksocknal_create_peer(ni, &peerid);
964 if (IS_ERR(peer_ni)) {
965 rc = PTR_ERR(peer_ni);
969 write_lock_bh(global_lock);
971 /* called with a ref on ni, so shutdown can't have started */
972 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
974 peer2 = ksocknal_find_peer_locked(ni, &peerid);
976 /* NB this puts an "empty" peer_ni in the peer_ni
977 * table (which takes my ref) */
978 hash_add(ksocknal_data.ksnd_peers,
979 &peer_ni->ksnp_list, nidhash(&peerid.nid));
981 ksocknal_peer_decref(peer_ni);
986 ksocknal_peer_addref(peer_ni);
987 peer_ni->ksnp_accepting++;
989 /* Am I already connecting to this guy? Resolve in
990 * favour of higher NID...
992 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
993 ksocknal_connecting(peer_ni->ksnp_conn_cb,
994 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
996 warn = "connection race resolution";
1001 if (peer_ni->ksnp_closing ||
1002 (active && conn_cb->ksnr_deleted)) {
1003 /* peer_ni/conn_cb got closed under me */
1005 warn = "peer_ni/conn_cb removed";
1009 if (peer_ni->ksnp_proto == NULL) {
1010 /* Never connected before.
1011 * NB recv_hello may have returned EPROTO to signal my peer_ni
1012 * wants a different protocol than the one I asked for.
1014 LASSERT(list_empty(&peer_ni->ksnp_conns));
1016 peer_ni->ksnp_proto = conn->ksnc_proto;
1017 peer_ni->ksnp_incarnation = incarnation;
1020 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1021 peer_ni->ksnp_incarnation != incarnation) {
1022 /* peer_ni rebooted or I've got the wrong protocol version */
1023 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1025 peer_ni->ksnp_proto = NULL;
1027 warn = peer_ni->ksnp_incarnation != incarnation ?
1028 "peer_ni rebooted" :
1029 "wrong proto version";
1039 warn = "lost conn race";
1042 warn = "retry with different protocol version";
1046 /* Refuse to duplicate an existing connection, unless this is a
1047 * loopback connection */
1048 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1049 (struct sockaddr *)&conn->ksnc_myaddr)) {
1050 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1052 (struct sockaddr *)&conn2->ksnc_peeraddr,
1053 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1055 (struct sockaddr *)&conn2->ksnc_myaddr,
1056 (struct sockaddr *)&conn->ksnc_myaddr) ||
1057 conn2->ksnc_type != conn->ksnc_type)
1061 /* If max conns per type is not registered in conn_cb
1062 * as ksnr_max_conns, use ni's conns_per_peer
1064 if ((peer_ni->ksnp_conn_cb &&
1065 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1066 (!peer_ni->ksnp_conn_cb &&
1067 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1070 /* Reply on a passive connection attempt so the peer_ni
1071 * realises we're connected.
1081 /* If the connection created by this route didn't bind to the IP
1082 * address the route connected to, the connection/route matching
1083 * code below probably isn't going to work.
1086 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1087 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1088 CERROR("Route %s %pIS connected to %pIS\n",
1089 libcfs_idstr(&peer_ni->ksnp_id),
1090 &conn_cb->ksnr_addr,
1091 &conn->ksnc_peeraddr);
1094 /* Search for a conn_cb corresponding to the new connection and
1095 * create an association. This allows incoming connections created
1096 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1097 * continually create duplicate conn_cbs.
1099 conn_cb = peer_ni->ksnp_conn_cb;
1101 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1102 (struct sockaddr *)&conn_cb->ksnr_addr))
1103 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1105 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1106 peer_ni->ksnp_last_alive = ktime_get_seconds();
1107 peer_ni->ksnp_send_keepalive = 0;
1108 peer_ni->ksnp_error = 0;
1110 sched = ksocknal_choose_scheduler_locked(cpt);
1112 CERROR("no schedulers available. node is unhealthy\n");
1116 * The cpt might have changed if we ended up selecting a non cpt
1117 * native scheduler. So use the scheduler's cpt instead.
1119 cpt = sched->kss_cpt;
1120 sched->kss_nconns++;
1121 conn->ksnc_scheduler = sched;
1123 conn->ksnc_tx_last_post = ktime_get_seconds();
1124 /* Set the deadline for the outgoing HELLO to drain */
1125 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1126 conn->ksnc_tx_deadline = ktime_get_seconds() +
1128 smp_mb(); /* order with adding to peer_ni's conn list */
1130 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1131 ksocknal_conn_addref(conn);
1133 ksocknal_new_packet(conn, 0);
1135 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1137 /* Take packets blocking for this connection. */
1138 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1139 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1143 list_del(&tx->tx_list);
1144 ksocknal_queue_tx_locked(tx, conn);
1147 write_unlock_bh(global_lock);
1148 /* We've now got a new connection. Any errors from here on are just
1149 * like "normal" comms errors and we close the connection normally.
