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>
42 static lnd_t the_ksocklnd;
43 ksock_nal_data_t ksocknal_data;
45 static ksock_interface_t *
46 ksocknal_ip2iface(lnet_ni_t *ni, __u32 ip)
48 ksock_net_t *net = ni->ni_data;
50 ksock_interface_t *iface;
52 for (i = 0; i < net->ksnn_ninterfaces; i++) {
53 LASSERT(i < LNET_MAX_INTERFACES);
54 iface = &net->ksnn_interfaces[i];
56 if (iface->ksni_ipaddr == ip)
63 static ksock_route_t *
64 ksocknal_create_route (__u32 ipaddr, int port)
68 LIBCFS_ALLOC (route, sizeof (*route));
72 atomic_set (&route->ksnr_refcount, 1);
73 route->ksnr_peer = NULL;
74 route->ksnr_retry_interval = 0; /* OK to connect at any time */
75 route->ksnr_ipaddr = ipaddr;
76 route->ksnr_port = port;
77 route->ksnr_scheduled = 0;
78 route->ksnr_connecting = 0;
79 route->ksnr_connected = 0;
80 route->ksnr_deleted = 0;
81 route->ksnr_conn_count = 0;
82 route->ksnr_share_count = 0;
88 ksocknal_destroy_route (ksock_route_t *route)
90 LASSERT (atomic_read(&route->ksnr_refcount) == 0);
92 if (route->ksnr_peer != NULL)
93 ksocknal_peer_decref(route->ksnr_peer);
95 LIBCFS_FREE (route, sizeof (*route));
99 ksocknal_create_peer(ksock_peer_t **peerp, lnet_ni_t *ni, lnet_process_id_t id)
101 int cpt = lnet_cpt_of_nid(id.nid);
102 ksock_net_t *net = ni->ni_data;
105 LASSERT(id.nid != LNET_NID_ANY);
106 LASSERT(id.pid != LNET_PID_ANY);
107 LASSERT(!in_interrupt());
109 LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
115 atomic_set(&peer->ksnp_refcount, 1); /* 1 ref for caller */
116 peer->ksnp_closing = 0;
117 peer->ksnp_accepting = 0;
118 peer->ksnp_proto = NULL;
119 peer->ksnp_last_alive = 0;
120 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
122 INIT_LIST_HEAD(&peer->ksnp_conns);
123 INIT_LIST_HEAD(&peer->ksnp_routes);
124 INIT_LIST_HEAD(&peer->ksnp_tx_queue);
125 INIT_LIST_HEAD(&peer->ksnp_zc_req_list);
126 spin_lock_init(&peer->ksnp_lock);
128 spin_lock_bh(&net->ksnn_lock);
130 if (net->ksnn_shutdown) {
131 spin_unlock_bh(&net->ksnn_lock);
133 LIBCFS_FREE(peer, sizeof(*peer));
134 CERROR("Can't create peer: network shutdown\n");
140 spin_unlock_bh(&net->ksnn_lock);
147 ksocknal_destroy_peer (ksock_peer_t *peer)
149 ksock_net_t *net = peer->ksnp_ni->ni_data;
151 CDEBUG (D_NET, "peer %s %p deleted\n",
152 libcfs_id2str(peer->ksnp_id), peer);
154 LASSERT(atomic_read(&peer->ksnp_refcount) == 0);
155 LASSERT(peer->ksnp_accepting == 0);
156 LASSERT(list_empty(&peer->ksnp_conns));
157 LASSERT(list_empty(&peer->ksnp_routes));
158 LASSERT(list_empty(&peer->ksnp_tx_queue));
159 LASSERT(list_empty(&peer->ksnp_zc_req_list));
161 LIBCFS_FREE(peer, sizeof(*peer));
163 /* NB a peer's connections and routes keep a reference on their peer
164 * until they are destroyed, so we can be assured that _all_ state to
165 * do with this peer has been cleaned up when its refcount drops to
167 spin_lock_bh(&net->ksnn_lock);
169 spin_unlock_bh(&net->ksnn_lock);
173 ksocknal_find_peer_locked (lnet_ni_t *ni, lnet_process_id_t id)
175 struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
176 struct list_head *tmp;
179 list_for_each(tmp, peer_list) {
181 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
183 LASSERT(!peer->ksnp_closing);
185 if (peer->ksnp_ni != ni)
188 if (peer->ksnp_id.nid != id.nid ||
189 peer->ksnp_id.pid != id.pid)
192 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
193 peer, libcfs_id2str(id),
194 atomic_read(&peer->ksnp_refcount));
201 ksocknal_find_peer (lnet_ni_t *ni, lnet_process_id_t id)
205 read_lock(&ksocknal_data.ksnd_global_lock);
206 peer = ksocknal_find_peer_locked(ni, id);
207 if (peer != NULL) /* +1 ref for caller? */
208 ksocknal_peer_addref(peer);
209 read_unlock(&ksocknal_data.ksnd_global_lock);
215 ksocknal_unlink_peer_locked (ksock_peer_t *peer)
219 ksock_interface_t *iface;
221 for (i = 0; i < peer->ksnp_n_passive_ips; i++) {
