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, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/socklnd/socklnd.c
34 * Author: Zach Brown <zab@zabbo.net>
35 * Author: Peter J. Braam <braam@clusterfs.com>
36 * Author: Phil Schwan <phil@clusterfs.com>
37 * Author: Eric Barton <eric@bartonsoftware.com>
40 #include <linux/pci.h>
43 static struct lnet_lnd the_ksocklnd;
44 ksock_nal_data_t ksocknal_data;
46 static ksock_interface_t *
47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
49 ksock_net_t *net = ni->ni_data;
51 ksock_interface_t *iface;
53 for (i = 0; i < net->ksnn_ninterfaces; i++) {
54 LASSERT(i < LNET_MAX_INTERFACES);
55 iface = &net->ksnn_interfaces[i];
57 if (iface->ksni_ipaddr == ip)
64 static ksock_route_t *
65 ksocknal_create_route (__u32 ipaddr, int port)
69 LIBCFS_ALLOC (route, sizeof (*route));
73 atomic_set (&route->ksnr_refcount, 1);
74 route->ksnr_peer = NULL;
75 route->ksnr_retry_interval = 0; /* OK to connect at any time */
76 route->ksnr_ipaddr = ipaddr;
77 route->ksnr_port = port;
78 route->ksnr_scheduled = 0;
79 route->ksnr_connecting = 0;
80 route->ksnr_connected = 0;
81 route->ksnr_deleted = 0;
82 route->ksnr_conn_count = 0;
83 route->ksnr_share_count = 0;
89 ksocknal_destroy_route (ksock_route_t *route)
91 LASSERT (atomic_read(&route->ksnr_refcount) == 0);
93 if (route->ksnr_peer != NULL)
94 ksocknal_peer_decref(route->ksnr_peer);
96 LIBCFS_FREE (route, sizeof (*route));
100 ksocknal_create_peer(ksock_peer_ni_t **peerp, struct lnet_ni *ni,
101 struct lnet_process_id id)
103 int cpt = lnet_cpt_of_nid(id.nid, ni);
104 ksock_net_t *net = ni->ni_data;
105 ksock_peer_ni_t *peer_ni;
107 LASSERT(id.nid != LNET_NID_ANY);
108 LASSERT(id.pid != LNET_PID_ANY);
109 LASSERT(!in_interrupt());
111 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
115 peer_ni->ksnp_ni = ni;
116 peer_ni->ksnp_id = id;
117 atomic_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
118 peer_ni->ksnp_closing = 0;
119 peer_ni->ksnp_accepting = 0;
120 peer_ni->ksnp_proto = NULL;
121 peer_ni->ksnp_last_alive = 0;
122 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
124 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
125 INIT_LIST_HEAD(&peer_ni->ksnp_routes);
126 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
127 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
128 spin_lock_init(&peer_ni->ksnp_lock);
130 spin_lock_bh(&net->ksnn_lock);
132 if (net->ksnn_shutdown) {
133 spin_unlock_bh(&net->ksnn_lock);
135 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
136 CERROR("Can't create peer_ni: network shutdown\n");
142 spin_unlock_bh(&net->ksnn_lock);
149 ksocknal_destroy_peer (ksock_peer_ni_t *peer_ni)
151 ksock_net_t *net = peer_ni->ksnp_ni->ni_data;
153 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
154 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
156 LASSERT(atomic_read(&peer_ni->ksnp_refcount) == 0);
157 LASSERT(peer_ni->ksnp_accepting == 0);
158 LASSERT(list_empty(&peer_ni->ksnp_conns));
159 LASSERT(list_empty(&peer_ni->ksnp_routes));
160 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
161 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
163 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
165 /* NB a peer_ni's connections and routes keep a reference on their peer_ni
166 * until they are destroyed, so we can be assured that _all_ state to
167 * do with this peer_ni has been cleaned up when its refcount drops to
169 spin_lock_bh(&net->ksnn_lock);
171 spin_unlock_bh(&net->ksnn_lock);
175 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
177 struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
178 struct list_head *tmp;
179 ksock_peer_ni_t *peer_ni;
181 list_for_each(tmp, peer_list) {
183 peer_ni = list_entry(tmp, ksock_peer_ni_t, ksnp_list);
185 LASSERT(!peer_ni->ksnp_closing);
187 if (peer_ni->ksnp_ni != ni)
190 if (peer_ni->ksnp_id.nid != id.nid ||
191 peer_ni->ksnp_id.pid != id.pid)
194 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
195 peer_ni, libcfs_id2str(id),
196 atomic_read(&peer_ni->ksnp_refcount));
203 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
205 ksock_peer_ni_t *peer_ni;
207 read_lock(&ksocknal_data.ksnd_global_lock);
208 peer_ni = ksocknal_find_peer_locked(ni, id);
209 if (peer_ni != NULL) /* +1 ref for caller? */
210 ksocknal_peer_addref(peer_ni);
211 read_unlock(&ksocknal_data.ksnd_global_lock);
217 ksocknal_unlink_peer_locked (ksock_peer_ni_t *peer_ni)
221 ksock_interface_t *iface;
223 for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
224 LASSERT (i < LNET_MAX_INTERFACES);
225 ip = peer_ni->ksnp_passive_ips[i];
227 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
228 /* All IPs in peer_ni->ksnp_passive_ips[] come from the
229 * interface list, therefore the call must succeed. */
230 LASSERT (iface != NULL);
232 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
233 peer_ni, iface, iface->ksni_nroutes);
234 iface->ksni_npeers--;
237 LASSERT(list_empty(&peer_ni->ksnp_conns));
238 LASSERT(list_empty(&peer_ni->ksnp_routes));
239 LASSERT(!peer_ni->ksnp_closing);
240 peer_ni->ksnp_closing = 1;
241 list_del(&peer_ni->ksnp_list);
242 /* lose peerlist's ref */
243 ksocknal_peer_decref(peer_ni);
247 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
248 struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
249 int *port, int *conn_count, int *share_count)
251 ksock_peer_ni_t *peer_ni;
252 struct list_head *ptmp;
253 ksock_route_t *route;
254 struct list_head *rtmp;
259 read_lock(&ksocknal_data.ksnd_global_lock);
261 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
262 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
263 peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
265 if (peer_ni->ksnp_ni != ni)
268 if (peer_ni->ksnp_n_passive_ips == 0 &&
269 list_empty(&peer_ni->ksnp_routes)) {
273 *id = peer_ni->ksnp_id;
283 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
287 *id = peer_ni->ksnp_id;
288 *myip = peer_ni->ksnp_passive_ips[j];
297 list_for_each(rtmp, &peer_ni->ksnp_routes) {
301 route = list_entry(rtmp, ksock_route_t,
304 *id = peer_ni->ksnp_id;
305 *myip = route->ksnr_myipaddr;
306 *peer_ip = route->ksnr_ipaddr;
307 *port = route->ksnr_port;
308 *conn_count = route->ksnr_conn_count;
309 *share_count = route->ksnr_share_count;
316 read_unlock(&ksocknal_data.ksnd_global_lock);
321 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
323 ksock_peer_ni_t *peer_ni = route->ksnr_peer;
324 int type = conn->ksnc_type;
325 ksock_interface_t *iface;
327 conn->ksnc_route = route;
328 ksocknal_route_addref(route);
330 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
331 if (route->ksnr_myipaddr == 0) {
332 /* route wasn't bound locally yet (the initial route) */
333 CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
334 libcfs_id2str(peer_ni->ksnp_id),
336 &conn->ksnc_myipaddr);
338 CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
339 "to %pI4h\n", libcfs_id2str(peer_ni->ksnp_id),
341 &route->ksnr_myipaddr,
342 &conn->ksnc_myipaddr);
344 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
345 route->ksnr_myipaddr);
347 iface->ksni_nroutes--;
349 route->ksnr_myipaddr = conn->ksnc_myipaddr;
350 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
351 route->ksnr_myipaddr);
353 iface->ksni_nroutes++;
356 route->ksnr_connected |= (1<<type);
357 route->ksnr_conn_count++;
359 /* Successful connection => further attempts can
360 * proceed immediately */
361 route->ksnr_retry_interval = 0;
365 ksocknal_add_route_locked (ksock_peer_ni_t *peer_ni, ksock_route_t *route)
367 struct list_head *tmp;
369 ksock_route_t *route2;
371 LASSERT(!peer_ni->ksnp_closing);
372 LASSERT(route->ksnr_peer == NULL);
373 LASSERT(!route->ksnr_scheduled);
374 LASSERT(!route->ksnr_connecting);
375 LASSERT(route->ksnr_connected == 0);
377 /* LASSERT(unique) */
378 list_for_each(tmp, &peer_ni->ksnp_routes) {
379 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
381 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
382 CERROR("Duplicate route %s %pI4h\n",
383 libcfs_id2str(peer_ni->ksnp_id),
384 &route->ksnr_ipaddr);
389 route->ksnr_peer = peer_ni;
390 ksocknal_peer_addref(peer_ni);
391 /* peer_ni's routelist takes over my ref on 'route' */
