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>
43 #include <net/addrconf.h>
45 static const struct lnet_lnd the_ksocklnd;
46 struct ksock_nal_data ksocknal_data;
48 static struct ksock_interface *
49 ksocknal_index2iface(struct lnet_ni *ni, int index)
51 struct ksock_net *net = ni->ni_data;
52 struct ksock_interface *iface;
54 iface = &net->ksnn_interface;
56 if (iface->ksni_index == index)
62 static int ksocknal_ip2index(struct sockaddr *addr, struct lnet_ni *ni)
64 struct net_device *dev;
66 DECLARE_CONST_IN_IFADDR(ifa);
68 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
72 for_each_netdev(ni->ni_net_ns, dev) {
73 int flags = dev_get_flags(dev);
74 struct in_device *in_dev;
76 if (flags & IFF_LOOPBACK) /* skip the loopback IF */
79 if (!(flags & IFF_UP))
82 switch (addr->sa_family) {
84 in_dev = __in_dev_get_rcu(dev);
88 in_dev_for_each_ifa_rcu(ifa, in_dev) {
90 ((struct sockaddr_in *)addr)->sin_addr.s_addr)
95 #if IS_ENABLED(CONFIG_IPV6)
97 struct inet6_dev *in6_dev;
98 const struct inet6_ifaddr *ifa6;
99 struct sockaddr_in6 *addr6 = (struct sockaddr_in6*)addr;
101 in6_dev = __in6_dev_get(dev);
105 list_for_each_entry_rcu(ifa6, &in6_dev->addr_list, if_list) {
106 if (ipv6_addr_cmp(&ifa6->addr,
107 &addr6->sin6_addr) == 0)
112 #endif /* IS_ENABLED(CONFIG_IPV6) */
122 static struct ksock_conn_cb *
123 ksocknal_create_conn_cb(struct sockaddr *addr)
125 struct ksock_conn_cb *conn_cb;
127 LIBCFS_ALLOC(conn_cb, sizeof(*conn_cb));
131 refcount_set(&conn_cb->ksnr_refcount, 1);
132 conn_cb->ksnr_peer = NULL;
133 conn_cb->ksnr_retry_interval = 0; /* OK to connect at any time */
134 rpc_copy_addr((struct sockaddr *)&conn_cb->ksnr_addr, addr);
135 rpc_set_port((struct sockaddr *)&conn_cb->ksnr_addr,
137 conn_cb->ksnr_myiface = -1;
138 conn_cb->ksnr_scheduled = 0;
139 conn_cb->ksnr_connecting = 0;
140 conn_cb->ksnr_connected = 0;
141 conn_cb->ksnr_deleted = 0;
142 conn_cb->ksnr_conn_count = 0;
143 conn_cb->ksnr_ctrl_conn_count = 0;
144 conn_cb->ksnr_blki_conn_count = 0;
145 conn_cb->ksnr_blko_conn_count = 0;
146 conn_cb->ksnr_max_conns = 0;
152 ksocknal_destroy_conn_cb(struct ksock_conn_cb *conn_cb)
154 LASSERT(refcount_read(&conn_cb->ksnr_refcount) == 0);
156 if (conn_cb->ksnr_peer)
157 ksocknal_peer_decref(conn_cb->ksnr_peer);
159 LIBCFS_FREE(conn_cb, sizeof(*conn_cb));
162 static struct ksock_peer_ni *
163 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_processid *id)
165 int cpt = lnet_nid2cpt(&id->nid, ni);
166 struct ksock_net *net = ni->ni_data;
167 struct ksock_peer_ni *peer_ni;
169 LASSERT(!LNET_NID_IS_ANY(&id->nid));
170 LASSERT(id->pid != LNET_PID_ANY);
171 LASSERT(!in_interrupt());
173 if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
174 CERROR("Can't create peer_ni: network shutdown\n");
175 return ERR_PTR(-ESHUTDOWN);
178 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
180 atomic_dec(&net->ksnn_npeers);
181 return ERR_PTR(-ENOMEM);
184 peer_ni->ksnp_ni = ni;
185 peer_ni->ksnp_id = *id;
186 refcount_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
187 peer_ni->ksnp_closing = 0;
188 peer_ni->ksnp_accepting = 0;
189 peer_ni->ksnp_proto = NULL;
190 peer_ni->ksnp_last_alive = 0;
191 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
192 peer_ni->ksnp_conn_cb = NULL;
194 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
195 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
196 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
197 spin_lock_init(&peer_ni->ksnp_lock);
203 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
205 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
207 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
208 libcfs_idstr(&peer_ni->ksnp_id), peer_ni);
210 LASSERT(refcount_read(&peer_ni->ksnp_refcount) == 0);
211 LASSERT(peer_ni->ksnp_accepting == 0);
212 LASSERT(list_empty(&peer_ni->ksnp_conns));
213 LASSERT(peer_ni->ksnp_conn_cb == NULL);
214 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
215 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
217 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
219 /* NB a peer_ni's connections and conn_cb keep a reference on their
220 * peer_ni until they are destroyed, so we can be assured that _all_
221 * state to do with this peer_ni has been cleaned up when its refcount
224 if (atomic_dec_and_test(&net->ksnn_npeers))
225 wake_up_var(&net->ksnn_npeers);
228 struct ksock_peer_ni *
229 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_processid *id)
231 struct ksock_peer_ni *peer_ni;
232 unsigned long hash = nidhash(&id->nid);
234 hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
236 LASSERT(!peer_ni->ksnp_closing);
238 if (peer_ni->ksnp_ni != ni)
241 if (!nid_same(&peer_ni->ksnp_id.nid, &id->nid) ||
242 peer_ni->ksnp_id.pid != id->pid)
245 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
246 peer_ni, libcfs_idstr(id),
247 refcount_read(&peer_ni->ksnp_refcount));
253 struct ksock_peer_ni *
254 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_processid *id)
256 struct ksock_peer_ni *peer_ni;
258 read_lock(&ksocknal_data.ksnd_global_lock);
259 peer_ni = ksocknal_find_peer_locked(ni, id);
260 if (peer_ni != NULL) /* +1 ref for caller? */
261 ksocknal_peer_addref(peer_ni);
262 read_unlock(&ksocknal_data.ksnd_global_lock);
268 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
270 LASSERT(list_empty(&peer_ni->ksnp_conns));
271 LASSERT(peer_ni->ksnp_conn_cb == NULL);
272 LASSERT(!peer_ni->ksnp_closing);
273 peer_ni->ksnp_closing = 1;
274 hlist_del(&peer_ni->ksnp_list);
275 /* lose peerlist's ref */
276 ksocknal_peer_decref(peer_ni);
280 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
281 struct lnet_processid *id, __u32 *myip, __u32 *peer_ip,
282 int *port, int *conn_count, int *share_count)
284 struct ksock_peer_ni *peer_ni;
285 struct ksock_conn_cb *conn_cb;
289 read_lock(&ksocknal_data.ksnd_global_lock);
291 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
293 if (peer_ni->ksnp_ni != ni)
298 conn_cb = peer_ni->ksnp_conn_cb;
299 if (conn_cb == NULL) {
300 *id = peer_ni->ksnp_id;
308 *id = peer_ni->ksnp_id;
309 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
310 struct sockaddr_in *sa =
311 (void *)&conn_cb->ksnr_addr;
313 rc = choose_ipv4_src(myip,
314 conn_cb->ksnr_myiface,
315 ntohl(sa->sin_addr.s_addr),
317 *peer_ip = ntohl(sa->sin_addr.s_addr);
318 *port = ntohs(sa->sin_port);
321 *peer_ip = 0xFFFFFFFF;
325 *conn_count = conn_cb->ksnr_conn_count;
330 read_unlock(&ksocknal_data.ksnd_global_lock);
335 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
338 unsigned int count = 0;
341 case SOCKLND_CONN_CONTROL:
342 count = conn_cb->ksnr_ctrl_conn_count;