222 LASSERT (i < LNET_MAX_INTERFACES);
223 ip = peer->ksnp_passive_ips[i];
225 iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
226 /* All IPs in peer->ksnp_passive_ips[] come from the
227 * interface list, therefore the call must succeed. */
228 LASSERT (iface != NULL);
230 CDEBUG(D_NET, "peer=%p iface=%p ksni_nroutes=%d\n",
231 peer, iface, iface->ksni_nroutes);
232 iface->ksni_npeers--;
235 LASSERT(list_empty(&peer->ksnp_conns));
236 LASSERT(list_empty(&peer->ksnp_routes));
237 LASSERT(!peer->ksnp_closing);
238 peer->ksnp_closing = 1;
239 list_del(&peer->ksnp_list);
240 /* lose peerlist's ref */
241 ksocknal_peer_decref(peer);
245 ksocknal_get_peer_info (lnet_ni_t *ni, int index,
246 lnet_process_id_t *id, __u32 *myip, __u32 *peer_ip,
247 int *port, int *conn_count, int *share_count)
250 struct list_head *ptmp;
251 ksock_route_t *route;
252 struct list_head *rtmp;
257 read_lock(&ksocknal_data.ksnd_global_lock);
259 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
260 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
261 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
263 if (peer->ksnp_ni != ni)
266 if (peer->ksnp_n_passive_ips == 0 &&
267 list_empty(&peer->ksnp_routes)) {
281 for (j = 0; j < peer->ksnp_n_passive_ips; j++) {
286 *myip = peer->ksnp_passive_ips[j];
295 list_for_each(rtmp, &peer->ksnp_routes) {
299 route = list_entry(rtmp, ksock_route_t,
303 *myip = route->ksnr_myipaddr;
304 *peer_ip = route->ksnr_ipaddr;
305 *port = route->ksnr_port;
306 *conn_count = route->ksnr_conn_count;
307 *share_count = route->ksnr_share_count;
314 read_unlock(&ksocknal_data.ksnd_global_lock);
319 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
321 ksock_peer_t *peer = route->ksnr_peer;
322 int type = conn->ksnc_type;
323 ksock_interface_t *iface;
325 conn->ksnc_route = route;
326 ksocknal_route_addref(route);
328 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
329 if (route->ksnr_myipaddr == 0) {
330 /* route wasn't bound locally yet (the initial route) */
331 CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
332 libcfs_id2str(peer->ksnp_id),
334 &conn->ksnc_myipaddr);
336 CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
337 "to %pI4h\n", libcfs_id2str(peer->ksnp_id),
339 &route->ksnr_myipaddr,
340 &conn->ksnc_myipaddr);
342 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
343 route->ksnr_myipaddr);
345 iface->ksni_nroutes--;
347 route->ksnr_myipaddr = conn->ksnc_myipaddr;
348 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
349 route->ksnr_myipaddr);
351 iface->ksni_nroutes++;
354 route->ksnr_connected |= (1<<type);
355 route->ksnr_conn_count++;
357 /* Successful connection => further attempts can
358 * proceed immediately */
359 route->ksnr_retry_interval = 0;
363 ksocknal_add_route_locked (ksock_peer_t *peer, ksock_route_t *route)
365 struct list_head *tmp;
367 ksock_route_t *route2;
369 LASSERT(!peer->ksnp_closing);
370 LASSERT(route->ksnr_peer == NULL);
371 LASSERT(!route->ksnr_scheduled);
372 LASSERT(!route->ksnr_connecting);
373 LASSERT(route->ksnr_connected == 0);
375 /* LASSERT(unique) */
376 list_for_each(tmp, &peer->ksnp_routes) {
377 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
379 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
380 CERROR("Duplicate route %s %pI4h\n",
381 libcfs_id2str(peer->ksnp_id),
382 &route->ksnr_ipaddr);
387 route->ksnr_peer = peer;
388 ksocknal_peer_addref(peer);
389 /* peer's routelist takes over my ref on 'route' */
390 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
392 list_for_each(tmp, &peer->ksnp_conns) {
393 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
395 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
398 ksocknal_associate_route_conn_locked(route, conn);
399 /* keep going (typed routes) */
404 ksocknal_del_route_locked (ksock_route_t *route)