392 list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
394 list_for_each(tmp, &peer_ni->ksnp_conns) {
395 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
397 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
400 ksocknal_associate_route_conn_locked(route, conn);
401 /* keep going (typed routes) */
406 ksocknal_del_route_locked (ksock_route_t *route)
408 ksock_peer_ni_t *peer_ni = route->ksnr_peer;
409 ksock_interface_t *iface;
411 struct list_head *ctmp;
412 struct list_head *cnxt;
414 LASSERT(!route->ksnr_deleted);
416 /* Close associated conns */
417 list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
418 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
420 if (conn->ksnc_route != route)
423 ksocknal_close_conn_locked(conn, 0);
426 if (route->ksnr_myipaddr != 0) {
427 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
428 route->ksnr_myipaddr);
430 iface->ksni_nroutes--;
433 route->ksnr_deleted = 1;
434 list_del(&route->ksnr_list);
435 ksocknal_route_decref(route); /* drop peer_ni's ref */
437 if (list_empty(&peer_ni->ksnp_routes) &&
438 list_empty(&peer_ni->ksnp_conns)) {
439 /* I've just removed the last route to a peer_ni with no active
441 ksocknal_unlink_peer_locked(peer_ni);
446 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
449 struct list_head *tmp;
450 ksock_peer_ni_t *peer_ni;
451 ksock_peer_ni_t *peer2;
452 ksock_route_t *route;
453 ksock_route_t *route2;
456 if (id.nid == LNET_NID_ANY ||
457 id.pid == LNET_PID_ANY)
460 /* Have a brand new peer_ni ready... */
461 rc = ksocknal_create_peer(&peer_ni, ni, id);
465 route = ksocknal_create_route (ipaddr, port);
467 ksocknal_peer_decref(peer_ni);
471 write_lock_bh(&ksocknal_data.ksnd_global_lock);
473 /* always called with a ref on ni, so shutdown can't have started */
474 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
476 peer2 = ksocknal_find_peer_locked(ni, id);
478 ksocknal_peer_decref(peer_ni);
481 /* peer_ni table takes my ref on peer_ni */
482 list_add_tail(&peer_ni->ksnp_list,
483 ksocknal_nid2peerlist(id.nid));
487 list_for_each(tmp, &peer_ni->ksnp_routes) {
488 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
490 if (route2->ksnr_ipaddr == ipaddr)
495 if (route2 == NULL) {
496 ksocknal_add_route_locked(peer_ni, route);
497 route->ksnr_share_count++;
499 ksocknal_route_decref(route);
500 route2->ksnr_share_count++;
503 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
509 ksocknal_del_peer_locked (ksock_peer_ni_t *peer_ni, __u32 ip)
512 ksock_route_t *route;
513 struct list_head *tmp;
514 struct list_head *nxt;
517 LASSERT(!peer_ni->ksnp_closing);
519 /* Extra ref prevents peer_ni disappearing until I'm done with it */
520 ksocknal_peer_addref(peer_ni);
522 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
523 route = list_entry(tmp, ksock_route_t, ksnr_list);
526 if (!(ip == 0 || route->ksnr_ipaddr == ip))
529 route->ksnr_share_count = 0;
530 /* This deletes associated conns too */
531 ksocknal_del_route_locked(route);
535 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
536 route = list_entry(tmp, ksock_route_t, ksnr_list);
537 nshared += route->ksnr_share_count;
541 /* remove everything else if there are no explicit entries
544 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
545 route = list_entry(tmp, ksock_route_t, ksnr_list);
547 /* we should only be removing auto-entries */
548 LASSERT(route->ksnr_share_count == 0);
549 ksocknal_del_route_locked(route);
552 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
553 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
555 ksocknal_close_conn_locked(conn, 0);
559 ksocknal_peer_decref(peer_ni);
560 /* NB peer_ni unlinks itself when last conn/route is removed */
564 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
566 struct list_head zombies = LIST_HEAD_INIT(zombies);
567 struct list_head *ptmp;
568 struct list_head *pnxt;
569 ksock_peer_ni_t *peer_ni;
575 write_lock_bh(&ksocknal_data.ksnd_global_lock);
577 if (id.nid != LNET_NID_ANY) {
578 hi = (int)(ksocknal_nid2peerlist(id.nid) -
579 ksocknal_data.ksnd_peers);
583 hi = ksocknal_data.ksnd_peer_hash_size - 1;
586 for (i = lo; i <= hi; i++) {
587 list_for_each_safe(ptmp, pnxt,
588 &ksocknal_data.ksnd_peers[i]) {
589 peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
591 if (peer_ni->ksnp_ni != ni)
594 if (!((id.nid == LNET_NID_ANY ||
595 peer_ni->ksnp_id.nid == id.nid) &&
596 (id.pid == LNET_PID_ANY ||
597 peer_ni->ksnp_id.pid == id.pid)))
600 ksocknal_peer_addref(peer_ni); /* a ref for me... */
602 ksocknal_del_peer_locked(peer_ni, ip);
604 if (peer_ni->ksnp_closing &&
605 !list_empty(&peer_ni->ksnp_tx_queue)) {
606 LASSERT(list_empty(&peer_ni->ksnp_conns));
607 LASSERT(list_empty(&peer_ni->ksnp_routes));
609 list_splice_init(&peer_ni->ksnp_tx_queue,
613 ksocknal_peer_decref(peer_ni); /* ...till here */
615 rc = 0; /* matched! */
619 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
621 ksocknal_txlist_done(ni, &zombies, 1);
626 static ksock_conn_t *
627 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
629 ksock_peer_ni_t *peer_ni;
630 struct list_head *ptmp;
632 struct list_head *ctmp;
635 read_lock(&ksocknal_data.ksnd_global_lock);
637 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
638 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
639 peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
641 LASSERT(!peer_ni->ksnp_closing);
643 if (peer_ni->ksnp_ni != ni)
646 list_for_each(ctmp, &peer_ni->ksnp_conns) {
650 conn = list_entry(ctmp, ksock_conn_t,
652 ksocknal_conn_addref(conn);
653 read_unlock(&ksocknal_data. \
660 read_unlock(&ksocknal_data.ksnd_global_lock);
664 static ksock_sched_t *
665 ksocknal_choose_scheduler_locked(unsigned int cpt)
667 struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
668 ksock_sched_t *sched;
671 LASSERT(info->ksi_nthreads > 0);
673 sched = &info->ksi_scheds[0];
675 * NB: it's safe so far, but info->ksi_nthreads could be changed
676 * at runtime when we have dynamic LNet configuration, then we
677 * need to take care of this.
679 for (i = 1; i < info->ksi_nthreads; i++) {
680 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
681 sched = &info->ksi_scheds[i];
688 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
690 ksock_net_t *net = ni->ni_data;
694 read_lock(&ksocknal_data.ksnd_global_lock);
696 nip = net->ksnn_ninterfaces;
697 LASSERT (nip <= LNET_MAX_INTERFACES);
699 /* Only offer interfaces for additional connections if I have
702 read_unlock(&ksocknal_data.ksnd_global_lock);
706 for (i = 0; i < nip; i++) {
707 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
708 LASSERT (ipaddrs[i] != 0);
711 read_unlock(&ksocknal_data.ksnd_global_lock);
716 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
718 int best_netmatch = 0;
725 for (i = 0; i < nips; i++) {
729 this_xor = (ips[i] ^ iface->ksni_ipaddr);
730 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
733 best_netmatch < this_netmatch ||
734 (best_netmatch == this_netmatch &&
735 best_xor > this_xor)))
739 best_netmatch = this_netmatch;
748 ksocknal_select_ips(ksock_peer_ni_t *peer_ni, __u32 *peerips, int n_peerips)
750 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
751 ksock_net_t *net = peer_ni->ksnp_ni->ni_data;
752 ksock_interface_t *iface;
753 ksock_interface_t *best_iface;
764 /* CAVEAT EMPTOR: We do all our interface matching with an
765 * exclusive hold of global lock at IRQ priority. We're only
766 * expecting to be dealing with small numbers of interfaces, so the
767 * O(n**3)-ness shouldn't matter */
769 /* Also note that I'm not going to return more than n_peerips
770 * interfaces, even if I have more myself */
772 write_lock_bh(global_lock);
774 LASSERT (n_peerips <= LNET_MAX_INTERFACES);
775 LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
777 /* Only match interfaces for additional connections
778 * if I have > 1 interface */
779 n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
780 MIN(n_peerips, net->ksnn_ninterfaces);
782 for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
783 /* ^ yes really... */
785 /* If we have any new interfaces, first tick off all the
786 * peer_ni IPs that match old interfaces, then choose new
787 * interfaces to match the remaining peer_ni IPS.
