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_processid *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 = peer_ni->ksnp_id;
325 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
329 *id = peer_ni->ksnp_id;
330 *myip = peer_ni->ksnp_passive_ips[j];
339 if (peer_ni->ksnp_conn_cb) {
343 conn_cb = peer_ni->ksnp_conn_cb;
345 *id = peer_ni->ksnp_id;
346 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
347 struct sockaddr_in *sa =
348 (void *)&conn_cb->ksnr_addr;
350 rc = choose_ipv4_src(myip,
351 conn_cb->ksnr_myiface,
352 ntohl(sa->sin_addr.s_addr),
354 *peer_ip = ntohl(sa->sin_addr.s_addr);
355 *port = ntohs(sa->sin_port);
358 *peer_ip = 0xFFFFFFFF;
362 *conn_count = conn_cb->ksnr_conn_count;
368 read_unlock(&ksocknal_data.ksnd_global_lock);
373 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
376 unsigned int count = 0;
379 case SOCKLND_CONN_CONTROL:
380 count = conn_cb->ksnr_ctrl_conn_count;
382 case SOCKLND_CONN_BULK_IN:
383 count = conn_cb->ksnr_blki_conn_count;
385 case SOCKLND_CONN_BULK_OUT:
386 count = conn_cb->ksnr_blko_conn_count;
388 case SOCKLND_CONN_ANY:
389 count = conn_cb->ksnr_conn_count;
400 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
402 struct lnet_ni *ni = peer_ni->ksnp_ni;
403 struct lnet_ioctl_config_socklnd_tunables *tunables;
407 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
409 return tunables->lnd_conns_per_peer;
413 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
416 conn_cb->ksnr_conn_count++;
418 /* check if all connections of the given type got created */
420 case SOCKLND_CONN_CONTROL:
421 conn_cb->ksnr_ctrl_conn_count++;
422 /* there's a single control connection per peer,
423 * two in case of loopback
425 conn_cb->ksnr_connected |= BIT(type);
427 case SOCKLND_CONN_BULK_IN:
428 conn_cb->ksnr_blki_conn_count++;
429 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
430 conn_cb->ksnr_connected |= BIT(type);
432 case SOCKLND_CONN_BULK_OUT:
433 conn_cb->ksnr_blko_conn_count++;
434 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
435 conn_cb->ksnr_connected |= BIT(type);
437 case SOCKLND_CONN_ANY:
438 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
439 conn_cb->ksnr_connected |= BIT(type);
446 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
447 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
452 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
455 conn_cb->ksnr_conn_count--;
457 /* check if all connections of the given type got created */
459 case SOCKLND_CONN_CONTROL:
460 conn_cb->ksnr_ctrl_conn_count--;
461 /* there's a single control connection per peer,
462 * two in case of loopback
464 if (conn_cb->ksnr_ctrl_conn_count == 0)
465 conn_cb->ksnr_connected &= ~BIT(type);
467 case SOCKLND_CONN_BULK_IN:
468 conn_cb->ksnr_blki_conn_count--;
469 if (conn_cb->ksnr_blki_conn_count < conn_cb->ksnr_max_conns)
470 conn_cb->ksnr_connected &= ~BIT(type);
472 case SOCKLND_CONN_BULK_OUT:
473 conn_cb->ksnr_blko_conn_count--;
474 if (conn_cb->ksnr_blko_conn_count < conn_cb->ksnr_max_conns)
475 conn_cb->ksnr_connected &= ~BIT(type);
477 case SOCKLND_CONN_ANY:
478 if (conn_cb->ksnr_conn_count < conn_cb->ksnr_max_conns)
479 conn_cb->ksnr_connected &= ~BIT(type);
486 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
487 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
491 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
492 struct ksock_conn *conn)
494 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
495 int type = conn->ksnc_type;
496 struct ksock_interface *iface;
499 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
501 conn->ksnc_conn_cb = conn_cb;
502 ksocknal_conn_cb_addref(conn_cb);
504 if (conn_cb->ksnr_myiface != conn_iface) {
505 if (conn_cb->ksnr_myiface < 0) {
506 /* route wasn't bound locally yet (the initial route) */
507 CDEBUG(D_NET, "Binding %s %pIS to interface %d\n",
508 libcfs_idstr(&peer_ni->ksnp_id),
513 "Rebinding %s %pIS from interface %d to %d\n",
514 libcfs_idstr(&peer_ni->ksnp_id),
516 conn_cb->ksnr_myiface,
519 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
520 conn_cb->ksnr_myiface);
522 iface->ksni_nroutes--;
524 conn_cb->ksnr_myiface = conn_iface;
525 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
526 conn_cb->ksnr_myiface);
528 iface->ksni_nroutes++;
531 ksocknal_incr_conn_count(conn_cb, type);
533 /* Successful connection => further attempts can
534 * proceed immediately
536 conn_cb->ksnr_retry_interval = 0;
540 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
541 struct ksock_conn_cb *conn_cb)
543 struct ksock_conn *conn;
544 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
546 LASSERT(!peer_ni->ksnp_closing);
547 LASSERT(!conn_cb->ksnr_peer);
548 LASSERT(!conn_cb->ksnr_scheduled);
549 LASSERT(!conn_cb->ksnr_connecting);
550 LASSERT(conn_cb->ksnr_connected == 0);
552 conn_cb->ksnr_peer = peer_ni;
553 ksocknal_peer_addref(peer_ni);
555 /* set the conn_cb's interface to the current net's interface */
556 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
557 net->ksnn_interface.ksni_nroutes++;
559 /* peer_ni's route list takes over my ref on 'route' */
560 peer_ni->ksnp_conn_cb = conn_cb;
562 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
563 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
564 (struct sockaddr *)&conn_cb->ksnr_addr))
567 ksocknal_associate_cb_conn_locked(conn_cb, conn);
568 /* keep going (typed conns) */
573 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
575 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
576 struct ksock_interface *iface;
577 struct ksock_conn *conn;
578 struct ksock_conn *cnxt;
580 LASSERT(!conn_cb->ksnr_deleted);
582 /* Close associated conns */
583 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
584 if (conn->ksnc_conn_cb != conn_cb)
587 ksocknal_close_conn_locked(conn, 0);
590 if (conn_cb->ksnr_myiface >= 0) {
591 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
592 conn_cb->ksnr_myiface);
594 iface->ksni_nroutes--;
597 conn_cb->ksnr_deleted = 1;
598 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
599 peer_ni->ksnp_conn_cb = NULL;
601 if (list_empty(&peer_ni->ksnp_conns)) {