1150 * NB (a) we still have to send the reply HELLO for passive
1152 * (b) normal I/O on the conn is blocked until I setup and call the
1156 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1157 " incarnation:%lld sched[%d]\n",
1158 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1159 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1163 hello->kshm_nips = 0;
1164 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1167 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1168 kshm_ips[LNET_INTERFACES_NUM]));
1170 /* setup the socket AFTER I've received hello (it disables
1171 * SO_LINGER). I might call back to the acceptor who may want
1172 * to send a protocol version response and then close the
1173 * socket; this ensures the socket only tears down after the
1174 * response has been sent.
1177 rc = ksocknal_lib_setup_sock(sock);
1179 write_lock_bh(global_lock);
1181 /* NB my callbacks block while I hold ksnd_global_lock */
1182 ksocknal_lib_set_callback(sock, conn);
1185 peer_ni->ksnp_accepting--;
1187 write_unlock_bh(global_lock);
1190 write_lock_bh(global_lock);
1191 if (!conn->ksnc_closing) {
1192 /* could be closed by another thread */
1193 ksocknal_close_conn_locked(conn, rc);
1195 write_unlock_bh(global_lock);
1196 } else if (ksocknal_connsock_addref(conn) == 0) {
1197 /* Allow I/O to proceed. */
1198 ksocknal_read_callback(conn);
1199 ksocknal_write_callback(conn);
1200 ksocknal_connsock_decref(conn);
1203 ksocknal_connsock_decref(conn);
1204 ksocknal_conn_decref(conn);
1209 if (!peer_ni->ksnp_closing &&
1210 list_empty(&peer_ni->ksnp_conns) &&
1211 peer_ni->ksnp_conn_cb == NULL) {
1212 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1213 ksocknal_unlink_peer_locked(peer_ni);
1216 write_unlock_bh(global_lock);
1220 CERROR("Not creating conn %s type %d: %s\n",
1221 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1223 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1224 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1229 /* Request retry by replying with CONN_NONE
1230 * ksnc_proto has been set already
1232 conn->ksnc_type = SOCKLND_CONN_NONE;
1233 hello->kshm_nips = 0;
1234 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1237 write_lock_bh(global_lock);
1238 peer_ni->ksnp_accepting--;
1239 write_unlock_bh(global_lock);
1243 * If we get here without an error code, just use -EALREADY.
1244 * Depending on how we got here, the error may be positive
1245 * or negative. Normalize the value for ksocknal_txlist_done().
1247 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1248 ksocknal_txlist_done(ni, &zombies, rc2);
1249 ksocknal_peer_decref(peer_ni);
1253 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1254 kshm_ips[LNET_INTERFACES_NUM]));
1256 LIBCFS_FREE(conn, sizeof(*conn));
1265 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1267 /* This just does the immmediate housekeeping, and queues the
1268 * connection for the reaper to terminate.
1269 * Caller holds ksnd_global_lock exclusively in irq context */
1270 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1271 struct ksock_conn_cb *conn_cb;
1272 struct ksock_conn *conn2;
1274 int duplicate_count = 0;
1276 LASSERT(peer_ni->ksnp_error == 0);
1277 LASSERT(!conn->ksnc_closing);
1278 conn->ksnc_closing = 1;
1280 /* ksnd_deathrow_conns takes over peer_ni's ref */
1281 list_del(&conn->ksnc_list);
1283 conn_cb = conn->ksnc_conn_cb;
1284 if (conn_cb != NULL) {
1285 /* dissociate conn from cb... */
1286 LASSERT(!conn_cb->ksnr_deleted);
1288 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1290 /* connected bit is set only if all connections
1291 * of the given type got created
1293 if (conn_count == conn_cb->ksnr_max_conns)
1294 LASSERT((conn_cb->ksnr_connected &
1295 BIT(conn->ksnc_type)) != 0);
1297 if (conn_count == 1) {
1298 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1300 if (conn2->ksnc_conn_cb == conn_cb &&
1301 conn2->ksnc_type == conn->ksnc_type)
1302 duplicate_count += 1;
1304 if (duplicate_count > 0)
1305 CERROR("Found %d duplicate conns type %d\n",
1309 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1311 conn->ksnc_conn_cb = NULL;
1313 /* drop conn's ref on conn_cb */
1314 ksocknal_conn_cb_decref(conn_cb);
1317 if (list_empty(&peer_ni->ksnp_conns)) {
1318 /* No more connections to this peer_ni */
1320 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1321 struct ksock_tx *tx;