788 * We don't forget interfaces we've stopped using; we might
789 * start using them again... */
791 if (i < peer_ni->ksnp_n_passive_ips) {
793 ip = peer_ni->ksnp_passive_ips[i];
794 best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
796 /* peer_ni passive ips are kept up to date */
797 LASSERT(best_iface != NULL);
799 /* choose a new interface */
800 LASSERT (i == peer_ni->ksnp_n_passive_ips);
806 for (j = 0; j < net->ksnn_ninterfaces; j++) {
807 iface = &net->ksnn_interfaces[j];
808 ip = iface->ksni_ipaddr;
810 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
811 if (peer_ni->ksnp_passive_ips[k] == ip)
814 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
817 k = ksocknal_match_peerip(iface, peerips, n_peerips);
818 xor = (ip ^ peerips[k]);
819 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
821 if (!(best_iface == NULL ||
822 best_netmatch < this_netmatch ||
823 (best_netmatch == this_netmatch &&
824 best_npeers > iface->ksni_npeers)))
828 best_netmatch = this_netmatch;
829 best_npeers = iface->ksni_npeers;
832 LASSERT(best_iface != NULL);
834 best_iface->ksni_npeers++;
835 ip = best_iface->ksni_ipaddr;
836 peer_ni->ksnp_passive_ips[i] = ip;
837 peer_ni->ksnp_n_passive_ips = i+1;
840 /* mark the best matching peer_ni IP used */
841 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
845 /* Overwrite input peer_ni IP addresses */
846 memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
848 write_unlock_bh(global_lock);
854 ksocknal_create_routes(ksock_peer_ni_t *peer_ni, int port,
855 __u32 *peer_ipaddrs, int npeer_ipaddrs)
857 ksock_route_t *newroute = NULL;
858 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
859 struct lnet_ni *ni = peer_ni->ksnp_ni;
860 ksock_net_t *net = ni->ni_data;
861 struct list_head *rtmp;
862 ksock_route_t *route;
863 ksock_interface_t *iface;
864 ksock_interface_t *best_iface;
871 /* CAVEAT EMPTOR: We do all our interface matching with an
872 * exclusive hold of global lock at IRQ priority. We're only
873 * expecting to be dealing with small numbers of interfaces, so the
874 * O(n**3)-ness here shouldn't matter */
876 write_lock_bh(global_lock);
878 if (net->ksnn_ninterfaces < 2) {
879 /* Only create additional connections
880 * if I have > 1 interface */
881 write_unlock_bh(global_lock);
885 LASSERT (npeer_ipaddrs <= LNET_MAX_INTERFACES);
887 for (i = 0; i < npeer_ipaddrs; i++) {
888 if (newroute != NULL) {
889 newroute->ksnr_ipaddr = peer_ipaddrs[i];
891 write_unlock_bh(global_lock);
893 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
894 if (newroute == NULL)
897 write_lock_bh(global_lock);
900 if (peer_ni->ksnp_closing) {
901 /* peer_ni got closed under me */
905 /* Already got a route? */
907 list_for_each(rtmp, &peer_ni->ksnp_routes) {
908 route = list_entry(rtmp, ksock_route_t, ksnr_list);
910 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
922 LASSERT(net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
924 /* Select interface to connect from */
925 for (j = 0; j < net->ksnn_ninterfaces; j++) {
926 iface = &net->ksnn_interfaces[j];
928 /* Using this interface already? */
929 list_for_each(rtmp, &peer_ni->ksnp_routes) {
930 route = list_entry(rtmp, ksock_route_t,
933 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
941 this_netmatch = (((iface->ksni_ipaddr ^
942 newroute->ksnr_ipaddr) &
943 iface->ksni_netmask) == 0) ? 1 : 0;
945 if (!(best_iface == NULL ||
946 best_netmatch < this_netmatch ||
947 (best_netmatch == this_netmatch &&
948 best_nroutes > iface->ksni_nroutes)))
952 best_netmatch = this_netmatch;
953 best_nroutes = iface->ksni_nroutes;
956 if (best_iface == NULL)
959 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
960 best_iface->ksni_nroutes++;
962 ksocknal_add_route_locked(peer_ni, newroute);
966 write_unlock_bh(global_lock);
967 if (newroute != NULL)
968 ksocknal_route_decref(newroute);
972 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
979 rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
980 LASSERT(rc == 0); /* we succeeded before */
982 LIBCFS_ALLOC(cr, sizeof(*cr));
984 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
985 "%pI4h: memory exhausted\n", &peer_ip);
991 cr->ksncr_sock = sock;
993 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
995 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
996 wake_up(&ksocknal_data.ksnd_connd_waitq);
998 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1003 ksocknal_connecting (ksock_peer_ni_t *peer_ni, __u32 ipaddr)
1005 ksock_route_t *route;
1007 list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
1008 if (route->ksnr_ipaddr == ipaddr)
1009 return route->ksnr_connecting;
1015 ksocknal_create_conn(struct lnet_ni *ni, ksock_route_t *route,
1016 struct socket *sock, int type)
1018 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1019 struct list_head zombies = LIST_HEAD_INIT(zombies);
1020 struct lnet_process_id peerid;
1021 struct list_head *tmp;
1024 ksock_conn_t *conn2;
1025 ksock_peer_ni_t *peer_ni = NULL;
1026 ksock_peer_ni_t *peer2;
1027 ksock_sched_t *sched;
1028 struct ksock_hello_msg *hello;
1036 active = (route != NULL);
1038 LASSERT (active == (type != SOCKLND_CONN_NONE));
1040 LIBCFS_ALLOC(conn, sizeof(*conn));
1046 conn->ksnc_peer = NULL;
1047 conn->ksnc_route = NULL;
1048 conn->ksnc_sock = sock;
1049 /* 2 ref, 1 for conn, another extra ref prevents socket
1050 * being closed before establishment of connection */
1051 atomic_set (&conn->ksnc_sock_refcount, 2);
1052 conn->ksnc_type = type;
1053 ksocknal_lib_save_callback(sock, conn);
1054 atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1056 conn->ksnc_rx_ready = 0;
1057 conn->ksnc_rx_scheduled = 0;
1059 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1060 conn->ksnc_tx_ready = 0;
1061 conn->ksnc_tx_scheduled = 0;
1062 conn->ksnc_tx_carrier = NULL;
1063 atomic_set (&conn->ksnc_tx_nob, 0);
1065 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1066 kshm_ips[LNET_MAX_INTERFACES]));
1067 if (hello == NULL) {
1072 /* stash conn's local and remote addrs */
1073 rc = ksocknal_lib_get_conn_addrs (conn);
1077 /* Find out/confirm peer_ni's NID and connection type and get the
1078 * vector of interfaces she's willing to let me connect to.