344 case SOCKLND_CONN_BULK_IN:
345 count = conn_cb->ksnr_blki_conn_count;
347 case SOCKLND_CONN_BULK_OUT:
348 count = conn_cb->ksnr_blko_conn_count;
350 case SOCKLND_CONN_ANY:
351 count = conn_cb->ksnr_conn_count;
362 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
364 struct lnet_ni *ni = peer_ni->ksnp_ni;
365 struct lnet_ioctl_config_socklnd_tunables *tunables;
369 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
371 return tunables->lnd_conns_per_peer;
375 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
378 conn_cb->ksnr_conn_count++;
380 /* check if all connections of the given type got created */
382 case SOCKLND_CONN_CONTROL:
383 conn_cb->ksnr_ctrl_conn_count++;
384 /* there's a single control connection per peer,
385 * two in case of loopback
387 conn_cb->ksnr_connected |= BIT(type);
389 case SOCKLND_CONN_BULK_IN:
390 conn_cb->ksnr_blki_conn_count++;
391 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
392 conn_cb->ksnr_connected |= BIT(type);
394 case SOCKLND_CONN_BULK_OUT:
395 conn_cb->ksnr_blko_conn_count++;
396 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
397 conn_cb->ksnr_connected |= BIT(type);
399 case SOCKLND_CONN_ANY:
400 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
401 conn_cb->ksnr_connected |= BIT(type);
408 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
409 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
414 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
417 conn_cb->ksnr_conn_count--;
419 /* check if all connections of the given type got created */
421 case SOCKLND_CONN_CONTROL:
422 conn_cb->ksnr_ctrl_conn_count--;
423 /* there's a single control connection per peer,
424 * two in case of loopback
426 if (conn_cb->ksnr_ctrl_conn_count == 0)
427 conn_cb->ksnr_connected &= ~BIT(type);
429 case SOCKLND_CONN_BULK_IN:
430 conn_cb->ksnr_blki_conn_count--;
431 if (conn_cb->ksnr_blki_conn_count < conn_cb->ksnr_max_conns)
432 conn_cb->ksnr_connected &= ~BIT(type);
434 case SOCKLND_CONN_BULK_OUT:
435 conn_cb->ksnr_blko_conn_count--;
436 if (conn_cb->ksnr_blko_conn_count < conn_cb->ksnr_max_conns)
437 conn_cb->ksnr_connected &= ~BIT(type);
439 case SOCKLND_CONN_ANY:
440 if (conn_cb->ksnr_conn_count < conn_cb->ksnr_max_conns)
441 conn_cb->ksnr_connected &= ~BIT(type);
448 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
449 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
453 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
454 struct ksock_conn *conn)
456 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
457 int type = conn->ksnc_type;
458 struct ksock_interface *iface;
461 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
463 conn->ksnc_conn_cb = conn_cb;
464 ksocknal_conn_cb_addref(conn_cb);
466 if (conn_cb->ksnr_myiface != conn_iface) {
467 if (conn_cb->ksnr_myiface < 0) {
468 /* route wasn't bound locally yet (the initial route) */
469 CDEBUG(D_NET, "Binding %s %pISc to interface %d\n",
470 libcfs_idstr(&peer_ni->ksnp_id),
475 "Rebinding %s %pISc from interface %d to %d\n",
476 libcfs_idstr(&peer_ni->ksnp_id),
478 conn_cb->ksnr_myiface,
481 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
482 conn_cb->ksnr_myiface);
484 iface->ksni_nroutes--;
486 conn_cb->ksnr_myiface = conn_iface;
487 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
488 conn_cb->ksnr_myiface);
490 iface->ksni_nroutes++;
493 ksocknal_incr_conn_count(conn_cb, type);
495 /* Successful connection => further attempts can
496 * proceed immediately
498 conn_cb->ksnr_retry_interval = 0;
502 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
503 struct ksock_conn_cb *conn_cb)
505 struct ksock_conn *conn;
506 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
508 LASSERT(!peer_ni->ksnp_closing);
509 LASSERT(!conn_cb->ksnr_peer);
510 LASSERT(!conn_cb->ksnr_scheduled);
511 LASSERT(!conn_cb->ksnr_connecting);
512 LASSERT(conn_cb->ksnr_connected == 0);
514 conn_cb->ksnr_peer = peer_ni;
515 ksocknal_peer_addref(peer_ni);
517 /* set the conn_cb's interface to the current net's interface */
518 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
519 net->ksnn_interface.ksni_nroutes++;
521 /* peer_ni's route list takes over my ref on 'route' */
522 peer_ni->ksnp_conn_cb = conn_cb;
524 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
525 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
526 (struct sockaddr *)&conn_cb->ksnr_addr))
529 ksocknal_associate_cb_conn_locked(conn_cb, conn);
530 /* keep going (typed conns) */
535 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
537 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
538 struct ksock_interface *iface;
539 struct ksock_conn *conn;
540 struct ksock_conn *cnxt;
542 LASSERT(!conn_cb->ksnr_deleted);
544 /* Close associated conns */
545 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
546 if (conn->ksnc_conn_cb != conn_cb)
549 ksocknal_close_conn_locked(conn, 0);
552 if (conn_cb->ksnr_myiface >= 0) {
553 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
554 conn_cb->ksnr_myiface);
556 iface->ksni_nroutes--;
559 conn_cb->ksnr_deleted = 1;
560 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
561 peer_ni->ksnp_conn_cb = NULL;
563 if (list_empty(&peer_ni->ksnp_conns)) {
564 /* I've just removed the last route to a peer_ni with no active
567 ksocknal_unlink_peer_locked(peer_ni);
572 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_processid *id,
573 struct sockaddr *addr)
575 struct ksock_peer_ni *peer_ni;
576 struct ksock_peer_ni *peer2;
577 struct ksock_conn_cb *conn_cb;
579 if (LNET_NID_IS_ANY(&id->nid) ||
580 id->pid == LNET_PID_ANY)
583 /* Have a brand new peer_ni ready... */
584 peer_ni = ksocknal_create_peer(ni, id);
586 return PTR_ERR(peer_ni);
588 conn_cb = ksocknal_create_conn_cb(addr);
590 ksocknal_peer_decref(peer_ni);
594 write_lock_bh(&ksocknal_data.ksnd_global_lock);
596 /* always called with a ref on ni, so shutdown can't have started */
597 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
600 peer2 = ksocknal_find_peer_locked(ni, id);
602 ksocknal_peer_decref(peer_ni);
605 /* peer_ni table takes my ref on peer_ni */
606 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
610 if (peer_ni->ksnp_conn_cb) {
611 ksocknal_conn_cb_decref(conn_cb);
613 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
614 /* Remember conns_per_peer setting at the time
615 * of connection initiation. It will define the
616 * max number of conns per type for this conn_cb
619 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
622 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
628 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni)
630 struct ksock_conn *conn;