1323 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1325 /* throw them to the last connection...,
1326 * these TXs will be send to /dev/null by scheduler */
1327 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1329 ksocknal_tx_prep(conn, tx);
1331 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1332 list_splice_init(&peer_ni->ksnp_tx_queue,
1333 &conn->ksnc_tx_queue);
1334 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1337 /* renegotiate protocol version */
1338 peer_ni->ksnp_proto = NULL;
1339 /* stash last conn close reason */
1340 peer_ni->ksnp_error = error;
1342 if (peer_ni->ksnp_conn_cb == NULL) {
1343 /* I've just closed last conn belonging to a
1344 * peer_ni with no connections to it
1346 ksocknal_unlink_peer_locked(peer_ni);
1350 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1352 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1353 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1355 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1359 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1361 bool notify = false;
1362 time64_t last_alive = 0;
1364 /* There has been a connection failure or comms error; but I'll only
1365 * tell LNET I think the peer_ni is dead if it's to another kernel and
1366 * there are no connections or connection attempts in existence. */
1368 read_lock(&ksocknal_data.ksnd_global_lock);
1370 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1371 list_empty(&peer_ni->ksnp_conns) &&
1372 peer_ni->ksnp_accepting == 0 &&
1373 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1375 last_alive = peer_ni->ksnp_last_alive;
1378 read_unlock(&ksocknal_data.ksnd_global_lock);
1381 lnet_notify(peer_ni->ksnp_ni,
1382 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1383 false, false, last_alive);
1387 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1389 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1390 struct ksock_tx *tx;
1391 struct ksock_tx *tmp;
1394 /* NB safe to finalize TXs because closing of socket will
1395 * abort all buffered data */
1396 LASSERT(conn->ksnc_sock == NULL);
1398 spin_lock(&peer_ni->ksnp_lock);
1400 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1402 if (tx->tx_conn != conn)
1405 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1407 tx->tx_msg.ksm_zc_cookies[0] = 0;
1408 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1409 list_move(&tx->tx_zc_list, &zlist);
1412 spin_unlock(&peer_ni->ksnp_lock);
1414 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1415 tx_zc_list)) != NULL) {
1416 list_del(&tx->tx_zc_list);
1417 ksocknal_tx_decref(tx);
1422 ksocknal_terminate_conn(struct ksock_conn *conn)
1424 /* This gets called by the reaper (guaranteed thread context) to
1425 * disengage the socket from its callbacks and close it.
1426 * ksnc_refcount will eventually hit zero, and then the reaper will
1429 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1430 struct ksock_sched *sched = conn->ksnc_scheduler;
1431 bool failed = false;
1433 LASSERT(conn->ksnc_closing);
1435 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1436 spin_lock_bh(&sched->kss_lock);
1438 /* a closing conn is always ready to tx */
1439 conn->ksnc_tx_ready = 1;
1441 if (!conn->ksnc_tx_scheduled &&
1442 !list_empty(&conn->ksnc_tx_queue)) {
1443 list_add_tail(&conn->ksnc_tx_list,
1444 &sched->kss_tx_conns);
1445 conn->ksnc_tx_scheduled = 1;
1446 /* extra ref for scheduler */
1447 ksocknal_conn_addref(conn);
1449 wake_up(&sched->kss_waitq);
1452 spin_unlock_bh(&sched->kss_lock);
1454 /* serialise with callbacks */
1455 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1457 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1459 /* OK, so this conn may not be completely disengaged from its
1460 * scheduler yet, but it _has_ committed to terminate...
1462 conn->ksnc_scheduler->kss_nconns--;
1464 if (peer_ni->ksnp_error != 0) {
1465 /* peer_ni's last conn closed in error */
1466 LASSERT(list_empty(&peer_ni->ksnp_conns));
1468 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1471 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1474 ksocknal_peer_failed(peer_ni);
1476 /* The socket is closed on the final put; either here, or in
1477 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1478 * when the connection was established, this will close the socket
1479 * immediately, aborting anything buffered in it. Any hung
1480 * zero-copy transmits will therefore complete in finite time.