602 /* I've just removed the last route to a peer_ni with no active
605 ksocknal_unlink_peer_locked(peer_ni);
610 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_processid *id,
611 struct sockaddr *addr)
613 struct ksock_peer_ni *peer_ni;
614 struct ksock_peer_ni *peer2;
615 struct ksock_conn_cb *conn_cb;
617 if (LNET_NID_IS_ANY(&id->nid) ||
618 id->pid == LNET_PID_ANY)
621 /* Have a brand new peer_ni ready... */
622 peer_ni = ksocknal_create_peer(ni, id);
624 return PTR_ERR(peer_ni);
626 conn_cb = ksocknal_create_conn_cb(addr);
628 ksocknal_peer_decref(peer_ni);
632 write_lock_bh(&ksocknal_data.ksnd_global_lock);
634 /* always called with a ref on ni, so shutdown can't have started */
635 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
638 peer2 = ksocknal_find_peer_locked(ni, id);
640 ksocknal_peer_decref(peer_ni);
643 /* peer_ni table takes my ref on peer_ni */
644 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
648 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
650 /* Remember conns_per_peer setting at the time
651 * of connection initiation. It will define the
652 * max number of conns per type for this conn_cb
655 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
657 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
663 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni)
665 struct ksock_conn *conn;
666 struct ksock_conn *cnxt;
667 struct ksock_conn_cb *conn_cb;
669 LASSERT(!peer_ni->ksnp_closing);
671 /* Extra ref prevents peer_ni disappearing until I'm done with it */
672 ksocknal_peer_addref(peer_ni);
673 conn_cb = peer_ni->ksnp_conn_cb;
675 ksocknal_del_conn_cb_locked(conn_cb);
677 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
679 ksocknal_close_conn_locked(conn, 0);
681 ksocknal_peer_decref(peer_ni);
682 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
686 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_processid *id)
689 struct hlist_node *pnxt;
690 struct ksock_peer_ni *peer_ni;
696 write_lock_bh(&ksocknal_data.ksnd_global_lock);
698 if (id && !LNET_NID_IS_ANY(&id->nid)) {
699 lo = hash_min(nidhash(&id->nid),
700 HASH_BITS(ksocknal_data.ksnd_peers));
704 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
707 for (i = lo; i <= hi; i++) {
708 hlist_for_each_entry_safe(peer_ni, pnxt,
709 &ksocknal_data.ksnd_peers[i],
711 if (peer_ni->ksnp_ni != ni)
714 if (!((!id || LNET_NID_IS_ANY(&id->nid) ||
715 nid_same(&peer_ni->ksnp_id.nid, &id->nid)) &&
716 (!id || id->pid == LNET_PID_ANY ||
717 peer_ni->ksnp_id.pid == id->pid)))
720 ksocknal_peer_addref(peer_ni); /* a ref for me... */
722 ksocknal_del_peer_locked(peer_ni);
724 if (peer_ni->ksnp_closing &&
725 !list_empty(&peer_ni->ksnp_tx_queue)) {
726 LASSERT(list_empty(&peer_ni->ksnp_conns));
727 LASSERT(peer_ni->ksnp_conn_cb == NULL);
729 list_splice_init(&peer_ni->ksnp_tx_queue,
733 ksocknal_peer_decref(peer_ni); /* ...till here */
735 rc = 0; /* matched! */
739 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
741 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
746 static struct ksock_conn *
747 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
749 struct ksock_peer_ni *peer_ni;
750 struct ksock_conn *conn;
753 read_lock(&ksocknal_data.ksnd_global_lock);
755 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
756 LASSERT(!peer_ni->ksnp_closing);
758 if (peer_ni->ksnp_ni != ni)
761 list_for_each_entry(conn, &peer_ni->ksnp_conns,
766 ksocknal_conn_addref(conn);
767 read_unlock(&ksocknal_data.ksnd_global_lock);
772 read_unlock(&ksocknal_data.ksnd_global_lock);
776 static struct ksock_sched *
777 ksocknal_choose_scheduler_locked(unsigned int cpt)
779 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
782 if (sched->kss_nthreads == 0) {
783 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
784 if (sched->kss_nthreads > 0) {
785 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
786 cpt, sched->kss_cpt);
797 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
799 struct ksock_connreq *cr;
801 struct sockaddr_storage peer;
803 rc = lnet_sock_getaddr(sock, true, &peer);
805 CERROR("Can't determine new connection's address\n");
809 LIBCFS_ALLOC(cr, sizeof(*cr));
811 LCONSOLE_ERROR_MSG(0x12f,
812 "Dropping connection request from %pIS: memory exhausted\n",
819 cr->ksncr_sock = sock;
821 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
823 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
824 wake_up(&ksocknal_data.ksnd_connd_waitq);
826 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
831 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
834 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
835 return conn_cb->ksnr_connecting;
840 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
841 struct socket *sock, int type)
843 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
845 struct lnet_processid peerid;
847 struct ksock_conn *conn;
848 struct ksock_conn *conn2;
849 struct ksock_peer_ni *peer_ni = NULL;
850 struct ksock_peer_ni *peer2;
851 struct ksock_sched *sched;
852 struct ksock_hello_msg *hello;
855 struct ksock_tx *txtmp;
862 active = (conn_cb != NULL);
864 LASSERT(active == (type != SOCKLND_CONN_NONE));
866 LIBCFS_ALLOC(conn, sizeof(*conn));
872 conn->ksnc_peer = NULL;
873 conn->ksnc_conn_cb = NULL;
874 conn->ksnc_sock = sock;
875 /* 2 ref, 1 for conn, another extra ref prevents socket
876 * being closed before establishment of connection */
877 refcount_set(&conn->ksnc_sock_refcount, 2);
878 conn->ksnc_type = type;
879 ksocknal_lib_save_callback(sock, conn);
880 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
882 conn->ksnc_rx_ready = 0;
883 conn->ksnc_rx_scheduled = 0;
885 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
886 conn->ksnc_tx_ready = 0;
887 conn->ksnc_tx_scheduled = 0;
888 conn->ksnc_tx_carrier = NULL;
889 atomic_set (&conn->ksnc_tx_nob, 0);
891 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
892 kshm_ips[LNET_INTERFACES_NUM]));
898 /* stash conn's local and remote addrs */
899 rc = ksocknal_lib_get_conn_addrs(conn);
903 /* Find out/confirm peer_ni's NID and connection type and get the
904 * vector of interfaces she's willing to let me connect to.