1079 * Passive connections use the listener timeout since the peer_ni sends
1083 peer_ni = route->ksnr_peer;
1084 LASSERT(ni == peer_ni->ksnp_ni);
1086 /* Active connection sends HELLO eagerly */
1087 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1088 peerid = peer_ni->ksnp_id;
1090 write_lock_bh(global_lock);
1091 conn->ksnc_proto = peer_ni->ksnp_proto;
1092 write_unlock_bh(global_lock);
1094 if (conn->ksnc_proto == NULL) {
1095 conn->ksnc_proto = &ksocknal_protocol_v3x;
1096 #if SOCKNAL_VERSION_DEBUG
1097 if (*ksocknal_tunables.ksnd_protocol == 2)
1098 conn->ksnc_proto = &ksocknal_protocol_v2x;
1099 else if (*ksocknal_tunables.ksnd_protocol == 1)
1100 conn->ksnc_proto = &ksocknal_protocol_v1x;
1104 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1108 peerid.nid = LNET_NID_ANY;
1109 peerid.pid = LNET_PID_ANY;
1111 /* Passive, get protocol from peer_ni */
1112 conn->ksnc_proto = NULL;
1115 rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1119 LASSERT (rc == 0 || active);
1120 LASSERT (conn->ksnc_proto != NULL);
1121 LASSERT (peerid.nid != LNET_NID_ANY);
1123 cpt = lnet_cpt_of_nid(peerid.nid, ni);
1126 ksocknal_peer_addref(peer_ni);
1127 write_lock_bh(global_lock);
1129 rc = ksocknal_create_peer(&peer_ni, ni, peerid);
1133 write_lock_bh(global_lock);
1135 /* called with a ref on ni, so shutdown can't have started */
1136 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1138 peer2 = ksocknal_find_peer_locked(ni, peerid);
1139 if (peer2 == NULL) {
1140 /* NB this puts an "empty" peer_ni in the peer_ni
1141 * table (which takes my ref) */
1142 list_add_tail(&peer_ni->ksnp_list,
1143 ksocknal_nid2peerlist(peerid.nid));
1145 ksocknal_peer_decref(peer_ni);
1150 ksocknal_peer_addref(peer_ni);
1151 peer_ni->ksnp_accepting++;
1153 /* Am I already connecting to this guy? Resolve in
1154 * favour of higher NID... */
1155 if (peerid.nid < ni->ni_nid &&
1156 ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1158 warn = "connection race resolution";
1163 if (peer_ni->ksnp_closing ||
1164 (active && route->ksnr_deleted)) {
1165 /* peer_ni/route got closed under me */
1167 warn = "peer_ni/route removed";
1171 if (peer_ni->ksnp_proto == NULL) {
1172 /* Never connected before.
1173 * NB recv_hello may have returned EPROTO to signal my peer_ni
1174 * wants a different protocol than the one I asked for.
1176 LASSERT(list_empty(&peer_ni->ksnp_conns));
1178 peer_ni->ksnp_proto = conn->ksnc_proto;
1179 peer_ni->ksnp_incarnation = incarnation;
1182 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1183 peer_ni->ksnp_incarnation != incarnation) {
1184 /* peer_ni rebooted or I've got the wrong protocol version */
1185 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1187 peer_ni->ksnp_proto = NULL;
1189 warn = peer_ni->ksnp_incarnation != incarnation ?
1190 "peer_ni rebooted" :
1191 "wrong proto version";
1201 warn = "lost conn race";
1204 warn = "retry with different protocol version";
1208 /* Refuse to duplicate an existing connection, unless this is a
1209 * loopback connection */
1210 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1211 list_for_each(tmp, &peer_ni->ksnp_conns) {
1212 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1214 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1215 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1216 conn2->ksnc_type != conn->ksnc_type)
1219 /* Reply on a passive connection attempt so the peer_ni
1220 * realises we're connected. */
1230 /* If the connection created by this route didn't bind to the IP
1231 * address the route connected to, the connection/route matching
1232 * code below probably isn't going to work. */
1234 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1235 CERROR("Route %s %pI4h connected to %pI4h\n",
1236 libcfs_id2str(peer_ni->ksnp_id),
1237 &route->ksnr_ipaddr,
1238 &conn->ksnc_ipaddr);
1241 /* Search for a route corresponding to the new connection and
1242 * create an association. This allows incoming connections created
1243 * by routes in my peer_ni to match my own route entries so I don't
1244 * continually create duplicate routes. */
1245 list_for_each(tmp, &peer_ni->ksnp_routes) {
1246 route = list_entry(tmp, ksock_route_t, ksnr_list);
1248 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1251 ksocknal_associate_route_conn_locked(route, conn);
1255 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1256 peer_ni->ksnp_last_alive = ktime_get_real_seconds();
1257 peer_ni->ksnp_send_keepalive = 0;
1258 peer_ni->ksnp_error = 0;
1260 sched = ksocknal_choose_scheduler_locked(cpt);
1261 sched->kss_nconns++;
1262 conn->ksnc_scheduler = sched;
1264 conn->ksnc_tx_last_post = ktime_get_real_seconds();
1265 /* Set the deadline for the outgoing HELLO to drain */
1266 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1267 conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1268 smp_mb(); /* order with adding to peer_ni's conn list */
1270 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1271 ksocknal_conn_addref(conn);
1273 ksocknal_new_packet(conn, 0);
1275 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1277 /* Take packets blocking for this connection. */
1278 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1279 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1283 list_del(&tx->tx_list);
1284 ksocknal_queue_tx_locked(tx, conn);
1287 write_unlock_bh(global_lock);
1289 /* We've now got a new connection. Any errors from here on are just
1290 * like "normal" comms errors and we close the connection normally.
1291 * NB (a) we still have to send the reply HELLO for passive
1293 * (b) normal I/O on the conn is blocked until I setup and call the
1297 CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1298 " incarnation:%lld sched[%d:%d]\n",
1299 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1300 &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1301 conn->ksnc_port, incarnation, cpt,
1302 (int)(sched - &sched->kss_info->ksi_scheds[0]));
1305 /* additional routes after interface exchange? */
1306 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1307 hello->kshm_ips, hello->kshm_nips);
1309 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1311 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1314 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1315 kshm_ips[LNET_MAX_INTERFACES]));
1317 /* setup the socket AFTER I've received hello (it disables
1318 * SO_LINGER). I might call back to the acceptor who may want
1319 * to send a protocol version response and then close the
1320 * socket; this ensures the socket only tears down after the
1321 * response has been sent. */
1323 rc = ksocknal_lib_setup_sock(sock);
1325 write_lock_bh(global_lock);
1327 /* NB my callbacks block while I hold ksnd_global_lock */
1328 ksocknal_lib_set_callback(sock, conn);
1331 peer_ni->ksnp_accepting--;
1333 write_unlock_bh(global_lock);
1336 write_lock_bh(global_lock);
1337 if (!conn->ksnc_closing) {
1338 /* could be closed by another thread */
1339 ksocknal_close_conn_locked(conn, rc);
1341 write_unlock_bh(global_lock);
1342 } else if (ksocknal_connsock_addref(conn) == 0) {
1343 /* Allow I/O to proceed. */
1344 ksocknal_read_callback(conn);
1345 ksocknal_write_callback(conn);
1346 ksocknal_connsock_decref(conn);
1349 ksocknal_connsock_decref(conn);
1350 ksocknal_conn_decref(conn);
1354 if (!peer_ni->ksnp_closing &&
1355 list_empty(&peer_ni->ksnp_conns) &&
1356 list_empty(&peer_ni->ksnp_routes)) {
1357 list_add(&zombies, &peer_ni->ksnp_tx_queue);
1358 list_del_init(&peer_ni->ksnp_tx_queue);
1359 ksocknal_unlink_peer_locked(peer_ni);
1362 write_unlock_bh(global_lock);
1366 CERROR("Not creating conn %s type %d: %s\n",
1367 libcfs_id2str(peerid), conn->ksnc_type, warn);
1369 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1370 libcfs_id2str(peerid), conn->ksnc_type, warn);
1375 /* Request retry by replying with CONN_NONE
1376 * ksnc_proto has been set already */
1377 conn->ksnc_type = SOCKLND_CONN_NONE;
1378 hello->kshm_nips = 0;
1379 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1382 write_lock_bh(global_lock);
1383 peer_ni->ksnp_accepting--;
1384 write_unlock_bh(global_lock);
1387 ksocknal_txlist_done(ni, &zombies, 1);
1388 ksocknal_peer_decref(peer_ni);
1392 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1393 kshm_ips[LNET_MAX_INTERFACES]));
1395 LIBCFS_FREE(conn, sizeof(*conn));
1403 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1405 /* This just does the immmediate housekeeping, and queues the
1406 * connection for the reaper to terminate.