1482 ksocknal_connsock_decref(conn);
1486 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1488 /* Queue the conn for the reaper to destroy */
1489 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1490 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1492 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1493 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1495 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1499 ksocknal_destroy_conn(struct ksock_conn *conn)
1503 /* Final coup-de-grace of the reaper */
1504 CDEBUG(D_NET, "connection %p\n", conn);
1506 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1507 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1508 LASSERT(conn->ksnc_sock == NULL);
1509 LASSERT(conn->ksnc_conn_cb == NULL);
1510 LASSERT(!conn->ksnc_tx_scheduled);
1511 LASSERT(!conn->ksnc_rx_scheduled);
1512 LASSERT(list_empty(&conn->ksnc_tx_queue));
1514 /* complete current receive if any */
1515 switch (conn->ksnc_rx_state) {
1516 case SOCKNAL_RX_LNET_PAYLOAD:
1517 last_rcv = conn->ksnc_rx_deadline -
1519 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1520 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1522 &conn->ksnc_peeraddr,
1523 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1524 ktime_get_seconds() - last_rcv);
1525 if (conn->ksnc_lnet_msg)
1526 conn->ksnc_lnet_msg->msg_health_status =
1527 LNET_MSG_STATUS_REMOTE_ERROR;
1528 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1530 case SOCKNAL_RX_LNET_HEADER:
1531 if (conn->ksnc_rx_started)
1532 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1533 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1534 &conn->ksnc_peeraddr,
1535 conn->ksnc_proto->pro_version);
1537 case SOCKNAL_RX_KSM_HEADER:
1538 if (conn->ksnc_rx_started)
1539 CERROR("Incomplete receive of ksock message from %s, ip %pISp, with error, protocol: %d.x.\n",
1540 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1541 &conn->ksnc_peeraddr,
1542 conn->ksnc_proto->pro_version);
1544 case SOCKNAL_RX_SLOP:
1545 if (conn->ksnc_rx_started)
1546 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1547 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1548 &conn->ksnc_peeraddr);
1555 ksocknal_peer_decref(conn->ksnc_peer);
1557 LIBCFS_FREE(conn, sizeof(*conn));
1561 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1562 struct sockaddr *addr, int why)
1564 struct ksock_conn *conn;
1565 struct ksock_conn *cnxt;
1568 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1571 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1573 ksocknal_close_conn_locked(conn, why);
1581 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1583 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1586 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1588 count = ksocknal_close_peer_conns_locked(
1589 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1591 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1597 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1599 struct ksock_peer_ni *peer_ni;
1600 struct hlist_node *pnxt;
1605 struct sockaddr_in sa = {.sin_family = AF_INET};
1607 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1609 if (!LNET_NID_IS_ANY(&id->nid)) {
1610 lo = hash_min(nidhash(&id->nid),
1611 HASH_BITS(ksocknal_data.ksnd_peers));
1615 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1618 sa.sin_addr.s_addr = htonl(ipaddr);
1619 for (i = lo; i <= hi; i++) {
1620 hlist_for_each_entry_safe(peer_ni, pnxt,
1621 &ksocknal_data.ksnd_peers[i],
1624 if (!((LNET_NID_IS_ANY(&id->nid) ||
1625 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1626 (id->pid == LNET_PID_ANY ||
1627 id->pid == peer_ni->ksnp_id.pid)))
1630 count += ksocknal_close_peer_conns_locked(
1632 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1636 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1638 /* wildcards always succeed */
1639 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1643 return (count == 0 ? -ENOENT : 0);
1647 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1649 /* The router is telling me she's been notified of a change in
1652 struct lnet_processid id = {
1653 .pid = LNET_PID_ANY,
1657 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1659 /* If the gateway crashed, close all open connections... */
1660 ksocknal_close_matching_conns(&id, 0);
1663 /* We can only establish new connections
1664 * if we have autroutes, and these connect on demand.