631 struct ksock_conn *cnxt;
632 struct ksock_conn_cb *conn_cb;
634 LASSERT(!peer_ni->ksnp_closing);
636 /* Extra ref prevents peer_ni disappearing until I'm done with it */
637 ksocknal_peer_addref(peer_ni);
638 conn_cb = peer_ni->ksnp_conn_cb;
640 ksocknal_del_conn_cb_locked(conn_cb);
642 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
644 ksocknal_close_conn_locked(conn, 0);
646 ksocknal_peer_decref(peer_ni);
647 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
651 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_processid *id)
654 struct hlist_node *pnxt;
655 struct ksock_peer_ni *peer_ni;
661 write_lock_bh(&ksocknal_data.ksnd_global_lock);
663 if (id && !LNET_NID_IS_ANY(&id->nid)) {
664 lo = hash_min(nidhash(&id->nid),
665 HASH_BITS(ksocknal_data.ksnd_peers));
669 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
672 for (i = lo; i <= hi; i++) {
673 hlist_for_each_entry_safe(peer_ni, pnxt,
674 &ksocknal_data.ksnd_peers[i],
676 if (peer_ni->ksnp_ni != ni)
679 if (!((!id || LNET_NID_IS_ANY(&id->nid) ||
680 nid_same(&peer_ni->ksnp_id.nid, &id->nid)) &&
681 (!id || id->pid == LNET_PID_ANY ||
682 peer_ni->ksnp_id.pid == id->pid)))
685 ksocknal_peer_addref(peer_ni); /* a ref for me... */
687 ksocknal_del_peer_locked(peer_ni);
689 if (peer_ni->ksnp_closing &&
690 !list_empty(&peer_ni->ksnp_tx_queue)) {
691 LASSERT(list_empty(&peer_ni->ksnp_conns));
692 LASSERT(peer_ni->ksnp_conn_cb == NULL);
694 list_splice_init(&peer_ni->ksnp_tx_queue,
698 ksocknal_peer_decref(peer_ni); /* ...till here */
700 rc = 0; /* matched! */
704 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
706 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
711 static struct ksock_conn *
712 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
714 struct ksock_peer_ni *peer_ni;
715 struct ksock_conn *conn;
718 read_lock(&ksocknal_data.ksnd_global_lock);
720 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
721 LASSERT(!peer_ni->ksnp_closing);
723 if (peer_ni->ksnp_ni != ni)
726 list_for_each_entry(conn, &peer_ni->ksnp_conns,
731 ksocknal_conn_addref(conn);
732 read_unlock(&ksocknal_data.ksnd_global_lock);
737 read_unlock(&ksocknal_data.ksnd_global_lock);
741 static struct ksock_sched *
742 ksocknal_choose_scheduler_locked(unsigned int cpt)
744 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
747 if (sched->kss_nthreads == 0) {
748 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
749 if (sched->kss_nthreads > 0) {
750 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
751 cpt, sched->kss_cpt);
762 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
764 struct ksock_connreq *cr;
766 struct sockaddr_storage peer;
768 rc = lnet_sock_getaddr(sock, true, &peer);
770 CERROR("Can't determine new connection's address\n");
774 LIBCFS_ALLOC(cr, sizeof(*cr));
776 LCONSOLE_ERROR_MSG(0x12f,
777 "Dropping connection request from %pISc: memory exhausted\n",
784 cr->ksncr_sock = sock;
786 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
788 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
789 wake_up(&ksocknal_data.ksnd_connd_waitq);
791 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
796 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
799 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
800 return conn_cb->ksnr_connecting;
805 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
806 struct socket *sock, int type)
808 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
810 struct lnet_processid peerid;
812 struct ksock_conn *conn;
813 struct ksock_conn *conn2;
814 struct ksock_peer_ni *peer_ni = NULL;
815 struct ksock_peer_ni *peer2;
816 struct ksock_sched *sched;
817 struct ksock_hello_msg *hello;
820 struct ksock_tx *txtmp;
827 active = (conn_cb != NULL);
829 LASSERT(active == (type != SOCKLND_CONN_NONE));
831 LIBCFS_ALLOC(conn, sizeof(*conn));
837 conn->ksnc_peer = NULL;
838 conn->ksnc_conn_cb = NULL;
839 conn->ksnc_sock = sock;
840 /* 2 ref, 1 for conn, another extra ref prevents socket
841 * being closed before establishment of connection */
842 refcount_set(&conn->ksnc_sock_refcount, 2);
843 conn->ksnc_type = type;
844 ksocknal_lib_save_callback(sock, conn);
845 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
847 conn->ksnc_rx_ready = 0;
848 conn->ksnc_rx_scheduled = 0;
850 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
851 conn->ksnc_tx_ready = 0;
852 conn->ksnc_tx_scheduled = 0;
853 conn->ksnc_tx_carrier = NULL;
854 atomic_set (&conn->ksnc_tx_nob, 0);
856 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
857 kshm_ips[LNET_INTERFACES_NUM]));
863 /* stash conn's local and remote addrs */
864 rc = ksocknal_lib_get_conn_addrs(conn);
868 /* Find out/confirm peer_ni's NID and connection type and get the
869 * vector of interfaces she's willing to let me connect to.
870 * Passive connections use the listener timeout since the peer_ni sends
875 peer_ni = conn_cb->ksnr_peer;
876 LASSERT(ni == peer_ni->ksnp_ni);
878 /* Active connection sends HELLO eagerly */
879 hello->kshm_nips = 0;
880 peerid = peer_ni->ksnp_id;
882 write_lock_bh(global_lock);
883 conn->ksnc_proto = peer_ni->ksnp_proto;
884 write_unlock_bh(global_lock);
886 if (conn->ksnc_proto == NULL) {
887 conn->ksnc_proto = &ksocknal_protocol_v3x;
888 #if SOCKNAL_VERSION_DEBUG
889 if (*ksocknal_tunables.ksnd_protocol == 2)
890 conn->ksnc_proto = &ksocknal_protocol_v2x;
891 else if (*ksocknal_tunables.ksnd_protocol == 1)
892 conn->ksnc_proto = &ksocknal_protocol_v1x;
896 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
900 peerid.nid = LNET_ANY_NID;
901 peerid.pid = LNET_PID_ANY;
903 /* Passive, get protocol from peer_ni */
904 conn->ksnc_proto = NULL;
907 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
911 LASSERT(rc == 0 || active);
912 LASSERT(conn->ksnc_proto != NULL);
913 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
915 cpt = lnet_nid2cpt(&peerid.nid, ni);
918 ksocknal_peer_addref(peer_ni);
919 write_lock_bh(global_lock);
921 peer_ni = ksocknal_create_peer(ni, &peerid);
922 if (IS_ERR(peer_ni)) {
923 rc = PTR_ERR(peer_ni);
927 write_lock_bh(global_lock);
929 /* called with a ref on ni, so shutdown can't have started */
930 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
932 peer2 = ksocknal_find_peer_locked(ni, &peerid);
934 /* NB this puts an "empty" peer_ni in the peer_ni
935 * table (which takes my ref) */
936 hash_add(ksocknal_data.ksnd_peers,
937 &peer_ni->ksnp_list, nidhash(&peerid.nid));
939 ksocknal_peer_decref(peer_ni);
944 ksocknal_peer_addref(peer_ni);
945 peer_ni->ksnp_accepting++;
947 /* Am I already connecting to this guy? Resolve in
948 * favour of higher NID...