406 ksock_peer_t *peer = route->ksnr_peer;
407 ksock_interface_t *iface;
409 struct list_head *ctmp;
410 struct list_head *cnxt;
412 LASSERT(!route->ksnr_deleted);
414 /* Close associated conns */
415 list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
416 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
418 if (conn->ksnc_route != route)
421 ksocknal_close_conn_locked(conn, 0);
424 if (route->ksnr_myipaddr != 0) {
425 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
426 route->ksnr_myipaddr);
428 iface->ksni_nroutes--;
431 route->ksnr_deleted = 1;
432 list_del(&route->ksnr_list);
433 ksocknal_route_decref(route); /* drop peer's ref */
435 if (list_empty(&peer->ksnp_routes) &&
436 list_empty(&peer->ksnp_conns)) {
437 /* I've just removed the last route to a peer with no active
439 ksocknal_unlink_peer_locked(peer);
444 ksocknal_add_peer(lnet_ni_t *ni, lnet_process_id_t id, __u32 ipaddr, int port)
446 struct list_head *tmp;
449 ksock_route_t *route;
450 ksock_route_t *route2;
453 if (id.nid == LNET_NID_ANY ||
454 id.pid == LNET_PID_ANY)
457 /* Have a brand new peer ready... */
458 rc = ksocknal_create_peer(&peer, ni, id);
462 route = ksocknal_create_route (ipaddr, port);
464 ksocknal_peer_decref(peer);
468 write_lock_bh(&ksocknal_data.ksnd_global_lock);
470 /* always called with a ref on ni, so shutdown can't have started */
471 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
473 peer2 = ksocknal_find_peer_locked(ni, id);
475 ksocknal_peer_decref(peer);
478 /* peer table takes my ref on peer */
479 list_add_tail(&peer->ksnp_list,
480 ksocknal_nid2peerlist(id.nid));
484 list_for_each(tmp, &peer->ksnp_routes) {
485 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
487 if (route2->ksnr_ipaddr == ipaddr)
492 if (route2 == NULL) {
493 ksocknal_add_route_locked(peer, route);
494 route->ksnr_share_count++;
496 ksocknal_route_decref(route);
497 route2->ksnr_share_count++;
500 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
506 ksocknal_del_peer_locked (ksock_peer_t *peer, __u32 ip)
509 ksock_route_t *route;
510 struct list_head *tmp;
511 struct list_head *nxt;
514 LASSERT(!peer->ksnp_closing);
516 /* Extra ref prevents peer disappearing until I'm done with it */
517 ksocknal_peer_addref(peer);
519 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
520 route = list_entry(tmp, ksock_route_t, ksnr_list);
523 if (!(ip == 0 || route->ksnr_ipaddr == ip))
526 route->ksnr_share_count = 0;
527 /* This deletes associated conns too */
528 ksocknal_del_route_locked(route);
532 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
533 route = list_entry(tmp, ksock_route_t, ksnr_list);
534 nshared += route->ksnr_share_count;
538 /* remove everything else if there are no explicit entries
541 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
542 route = list_entry(tmp, ksock_route_t, ksnr_list);
544 /* we should only be removing auto-entries */
545 LASSERT(route->ksnr_share_count == 0);
546 ksocknal_del_route_locked(route);
549 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
550 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
552 ksocknal_close_conn_locked(conn, 0);
556 ksocknal_peer_decref(peer);
557 /* NB peer unlinks itself when last conn/route is removed */
561 ksocknal_del_peer (lnet_ni_t *ni, lnet_process_id_t id, __u32 ip)
563 struct list_head zombies = LIST_HEAD_INIT(zombies);
564 struct list_head *ptmp;
565 struct list_head *pnxt;
572 write_lock_bh(&ksocknal_data.ksnd_global_lock);
574 if (id.nid != LNET_NID_ANY) {
575 hi = (int)(ksocknal_nid2peerlist(id.nid) -
576 ksocknal_data.ksnd_peers);
580 hi = ksocknal_data.ksnd_peer_hash_size - 1;
583 for (i = lo; i <= hi; i++) {
584 list_for_each_safe(ptmp, pnxt,
585 &ksocknal_data.ksnd_peers[i]) {
586 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