1669 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1673 struct ksock_conn *conn;
1675 for (index = 0; ; index++) {
1676 read_lock(&ksocknal_data.ksnd_global_lock);
1681 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1683 ksocknal_conn_addref(conn);
1688 read_unlock(&ksocknal_data.ksnd_global_lock);
1693 ksocknal_lib_push_conn (conn);
1694 ksocknal_conn_decref(conn);
1699 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1706 if (!LNET_NID_IS_ANY(&id->nid)) {
1707 lo = hash_min(nidhash(&id->nid),
1708 HASH_BITS(ksocknal_data.ksnd_peers));
1712 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1715 for (bkt = lo; bkt <= hi; bkt++) {
1716 int peer_off; /* searching offset in peer_ni hash table */
1718 for (peer_off = 0; ; peer_off++) {
1719 struct ksock_peer_ni *peer_ni;
1722 read_lock(&ksocknal_data.ksnd_global_lock);
1723 hlist_for_each_entry(peer_ni,
1724 &ksocknal_data.ksnd_peers[bkt],
1726 if (!((LNET_NID_IS_ANY(&id->nid) ||
1728 &peer_ni->ksnp_id.nid)) &&
1729 (id->pid == LNET_PID_ANY ||
1730 id->pid == peer_ni->ksnp_id.pid)))
1733 if (i++ == peer_off) {
1734 ksocknal_peer_addref(peer_ni);
1738 read_unlock(&ksocknal_data.ksnd_global_lock);
1740 if (i <= peer_off) /* no match */
1744 ksocknal_push_peer(peer_ni);
1745 ksocknal_peer_decref(peer_ni);
1752 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1754 struct lnet_process_id id4 = {};
1755 struct lnet_processid id = {};
1756 struct libcfs_ioctl_data *data = arg;
1760 case IOC_LIBCFS_GET_INTERFACE: {
1761 struct ksock_net *net = ni->ni_data;
1762 struct ksock_interface *iface;
1763 struct sockaddr_in *sa;
1765 read_lock(&ksocknal_data.ksnd_global_lock);
1767 if (data->ioc_count >= 1) {
1771 iface = &net->ksnn_interface;
1773 sa = (void *)&iface->ksni_addr;
1774 if (sa->sin_family == AF_INET)
1775 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1777 data->ioc_u32[0] = 0xFFFFFFFF;
1778 data->ioc_u32[1] = iface->ksni_netmask;
1779 data->ioc_u32[2] = iface->ksni_npeers;
1780 data->ioc_u32[3] = iface->ksni_nroutes;
1783 read_unlock(&ksocknal_data.ksnd_global_lock);
1787 case IOC_LIBCFS_GET_PEER: {
1792 int share_count = 0;
1794 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1795 &id4, &myip, &ip, &port,
1796 &conn_count, &share_count);
1800 data->ioc_nid = id4.nid;
1801 data->ioc_count = share_count;
1802 data->ioc_u32[0] = ip;
1803 data->ioc_u32[1] = port;
1804 data->ioc_u32[2] = myip;
1805 data->ioc_u32[3] = conn_count;
1806 data->ioc_u32[4] = id4.pid;
1810 case IOC_LIBCFS_ADD_PEER: {
1811 struct sockaddr_in sa = {.sin_family = AF_INET};
1813 id.pid = LNET_PID_LUSTRE;
1814 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1815 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1816 sa.sin_port = htons(data->ioc_u32[1]);
1817 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1819 case IOC_LIBCFS_DEL_PEER:
1820 id4.nid = data->ioc_nid;
1821 id4.pid = LNET_PID_ANY;
1822 return ksocknal_del_peer(ni, id4,
1823 data->ioc_u32[0]); /* IP */
1825 case IOC_LIBCFS_GET_CONN: {
1829 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1830 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1831 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1836 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1838 data->ioc_count = txmem;
1839 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1840 data->ioc_flags = nagle;
1841 if (psa->sin_family == AF_INET)
1842 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1844 data->ioc_u32[0] = 0xFFFFFFFF;
1845 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1846 &conn->ksnc_peeraddr);
1847 if (mysa->sin_family == AF_INET)
1848 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1850 data->ioc_u32[2] = 0xFFFFFFFF;
1851 data->ioc_u32[3] = conn->ksnc_type;
1852 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1853 data->ioc_u32[5] = rxmem;
1854 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1855 ksocknal_conn_decref(conn);
1859 case IOC_LIBCFS_CLOSE_CONNECTION:
1860 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1861 id.pid = LNET_PID_ANY;
1862 return ksocknal_close_matching_conns(&id,
1865 case IOC_LIBCFS_REGISTER_MYNID:
1866 /* Ignore if this is a noop */
1867 if (nid_is_nid4(&ni->ni_nid) &&
1868 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1871 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1872 libcfs_nid2str(data->ioc_nid),
1873 libcfs_nidstr(&ni->ni_nid));
1876 case IOC_LIBCFS_PUSH_CONNECTION:
1877 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1878 id.pid = LNET_PID_ANY;
1879 return ksocknal_push(ni, &id);
1888 ksocknal_free_buffers (void)
1890 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1892 if (ksocknal_data.ksnd_schedulers != NULL)
1893 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1895 spin_lock(&ksocknal_data.ksnd_tx_lock);
1897 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1899 struct ksock_tx *tx;
1901 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1902 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1904 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1905 tx_list)) != NULL) {
1906 list_del(&tx->tx_list);
1907 LIBCFS_FREE(tx, tx->tx_desc_size);
1910 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1914 static int ksocknal_get_link_status(struct net_device *dev)