905 * Passive connections use the listener timeout since the peer_ni sends
910 peer_ni = conn_cb->ksnr_peer;
911 LASSERT(ni == peer_ni->ksnp_ni);
913 /* Active connection sends HELLO eagerly */
914 hello->kshm_nips = 0;
915 peerid = peer_ni->ksnp_id;
917 write_lock_bh(global_lock);
918 conn->ksnc_proto = peer_ni->ksnp_proto;
919 write_unlock_bh(global_lock);
921 if (conn->ksnc_proto == NULL) {
922 conn->ksnc_proto = &ksocknal_protocol_v3x;
923 #if SOCKNAL_VERSION_DEBUG
924 if (*ksocknal_tunables.ksnd_protocol == 2)
925 conn->ksnc_proto = &ksocknal_protocol_v2x;
926 else if (*ksocknal_tunables.ksnd_protocol == 1)
927 conn->ksnc_proto = &ksocknal_protocol_v1x;
931 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
935 peerid.nid = LNET_ANY_NID;
936 peerid.pid = LNET_PID_ANY;
938 /* Passive, get protocol from peer_ni */
939 conn->ksnc_proto = NULL;
942 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
946 LASSERT(rc == 0 || active);
947 LASSERT(conn->ksnc_proto != NULL);
948 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
950 cpt = lnet_nid2cpt(&peerid.nid, ni);
953 ksocknal_peer_addref(peer_ni);
954 write_lock_bh(global_lock);
956 peer_ni = ksocknal_create_peer(ni, &peerid);
957 if (IS_ERR(peer_ni)) {
958 rc = PTR_ERR(peer_ni);
962 write_lock_bh(global_lock);
964 /* called with a ref on ni, so shutdown can't have started */
965 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
967 peer2 = ksocknal_find_peer_locked(ni, &peerid);
969 /* NB this puts an "empty" peer_ni in the peer_ni
970 * table (which takes my ref) */
971 hash_add(ksocknal_data.ksnd_peers,
972 &peer_ni->ksnp_list, nidhash(&peerid.nid));
974 ksocknal_peer_decref(peer_ni);
979 ksocknal_peer_addref(peer_ni);
980 peer_ni->ksnp_accepting++;
982 /* Am I already connecting to this guy? Resolve in
983 * favour of higher NID...
985 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
986 ksocknal_connecting(peer_ni->ksnp_conn_cb,
987 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
989 warn = "connection race resolution";
994 if (peer_ni->ksnp_closing ||
995 (active && conn_cb->ksnr_deleted)) {
996 /* peer_ni/conn_cb got closed under me */
998 warn = "peer_ni/conn_cb removed";
1002 if (peer_ni->ksnp_proto == NULL) {
1003 /* Never connected before.
1004 * NB recv_hello may have returned EPROTO to signal my peer_ni
1005 * wants a different protocol than the one I asked for.
1007 LASSERT(list_empty(&peer_ni->ksnp_conns));
1009 peer_ni->ksnp_proto = conn->ksnc_proto;
1010 peer_ni->ksnp_incarnation = incarnation;
1013 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1014 peer_ni->ksnp_incarnation != incarnation) {
1015 /* peer_ni rebooted or I've got the wrong protocol version */
1016 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1018 peer_ni->ksnp_proto = NULL;
1020 warn = peer_ni->ksnp_incarnation != incarnation ?
1021 "peer_ni rebooted" :
1022 "wrong proto version";
1032 warn = "lost conn race";
1035 warn = "retry with different protocol version";
1039 /* Refuse to duplicate an existing connection, unless this is a
1040 * loopback connection */
1041 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1042 (struct sockaddr *)&conn->ksnc_myaddr)) {
1043 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1045 (struct sockaddr *)&conn2->ksnc_peeraddr,
1046 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1048 (struct sockaddr *)&conn2->ksnc_myaddr,
1049 (struct sockaddr *)&conn->ksnc_myaddr) ||
1050 conn2->ksnc_type != conn->ksnc_type)
1054 /* If max conns per type is not registered in conn_cb
1055 * as ksnr_max_conns, use ni's conns_per_peer
1057 if ((peer_ni->ksnp_conn_cb &&
1058 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1059 (!peer_ni->ksnp_conn_cb &&
1060 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1063 /* Reply on a passive connection attempt so the peer_ni
1064 * realises we're connected.
1074 /* If the connection created by this route didn't bind to the IP
1075 * address the route connected to, the connection/route matching
1076 * code below probably isn't going to work.
1079 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1080 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1081 CERROR("Route %s %pIS connected to %pIS\n",
1082 libcfs_idstr(&peer_ni->ksnp_id),
1083 &conn_cb->ksnr_addr,
1084 &conn->ksnc_peeraddr);
1087 /* Search for a conn_cb corresponding to the new connection and
1088 * create an association. This allows incoming connections created
1089 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1090 * continually create duplicate conn_cbs.
1092 conn_cb = peer_ni->ksnp_conn_cb;
1094 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1095 (struct sockaddr *)&conn_cb->ksnr_addr))
1096 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1098 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1099 peer_ni->ksnp_last_alive = ktime_get_seconds();
1100 peer_ni->ksnp_send_keepalive = 0;
1101 peer_ni->ksnp_error = 0;
1103 sched = ksocknal_choose_scheduler_locked(cpt);
1105 CERROR("no schedulers available. node is unhealthy\n");
1109 * The cpt might have changed if we ended up selecting a non cpt
1110 * native scheduler. So use the scheduler's cpt instead.
1112 cpt = sched->kss_cpt;
1113 sched->kss_nconns++;
1114 conn->ksnc_scheduler = sched;
1116 conn->ksnc_tx_last_post = ktime_get_seconds();
1117 /* Set the deadline for the outgoing HELLO to drain */
1118 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1119 conn->ksnc_tx_deadline = ktime_get_seconds() +
1121 smp_mb(); /* order with adding to peer_ni's conn list */
1123 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1124 ksocknal_conn_addref(conn);
1126 ksocknal_new_packet(conn, 0);
1128 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1130 /* Take packets blocking for this connection. */
1131 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1132 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1136 list_del(&tx->tx_list);
1137 ksocknal_queue_tx_locked(tx, conn);
1140 write_unlock_bh(global_lock);
1141 /* We've now got a new connection. Any errors from here on are just
1142 * like "normal" comms errors and we close the connection normally.