588 if (peer->ksnp_ni != ni)
591 if (!((id.nid == LNET_NID_ANY ||
592 peer->ksnp_id.nid == id.nid) &&
593 (id.pid == LNET_PID_ANY ||
594 peer->ksnp_id.pid == id.pid)))
597 ksocknal_peer_addref(peer); /* a ref for me... */
599 ksocknal_del_peer_locked(peer, ip);
601 if (peer->ksnp_closing &&
602 !list_empty(&peer->ksnp_tx_queue)) {
603 LASSERT(list_empty(&peer->ksnp_conns));
604 LASSERT(list_empty(&peer->ksnp_routes));
606 list_splice_init(&peer->ksnp_tx_queue,
610 ksocknal_peer_decref(peer); /* ...till here */
612 rc = 0; /* matched! */
616 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
618 ksocknal_txlist_done(ni, &zombies, 1);
623 static ksock_conn_t *
624 ksocknal_get_conn_by_idx (lnet_ni_t *ni, int index)
627 struct list_head *ptmp;
629 struct list_head *ctmp;
632 read_lock(&ksocknal_data.ksnd_global_lock);
634 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
635 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
636 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
638 LASSERT(!peer->ksnp_closing);
640 if (peer->ksnp_ni != ni)
643 list_for_each(ctmp, &peer->ksnp_conns) {
647 conn = list_entry(ctmp, ksock_conn_t,
649 ksocknal_conn_addref(conn);
650 read_unlock(&ksocknal_data. \
657 read_unlock(&ksocknal_data.ksnd_global_lock);
661 static ksock_sched_t *
662 ksocknal_choose_scheduler_locked(unsigned int cpt)
664 struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
665 ksock_sched_t *sched;
668 LASSERT(info->ksi_nthreads > 0);
670 sched = &info->ksi_scheds[0];
672 * NB: it's safe so far, but info->ksi_nthreads could be changed
673 * at runtime when we have dynamic LNet configuration, then we
674 * need to take care of this.
676 for (i = 1; i < info->ksi_nthreads; i++) {
677 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
678 sched = &info->ksi_scheds[i];
685 ksocknal_local_ipvec (lnet_ni_t *ni, __u32 *ipaddrs)
687 ksock_net_t *net = ni->ni_data;
691 read_lock(&ksocknal_data.ksnd_global_lock);
693 nip = net->ksnn_ninterfaces;
694 LASSERT (nip <= LNET_MAX_INTERFACES);
696 /* Only offer interfaces for additional connections if I have
699 read_unlock(&ksocknal_data.ksnd_global_lock);
703 for (i = 0; i < nip; i++) {
704 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
705 LASSERT (ipaddrs[i] != 0);
708 read_unlock(&ksocknal_data.ksnd_global_lock);
713 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
715 int best_netmatch = 0;
722 for (i = 0; i < nips; i++) {
726 this_xor = (ips[i] ^ iface->ksni_ipaddr);
727 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
730 best_netmatch < this_netmatch ||
731 (best_netmatch == this_netmatch &&
732 best_xor > this_xor)))
736 best_netmatch = this_netmatch;
745 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
747 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
748 ksock_net_t *net = peer->ksnp_ni->ni_data;
749 ksock_interface_t *iface;
750 ksock_interface_t *best_iface;
761 /* CAVEAT EMPTOR: We do all our interface matching with an
762 * exclusive hold of global lock at IRQ priority. We're only
763 * expecting to be dealing with small numbers of interfaces, so the
764 * O(n**3)-ness shouldn't matter */
766 /* Also note that I'm not going to return more than n_peerips
767 * interfaces, even if I have more myself */
769 write_lock_bh(global_lock);
771 LASSERT (n_peerips <= LNET_MAX_INTERFACES);
772 LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
774 /* Only match interfaces for additional connections
775 * if I have > 1 interface */
776 n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
777 MIN(n_peerips, net->ksnn_ninterfaces);
779 for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
780 /* ^ yes really... */
782 /* If we have any new interfaces, first tick off all the
783 * peer IPs that match old interfaces, then choose new
784 * interfaces to match the remaining peer IPS.