1920 if (!netif_running(dev)) {
1922 CDEBUG(D_NET, "device not running\n");
1924 /* Some devices may not be providing link settings */
1925 else if (dev->ethtool_ops->get_link) {
1926 ret = dev->ethtool_ops->get_link(dev);
1927 CDEBUG(D_NET, "get_link returns %u\n", ret);
1934 ksocknal_handle_link_state_change(struct net_device *dev,
1935 unsigned char operstate)
1937 struct lnet_ni *ni = NULL;
1938 struct ksock_net *net;
1939 struct ksock_net *cnxt;
1941 unsigned char link_down = !(operstate == IF_OPER_UP);
1942 struct in_device *in_dev;
1943 bool found_ip = false;
1944 struct ksock_interface *ksi = NULL;
1945 struct sockaddr_in *sa;
1946 DECLARE_CONST_IN_IFADDR(ifa);
1948 ifindex = dev->ifindex;
1950 if (!ksocknal_data.ksnd_nnets)
1953 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1956 ksi = &net->ksnn_interface;
1957 sa = (void *)&ksi->ksni_addr;
1960 if (ksi->ksni_index != ifindex ||
1961 strcmp(ksi->ksni_name, dev->name))
1966 in_dev = __in_dev_get_rtnl(dev);
1968 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1970 CDEBUG(D_NET, "set link fatal state to 1\n");
1971 atomic_set(&ni->ni_fatal_error_on, 1);
1974 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1975 if (sa->sin_addr.s_addr == ifa->ifa_local)
1981 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1983 CDEBUG(D_NET, "set link fatal state to 1\n");
1984 atomic_set(&ni->ni_fatal_error_on, 1);
1989 CDEBUG(D_NET, "set link fatal state to 1\n");
1990 atomic_set(&ni->ni_fatal_error_on, link_down);
1992 CDEBUG(D_NET, "set link fatal state to %u\n",
1993 (ksocknal_get_link_status(dev) == 0));
1994 atomic_set(&ni->ni_fatal_error_on,
1995 (ksocknal_get_link_status(dev) == 0));
2004 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2007 struct ksock_net *net;
2008 struct ksock_net *cnxt;
2009 struct net_device *event_netdev = ifa->ifa_dev->dev;
2011 struct ksock_interface *ksi = NULL;
2012 struct sockaddr_in *sa;
2014 if (!ksocknal_data.ksnd_nnets)
2017 ifindex = event_netdev->ifindex;
2019 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2022 ksi = &net->ksnn_interface;
2023 sa = (void *)&ksi->ksni_addr;
2025 if (ksi->ksni_index != ifindex ||
2026 strcmp(ksi->ksni_name, event_netdev->name))
2029 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2030 CDEBUG(D_NET, "set link fatal state to %u\n",
2031 (event == NETDEV_DOWN));
2033 atomic_set(&ni->ni_fatal_error_on,
2034 (event == NETDEV_DOWN));
2041 /************************************
2042 * Net device notifier event handler
2043 ************************************/
2044 static int ksocknal_device_event(struct notifier_block *unused,
2045 unsigned long event, void *ptr)
2047 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2048 unsigned char operstate;
2050 operstate = dev->operstate;
2052 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2053 event, dev->name, dev->ifindex, operstate);
2059 ksocknal_handle_link_state_change(dev, operstate);
2066 /************************************
2067 * Inetaddr notifier event handler
2068 ************************************/
2069 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2070 unsigned long event, void *ptr)
2072 struct in_ifaddr *ifa = ptr;
2074 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2075 event, &ifa->ifa_address, &ifa->ifa_mask);
2081 ksocknal_handle_inetaddr_change(ifa, event);
2088 static struct notifier_block ksocknal_dev_notifier_block = {
2089 .notifier_call = ksocknal_device_event,
2092 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2093 .notifier_call = ksocknal_inetaddr_event,
2097 ksocknal_base_shutdown(void)
2099 struct ksock_sched *sched;
2100 struct ksock_peer_ni *peer_ni;
2103 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2104 libcfs_kmem_read());
2105 LASSERT (ksocknal_data.ksnd_nnets == 0);
2107 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2108 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2109 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2112 switch (ksocknal_data.ksnd_init) {
2117 case SOCKNAL_INIT_ALL:
2118 case SOCKNAL_INIT_DATA:
2119 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2122 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2123 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2124 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2125 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2126 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2128 if (ksocknal_data.ksnd_schedulers != NULL) {
2129 cfs_percpt_for_each(sched, i,
2130 ksocknal_data.ksnd_schedulers) {
2132 LASSERT(list_empty(&sched->kss_tx_conns));
2133 LASSERT(list_empty(&sched->kss_rx_conns));
2134 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2135 LASSERT(sched->kss_nconns == 0);
2139 /* flag threads to terminate; wake and wait for them to die */
2140 ksocknal_data.ksnd_shuttingdown = 1;
2141 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2142 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2144 if (ksocknal_data.ksnd_schedulers != NULL) {
2145 cfs_percpt_for_each(sched, i,
2146 ksocknal_data.ksnd_schedulers)
2147 wake_up_all(&sched->kss_waitq);
2150 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2151 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2152 "waiting for %d threads to terminate\n",