950 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
951 ksocknal_connecting(peer_ni->ksnp_conn_cb,
952 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
954 warn = "connection race resolution";
959 if (peer_ni->ksnp_closing ||
960 (active && conn_cb->ksnr_deleted)) {
961 /* peer_ni/conn_cb got closed under me */
963 warn = "peer_ni/conn_cb removed";
967 if (peer_ni->ksnp_proto == NULL) {
968 /* Never connected before.
969 * NB recv_hello may have returned EPROTO to signal my peer_ni
970 * wants a different protocol than the one I asked for.
972 LASSERT(list_empty(&peer_ni->ksnp_conns));
974 peer_ni->ksnp_proto = conn->ksnc_proto;
975 peer_ni->ksnp_incarnation = incarnation;
978 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
979 peer_ni->ksnp_incarnation != incarnation) {
980 /* peer_ni rebooted or I've got the wrong protocol version */
981 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
983 peer_ni->ksnp_proto = NULL;
985 warn = peer_ni->ksnp_incarnation != incarnation ?
987 "wrong proto version";
997 warn = "lost conn race";
1000 warn = "retry with different protocol version";
1004 /* Refuse to duplicate an existing connection, unless this is a
1005 * loopback connection */
1006 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1007 (struct sockaddr *)&conn->ksnc_myaddr)) {
1008 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1010 (struct sockaddr *)&conn2->ksnc_peeraddr,
1011 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1013 (struct sockaddr *)&conn2->ksnc_myaddr,
1014 (struct sockaddr *)&conn->ksnc_myaddr) ||
1015 conn2->ksnc_type != conn->ksnc_type)
1019 /* If max conns per type is not registered in conn_cb
1020 * as ksnr_max_conns, use ni's conns_per_peer
1022 if ((peer_ni->ksnp_conn_cb &&
1023 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1024 (!peer_ni->ksnp_conn_cb &&
1025 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1028 /* Reply on a passive connection attempt so the peer_ni
1029 * realises we're connected.
1039 /* If the connection created by this route didn't bind to the IP
1040 * address the route connected to, the connection/route matching
1041 * code below probably isn't going to work.
1044 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1045 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1046 CERROR("Route %s %pISc connected to %pISc\n",
1047 libcfs_idstr(&peer_ni->ksnp_id),
1048 &conn_cb->ksnr_addr,
1049 &conn->ksnc_peeraddr);
1052 /* Search for a conn_cb corresponding to the new connection and
1053 * create an association. This allows incoming connections created
1054 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1055 * continually create duplicate conn_cbs.
1057 conn_cb = peer_ni->ksnp_conn_cb;
1059 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1060 (struct sockaddr *)&conn_cb->ksnr_addr))
1061 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1063 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1064 peer_ni->ksnp_last_alive = ktime_get_seconds();
1065 peer_ni->ksnp_send_keepalive = 0;
1066 peer_ni->ksnp_error = 0;
1068 sched = ksocknal_choose_scheduler_locked(cpt);
1070 CERROR("no schedulers available. node is unhealthy\n");
1074 * The cpt might have changed if we ended up selecting a non cpt
1075 * native scheduler. So use the scheduler's cpt instead.
1077 cpt = sched->kss_cpt;
1078 sched->kss_nconns++;
1079 conn->ksnc_scheduler = sched;
1081 conn->ksnc_tx_last_post = ktime_get_seconds();
1082 /* Set the deadline for the outgoing HELLO to drain */
1083 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1084 conn->ksnc_tx_deadline = ktime_get_seconds() +
1086 smp_mb(); /* order with adding to peer_ni's conn list */
1088 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1089 ksocknal_conn_addref(conn);
1091 ksocknal_new_packet(conn, 0);
1093 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1095 /* Take packets blocking for this connection. */
1096 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1097 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1101 list_del(&tx->tx_list);
1102 ksocknal_queue_tx_locked(tx, conn);
1105 write_unlock_bh(global_lock);
1106 /* We've now got a new connection. Any errors from here on are just
1107 * like "normal" comms errors and we close the connection normally.
1108 * NB (a) we still have to send the reply HELLO for passive
1110 * (b) normal I/O on the conn is blocked until I setup and call the
1114 CDEBUG(D_NET, "New conn %s p %d.x %pISc -> %pIScp"
1115 " incarnation:%lld sched[%d]\n",
1116 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1117 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1121 hello->kshm_nips = 0;
1122 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1125 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1126 kshm_ips[LNET_INTERFACES_NUM]));
1128 /* setup the socket AFTER I've received hello (it disables
1129 * SO_LINGER). I might call back to the acceptor who may want
1130 * to send a protocol version response and then close the
1131 * socket; this ensures the socket only tears down after the
1132 * response has been sent.
1135 rc = ksocknal_lib_setup_sock(sock);
1137 write_lock_bh(global_lock);
1139 /* NB my callbacks block while I hold ksnd_global_lock */
1140 ksocknal_lib_set_callback(sock, conn);
1143 peer_ni->ksnp_accepting--;
1145 write_unlock_bh(global_lock);
1148 write_lock_bh(global_lock);
1149 if (!conn->ksnc_closing) {
1150 /* could be closed by another thread */
1151 ksocknal_close_conn_locked(conn, rc);
1153 write_unlock_bh(global_lock);
1154 } else if (ksocknal_connsock_addref(conn) == 0) {
1155 /* Allow I/O to proceed. */
1156 ksocknal_read_callback(conn);
1157 ksocknal_write_callback(conn);
1158 ksocknal_connsock_decref(conn);
1161 ksocknal_connsock_decref(conn);
1162 ksocknal_conn_decref(conn);
1167 if (!peer_ni->ksnp_closing &&
1168 list_empty(&peer_ni->ksnp_conns) &&
1169 peer_ni->ksnp_conn_cb == NULL) {
1170 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1171 ksocknal_unlink_peer_locked(peer_ni);
1174 write_unlock_bh(global_lock);
1178 CERROR("Not creating conn %s type %d: %s\n",
1179 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1181 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1182 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1187 /* Request retry by replying with CONN_NONE
1188 * ksnc_proto has been set already
1190 conn->ksnc_type = SOCKLND_CONN_NONE;
1191 hello->kshm_nips = 0;
1192 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1195 write_lock_bh(global_lock);
1196 peer_ni->ksnp_accepting--;
1197 write_unlock_bh(global_lock);
1201 * If we get here without an error code, just use -EALREADY.
1202 * Depending on how we got here, the error may be positive
1203 * or negative. Normalize the value for ksocknal_txlist_done().
1205 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1206 ksocknal_txlist_done(ni, &zombies, rc2);
1207 ksocknal_peer_decref(peer_ni);
1211 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1212 kshm_ips[LNET_INTERFACES_NUM]));
1214 LIBCFS_FREE(conn, sizeof(*conn));
1223 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1225 /* This just does the immmediate housekeeping, and queues the
1226 * connection for the reaper to terminate.