1407 * Caller holds ksnd_global_lock exclusively in irq context */
1408 ksock_peer_ni_t *peer_ni = conn->ksnc_peer;
1409 ksock_route_t *route;
1410 ksock_conn_t *conn2;
1411 struct list_head *tmp;
1413 LASSERT(peer_ni->ksnp_error == 0);
1414 LASSERT(!conn->ksnc_closing);
1415 conn->ksnc_closing = 1;
1417 /* ksnd_deathrow_conns takes over peer_ni's ref */
1418 list_del(&conn->ksnc_list);
1420 route = conn->ksnc_route;
1421 if (route != NULL) {
1422 /* dissociate conn from route... */
1423 LASSERT(!route->ksnr_deleted);
1424 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1427 list_for_each(tmp, &peer_ni->ksnp_conns) {
1428 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1430 if (conn2->ksnc_route == route &&
1431 conn2->ksnc_type == conn->ksnc_type)
1437 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1439 conn->ksnc_route = NULL;
1441 ksocknal_route_decref(route); /* drop conn's ref on route */
1444 if (list_empty(&peer_ni->ksnp_conns)) {
1445 /* No more connections to this peer_ni */
1447 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1450 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1452 /* throw them to the last connection...,
1453 * these TXs will be send to /dev/null by scheduler */
1454 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1456 ksocknal_tx_prep(conn, tx);
1458 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1459 list_splice_init(&peer_ni->ksnp_tx_queue,
1460 &conn->ksnc_tx_queue);
1461 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1464 /* renegotiate protocol version */
1465 peer_ni->ksnp_proto = NULL;
1466 /* stash last conn close reason */
1467 peer_ni->ksnp_error = error;
1469 if (list_empty(&peer_ni->ksnp_routes)) {
1470 /* I've just closed last conn belonging to a
1471 * peer_ni with no routes to it */
1472 ksocknal_unlink_peer_locked(peer_ni);
1476 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1478 list_add_tail(&conn->ksnc_list,
1479 &ksocknal_data.ksnd_deathrow_conns);
1480 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1482 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1486 ksocknal_peer_failed (ksock_peer_ni_t *peer_ni)
1489 cfs_time_t last_alive = 0;
1491 /* There has been a connection failure or comms error; but I'll only
1492 * tell LNET I think the peer_ni is dead if it's to another kernel and
1493 * there are no connections or connection attempts in existence. */
1495 read_lock(&ksocknal_data.ksnd_global_lock);
1497 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1498 list_empty(&peer_ni->ksnp_conns) &&
1499 peer_ni->ksnp_accepting == 0 &&
1500 ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1502 last_alive = peer_ni->ksnp_last_alive;
1505 read_unlock(&ksocknal_data.ksnd_global_lock);
1508 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid, 0,
1513 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1515 ksock_peer_ni_t *peer_ni = conn->ksnc_peer;
1518 struct list_head zlist = LIST_HEAD_INIT(zlist);
1520 /* NB safe to finalize TXs because closing of socket will
1521 * abort all buffered data */
1522 LASSERT(conn->ksnc_sock == NULL);
1524 spin_lock(&peer_ni->ksnp_lock);
1526 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1527 if (tx->tx_conn != conn)
1530 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1532 tx->tx_msg.ksm_zc_cookies[0] = 0;
1533 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1534 list_del(&tx->tx_zc_list);
1535 list_add(&tx->tx_zc_list, &zlist);
1538 spin_unlock(&peer_ni->ksnp_lock);
1540 while (!list_empty(&zlist)) {
1541 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1543 list_del(&tx->tx_zc_list);
1544 ksocknal_tx_decref(tx);
1549 ksocknal_terminate_conn(ksock_conn_t *conn)
1551 /* This gets called by the reaper (guaranteed thread context) to
1552 * disengage the socket from its callbacks and close it.
1553 * ksnc_refcount will eventually hit zero, and then the reaper will
1555 ksock_peer_ni_t *peer_ni = conn->ksnc_peer;
1556 ksock_sched_t *sched = conn->ksnc_scheduler;
1559 LASSERT(conn->ksnc_closing);
1561 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1562 spin_lock_bh(&sched->kss_lock);
1564 /* a closing conn is always ready to tx */
1565 conn->ksnc_tx_ready = 1;
1567 if (!conn->ksnc_tx_scheduled &&
1568 !list_empty(&conn->ksnc_tx_queue)) {
1569 list_add_tail(&conn->ksnc_tx_list,
1570 &sched->kss_tx_conns);
1571 conn->ksnc_tx_scheduled = 1;
1572 /* extra ref for scheduler */
1573 ksocknal_conn_addref(conn);
1575 wake_up (&sched->kss_waitq);
1578 spin_unlock_bh(&sched->kss_lock);
1580 /* serialise with callbacks */
1581 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1583 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1585 /* OK, so this conn may not be completely disengaged from its
1586 * scheduler yet, but it _has_ committed to terminate... */
1587 conn->ksnc_scheduler->kss_nconns--;
1589 if (peer_ni->ksnp_error != 0) {
1590 /* peer_ni's last conn closed in error */
1591 LASSERT(list_empty(&peer_ni->ksnp_conns));
1593 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1596 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1599 ksocknal_peer_failed(peer_ni);
1601 /* The socket is closed on the final put; either here, or in
1602 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1603 * when the connection was established, this will close the socket
1604 * immediately, aborting anything buffered in it. Any hung
1605 * zero-copy transmits will therefore complete in finite time. */
1606 ksocknal_connsock_decref(conn);
1610 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1612 /* Queue the conn for the reaper to destroy */
1614 LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1615 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1617 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1618 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1620 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1624 ksocknal_destroy_conn (ksock_conn_t *conn)
1626 cfs_time_t last_rcv;
1628 /* Final coup-de-grace of the reaper */
1629 CDEBUG (D_NET, "connection %p\n", conn);
1631 LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1632 LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1633 LASSERT (conn->ksnc_sock == NULL);
1634 LASSERT (conn->ksnc_route == NULL);
1635 LASSERT (!conn->ksnc_tx_scheduled);
1636 LASSERT (!conn->ksnc_rx_scheduled);
1637 LASSERT(list_empty(&conn->ksnc_tx_queue));
1639 /* complete current receive if any */
1640 switch (conn->ksnc_rx_state) {
1641 case SOCKNAL_RX_LNET_PAYLOAD:
1642 last_rcv = conn->ksnc_rx_deadline -
1643 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1644 CERROR("Completing partial receive from %s[%d], "
1645 "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1646 "last alive is %ld secs ago\n",
1647 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1648 &conn->ksnc_ipaddr, conn->ksnc_port,
1649 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1650 cfs_duration_sec(cfs_time_sub(ktime_get_real_seconds(),
1652 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1653 conn->ksnc_cookie, -EIO);
1655 case SOCKNAL_RX_LNET_HEADER:
1656 if (conn->ksnc_rx_started)
1657 CERROR("Incomplete receive of lnet header from %s, "
1658 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1659 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1660 &conn->ksnc_ipaddr, conn->ksnc_port,
1661 conn->ksnc_proto->pro_version);
1663 case SOCKNAL_RX_KSM_HEADER:
1664 if (conn->ksnc_rx_started)
1665 CERROR("Incomplete receive of ksock message from %s, "
1666 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1667 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1668 &conn->ksnc_ipaddr, conn->ksnc_port,
1669 conn->ksnc_proto->pro_version);
1671 case SOCKNAL_RX_SLOP:
1672 if (conn->ksnc_rx_started)
1673 CERROR("Incomplete receive of slops from %s, "
1674 "ip %pI4h:%d, with error\n",
1675 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1676 &conn->ksnc_ipaddr, conn->ksnc_port);
1683 ksocknal_peer_decref(conn->ksnc_peer);
1685 LIBCFS_FREE (conn, sizeof (*conn));
1689 ksocknal_close_peer_conns_locked (ksock_peer_ni_t *peer_ni, __u32 ipaddr, int why)
1692 struct list_head *ctmp;
1693 struct list_head *cnxt;
1696 list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1697 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1700 conn->ksnc_ipaddr == ipaddr) {
1702 ksocknal_close_conn_locked (conn, why);
1710 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1712 ksock_peer_ni_t *peer_ni = conn->ksnc_peer;
1713 __u32 ipaddr = conn->ksnc_ipaddr;
1716 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1718 count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1720 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1726 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1728 ksock_peer_ni_t *peer_ni;
1729 struct list_head *ptmp;
1730 struct list_head *pnxt;