2153 atomic_read(&ksocknal_data.ksnd_nthreads));
2155 ksocknal_free_buffers();
2157 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2161 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2162 libcfs_kmem_read());
2164 module_put(THIS_MODULE);
2168 ksocknal_base_startup(void)
2170 struct ksock_sched *sched;
2174 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2175 LASSERT(ksocknal_data.ksnd_nnets == 0);
2177 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2179 hash_init(ksocknal_data.ksnd_peers);
2181 rwlock_init(&ksocknal_data.ksnd_global_lock);
2182 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2184 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2185 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2186 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2187 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2188 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2190 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2191 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2192 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2193 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2195 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2196 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2198 /* NB memset above zeros whole of ksocknal_data */
2200 /* flag lists/ptrs/locks initialised */
2201 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2202 if (!try_module_get(THIS_MODULE))
2205 /* Create a scheduler block per available CPT */
2206 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2208 if (ksocknal_data.ksnd_schedulers == NULL)
2211 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2215 * make sure not to allocate more threads than there are
2216 * cores/CPUs in teh CPT
2218 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2219 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2220 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2223 * max to half of CPUs, assume another half should be
2224 * reserved for upper layer modules
2226 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2229 sched->kss_nthreads_max = nthrs;
2232 spin_lock_init(&sched->kss_lock);
2233 INIT_LIST_HEAD(&sched->kss_rx_conns);
2234 INIT_LIST_HEAD(&sched->kss_tx_conns);
2235 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2236 init_waitqueue_head(&sched->kss_waitq);
2239 ksocknal_data.ksnd_connd_starting = 0;
2240 ksocknal_data.ksnd_connd_failed_stamp = 0;
2241 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2242 /* must have at least 2 connds to remain responsive to accepts while
2244 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2245 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2247 if (*ksocknal_tunables.ksnd_nconnds_max <
2248 *ksocknal_tunables.ksnd_nconnds) {
2249 ksocknal_tunables.ksnd_nconnds_max =
2250 ksocknal_tunables.ksnd_nconnds;
2253 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2254 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2255 ksocknal_data.ksnd_connd_starting++;
2256 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2258 rc = ksocknal_thread_start(ksocknal_connd,
2259 (void *)((uintptr_t)i),
2260 "socknal_cd%02d", i);
2262 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2263 ksocknal_data.ksnd_connd_starting--;
2264 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2265 CERROR("Can't spawn socknal connd: %d\n", rc);
2270 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2272 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2276 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2277 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2279 /* flag everything initialised */
2280 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2285 ksocknal_base_shutdown();
2290 ksocknal_debug_peerhash(struct lnet_ni *ni)
2292 struct ksock_peer_ni *peer_ni;
2295 read_lock(&ksocknal_data.ksnd_global_lock);
2297 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2298 struct ksock_conn_cb *conn_cb;
2299 struct ksock_conn *conn;
2301 if (peer_ni->ksnp_ni != ni)
2304 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2305 libcfs_idstr(&peer_ni->ksnp_id),
2306 refcount_read(&peer_ni->ksnp_refcount),
2307 peer_ni->ksnp_closing,
2308 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2309 peer_ni->ksnp_zc_next_cookie,
2310 !list_empty(&peer_ni->ksnp_tx_queue),
2311 !list_empty(&peer_ni->ksnp_zc_req_list));
2313 conn_cb = peer_ni->ksnp_conn_cb;
2315 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2316 refcount_read(&conn_cb->ksnr_refcount),
2317 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2318 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2321 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2322 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2323 refcount_read(&conn->ksnc_conn_refcount),
2324 refcount_read(&conn->ksnc_sock_refcount),
2325 conn->ksnc_type, conn->ksnc_closing);
2330 read_unlock(&ksocknal_data.ksnd_global_lock);
2335 ksocknal_shutdown(struct lnet_ni *ni)
2337 struct ksock_net *net = ni->ni_data;
2338 struct lnet_process_id anyid = {
2339 .nid = LNET_NID_ANY,
2340 .pid = LNET_PID_ANY,