1227 * Caller holds ksnd_global_lock exclusively in irq context */
1228 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1229 struct ksock_conn_cb *conn_cb;
1230 struct ksock_conn *conn2;
1232 int duplicate_count = 0;
1234 LASSERT(peer_ni->ksnp_error == 0);
1235 LASSERT(!conn->ksnc_closing);
1236 conn->ksnc_closing = 1;
1238 /* ksnd_deathrow_conns takes over peer_ni's ref */
1239 list_del(&conn->ksnc_list);
1241 conn_cb = conn->ksnc_conn_cb;
1242 if (conn_cb != NULL) {
1243 /* dissociate conn from cb... */
1244 LASSERT(!conn_cb->ksnr_deleted);
1246 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1248 /* connected bit is set only if all connections
1249 * of the given type got created
1251 if (conn_count == conn_cb->ksnr_max_conns)
1252 LASSERT((conn_cb->ksnr_connected &
1253 BIT(conn->ksnc_type)) != 0);
1255 if (conn_count == 1) {
1256 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1258 if (conn2->ksnc_conn_cb == conn_cb &&
1259 conn2->ksnc_type == conn->ksnc_type)
1260 duplicate_count += 1;
1262 if (duplicate_count > 0)
1263 CERROR("Found %d duplicate conns type %d\n",
1267 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1269 conn->ksnc_conn_cb = NULL;
1271 /* drop conn's ref on conn_cb */
1272 ksocknal_conn_cb_decref(conn_cb);
1275 if (list_empty(&peer_ni->ksnp_conns)) {
1276 /* No more connections to this peer_ni */
1278 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1279 struct ksock_tx *tx;
1281 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1283 /* throw them to the last connection...,
1284 * these TXs will be send to /dev/null by scheduler */
1285 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1287 ksocknal_tx_prep(conn, tx);
1289 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1290 list_splice_init(&peer_ni->ksnp_tx_queue,
1291 &conn->ksnc_tx_queue);
1292 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1295 /* renegotiate protocol version */
1296 peer_ni->ksnp_proto = NULL;
1297 /* stash last conn close reason */
1298 peer_ni->ksnp_error = error;
1300 if (peer_ni->ksnp_conn_cb == NULL) {
1301 /* I've just closed last conn belonging to a
1302 * peer_ni with no connections to it
1304 ksocknal_unlink_peer_locked(peer_ni);
1308 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1310 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1311 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1313 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1317 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1319 bool notify = false;
1320 time64_t last_alive = 0;
1322 /* There has been a connection failure or comms error; but I'll only
1323 * tell LNET I think the peer_ni is dead if it's to another kernel and
1324 * there are no connections or connection attempts in existence. */
1326 read_lock(&ksocknal_data.ksnd_global_lock);
1328 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1329 list_empty(&peer_ni->ksnp_conns) &&
1330 peer_ni->ksnp_accepting == 0 &&
1331 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1333 last_alive = peer_ni->ksnp_last_alive;
1336 read_unlock(&ksocknal_data.ksnd_global_lock);
1339 lnet_notify(peer_ni->ksnp_ni,
1340 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1341 false, false, last_alive);
1345 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1347 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1348 struct ksock_tx *tx;
1349 struct ksock_tx *tmp;
1352 /* NB safe to finalize TXs because closing of socket will
1353 * abort all buffered data */
1354 LASSERT(conn->ksnc_sock == NULL);
1356 spin_lock(&peer_ni->ksnp_lock);
1358 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1360 if (tx->tx_conn != conn)
1363 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1365 tx->tx_msg.ksm_zc_cookies[0] = 0;
1366 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1367 list_move(&tx->tx_zc_list, &zlist);
1370 spin_unlock(&peer_ni->ksnp_lock);
1372 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1373 tx_zc_list)) != NULL) {
1374 list_del(&tx->tx_zc_list);
1375 ksocknal_tx_decref(tx);
1380 ksocknal_terminate_conn(struct ksock_conn *conn)
1382 /* This gets called by the reaper (guaranteed thread context) to
1383 * disengage the socket from its callbacks and close it.
1384 * ksnc_refcount will eventually hit zero, and then the reaper will
1387 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1388 struct ksock_sched *sched = conn->ksnc_scheduler;
1389 bool failed = false;
1391 LASSERT(conn->ksnc_closing);
1393 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1394 spin_lock_bh(&sched->kss_lock);
1396 /* a closing conn is always ready to tx */
1397 conn->ksnc_tx_ready = 1;
1399 if (!conn->ksnc_tx_scheduled &&
1400 !list_empty(&conn->ksnc_tx_queue)) {
1401 list_add_tail(&conn->ksnc_tx_list,
1402 &sched->kss_tx_conns);
1403 conn->ksnc_tx_scheduled = 1;
1404 /* extra ref for scheduler */
1405 ksocknal_conn_addref(conn);
1407 wake_up(&sched->kss_waitq);
1410 spin_unlock_bh(&sched->kss_lock);
1412 /* serialise with callbacks */
1413 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1415 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1417 /* OK, so this conn may not be completely disengaged from its
1418 * scheduler yet, but it _has_ committed to terminate...
1420 conn->ksnc_scheduler->kss_nconns--;
1422 if (peer_ni->ksnp_error != 0) {
1423 /* peer_ni's last conn closed in error */
1424 LASSERT(list_empty(&peer_ni->ksnp_conns));
1426 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1429 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1432 ksocknal_peer_failed(peer_ni);
1434 /* The socket is closed on the final put; either here, or in
1435 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1436 * when the connection was established, this will close the socket
1437 * immediately, aborting anything buffered in it. Any hung
1438 * zero-copy transmits will therefore complete in finite time.