1736 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1738 if (id.nid != LNET_NID_ANY)
1739 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1742 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1745 for (i = lo; i <= hi; i++) {
1746 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1748 peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
1750 if (!((id.nid == LNET_NID_ANY || id.nid == peer_ni->ksnp_id.nid) &&
1751 (id.pid == LNET_PID_ANY || id.pid == peer_ni->ksnp_id.pid)))
1754 count += ksocknal_close_peer_conns_locked (peer_ni, ipaddr, 0);
1758 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1760 /* wildcards always succeed */
1761 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1764 return (count == 0 ? -ENOENT : 0);
1768 ksocknal_notify(struct lnet_ni *ni, lnet_nid_t gw_nid, int alive)
1770 /* The router is telling me she's been notified of a change in
1773 struct lnet_process_id id = {
1775 .pid = LNET_PID_ANY,
1778 CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1779 alive ? "up" : "down");
1782 /* If the gateway crashed, close all open connections... */
1783 ksocknal_close_matching_conns (id, 0);
1787 /* ...otherwise do nothing. We can only establish new connections
1788 * if we have autroutes, and these connect on demand. */
1792 ksocknal_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1795 time64_t last_alive = 0;
1796 time64_t now = ktime_get_real_seconds();
1797 ksock_peer_ni_t *peer_ni = NULL;
1798 rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1799 struct lnet_process_id id = {
1801 .pid = LNET_PID_LUSTRE,
1806 peer_ni = ksocknal_find_peer_locked(ni, id);
1807 if (peer_ni != NULL) {
1808 struct list_head *tmp;
1812 list_for_each(tmp, &peer_ni->ksnp_conns) {
1813 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1814 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1816 if (bufnob < conn->ksnc_tx_bufnob) {
1817 /* something got ACKed */
1818 conn->ksnc_tx_deadline =
1819 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1820 peer_ni->ksnp_last_alive = now;
1821 conn->ksnc_tx_bufnob = bufnob;
1825 last_alive = peer_ni->ksnp_last_alive;
1826 if (ksocknal_find_connectable_route_locked(peer_ni) == NULL)
1832 if (last_alive != 0)
1835 CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago, connect %d\n",
1836 libcfs_nid2str(nid), peer_ni,
1837 last_alive ? cfs_duration_sec(now - last_alive) : -1,
1843 ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1845 write_lock_bh(glock);
1847 peer_ni = ksocknal_find_peer_locked(ni, id);
1848 if (peer_ni != NULL)
1849 ksocknal_launch_all_connections_locked(peer_ni);
1851 write_unlock_bh(glock);
1856 ksocknal_push_peer (ksock_peer_ni_t *peer_ni)
1860 struct list_head *tmp;
1863 for (index = 0; ; index++) {
1864 read_lock(&ksocknal_data.ksnd_global_lock);
1869 list_for_each(tmp, &peer_ni->ksnp_conns) {
1871 conn = list_entry(tmp, ksock_conn_t,
1873 ksocknal_conn_addref(conn);
1878 read_unlock(&ksocknal_data.ksnd_global_lock);
1883 ksocknal_lib_push_conn (conn);
1884 ksocknal_conn_decref(conn);
1889 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1891 struct list_head *start;
1892 struct list_head *end;
1893 struct list_head *tmp;
1895 unsigned int hsize = ksocknal_data.ksnd_peer_hash_size;
1897 if (id.nid == LNET_NID_ANY) {
1898 start = &ksocknal_data.ksnd_peers[0];
1899 end = &ksocknal_data.ksnd_peers[hsize - 1];
1901 start = end = ksocknal_nid2peerlist(id.nid);
1904 for (tmp = start; tmp <= end; tmp++) {
1905 int peer_off; /* searching offset in peer_ni hash table */
1907 for (peer_off = 0; ; peer_off++) {
1908 ksock_peer_ni_t *peer_ni;
1911 read_lock(&ksocknal_data.ksnd_global_lock);
1912 list_for_each_entry(peer_ni, tmp, ksnp_list) {
1913 if (!((id.nid == LNET_NID_ANY ||
1914 id.nid == peer_ni->ksnp_id.nid) &&
1915 (id.pid == LNET_PID_ANY ||
1916 id.pid == peer_ni->ksnp_id.pid)))
1919 if (i++ == peer_off) {
1920 ksocknal_peer_addref(peer_ni);
1924 read_unlock(&ksocknal_data.ksnd_global_lock);
1926 if (i == 0) /* no match */
1930 ksocknal_push_peer(peer_ni);
1931 ksocknal_peer_decref(peer_ni);
1938 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1940 ksock_net_t *net = ni->ni_data;
1941 ksock_interface_t *iface;
1945 struct list_head *ptmp;
1946 ksock_peer_ni_t *peer_ni;
1947 struct list_head *rtmp;
1948 ksock_route_t *route;
1950 if (ipaddress == 0 ||
1954 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1956 iface = ksocknal_ip2iface(ni, ipaddress);
1957 if (iface != NULL) {
1958 /* silently ignore dups */
1960 } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1963 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1965 iface->ksni_ipaddr = ipaddress;
1966 iface->ksni_netmask = netmask;
1967 iface->ksni_nroutes = 0;
1968 iface->ksni_npeers = 0;
1970 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1971 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1972 peer_ni = list_entry(ptmp, ksock_peer_ni_t,
1975 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
1976 if (peer_ni->ksnp_passive_ips[j] == ipaddress)
1977 iface->ksni_npeers++;
1979 list_for_each(rtmp, &peer_ni->ksnp_routes) {
1980 route = list_entry(rtmp,
1984 if (route->ksnr_myipaddr == ipaddress)
1985 iface->ksni_nroutes++;
1991 /* NB only new connections will pay attention to the new interface! */
1994 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2000 ksocknal_peer_del_interface_locked(ksock_peer_ni_t *peer_ni, __u32 ipaddr)
2002 struct list_head *tmp;
2003 struct list_head *nxt;
2004 ksock_route_t *route;
2009 for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
2010 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
2011 for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
2012 peer_ni->ksnp_passive_ips[j-1] =
2013 peer_ni->ksnp_passive_ips[j];
2014 peer_ni->ksnp_n_passive_ips--;
2018 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
2019 route = list_entry(tmp, ksock_route_t, ksnr_list);
2021 if (route->ksnr_myipaddr != ipaddr)
2024 if (route->ksnr_share_count != 0) {
2025 /* Manually created; keep, but unbind */
2026 route->ksnr_myipaddr = 0;
2028 ksocknal_del_route_locked(route);
2032 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
2033 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2035 if (conn->ksnc_myipaddr == ipaddr)
2036 ksocknal_close_conn_locked (conn, 0);
2041 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
2043 ksock_net_t *net = ni->ni_data;
2045 struct list_head *tmp;
2046 struct list_head *nxt;
2047 ksock_peer_ni_t *peer_ni;
2052 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2054 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2055 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2057 if (!(ipaddress == 0 ||
2058 ipaddress == this_ip))
2063 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2064 net->ksnn_interfaces[j-1] =
2065 net->ksnn_interfaces[j];
2067 net->ksnn_ninterfaces--;
2069 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2070 list_for_each_safe(tmp, nxt,
2071 &ksocknal_data.ksnd_peers[j]) {
2072 peer_ni = list_entry(tmp, ksock_peer_ni_t,
2075 if (peer_ni->ksnp_ni != ni)
2078 ksocknal_peer_del_interface_locked(peer_ni, this_ip);
2083 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2089 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2091 struct lnet_process_id id = {0};
2092 struct libcfs_ioctl_data *data = arg;
2096 case IOC_LIBCFS_GET_INTERFACE: {
2097 ksock_net_t *net = ni->ni_data;
2098 ksock_interface_t *iface;
2100 read_lock(&ksocknal_data.ksnd_global_lock);
2102 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2106 iface = &net->ksnn_interfaces[data->ioc_count];
2108 data->ioc_u32[0] = iface->ksni_ipaddr;
2109 data->ioc_u32[1] = iface->ksni_netmask;
2110 data->ioc_u32[2] = iface->ksni_npeers;
2111 data->ioc_u32[3] = iface->ksni_nroutes;
2114 read_unlock(&ksocknal_data.ksnd_global_lock);
2118 case IOC_LIBCFS_ADD_INTERFACE:
2119 return ksocknal_add_interface(ni,
2120 data->ioc_u32[0], /* IP address */
2121 data->ioc_u32[1]); /* net mask */
2123 case IOC_LIBCFS_DEL_INTERFACE:
2124 return ksocknal_del_interface(ni,
2125 data->ioc_u32[0]); /* IP address */
2127 case IOC_LIBCFS_GET_PEER: {
2132 int share_count = 0;
2134 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2135 &id, &myip, &ip, &port,
2136 &conn_count, &share_count);
2140 data->ioc_nid = id.nid;
2141 data->ioc_count = share_count;
2142 data->ioc_u32[0] = ip;
2143 data->ioc_u32[1] = port;
2144 data->ioc_u32[2] = myip;
2145 data->ioc_u32[3] = conn_count;
2146 data->ioc_u32[4] = id.pid;
2150 case IOC_LIBCFS_ADD_PEER:
2151 id.nid = data->ioc_nid;
2152 id.pid = LNET_PID_LUSTRE;
2153 return ksocknal_add_peer (ni, id,
2154 data->ioc_u32[0], /* IP */
2155 data->ioc_u32[1]); /* port */
2157 case IOC_LIBCFS_DEL_PEER:
2158 id.nid = data->ioc_nid;
2159 id.pid = LNET_PID_ANY;
2160 return ksocknal_del_peer (ni, id,
2161 data->ioc_u32[0]); /* IP */
2163 case IOC_LIBCFS_GET_CONN: {