1143 * NB (a) we still have to send the reply HELLO for passive
1145 * (b) normal I/O on the conn is blocked until I setup and call the
1149 CDEBUG(D_NET, "New conn %s p %d.x %pIS -> %pISp"
1150 " incarnation:%lld sched[%d]\n",
1151 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1152 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1156 hello->kshm_nips = 0;
1157 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1160 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1161 kshm_ips[LNET_INTERFACES_NUM]));
1163 /* setup the socket AFTER I've received hello (it disables
1164 * SO_LINGER). I might call back to the acceptor who may want
1165 * to send a protocol version response and then close the
1166 * socket; this ensures the socket only tears down after the
1167 * response has been sent.
1170 rc = ksocknal_lib_setup_sock(sock);
1172 write_lock_bh(global_lock);
1174 /* NB my callbacks block while I hold ksnd_global_lock */
1175 ksocknal_lib_set_callback(sock, conn);
1178 peer_ni->ksnp_accepting--;
1180 write_unlock_bh(global_lock);
1183 write_lock_bh(global_lock);
1184 if (!conn->ksnc_closing) {
1185 /* could be closed by another thread */
1186 ksocknal_close_conn_locked(conn, rc);
1188 write_unlock_bh(global_lock);
1189 } else if (ksocknal_connsock_addref(conn) == 0) {
1190 /* Allow I/O to proceed. */
1191 ksocknal_read_callback(conn);
1192 ksocknal_write_callback(conn);
1193 ksocknal_connsock_decref(conn);
1196 ksocknal_connsock_decref(conn);
1197 ksocknal_conn_decref(conn);
1202 if (!peer_ni->ksnp_closing &&
1203 list_empty(&peer_ni->ksnp_conns) &&
1204 peer_ni->ksnp_conn_cb == NULL) {
1205 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1206 ksocknal_unlink_peer_locked(peer_ni);
1209 write_unlock_bh(global_lock);
1213 CERROR("Not creating conn %s type %d: %s\n",
1214 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1216 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1217 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1222 /* Request retry by replying with CONN_NONE
1223 * ksnc_proto has been set already
1225 conn->ksnc_type = SOCKLND_CONN_NONE;
1226 hello->kshm_nips = 0;
1227 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1230 write_lock_bh(global_lock);
1231 peer_ni->ksnp_accepting--;
1232 write_unlock_bh(global_lock);
1236 * If we get here without an error code, just use -EALREADY.
1237 * Depending on how we got here, the error may be positive
1238 * or negative. Normalize the value for ksocknal_txlist_done().
1240 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1241 ksocknal_txlist_done(ni, &zombies, rc2);
1242 ksocknal_peer_decref(peer_ni);
1246 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1247 kshm_ips[LNET_INTERFACES_NUM]));
1249 LIBCFS_FREE(conn, sizeof(*conn));
1258 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1260 /* This just does the immmediate housekeeping, and queues the
1261 * connection for the reaper to terminate.
1262 * Caller holds ksnd_global_lock exclusively in irq context */
1263 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1264 struct ksock_conn_cb *conn_cb;
1265 struct ksock_conn *conn2;
1267 int duplicate_count = 0;
1269 LASSERT(peer_ni->ksnp_error == 0);
1270 LASSERT(!conn->ksnc_closing);
1271 conn->ksnc_closing = 1;
1273 /* ksnd_deathrow_conns takes over peer_ni's ref */
1274 list_del(&conn->ksnc_list);
1276 conn_cb = conn->ksnc_conn_cb;
1277 if (conn_cb != NULL) {
1278 /* dissociate conn from cb... */
1279 LASSERT(!conn_cb->ksnr_deleted);
1281 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1283 /* connected bit is set only if all connections
1284 * of the given type got created
1286 if (conn_count == conn_cb->ksnr_max_conns)
1287 LASSERT((conn_cb->ksnr_connected &
1288 BIT(conn->ksnc_type)) != 0);
1290 if (conn_count == 1) {
1291 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1293 if (conn2->ksnc_conn_cb == conn_cb &&
1294 conn2->ksnc_type == conn->ksnc_type)
1295 duplicate_count += 1;
1297 if (duplicate_count > 0)
1298 CERROR("Found %d duplicate conns type %d\n",
1302 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1304 conn->ksnc_conn_cb = NULL;
1306 /* drop conn's ref on conn_cb */
1307 ksocknal_conn_cb_decref(conn_cb);
1310 if (list_empty(&peer_ni->ksnp_conns)) {
1311 /* No more connections to this peer_ni */
1313 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1314 struct ksock_tx *tx;
1316 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1318 /* throw them to the last connection...,
1319 * these TXs will be send to /dev/null by scheduler */
1320 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1322 ksocknal_tx_prep(conn, tx);
1324 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1325 list_splice_init(&peer_ni->ksnp_tx_queue,
1326 &conn->ksnc_tx_queue);
1327 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1330 /* renegotiate protocol version */
1331 peer_ni->ksnp_proto = NULL;
1332 /* stash last conn close reason */
1333 peer_ni->ksnp_error = error;
1335 if (peer_ni->ksnp_conn_cb == NULL) {
1336 /* I've just closed last conn belonging to a
1337 * peer_ni with no connections to it
1339 ksocknal_unlink_peer_locked(peer_ni);
1343 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1345 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1346 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1348 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1352 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1354 bool notify = false;
1355 time64_t last_alive = 0;
1357 /* There has been a connection failure or comms error; but I'll only
1358 * tell LNET I think the peer_ni is dead if it's to another kernel and
1359 * there are no connections or connection attempts in existence. */
1361 read_lock(&ksocknal_data.ksnd_global_lock);
1363 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1364 list_empty(&peer_ni->ksnp_conns) &&
1365 peer_ni->ksnp_accepting == 0 &&
1366 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1368 last_alive = peer_ni->ksnp_last_alive;
1371 read_unlock(&ksocknal_data.ksnd_global_lock);
1374 lnet_notify(peer_ni->ksnp_ni,
1375 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1376 false, false, last_alive);
1380 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1382 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1383 struct ksock_tx *tx;
1384 struct ksock_tx *tmp;
1387 /* NB safe to finalize TXs because closing of socket will
1388 * abort all buffered data */
1389 LASSERT(conn->ksnc_sock == NULL);
1391 spin_lock(&peer_ni->ksnp_lock);
1393 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1395 if (tx->tx_conn != conn)
1398 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1400 tx->tx_msg.ksm_zc_cookies[0] = 0;
1401 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1402 list_move(&tx->tx_zc_list, &zlist);
1405 spin_unlock(&peer_ni->ksnp_lock);
1407 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1408 tx_zc_list)) != NULL) {
1409 list_del(&tx->tx_zc_list);
1410 ksocknal_tx_decref(tx);
1415 ksocknal_terminate_conn(struct ksock_conn *conn)
1417 /* This gets called by the reaper (guaranteed thread context) to
1418 * disengage the socket from its callbacks and close it.