1440 ksocknal_connsock_decref(conn);
1444 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1446 /* Queue the conn for the reaper to destroy */
1447 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1448 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1450 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1451 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1453 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1457 ksocknal_destroy_conn(struct ksock_conn *conn)
1461 /* Final coup-de-grace of the reaper */
1462 CDEBUG(D_NET, "connection %p\n", conn);
1464 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1465 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1466 LASSERT(conn->ksnc_sock == NULL);
1467 LASSERT(conn->ksnc_conn_cb == NULL);
1468 LASSERT(!conn->ksnc_tx_scheduled);
1469 LASSERT(!conn->ksnc_rx_scheduled);
1470 LASSERT(list_empty(&conn->ksnc_tx_queue));
1472 /* complete current receive if any */
1473 switch (conn->ksnc_rx_state) {
1474 case SOCKNAL_RX_LNET_PAYLOAD:
1475 last_rcv = conn->ksnc_rx_deadline -
1477 CERROR("Completing partial receive from %s[%d], ip %pIScp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1478 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1480 &conn->ksnc_peeraddr,
1481 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1482 ktime_get_seconds() - last_rcv);
1483 if (conn->ksnc_lnet_msg)
1484 conn->ksnc_lnet_msg->msg_health_status =
1485 LNET_MSG_STATUS_REMOTE_ERROR;
1486 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1488 case SOCKNAL_RX_LNET_HEADER:
1489 if (conn->ksnc_rx_started)
1490 CERROR("Incomplete receive of lnet header from %s, ip %pIScp, with error, protocol: %d.x.\n",
1491 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1492 &conn->ksnc_peeraddr,
1493 conn->ksnc_proto->pro_version);
1495 case SOCKNAL_RX_KSM_HEADER:
1496 if (conn->ksnc_rx_started)
1497 CERROR("Incomplete receive of ksock message from %s, ip %pIScp, with error, protocol: %d.x.\n",
1498 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1499 &conn->ksnc_peeraddr,
1500 conn->ksnc_proto->pro_version);
1502 case SOCKNAL_RX_SLOP:
1503 if (conn->ksnc_rx_started)
1504 CERROR("Incomplete receive of slops from %s, ip %pIScp, with error\n",
1505 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1506 &conn->ksnc_peeraddr);
1513 ksocknal_peer_decref(conn->ksnc_peer);
1515 LIBCFS_FREE(conn, sizeof(*conn));
1519 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1520 struct sockaddr *addr, int why)
1522 struct ksock_conn *conn;
1523 struct ksock_conn *cnxt;
1526 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1529 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1531 ksocknal_close_conn_locked(conn, why);
1539 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1541 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1544 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1546 count = ksocknal_close_peer_conns_locked(
1547 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1549 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1555 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1557 struct ksock_peer_ni *peer_ni;
1558 struct hlist_node *pnxt;
1563 struct sockaddr_in sa = {.sin_family = AF_INET};
1565 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1567 if (!LNET_NID_IS_ANY(&id->nid)) {
1568 lo = hash_min(nidhash(&id->nid),
1569 HASH_BITS(ksocknal_data.ksnd_peers));
1573 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1576 sa.sin_addr.s_addr = htonl(ipaddr);
1577 for (i = lo; i <= hi; i++) {
1578 hlist_for_each_entry_safe(peer_ni, pnxt,
1579 &ksocknal_data.ksnd_peers[i],
1582 if (!((LNET_NID_IS_ANY(&id->nid) ||
1583 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1584 (id->pid == LNET_PID_ANY ||
1585 id->pid == peer_ni->ksnp_id.pid)))
1588 count += ksocknal_close_peer_conns_locked(
1590 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1594 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1596 /* wildcards always succeed */
1597 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1601 return (count == 0 ? -ENOENT : 0);
1605 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1607 /* The router is telling me she's been notified of a change in
1610 struct lnet_processid id = {
1611 .pid = LNET_PID_ANY,
1615 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1617 /* If the gateway crashed, close all open connections... */
1618 ksocknal_close_matching_conns(&id, 0);
1621 /* We can only establish new connections
1622 * if we have autroutes, and these connect on demand.
1627 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1631 struct ksock_conn *conn;
1633 for (index = 0; ; index++) {
1634 read_lock(&ksocknal_data.ksnd_global_lock);
1639 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1641 ksocknal_conn_addref(conn);
1646 read_unlock(&ksocknal_data.ksnd_global_lock);
1651 ksocknal_lib_push_conn (conn);
1652 ksocknal_conn_decref(conn);
1657 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1664 if (!LNET_NID_IS_ANY(&id->nid)) {
1665 lo = hash_min(nidhash(&id->nid),
1666 HASH_BITS(ksocknal_data.ksnd_peers));
1670 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1673 for (bkt = lo; bkt <= hi; bkt++) {
1674 int peer_off; /* searching offset in peer_ni hash table */
1676 for (peer_off = 0; ; peer_off++) {
1677 struct ksock_peer_ni *peer_ni;
1680 read_lock(&ksocknal_data.ksnd_global_lock);
1681 hlist_for_each_entry(peer_ni,
1682 &ksocknal_data.ksnd_peers[bkt],
1684 if (!((LNET_NID_IS_ANY(&id->nid) ||
1686 &peer_ni->ksnp_id.nid)) &&
1687 (id->pid == LNET_PID_ANY ||
1688 id->pid == peer_ni->ksnp_id.pid)))
1691 if (i++ == peer_off) {
1692 ksocknal_peer_addref(peer_ni);
1696 read_unlock(&ksocknal_data.ksnd_global_lock);
1698 if (i <= peer_off) /* no match */
1702 ksocknal_push_peer(peer_ni);
1703 ksocknal_peer_decref(peer_ni);
1710 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1712 struct lnet_processid id = {};
1713 struct libcfs_ioctl_data *data = arg;
1717 case IOC_LIBCFS_GET_INTERFACE: {
1718 struct ksock_net *net = ni->ni_data;
1719 struct ksock_interface *iface;
1720 struct sockaddr_in *sa;
1722 read_lock(&ksocknal_data.ksnd_global_lock);
1724 if (data->ioc_count >= 1) {
1728 iface = &net->ksnn_interface;
1730 sa = (void *)&iface->ksni_addr;
1731 if (sa->sin_family == AF_INET)
1732 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1734 data->ioc_u32[0] = 0xFFFFFFFF;
1735 data->ioc_u32[1] = iface->ksni_netmask;
1736 data->ioc_u32[2] = iface->ksni_npeers;
1737 data->ioc_u32[3] = iface->ksni_nroutes;
1740 read_unlock(&ksocknal_data.ksnd_global_lock);
1744 case IOC_LIBCFS_GET_PEER: {
1749 int share_count = 0;
1751 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1752 &id, &myip, &ip, &port,
1753 &conn_count, &share_count);
1756 if (!nid_is_nid4(&id.nid))
1758 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
1759 data->ioc_count = share_count;
1760 data->ioc_u32[0] = ip;
1761 data->ioc_u32[1] = port;
1762 data->ioc_u32[2] = myip;
1763 data->ioc_u32[3] = conn_count;
1764 data->ioc_u32[4] = id.pid;
1768 case IOC_LIBCFS_ADD_PEER: {
1769 struct sockaddr_in sa = {.sin_family = AF_INET};
1771 id.pid = LNET_PID_LUSTRE;
1772 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1773 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1774 sa.sin_port = htons(data->ioc_u32[1]);
1775 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1777 case IOC_LIBCFS_DEL_PEER:
1778 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1779 id.pid = LNET_PID_ANY;
1780 return ksocknal_del_peer(ni, &id);
1782 case IOC_LIBCFS_GET_CONN: {
1786 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1787 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1788 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1793 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1795 data->ioc_count = txmem;
1796 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1797 data->ioc_flags = nagle;
1798 if (psa->sin_family == AF_INET)
1799 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1801 data->ioc_u32[0] = 0xFFFFFFFF;
1802 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1803 &conn->ksnc_peeraddr);
1804 if (mysa->sin_family == AF_INET)
1805 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1807 data->ioc_u32[2] = 0xFFFFFFFF;
1808 data->ioc_u32[3] = conn->ksnc_type;
1809 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1810 data->ioc_u32[5] = rxmem;
1811 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1812 ksocknal_conn_decref(conn);