2343 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2344 LASSERT(ksocknal_data.ksnd_nnets > 0);
2346 /* prevent new peers */
2347 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2349 /* Delete all peers */
2350 ksocknal_del_peer(ni, anyid, 0);
2352 /* Wait for all peer_ni state to clean up */
2353 wait_var_event_warning(&net->ksnn_npeers,
2354 atomic_read(&net->ksnn_npeers) ==
2355 SOCKNAL_SHUTDOWN_BIAS,
2356 "waiting for %d peers to disconnect\n",
2357 ksocknal_debug_peerhash(ni) +
2358 atomic_read(&net->ksnn_npeers) -
2359 SOCKNAL_SHUTDOWN_BIAS);
2361 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2362 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2364 list_del(&net->ksnn_list);
2365 LIBCFS_FREE(net, sizeof(*net));
2367 ksocknal_data.ksnd_nnets--;
2368 if (ksocknal_data.ksnd_nnets == 0)
2369 ksocknal_base_shutdown();
2373 ksocknal_search_new_ipif(struct ksock_net *net)
2376 char *ifnam = &net->ksnn_interface.ksni_name[0];
2377 char *colon = strchr(ifnam, ':');
2379 struct ksock_net *tmp;
2384 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2385 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2386 char *colon2 = strchr(ifnam2, ':');
2391 found = strcmp(ifnam, ifnam2) == 0;
2404 ksocknal_start_schedulers(struct ksock_sched *sched)
2410 if (sched->kss_nthreads == 0) {
2411 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2412 nthrs = sched->kss_nthreads_max;
2414 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2416 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2417 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2419 nthrs = min(nthrs, sched->kss_nthreads_max);
2421 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2422 /* increase two threads if there is new interface */
2423 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2426 for (i = 0; i < nthrs; i++) {
2429 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2430 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2431 "socknal_sd%02d_%02d",
2433 (int)KSOCK_THREAD_SID(id));
2437 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2438 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2442 sched->kss_nthreads += i;
2447 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2449 int newif = ksocknal_search_new_ipif(net);
2453 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2456 for (i = 0; i < ncpts; i++) {
2457 struct ksock_sched *sched;
2458 int cpt = (cpts == NULL) ? i : cpts[i];
2460 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2461 sched = ksocknal_data.ksnd_schedulers[cpt];
2463 if (!newif && sched->kss_nthreads > 0)
2466 rc = ksocknal_start_schedulers(sched);
2474 ksocknal_startup(struct lnet_ni *ni)
2476 struct ksock_net *net;
2477 struct ksock_interface *ksi = NULL;
2478 struct lnet_inetdev *ifaces = NULL;
2479 struct sockaddr_in *sa;
2483 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2484 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2485 rc = ksocknal_base_startup();
2489 LIBCFS_ALLOC(net, sizeof(*net));
2492 net->ksnn_incarnation = ktime_get_real_ns();
2495 ksocknal_tunables_setup(ni);
2497 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2501 ksi = &net->ksnn_interface;
2503 /* Use the first discovered interface or look in the list */
2504 if (ni->ni_interface) {
2505 for (i = 0; i < rc; i++)
2506 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2509 /* ni_interfaces doesn't contain the interface we want */
2511 CERROR("ksocklnd: failed to find interface %s\n",
2517 ni->ni_dev_cpt = ifaces[i].li_cpt;
2518 sa = (void *)&ksi->ksni_addr;
2519 memset(sa, 0, sizeof(*sa));
2520 sa->sin_family = AF_INET;
2521 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2522 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2523 ksi->ksni_netmask = ifaces[i].li_netmask;
2524 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2526 /* call it before add it to ksocknal_data.ksnd_nets */
2527 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2532 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2533 ni->ni_nid.nid_addr[0] =
2534 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2535 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2537 ksocknal_data.ksnd_nnets++;
2542 LIBCFS_FREE(net, sizeof(*net));
2544 if (ksocknal_data.ksnd_nnets == 0)
2545 ksocknal_base_shutdown();
2550 static void __exit ksocklnd_exit(void)
2552 lnet_unregister_lnd(&the_ksocklnd);
2555 static const struct lnet_lnd the_ksocklnd = {
2556 .lnd_type = SOCKLND,
2557 .lnd_startup = ksocknal_startup,
2558 .lnd_shutdown = ksocknal_shutdown,
2559 .lnd_ctl = ksocknal_ctl,
2560 .lnd_send = ksocknal_send,
2561 .lnd_recv = ksocknal_recv,
2562 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2563 .lnd_accept = ksocknal_accept,
2566 static int __init ksocklnd_init(void)
2570 /* check ksnr_connected/connecting field large enough */
2571 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2572 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2574 rc = ksocknal_tunables_init();
2578 lnet_register_lnd(&the_ksocklnd);
2583 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2584 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2585 MODULE_VERSION("2.8.0");
2586 MODULE_LICENSE("GPL");
2588 module_init(ksocklnd_init);
2589 module_exit(ksocklnd_exit);