1419 * ksnc_refcount will eventually hit zero, and then the reaper will
1422 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1423 struct ksock_sched *sched = conn->ksnc_scheduler;
1424 bool failed = false;
1426 LASSERT(conn->ksnc_closing);
1428 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1429 spin_lock_bh(&sched->kss_lock);
1431 /* a closing conn is always ready to tx */
1432 conn->ksnc_tx_ready = 1;
1434 if (!conn->ksnc_tx_scheduled &&
1435 !list_empty(&conn->ksnc_tx_queue)) {
1436 list_add_tail(&conn->ksnc_tx_list,
1437 &sched->kss_tx_conns);
1438 conn->ksnc_tx_scheduled = 1;
1439 /* extra ref for scheduler */
1440 ksocknal_conn_addref(conn);
1442 wake_up(&sched->kss_waitq);
1445 spin_unlock_bh(&sched->kss_lock);
1447 /* serialise with callbacks */
1448 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1450 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1452 /* OK, so this conn may not be completely disengaged from its
1453 * scheduler yet, but it _has_ committed to terminate...
1455 conn->ksnc_scheduler->kss_nconns--;
1457 if (peer_ni->ksnp_error != 0) {
1458 /* peer_ni's last conn closed in error */
1459 LASSERT(list_empty(&peer_ni->ksnp_conns));
1461 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1464 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1467 ksocknal_peer_failed(peer_ni);
1469 /* The socket is closed on the final put; either here, or in
1470 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1471 * when the connection was established, this will close the socket
1472 * immediately, aborting anything buffered in it. Any hung
1473 * zero-copy transmits will therefore complete in finite time.
1475 ksocknal_connsock_decref(conn);
1479 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1481 /* Queue the conn for the reaper to destroy */
1482 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1483 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1485 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1486 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1488 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1492 ksocknal_destroy_conn(struct ksock_conn *conn)
1496 /* Final coup-de-grace of the reaper */
1497 CDEBUG(D_NET, "connection %p\n", conn);
1499 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1500 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1501 LASSERT(conn->ksnc_sock == NULL);
1502 LASSERT(conn->ksnc_conn_cb == NULL);
1503 LASSERT(!conn->ksnc_tx_scheduled);
1504 LASSERT(!conn->ksnc_rx_scheduled);
1505 LASSERT(list_empty(&conn->ksnc_tx_queue));
1507 /* complete current receive if any */
1508 switch (conn->ksnc_rx_state) {
1509 case SOCKNAL_RX_LNET_PAYLOAD:
1510 last_rcv = conn->ksnc_rx_deadline -
1512 CERROR("Completing partial receive from %s[%d], ip %pISp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1513 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1515 &conn->ksnc_peeraddr,
1516 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1517 ktime_get_seconds() - last_rcv);
1518 if (conn->ksnc_lnet_msg)
1519 conn->ksnc_lnet_msg->msg_health_status =
1520 LNET_MSG_STATUS_REMOTE_ERROR;
1521 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1523 case SOCKNAL_RX_LNET_HEADER:
1524 if (conn->ksnc_rx_started)
1525 CERROR("Incomplete receive of lnet header from %s, ip %pISp, with error, protocol: %d.x.\n",
1526 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1527 &conn->ksnc_peeraddr,
1528 conn->ksnc_proto->pro_version);
1530 case SOCKNAL_RX_KSM_HEADER:
1531 if (conn->ksnc_rx_started)
1532 CERROR("Incomplete receive of ksock message 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_SLOP:
1538 if (conn->ksnc_rx_started)
1539 CERROR("Incomplete receive of slops from %s, ip %pISp, with error\n",
1540 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1541 &conn->ksnc_peeraddr);
1548 ksocknal_peer_decref(conn->ksnc_peer);
1550 LIBCFS_FREE(conn, sizeof(*conn));
1554 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1555 struct sockaddr *addr, int why)
1557 struct ksock_conn *conn;
1558 struct ksock_conn *cnxt;
1561 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1564 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1566 ksocknal_close_conn_locked(conn, why);
1574 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1576 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1579 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1581 count = ksocknal_close_peer_conns_locked(
1582 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1584 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1590 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1592 struct ksock_peer_ni *peer_ni;
1593 struct hlist_node *pnxt;
1598 struct sockaddr_in sa = {.sin_family = AF_INET};
1600 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1602 if (!LNET_NID_IS_ANY(&id->nid)) {
1603 lo = hash_min(nidhash(&id->nid),
1604 HASH_BITS(ksocknal_data.ksnd_peers));
1608 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1611 sa.sin_addr.s_addr = htonl(ipaddr);
1612 for (i = lo; i <= hi; i++) {
1613 hlist_for_each_entry_safe(peer_ni, pnxt,
1614 &ksocknal_data.ksnd_peers[i],
1617 if (!((LNET_NID_IS_ANY(&id->nid) ||
1618 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1619 (id->pid == LNET_PID_ANY ||
1620 id->pid == peer_ni->ksnp_id.pid)))
1623 count += ksocknal_close_peer_conns_locked(
1625 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1629 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1631 /* wildcards always succeed */
1632 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1636 return (count == 0 ? -ENOENT : 0);
1640 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1642 /* The router is telling me she's been notified of a change in
1645 struct lnet_processid id = {
1646 .pid = LNET_PID_ANY,
1650 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1652 /* If the gateway crashed, close all open connections... */
1653 ksocknal_close_matching_conns(&id, 0);
1656 /* We can only establish new connections
1657 * if we have autroutes, and these connect on demand.