1816 case IOC_LIBCFS_CLOSE_CONNECTION:
1817 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1818 id.pid = LNET_PID_ANY;
1819 return ksocknal_close_matching_conns(&id,
1822 case IOC_LIBCFS_REGISTER_MYNID:
1823 /* Ignore if this is a noop */
1824 if (nid_is_nid4(&ni->ni_nid) &&
1825 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1828 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1829 libcfs_nid2str(data->ioc_nid),
1830 libcfs_nidstr(&ni->ni_nid));
1833 case IOC_LIBCFS_PUSH_CONNECTION:
1834 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1835 id.pid = LNET_PID_ANY;
1836 return ksocknal_push(ni, &id);
1845 ksocknal_free_buffers (void)
1847 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1849 if (ksocknal_data.ksnd_schedulers != NULL)
1850 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1852 spin_lock(&ksocknal_data.ksnd_tx_lock);
1854 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1856 struct ksock_tx *tx;
1858 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1859 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1861 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1862 tx_list)) != NULL) {
1863 list_del(&tx->tx_list);
1864 LIBCFS_FREE(tx, tx->tx_desc_size);
1867 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1871 static int ksocknal_get_link_status(struct net_device *dev)
1877 if (!netif_running(dev)) {
1879 CDEBUG(D_NET, "device not running\n");
1881 /* Some devices may not be providing link settings */
1882 else if (dev->ethtool_ops->get_link) {
1883 ret = dev->ethtool_ops->get_link(dev);
1884 CDEBUG(D_NET, "get_link returns %u\n", ret);
1891 ksocknal_handle_link_state_change(struct net_device *dev,
1892 unsigned char operstate)
1894 struct lnet_ni *ni = NULL;
1895 struct ksock_net *net;
1896 struct ksock_net *cnxt;
1898 unsigned char link_down = !(operstate == IF_OPER_UP);
1899 struct in_device *in_dev;
1900 bool found_ip = false;
1901 struct ksock_interface *ksi = NULL;
1902 struct sockaddr_in *sa;
1903 DECLARE_CONST_IN_IFADDR(ifa);
1905 ifindex = dev->ifindex;
1907 if (!ksocknal_data.ksnd_nnets)
1910 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1913 ksi = &net->ksnn_interface;
1914 sa = (void *)&ksi->ksni_addr;
1917 if (ksi->ksni_index != ifindex ||
1918 strcmp(ksi->ksni_name, dev->name))
1923 in_dev = __in_dev_get_rtnl(dev);
1925 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1927 CDEBUG(D_NET, "set link fatal state to 1\n");
1928 atomic_set(&ni->ni_fatal_error_on, 1);
1931 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1932 if (sa->sin_addr.s_addr == ifa->ifa_local)
1938 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1940 CDEBUG(D_NET, "set link fatal state to 1\n");
1941 atomic_set(&ni->ni_fatal_error_on, 1);
1946 CDEBUG(D_NET, "set link fatal state to 1\n");
1947 atomic_set(&ni->ni_fatal_error_on, link_down);
1949 CDEBUG(D_NET, "set link fatal state to %u\n",
1950 (ksocknal_get_link_status(dev) == 0));
1951 atomic_set(&ni->ni_fatal_error_on,
1952 (ksocknal_get_link_status(dev) == 0));
1961 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
1964 struct ksock_net *net;
1965 struct ksock_net *cnxt;
1966 struct net_device *event_netdev = ifa->ifa_dev->dev;
1968 struct ksock_interface *ksi = NULL;
1969 struct sockaddr_in *sa;
1971 if (!ksocknal_data.ksnd_nnets)
1974 ifindex = event_netdev->ifindex;
1976 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1979 ksi = &net->ksnn_interface;
1980 sa = (void *)&ksi->ksni_addr;
1982 if (ksi->ksni_index != ifindex ||
1983 strcmp(ksi->ksni_name, event_netdev->name))
1986 if (sa->sin_addr.s_addr == ifa->ifa_local) {
1987 CDEBUG(D_NET, "set link fatal state to %u\n",
1988 (event == NETDEV_DOWN));
1990 atomic_set(&ni->ni_fatal_error_on,
1991 (event == NETDEV_DOWN));
1998 /************************************
1999 * Net device notifier event handler
2000 ************************************/
2001 static int ksocknal_device_event(struct notifier_block *unused,
2002 unsigned long event, void *ptr)
2004 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2005 unsigned char operstate;
2007 operstate = dev->operstate;
2009 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2010 event, dev->name, dev->ifindex, operstate);
2016 ksocknal_handle_link_state_change(dev, operstate);
2023 /************************************
2024 * Inetaddr notifier event handler
2025 ************************************/
2026 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2027 unsigned long event, void *ptr)
2029 struct in_ifaddr *ifa = ptr;
2031 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2032 event, &ifa->ifa_address, &ifa->ifa_mask);
2038 ksocknal_handle_inetaddr_change(ifa, event);
2045 static struct notifier_block ksocknal_dev_notifier_block = {
2046 .notifier_call = ksocknal_device_event,
2049 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2050 .notifier_call = ksocknal_inetaddr_event,
2054 ksocknal_base_shutdown(void)
2056 struct ksock_sched *sched;
2057 struct ksock_peer_ni *peer_ni;
2060 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2061 libcfs_kmem_read());
2062 LASSERT (ksocknal_data.ksnd_nnets == 0);
2064 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2065 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2066 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2069 switch (ksocknal_data.ksnd_init) {
2074 case SOCKNAL_INIT_ALL:
2075 case SOCKNAL_INIT_DATA:
2076 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2079 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2080 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2081 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2082 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2083 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2085 if (ksocknal_data.ksnd_schedulers != NULL) {
2086 cfs_percpt_for_each(sched, i,
2087 ksocknal_data.ksnd_schedulers) {
2089 LASSERT(list_empty(&sched->kss_tx_conns));
2090 LASSERT(list_empty(&sched->kss_rx_conns));
2091 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2092 LASSERT(sched->kss_nconns == 0);
2096 /* flag threads to terminate; wake and wait for them to die */
2097 ksocknal_data.ksnd_shuttingdown = 1;
2098 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2099 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2101 if (ksocknal_data.ksnd_schedulers != NULL) {
2102 cfs_percpt_for_each(sched, i,
2103 ksocknal_data.ksnd_schedulers)
2104 wake_up_all(&sched->kss_waitq);
2107 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2108 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2109 "waiting for %d threads to terminate\n",
2110 atomic_read(&ksocknal_data.ksnd_nthreads));
2112 ksocknal_free_buffers();
2114 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2118 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2119 libcfs_kmem_read());
2121 module_put(THIS_MODULE);
2125 ksocknal_base_startup(void)
2127 struct ksock_sched *sched;
2131 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2132 LASSERT(ksocknal_data.ksnd_nnets == 0);
2134 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2136 hash_init(ksocknal_data.ksnd_peers);
2138 rwlock_init(&ksocknal_data.ksnd_global_lock);
2139 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2141 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2142 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2143 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2144 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2145 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2147 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2148 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2149 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2150 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2152 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2153 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2155 /* NB memset above zeros whole of ksocknal_data */
2157 /* flag lists/ptrs/locks initialised */
2158 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2159 if (!try_module_get(THIS_MODULE))
2162 /* Create a scheduler block per available CPT */
2163 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2165 if (ksocknal_data.ksnd_schedulers == NULL)
2168 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2172 * make sure not to allocate more threads than there are
2173 * cores/CPUs in teh CPT