2167 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2172 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2174 data->ioc_count = txmem;
2175 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
2176 data->ioc_flags = nagle;
2177 data->ioc_u32[0] = conn->ksnc_ipaddr;
2178 data->ioc_u32[1] = conn->ksnc_port;
2179 data->ioc_u32[2] = conn->ksnc_myipaddr;
2180 data->ioc_u32[3] = conn->ksnc_type;
2181 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2182 data->ioc_u32[5] = rxmem;
2183 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2184 ksocknal_conn_decref(conn);
2188 case IOC_LIBCFS_CLOSE_CONNECTION:
2189 id.nid = data->ioc_nid;
2190 id.pid = LNET_PID_ANY;
2191 return ksocknal_close_matching_conns (id,
2194 case IOC_LIBCFS_REGISTER_MYNID:
2195 /* Ignore if this is a noop */
2196 if (data->ioc_nid == ni->ni_nid)
2199 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2200 libcfs_nid2str(data->ioc_nid),
2201 libcfs_nid2str(ni->ni_nid));
2204 case IOC_LIBCFS_PUSH_CONNECTION:
2205 id.nid = data->ioc_nid;
2206 id.pid = LNET_PID_ANY;
2207 return ksocknal_push(ni, id);
2216 ksocknal_free_buffers (void)
2218 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2220 if (ksocknal_data.ksnd_sched_info != NULL) {
2221 struct ksock_sched_info *info;
2224 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2225 if (info->ksi_scheds != NULL) {
2226 LIBCFS_FREE(info->ksi_scheds,
2227 info->ksi_nthreads_max *
2228 sizeof(info->ksi_scheds[0]));
2231 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2234 LIBCFS_FREE (ksocknal_data.ksnd_peers,
2235 sizeof(struct list_head) *
2236 ksocknal_data.ksnd_peer_hash_size);
2238 spin_lock(&ksocknal_data.ksnd_tx_lock);
2240 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2241 struct list_head zlist;
2244 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2245 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2246 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2248 while (!list_empty(&zlist)) {
2249 tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2250 list_del(&tx->tx_list);
2251 LIBCFS_FREE(tx, tx->tx_desc_size);
2254 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2259 ksocknal_base_shutdown(void)
2261 struct ksock_sched_info *info;
2262 ksock_sched_t *sched;
2266 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2267 atomic_read (&libcfs_kmemory));
2268 LASSERT (ksocknal_data.ksnd_nnets == 0);
2270 switch (ksocknal_data.ksnd_init) {
2274 case SOCKNAL_INIT_ALL:
2275 case SOCKNAL_INIT_DATA:
2276 LASSERT (ksocknal_data.ksnd_peers != NULL);
2277 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2278 LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2281 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2282 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2283 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2284 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2285 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2287 if (ksocknal_data.ksnd_sched_info != NULL) {
2288 cfs_percpt_for_each(info, i,
2289 ksocknal_data.ksnd_sched_info) {
2290 if (info->ksi_scheds == NULL)
2293 for (j = 0; j < info->ksi_nthreads_max; j++) {
2295 sched = &info->ksi_scheds[j];
2296 LASSERT(list_empty(&sched->\
2298 LASSERT(list_empty(&sched->\
2300 LASSERT(list_empty(&sched-> \
2301 kss_zombie_noop_txs));
2302 LASSERT(sched->kss_nconns == 0);
2307 /* flag threads to terminate; wake and wait for them to die */
2308 ksocknal_data.ksnd_shuttingdown = 1;
2309 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2310 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2312 if (ksocknal_data.ksnd_sched_info != NULL) {
2313 cfs_percpt_for_each(info, i,
2314 ksocknal_data.ksnd_sched_info) {
2315 if (info->ksi_scheds == NULL)
2318 for (j = 0; j < info->ksi_nthreads_max; j++) {
2319 sched = &info->ksi_scheds[j];
2320 wake_up_all(&sched->kss_waitq);
2326 read_lock(&ksocknal_data.ksnd_global_lock);
2327 while (ksocknal_data.ksnd_nthreads != 0) {
2330 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2331 "waiting for %d threads to terminate\n",
2332 ksocknal_data.ksnd_nthreads);
2333 read_unlock(&ksocknal_data.ksnd_global_lock);
2334 set_current_state(TASK_UNINTERRUPTIBLE);
2335 schedule_timeout(cfs_time_seconds(1));
2336 read_lock(&ksocknal_data.ksnd_global_lock);
2338 read_unlock(&ksocknal_data.ksnd_global_lock);
2340 ksocknal_free_buffers();
2342 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2346 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2347 atomic_read (&libcfs_kmemory));
2349 module_put(THIS_MODULE);
2352 static __u64 ksocknal_new_incarnation(void)
2356 /* The incarnation number is the time this module loaded and it
2357 * identifies this particular instance of the socknal. Hopefully
2358 * we won't be able to reboot more frequently than 1MHz for the
2359 * forseeable future :) */
2361 do_gettimeofday(&tv);
2363 return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2367 ksocknal_base_startup(void)
2369 struct ksock_sched_info *info;
2373 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2374 LASSERT (ksocknal_data.ksnd_nnets == 0);
2376 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2378 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2379 LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2380 sizeof(struct list_head) *
2381 ksocknal_data.ksnd_peer_hash_size);
2382 if (ksocknal_data.ksnd_peers == NULL)
2385 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2386 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2388 rwlock_init(&ksocknal_data.ksnd_global_lock);
2389 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2391 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2392 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2393 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2394 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2395 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2397 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2398 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2399 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2400 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2402 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2403 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2405 /* NB memset above zeros whole of ksocknal_data */
2407 /* flag lists/ptrs/locks initialised */
2408 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2409 try_module_get(THIS_MODULE);
2411 ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2413 if (ksocknal_data.ksnd_sched_info == NULL)
2416 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2417 ksock_sched_t *sched;
2420 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2421 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2422 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2424 /* max to half of CPUs, assume another half should be
2425 * reserved for upper layer modules */
2426 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2429 info->ksi_nthreads_max = nthrs;
2432 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2433 info->ksi_nthreads_max * sizeof(*sched));
2434 if (info->ksi_scheds == NULL)
2437 for (; nthrs > 0; nthrs--) {
2438 sched = &info->ksi_scheds[nthrs - 1];
2440 sched->kss_info = info;
2441 spin_lock_init(&sched->kss_lock);
2442 INIT_LIST_HEAD(&sched->kss_rx_conns);
2443 INIT_LIST_HEAD(&sched->kss_tx_conns);
2444 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2445 init_waitqueue_head(&sched->kss_waitq);
2449 ksocknal_data.ksnd_connd_starting = 0;
2450 ksocknal_data.ksnd_connd_failed_stamp = 0;
2451 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2452 /* must have at least 2 connds to remain responsive to accepts while
2454 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2455 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2457 if (*ksocknal_tunables.ksnd_nconnds_max <
2458 *ksocknal_tunables.ksnd_nconnds) {
2459 ksocknal_tunables.ksnd_nconnds_max =
2460 ksocknal_tunables.ksnd_nconnds;
2463 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2465 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2466 ksocknal_data.ksnd_connd_starting++;
2467 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2470 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2471 rc = ksocknal_thread_start(ksocknal_connd,
2472 (void *)((uintptr_t)i), name);
2474 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2475 ksocknal_data.ksnd_connd_starting--;
2476 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2477 CERROR("Can't spawn socknal connd: %d\n", rc);
2482 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2484 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2488 /* flag everything initialised */
2489 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2494 ksocknal_base_shutdown();
2499 ksocknal_debug_peerhash(struct lnet_ni *ni)