1662 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1666 struct ksock_conn *conn;
1668 for (index = 0; ; index++) {
1669 read_lock(&ksocknal_data.ksnd_global_lock);
1674 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1676 ksocknal_conn_addref(conn);
1681 read_unlock(&ksocknal_data.ksnd_global_lock);
1686 ksocknal_lib_push_conn (conn);
1687 ksocknal_conn_decref(conn);
1692 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1699 if (!LNET_NID_IS_ANY(&id->nid)) {
1700 lo = hash_min(nidhash(&id->nid),
1701 HASH_BITS(ksocknal_data.ksnd_peers));
1705 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1708 for (bkt = lo; bkt <= hi; bkt++) {
1709 int peer_off; /* searching offset in peer_ni hash table */
1711 for (peer_off = 0; ; peer_off++) {
1712 struct ksock_peer_ni *peer_ni;
1715 read_lock(&ksocknal_data.ksnd_global_lock);
1716 hlist_for_each_entry(peer_ni,
1717 &ksocknal_data.ksnd_peers[bkt],
1719 if (!((LNET_NID_IS_ANY(&id->nid) ||
1721 &peer_ni->ksnp_id.nid)) &&
1722 (id->pid == LNET_PID_ANY ||
1723 id->pid == peer_ni->ksnp_id.pid)))
1726 if (i++ == peer_off) {
1727 ksocknal_peer_addref(peer_ni);
1731 read_unlock(&ksocknal_data.ksnd_global_lock);
1733 if (i <= peer_off) /* no match */
1737 ksocknal_push_peer(peer_ni);
1738 ksocknal_peer_decref(peer_ni);
1745 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1747 struct lnet_processid id = {};
1748 struct libcfs_ioctl_data *data = arg;
1752 case IOC_LIBCFS_GET_INTERFACE: {
1753 struct ksock_net *net = ni->ni_data;
1754 struct ksock_interface *iface;
1755 struct sockaddr_in *sa;
1757 read_lock(&ksocknal_data.ksnd_global_lock);
1759 if (data->ioc_count >= 1) {
1763 iface = &net->ksnn_interface;
1765 sa = (void *)&iface->ksni_addr;
1766 if (sa->sin_family == AF_INET)
1767 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1769 data->ioc_u32[0] = 0xFFFFFFFF;
1770 data->ioc_u32[1] = iface->ksni_netmask;
1771 data->ioc_u32[2] = iface->ksni_npeers;
1772 data->ioc_u32[3] = iface->ksni_nroutes;
1775 read_unlock(&ksocknal_data.ksnd_global_lock);
1779 case IOC_LIBCFS_GET_PEER: {
1784 int share_count = 0;
1786 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1787 &id, &myip, &ip, &port,
1788 &conn_count, &share_count);
1791 if (!nid_is_nid4(&id.nid))
1793 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
1794 data->ioc_count = share_count;
1795 data->ioc_u32[0] = ip;
1796 data->ioc_u32[1] = port;
1797 data->ioc_u32[2] = myip;
1798 data->ioc_u32[3] = conn_count;
1799 data->ioc_u32[4] = id.pid;
1803 case IOC_LIBCFS_ADD_PEER: {
1804 struct sockaddr_in sa = {.sin_family = AF_INET};
1806 id.pid = LNET_PID_LUSTRE;
1807 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1808 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1809 sa.sin_port = htons(data->ioc_u32[1]);
1810 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1812 case IOC_LIBCFS_DEL_PEER:
1813 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1814 id.pid = LNET_PID_ANY;
1815 return ksocknal_del_peer(ni, &id);
1817 case IOC_LIBCFS_GET_CONN: {
1821 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1822 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1823 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1828 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1830 data->ioc_count = txmem;
1831 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1832 data->ioc_flags = nagle;
1833 if (psa->sin_family == AF_INET)
1834 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1836 data->ioc_u32[0] = 0xFFFFFFFF;
1837 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1838 &conn->ksnc_peeraddr);
1839 if (mysa->sin_family == AF_INET)
1840 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1842 data->ioc_u32[2] = 0xFFFFFFFF;
1843 data->ioc_u32[3] = conn->ksnc_type;
1844 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1845 data->ioc_u32[5] = rxmem;
1846 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1847 ksocknal_conn_decref(conn);
1851 case IOC_LIBCFS_CLOSE_CONNECTION:
1852 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1853 id.pid = LNET_PID_ANY;
1854 return ksocknal_close_matching_conns(&id,
1857 case IOC_LIBCFS_REGISTER_MYNID:
1858 /* Ignore if this is a noop */
1859 if (nid_is_nid4(&ni->ni_nid) &&
1860 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1863 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1864 libcfs_nid2str(data->ioc_nid),
1865 libcfs_nidstr(&ni->ni_nid));
1868 case IOC_LIBCFS_PUSH_CONNECTION:
1869 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1870 id.pid = LNET_PID_ANY;
1871 return ksocknal_push(ni, &id);
1880 ksocknal_free_buffers (void)
1882 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1884 if (ksocknal_data.ksnd_schedulers != NULL)
1885 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1887 spin_lock(&ksocknal_data.ksnd_tx_lock);
1889 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1891 struct ksock_tx *tx;
1893 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1894 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1896 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1897 tx_list)) != NULL) {
1898 list_del(&tx->tx_list);
1899 LIBCFS_FREE(tx, tx->tx_desc_size);
1902 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1906 static int ksocknal_get_link_status(struct net_device *dev)
1912 if (!netif_running(dev)) {
1914 CDEBUG(D_NET, "device not running\n");
1916 /* Some devices may not be providing link settings */
1917 else if (dev->ethtool_ops->get_link) {
1918 ret = dev->ethtool_ops->get_link(dev);
1919 CDEBUG(D_NET, "get_link returns %u\n", ret);
1926 ksocknal_handle_link_state_change(struct net_device *dev,
1927 unsigned char operstate)
1929 struct lnet_ni *ni = NULL;
1930 struct ksock_net *net;
1931 struct ksock_net *cnxt;
1933 unsigned char link_down = !(operstate == IF_OPER_UP);
1934 struct in_device *in_dev;
1935 bool found_ip = false;
1936 struct ksock_interface *ksi = NULL;
1937 struct sockaddr_in *sa;
1938 DECLARE_CONST_IN_IFADDR(ifa);
1940 ifindex = dev->ifindex;
1942 if (!ksocknal_data.ksnd_nnets)
1945 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1948 ksi = &net->ksnn_interface;
1949 sa = (void *)&ksi->ksni_addr;
1952 if (ksi->ksni_index != ifindex ||
1953 strcmp(ksi->ksni_name, dev->name))
1958 in_dev = __in_dev_get_rtnl(dev);
1960 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1962 CDEBUG(D_NET, "set link fatal state to 1\n");
1963 atomic_set(&ni->ni_fatal_error_on, 1);
1966 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1967 if (sa->sin_addr.s_addr == ifa->ifa_local)
1973 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1975 CDEBUG(D_NET, "set link fatal state to 1\n");
1976 atomic_set(&ni->ni_fatal_error_on, 1);
1981 CDEBUG(D_NET, "set link fatal state to 1\n");
1982 atomic_set(&ni->ni_fatal_error_on, link_down);
1984 CDEBUG(D_NET, "set link fatal state to %u\n",
1985 (ksocknal_get_link_status(dev) == 0));
1986 atomic_set(&ni->ni_fatal_error_on,