2175 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2176 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2177 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2180 * max to half of CPUs, assume another half should be
2181 * reserved for upper layer modules
2183 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2186 sched->kss_nthreads_max = nthrs;
2189 spin_lock_init(&sched->kss_lock);
2190 INIT_LIST_HEAD(&sched->kss_rx_conns);
2191 INIT_LIST_HEAD(&sched->kss_tx_conns);
2192 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2193 init_waitqueue_head(&sched->kss_waitq);
2196 ksocknal_data.ksnd_connd_starting = 0;
2197 ksocknal_data.ksnd_connd_failed_stamp = 0;
2198 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2199 /* must have at least 2 connds to remain responsive to accepts while
2201 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2202 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2204 if (*ksocknal_tunables.ksnd_nconnds_max <
2205 *ksocknal_tunables.ksnd_nconnds) {
2206 ksocknal_tunables.ksnd_nconnds_max =
2207 ksocknal_tunables.ksnd_nconnds;
2210 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2211 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2212 ksocknal_data.ksnd_connd_starting++;
2213 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2215 rc = ksocknal_thread_start(ksocknal_connd,
2216 (void *)((uintptr_t)i),
2217 "socknal_cd%02d", i);
2219 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2220 ksocknal_data.ksnd_connd_starting--;
2221 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2222 CERROR("Can't spawn socknal connd: %d\n", rc);
2227 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2229 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2233 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2234 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2236 /* flag everything initialised */
2237 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2242 ksocknal_base_shutdown();
2247 ksocknal_debug_peerhash(struct lnet_ni *ni)
2249 struct ksock_peer_ni *peer_ni;
2252 read_lock(&ksocknal_data.ksnd_global_lock);
2254 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2255 struct ksock_conn_cb *conn_cb;
2256 struct ksock_conn *conn;
2258 if (peer_ni->ksnp_ni != ni)
2261 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2262 libcfs_idstr(&peer_ni->ksnp_id),
2263 refcount_read(&peer_ni->ksnp_refcount),
2264 peer_ni->ksnp_closing,
2265 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2266 peer_ni->ksnp_zc_next_cookie,
2267 !list_empty(&peer_ni->ksnp_tx_queue),
2268 !list_empty(&peer_ni->ksnp_zc_req_list));
2270 conn_cb = peer_ni->ksnp_conn_cb;
2272 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2273 refcount_read(&conn_cb->ksnr_refcount),
2274 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2275 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2278 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2279 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2280 refcount_read(&conn->ksnc_conn_refcount),
2281 refcount_read(&conn->ksnc_sock_refcount),
2282 conn->ksnc_type, conn->ksnc_closing);
2287 read_unlock(&ksocknal_data.ksnd_global_lock);
2292 ksocknal_shutdown(struct lnet_ni *ni)
2294 struct ksock_net *net = ni->ni_data;
2296 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2297 LASSERT(ksocknal_data.ksnd_nnets > 0);
2299 /* prevent new peers */
2300 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2302 /* Delete all peers */
2303 ksocknal_del_peer(ni, NULL);
2305 /* Wait for all peer_ni state to clean up */
2306 wait_var_event_warning(&net->ksnn_npeers,
2307 atomic_read(&net->ksnn_npeers) ==
2308 SOCKNAL_SHUTDOWN_BIAS,
2309 "waiting for %d peers to disconnect\n",
2310 ksocknal_debug_peerhash(ni) +
2311 atomic_read(&net->ksnn_npeers) -
2312 SOCKNAL_SHUTDOWN_BIAS);
2314 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2315 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2317 list_del(&net->ksnn_list);
2318 LIBCFS_FREE(net, sizeof(*net));
2320 ksocknal_data.ksnd_nnets--;
2321 if (ksocknal_data.ksnd_nnets == 0)
2322 ksocknal_base_shutdown();
2326 ksocknal_search_new_ipif(struct ksock_net *net)
2329 char *ifnam = &net->ksnn_interface.ksni_name[0];
2330 char *colon = strchr(ifnam, ':');
2332 struct ksock_net *tmp;
2337 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2338 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2339 char *colon2 = strchr(ifnam2, ':');
2344 found = strcmp(ifnam, ifnam2) == 0;
2357 ksocknal_start_schedulers(struct ksock_sched *sched)
2363 if (sched->kss_nthreads == 0) {
2364 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2365 nthrs = sched->kss_nthreads_max;
2367 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2369 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2370 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2372 nthrs = min(nthrs, sched->kss_nthreads_max);
2374 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2375 /* increase two threads if there is new interface */
2376 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2379 for (i = 0; i < nthrs; i++) {
2382 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2383 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2384 "socknal_sd%02d_%02d",
2386 (int)KSOCK_THREAD_SID(id));
2390 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2391 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2395 sched->kss_nthreads += i;
2400 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2402 int newif = ksocknal_search_new_ipif(net);
2406 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2409 for (i = 0; i < ncpts; i++) {
2410 struct ksock_sched *sched;
2411 int cpt = (cpts == NULL) ? i : cpts[i];
2413 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2414 sched = ksocknal_data.ksnd_schedulers[cpt];
2416 if (!newif && sched->kss_nthreads > 0)
2419 rc = ksocknal_start_schedulers(sched);
2427 ksocknal_startup(struct lnet_ni *ni)
2429 struct ksock_net *net;
2430 struct ksock_interface *ksi = NULL;
2431 struct lnet_inetdev *ifaces = NULL;
2432 struct sockaddr_in *sa;
2436 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2437 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2438 rc = ksocknal_base_startup();
2442 LIBCFS_ALLOC(net, sizeof(*net));
2445 net->ksnn_incarnation = ktime_get_real_ns();
2448 ksocknal_tunables_setup(ni);
2450 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2454 ksi = &net->ksnn_interface;
2456 /* Use the first discovered interface or look in the list */
2457 if (ni->ni_interface) {
2458 for (i = 0; i < rc; i++)
2459 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2462 /* ni_interfaces doesn't contain the interface we want */
2464 CERROR("ksocklnd: failed to find interface %s\n",
2470 ni->ni_dev_cpt = ifaces[i].li_cpt;
2471 ksi->ksni_index = ifaces[i].li_index;
2472 sa = (void *)&ksi->ksni_addr;
2473 memset(sa, 0, sizeof(*sa));
2474 sa->sin_family = AF_INET;
2475 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2476 ksi->ksni_netmask = ifaces[i].li_netmask;
2477 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2479 /* call it before add it to ksocknal_data.ksnd_nets */
2480 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2485 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2486 ni->ni_nid.nid_addr[0] =
2487 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2488 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2490 ksocknal_data.ksnd_nnets++;
2495 LIBCFS_FREE(net, sizeof(*net));
2497 if (ksocknal_data.ksnd_nnets == 0)
2498 ksocknal_base_shutdown();
2503 static void __exit ksocklnd_exit(void)
2505 lnet_unregister_lnd(&the_ksocklnd);
2508 static const struct lnet_lnd the_ksocklnd = {
2509 .lnd_type = SOCKLND,
2510 .lnd_startup = ksocknal_startup,
2511 .lnd_shutdown = ksocknal_shutdown,
2512 .lnd_ctl = ksocknal_ctl,
2513 .lnd_send = ksocknal_send,
2514 .lnd_recv = ksocknal_recv,
2515 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2516 .lnd_accept = ksocknal_accept,
2519 static int __init ksocklnd_init(void)
2523 /* check ksnr_connected/connecting field large enough */
2524 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2525 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2527 rc = ksocknal_tunables_init();
2531 lnet_register_lnd(&the_ksocklnd);
2536 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2537 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2538 MODULE_VERSION("2.8.0");
2539 MODULE_LICENSE("GPL");
2541 module_init(ksocklnd_init);
2542 module_exit(ksocklnd_exit);