2501 ksock_peer_ni_t *peer_ni = NULL;
2502 struct list_head *tmp;
2505 read_lock(&ksocknal_data.ksnd_global_lock);
2507 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2508 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2509 peer_ni = list_entry(tmp, ksock_peer_ni_t, ksnp_list);
2511 if (peer_ni->ksnp_ni == ni) break;
2517 if (peer_ni != NULL) {
2518 ksock_route_t *route;
2521 CWARN ("Active peer_ni on shutdown: %s, ref %d, scnt %d, "
2522 "closing %d, accepting %d, err %d, zcookie %llu, "
2523 "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2524 atomic_read(&peer_ni->ksnp_refcount),
2525 peer_ni->ksnp_sharecount, peer_ni->ksnp_closing,
2526 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2527 peer_ni->ksnp_zc_next_cookie,
2528 !list_empty(&peer_ni->ksnp_tx_queue),
2529 !list_empty(&peer_ni->ksnp_zc_req_list));
2531 list_for_each(tmp, &peer_ni->ksnp_routes) {
2532 route = list_entry(tmp, ksock_route_t, ksnr_list);
2533 CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2534 "del %d\n", atomic_read(&route->ksnr_refcount),
2535 route->ksnr_scheduled, route->ksnr_connecting,
2536 route->ksnr_connected, route->ksnr_deleted);
2539 list_for_each(tmp, &peer_ni->ksnp_conns) {
2540 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2541 CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2542 atomic_read(&conn->ksnc_conn_refcount),
2543 atomic_read(&conn->ksnc_sock_refcount),
2544 conn->ksnc_type, conn->ksnc_closing);
2548 read_unlock(&ksocknal_data.ksnd_global_lock);
2553 ksocknal_shutdown(struct lnet_ni *ni)
2555 ksock_net_t *net = ni->ni_data;
2556 struct lnet_process_id anyid = {
2557 .nid = LNET_NID_ANY,
2558 .pid = LNET_PID_ANY,
2562 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2563 LASSERT(ksocknal_data.ksnd_nnets > 0);
2565 spin_lock_bh(&net->ksnn_lock);
2566 net->ksnn_shutdown = 1; /* prevent new peers */
2567 spin_unlock_bh(&net->ksnn_lock);
2569 /* Delete all peers */
2570 ksocknal_del_peer(ni, anyid, 0);
2572 /* Wait for all peer_ni state to clean up */
2574 spin_lock_bh(&net->ksnn_lock);
2575 while (net->ksnn_npeers != 0) {
2576 spin_unlock_bh(&net->ksnn_lock);
2579 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2580 "waiting for %d peers to disconnect\n",
2582 set_current_state(TASK_UNINTERRUPTIBLE);
2583 schedule_timeout(cfs_time_seconds(1));
2585 ksocknal_debug_peerhash(ni);
2587 spin_lock_bh(&net->ksnn_lock);
2589 spin_unlock_bh(&net->ksnn_lock);
2591 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2592 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2593 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2596 list_del(&net->ksnn_list);
2597 LIBCFS_FREE(net, sizeof(*net));
2599 ksocknal_data.ksnd_nnets--;
2600 if (ksocknal_data.ksnd_nnets == 0)
2601 ksocknal_base_shutdown();
2605 ksocknal_enumerate_interfaces(ksock_net_t *net)
2613 n = lnet_ipif_enumerate(&names);
2615 CERROR("Can't enumerate interfaces: %d\n", n);
2619 for (i = j = 0; i < n; i++) {
2624 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2627 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2629 CWARN("Can't get interface %s info: %d\n",
2635 CWARN("Ignoring interface %s (down)\n",
2640 if (j == LNET_MAX_INTERFACES) {
2641 CWARN("Ignoring interface %s (too many interfaces)\n",
2646 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2647 net->ksnn_interfaces[j].ksni_netmask = mask;
2648 strlcpy(net->ksnn_interfaces[j].ksni_name,
2649 names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2653 lnet_ipif_free_enumeration(names, n);
2656 CERROR("Can't find any usable interfaces\n");
2662 ksocknal_search_new_ipif(ksock_net_t *net)
2667 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2668 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2669 char *colon = strchr(ifnam, ':');
2674 if (colon != NULL) /* ignore alias device */
2677 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2679 for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2680 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2682 char *colon2 = strchr(ifnam2, ':');
2687 found = strcmp(ifnam, ifnam2) == 0;
2704 ksocknal_start_schedulers(struct ksock_sched_info *info)
2710 if (info->ksi_nthreads == 0) {
2711 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2712 nthrs = info->ksi_nthreads_max;
2714 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2716 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2717 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2719 nthrs = min(nthrs, info->ksi_nthreads_max);
2721 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2722 /* increase two threads if there is new interface */
2723 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2726 for (i = 0; i < nthrs; i++) {
2729 ksock_sched_t *sched;
2730 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2731 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2732 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2733 info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2735 rc = ksocknal_thread_start(ksocknal_scheduler,
2740 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2741 info->ksi_cpt, info->ksi_nthreads + i, rc);
2745 info->ksi_nthreads += i;
2750 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2752 int newif = ksocknal_search_new_ipif(net);
2756 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2759 for (i = 0; i < ncpts; i++) {
2760 struct ksock_sched_info *info;
2761 int cpt = (cpts == NULL) ? i : cpts[i];
2763 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2764 info = ksocknal_data.ksnd_sched_info[cpt];
2766 if (!newif && info->ksi_nthreads > 0)
2769 rc = ksocknal_start_schedulers(info);
2777 ksocknal_startup(struct lnet_ni *ni)
2782 struct net_device *net_dev;
2785 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2787 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2788 rc = ksocknal_base_startup();
2793 LIBCFS_ALLOC(net, sizeof(*net));
2797 spin_lock_init(&net->ksnn_lock);
2798 net->ksnn_incarnation = ksocknal_new_incarnation();
2800 if (!ni->ni_net->net_tunables_set) {
2801 ni->ni_net->net_tunables.lct_peer_timeout =
2802 *ksocknal_tunables.ksnd_peertimeout;
2803 ni->ni_net->net_tunables.lct_max_tx_credits =
2804 *ksocknal_tunables.ksnd_credits;
2805 ni->ni_net->net_tunables.lct_peer_tx_credits =
2806 *ksocknal_tunables.ksnd_peertxcredits;
2807 ni->ni_net->net_tunables.lct_peer_rtr_credits =
2808 *ksocknal_tunables.ksnd_peerrtrcredits;
2809 ni->ni_net->net_tunables_set = true;
2813 if (ni->ni_interfaces[0] == NULL) {
2814 rc = ksocknal_enumerate_interfaces(net);
2818 net->ksnn_ninterfaces = 1;
2820 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2823 if (ni->ni_interfaces[i] == NULL)
2826 rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2827 &net->ksnn_interfaces[i].ksni_ipaddr,
2828 &net->ksnn_interfaces[i].ksni_netmask);
2831 CERROR("Can't get interface %s info: %d\n",
2832 ni->ni_interfaces[i], rc);
2837 CERROR("Interface %s is down\n",
2838 ni->ni_interfaces[i]);
2842 strlcpy(net->ksnn_interfaces[i].ksni_name,
2843 ni->ni_interfaces[i],
2844 sizeof(net->ksnn_interfaces[i].ksni_name));
2847 net->ksnn_ninterfaces = i;
2850 net_dev = dev_get_by_name(&init_net,
2851 net->ksnn_interfaces[0].ksni_name);
2852 if (net_dev != NULL) {
2853 node_id = dev_to_node(&net_dev->dev);
2854 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
2857 ni->ni_dev_cpt = CFS_CPT_ANY;
2860 /* call it before add it to ksocknal_data.ksnd_nets */
2861 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2865 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2866 net->ksnn_interfaces[0].ksni_ipaddr);
2867 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2869 ksocknal_data.ksnd_nnets++;
2874 LIBCFS_FREE(net, sizeof(*net));
2876 if (ksocknal_data.ksnd_nnets == 0)
2877 ksocknal_base_shutdown();
2883 static void __exit ksocklnd_exit(void)
2885 lnet_unregister_lnd(&the_ksocklnd);
2888 static int __init ksocklnd_init(void)
2892 /* check ksnr_connected/connecting field large enough */
2893 CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2894 CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2896 /* initialize the_ksocklnd */
2897 the_ksocklnd.lnd_type = SOCKLND;
2898 the_ksocklnd.lnd_startup = ksocknal_startup;
2899 the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2900 the_ksocklnd.lnd_ctl = ksocknal_ctl;
2901 the_ksocklnd.lnd_send = ksocknal_send;
2902 the_ksocklnd.lnd_recv = ksocknal_recv;
2903 the_ksocklnd.lnd_notify = ksocknal_notify;
2904 the_ksocklnd.lnd_query = ksocknal_query;
2905 the_ksocklnd.lnd_accept = ksocknal_accept;
2907 rc = ksocknal_tunables_init();
2911 lnet_register_lnd(&the_ksocklnd);
2916 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2917 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2918 MODULE_VERSION("2.8.0");
2919 MODULE_LICENSE("GPL");
2921 module_init(ksocklnd_init);
2922 module_exit(ksocklnd_exit);