1987 (ksocknal_get_link_status(dev) == 0));
1996 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
1999 struct ksock_net *net;
2000 struct ksock_net *cnxt;
2001 struct net_device *event_netdev = ifa->ifa_dev->dev;
2003 struct ksock_interface *ksi = NULL;
2004 struct sockaddr_in *sa;
2006 if (!ksocknal_data.ksnd_nnets)
2009 ifindex = event_netdev->ifindex;
2011 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2014 ksi = &net->ksnn_interface;
2015 sa = (void *)&ksi->ksni_addr;
2017 if (ksi->ksni_index != ifindex ||
2018 strcmp(ksi->ksni_name, event_netdev->name))
2021 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2022 CDEBUG(D_NET, "set link fatal state to %u\n",
2023 (event == NETDEV_DOWN));
2025 atomic_set(&ni->ni_fatal_error_on,
2026 (event == NETDEV_DOWN));
2033 /************************************
2034 * Net device notifier event handler
2035 ************************************/
2036 static int ksocknal_device_event(struct notifier_block *unused,
2037 unsigned long event, void *ptr)
2039 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2040 unsigned char operstate;
2042 operstate = dev->operstate;
2044 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2045 event, dev->name, dev->ifindex, operstate);
2051 ksocknal_handle_link_state_change(dev, operstate);
2058 /************************************
2059 * Inetaddr notifier event handler
2060 ************************************/
2061 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2062 unsigned long event, void *ptr)
2064 struct in_ifaddr *ifa = ptr;
2066 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2067 event, &ifa->ifa_address, &ifa->ifa_mask);
2073 ksocknal_handle_inetaddr_change(ifa, event);
2080 static struct notifier_block ksocknal_dev_notifier_block = {
2081 .notifier_call = ksocknal_device_event,
2084 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2085 .notifier_call = ksocknal_inetaddr_event,
2089 ksocknal_base_shutdown(void)
2091 struct ksock_sched *sched;
2092 struct ksock_peer_ni *peer_ni;
2095 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2096 libcfs_kmem_read());
2097 LASSERT (ksocknal_data.ksnd_nnets == 0);
2099 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2100 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2101 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2104 switch (ksocknal_data.ksnd_init) {
2109 case SOCKNAL_INIT_ALL:
2110 case SOCKNAL_INIT_DATA:
2111 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2114 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2115 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2116 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2117 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2118 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2120 if (ksocknal_data.ksnd_schedulers != NULL) {
2121 cfs_percpt_for_each(sched, i,
2122 ksocknal_data.ksnd_schedulers) {
2124 LASSERT(list_empty(&sched->kss_tx_conns));
2125 LASSERT(list_empty(&sched->kss_rx_conns));
2126 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2127 LASSERT(sched->kss_nconns == 0);
2131 /* flag threads to terminate; wake and wait for them to die */
2132 ksocknal_data.ksnd_shuttingdown = 1;
2133 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2134 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2136 if (ksocknal_data.ksnd_schedulers != NULL) {
2137 cfs_percpt_for_each(sched, i,
2138 ksocknal_data.ksnd_schedulers)
2139 wake_up_all(&sched->kss_waitq);
2142 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2143 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2144 "waiting for %d threads to terminate\n",
2145 atomic_read(&ksocknal_data.ksnd_nthreads));
2147 ksocknal_free_buffers();
2149 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2153 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2154 libcfs_kmem_read());
2156 module_put(THIS_MODULE);
2160 ksocknal_base_startup(void)
2162 struct ksock_sched *sched;
2166 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2167 LASSERT(ksocknal_data.ksnd_nnets == 0);
2169 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2171 hash_init(ksocknal_data.ksnd_peers);
2173 rwlock_init(&ksocknal_data.ksnd_global_lock);
2174 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2176 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2177 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2178 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2179 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2180 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2182 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2183 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2184 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2185 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2187 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2188 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2190 /* NB memset above zeros whole of ksocknal_data */
2192 /* flag lists/ptrs/locks initialised */
2193 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2194 if (!try_module_get(THIS_MODULE))
2197 /* Create a scheduler block per available CPT */
2198 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2200 if (ksocknal_data.ksnd_schedulers == NULL)
2203 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2207 * make sure not to allocate more threads than there are
2208 * cores/CPUs in teh CPT
2210 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2211 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2212 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2215 * max to half of CPUs, assume another half should be
2216 * reserved for upper layer modules
2218 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2221 sched->kss_nthreads_max = nthrs;
2224 spin_lock_init(&sched->kss_lock);
2225 INIT_LIST_HEAD(&sched->kss_rx_conns);
2226 INIT_LIST_HEAD(&sched->kss_tx_conns);
2227 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2228 init_waitqueue_head(&sched->kss_waitq);
2231 ksocknal_data.ksnd_connd_starting = 0;
2232 ksocknal_data.ksnd_connd_failed_stamp = 0;
2233 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2234 /* must have at least 2 connds to remain responsive to accepts while
2236 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2237 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2239 if (*ksocknal_tunables.ksnd_nconnds_max <
2240 *ksocknal_tunables.ksnd_nconnds) {
2241 ksocknal_tunables.ksnd_nconnds_max =
2242 ksocknal_tunables.ksnd_nconnds;
2245 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2246 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2247 ksocknal_data.ksnd_connd_starting++;
2248 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2250 rc = ksocknal_thread_start(ksocknal_connd,
2251 (void *)((uintptr_t)i),
2252 "socknal_cd%02d", 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);
2257 CERROR("Can't spawn socknal connd: %d\n", rc);
2262 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2264 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2268 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2269 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2271 /* flag everything initialised */
2272 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2277 ksocknal_base_shutdown();
2282 ksocknal_debug_peerhash(struct lnet_ni *ni)
2284 struct ksock_peer_ni *peer_ni;
2287 read_lock(&ksocknal_data.ksnd_global_lock);
2289 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2290 struct ksock_conn_cb *conn_cb;
2291 struct ksock_conn *conn;
2293 if (peer_ni->ksnp_ni != ni)
2296 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2297 libcfs_idstr(&peer_ni->ksnp_id),
2298 refcount_read(&peer_ni->ksnp_refcount),
2299 peer_ni->ksnp_closing,
2300 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2301 peer_ni->ksnp_zc_next_cookie,
2302 !list_empty(&peer_ni->ksnp_tx_queue),
2303 !list_empty(&peer_ni->ksnp_zc_req_list));
2305 conn_cb = peer_ni->ksnp_conn_cb;
2307 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2308 refcount_read(&conn_cb->ksnr_refcount),
2309 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2310 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2313 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2314 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2315 refcount_read(&conn->ksnc_conn_refcount),
2316 refcount_read(&conn->ksnc_sock_refcount),
2317 conn->ksnc_type, conn->ksnc_closing);
2322 read_unlock(&ksocknal_data.ksnd_global_lock);
2327 ksocknal_shutdown(struct lnet_ni *ni)
2329 struct ksock_net *net = ni->ni_data;
2331 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2332 LASSERT(ksocknal_data.ksnd_nnets > 0);
2334 /* prevent new peers */
2335 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2337 /* Delete all peers */
2338 ksocknal_del_peer(ni, NULL);
2340 /* Wait for all peer_ni state to clean up */
2341 wait_var_event_warning(&net->ksnn_npeers,
2342 atomic_read(&net->ksnn_npeers) ==
2343 SOCKNAL_SHUTDOWN_BIAS,
2344 "waiting for %d peers to disconnect\n",
2345 ksocknal_debug_peerhash(ni) +
2346 atomic_read(&net->ksnn_npeers) -
2347 SOCKNAL_SHUTDOWN_BIAS);
2349 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2350 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2352 list_del(&net->ksnn_list);
2353 LIBCFS_FREE(net, sizeof(*net));
2355 ksocknal_data.ksnd_nnets--;
2356 if (ksocknal_data.ksnd_nnets == 0)
2357 ksocknal_base_shutdown();
2361 ksocknal_search_new_ipif(struct ksock_net *net)
2364 char *ifnam = &net->ksnn_interface.ksni_name[0];
2365 char *colon = strchr(ifnam, ':');
2367 struct ksock_net *tmp;
2372 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2373 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2374 char *colon2 = strchr(ifnam2, ':');
2379 found = strcmp(ifnam, ifnam2) == 0;
2392 ksocknal_start_schedulers(struct ksock_sched *sched)
2398 if (sched->kss_nthreads == 0) {
2399 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2400 nthrs = sched->kss_nthreads_max;
2402 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2404 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2405 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2407 nthrs = min(nthrs, sched->kss_nthreads_max);
2409 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2410 /* increase two threads if there is new interface */
2411 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2414 for (i = 0; i < nthrs; i++) {
2417 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2418 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2419 "socknal_sd%02d_%02d",
2421 (int)KSOCK_THREAD_SID(id));
2425 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2426 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2430 sched->kss_nthreads += i;
2435 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2437 int newif = ksocknal_search_new_ipif(net);
2441 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2444 for (i = 0; i < ncpts; i++) {
2445 struct ksock_sched *sched;
2446 int cpt = (cpts == NULL) ? i : cpts[i];
2448 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2449 sched = ksocknal_data.ksnd_schedulers[cpt];
2451 if (!newif && sched->kss_nthreads > 0)
2454 rc = ksocknal_start_schedulers(sched);
2462 ksocknal_startup(struct lnet_ni *ni)
2464 struct ksock_net *net;
2465 struct ksock_interface *ksi = NULL;
2466 struct lnet_inetdev *ifaces = NULL;
2467 struct sockaddr_in *sa;
2471 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2472 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2473 rc = ksocknal_base_startup();
2477 LIBCFS_ALLOC(net, sizeof(*net));
2480 net->ksnn_incarnation = ktime_get_real_ns();
2483 ksocknal_tunables_setup(ni);
2485 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2489 ksi = &net->ksnn_interface;
2491 /* Use the first discovered interface or look in the list */
2492 if (ni->ni_interface) {
2493 for (i = 0; i < rc; i++)
2494 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2497 /* ni_interfaces doesn't contain the interface we want */
2499 CERROR("ksocklnd: failed to find interface %s\n",
2505 ni->ni_dev_cpt = ifaces[i].li_cpt;
2506 sa = (void *)&ksi->ksni_addr;
2507 memset(sa, 0, sizeof(*sa));
2508 sa->sin_family = AF_INET;
2509 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2510 ksi->ksni_index = ksocknal_ip2index((struct sockaddr *)sa, ni);
2511 ksi->ksni_netmask = ifaces[i].li_netmask;
2512 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2514 /* call it before add it to ksocknal_data.ksnd_nets */
2515 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2520 LASSERT(ksi->ksni_addr.ss_family == AF_INET);
2521 ni->ni_nid.nid_addr[0] =
2522 ((struct sockaddr_in *)&ksi->ksni_addr)->sin_addr.s_addr;
2523 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2525 ksocknal_data.ksnd_nnets++;
2530 LIBCFS_FREE(net, sizeof(*net));
2532 if (ksocknal_data.ksnd_nnets == 0)
2533 ksocknal_base_shutdown();
2538 static void __exit ksocklnd_exit(void)
2540 lnet_unregister_lnd(&the_ksocklnd);
2543 static const struct lnet_lnd the_ksocklnd = {
2544 .lnd_type = SOCKLND,
2545 .lnd_startup = ksocknal_startup,
2546 .lnd_shutdown = ksocknal_shutdown,
2547 .lnd_ctl = ksocknal_ctl,
2548 .lnd_send = ksocknal_send,
2549 .lnd_recv = ksocknal_recv,
2550 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2551 .lnd_accept = ksocknal_accept,
2554 static int __init ksocklnd_init(void)
2558 /* check ksnr_connected/connecting field large enough */
2559 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2560 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2562 rc = ksocknal_tunables_init();
2566 lnet_register_lnd(&the_ksocklnd);
2571 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2572 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2573 MODULE_VERSION("2.8.0");
2574 MODULE_LICENSE("GPL");
2576 module_init(ksocklnd_init);
2577 module_exit(ksocklnd_exit);