785 * We don't forget interfaces we've stopped using; we might
786 * start using them again... */
788 if (i < peer->ksnp_n_passive_ips) {
790 ip = peer->ksnp_passive_ips[i];
791 best_iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
793 /* peer passive ips are kept up to date */
794 LASSERT(best_iface != NULL);
796 /* choose a new interface */
797 LASSERT (i == peer->ksnp_n_passive_ips);
803 for (j = 0; j < net->ksnn_ninterfaces; j++) {
804 iface = &net->ksnn_interfaces[j];
805 ip = iface->ksni_ipaddr;
807 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
808 if (peer->ksnp_passive_ips[k] == ip)
811 if (k < peer->ksnp_n_passive_ips) /* using it already */
814 k = ksocknal_match_peerip(iface, peerips, n_peerips);
815 xor = (ip ^ peerips[k]);
816 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
818 if (!(best_iface == NULL ||
819 best_netmatch < this_netmatch ||
820 (best_netmatch == this_netmatch &&
821 best_npeers > iface->ksni_npeers)))
825 best_netmatch = this_netmatch;
826 best_npeers = iface->ksni_npeers;
829 LASSERT(best_iface != NULL);
831 best_iface->ksni_npeers++;
832 ip = best_iface->ksni_ipaddr;
833 peer->ksnp_passive_ips[i] = ip;
834 peer->ksnp_n_passive_ips = i+1;
837 /* mark the best matching peer IP used */
838 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
842 /* Overwrite input peer IP addresses */
843 memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
845 write_unlock_bh(global_lock);
851 ksocknal_create_routes(ksock_peer_t *peer, int port,
852 __u32 *peer_ipaddrs, int npeer_ipaddrs)
854 ksock_route_t *newroute = NULL;
855 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
856 lnet_ni_t *ni = peer->ksnp_ni;
857 ksock_net_t *net = ni->ni_data;
858 struct list_head *rtmp;
859 ksock_route_t *route;
860 ksock_interface_t *iface;
861 ksock_interface_t *best_iface;
868 /* CAVEAT EMPTOR: We do all our interface matching with an
869 * exclusive hold of global lock at IRQ priority. We're only
870 * expecting to be dealing with small numbers of interfaces, so the
871 * O(n**3)-ness here shouldn't matter */
873 write_lock_bh(global_lock);
875 if (net->ksnn_ninterfaces < 2) {
876 /* Only create additional connections
877 * if I have > 1 interface */
878 write_unlock_bh(global_lock);
882 LASSERT (npeer_ipaddrs <= LNET_MAX_INTERFACES);
884 for (i = 0; i < npeer_ipaddrs; i++) {
885 if (newroute != NULL) {
886 newroute->ksnr_ipaddr = peer_ipaddrs[i];
888 write_unlock_bh(global_lock);
890 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
891 if (newroute == NULL)
894 write_lock_bh(global_lock);
897 if (peer->ksnp_closing) {
898 /* peer got closed under me */
902 /* Already got a route? */
904 list_for_each(rtmp, &peer->ksnp_routes) {
905 route = list_entry(rtmp, ksock_route_t, ksnr_list);
907 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
919 LASSERT(net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
921 /* Select interface to connect from */
922 for (j = 0; j < net->ksnn_ninterfaces; j++) {
923 iface = &net->ksnn_interfaces[j];
925 /* Using this interface already? */
926 list_for_each(rtmp, &peer->ksnp_routes) {
927 route = list_entry(rtmp, ksock_route_t,
930 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
938 this_netmatch = (((iface->ksni_ipaddr ^
939 newroute->ksnr_ipaddr) &
940 iface->ksni_netmask) == 0) ? 1 : 0;
942 if (!(best_iface == NULL ||
943 best_netmatch < this_netmatch ||
944 (best_netmatch == this_netmatch &&
945 best_nroutes > iface->ksni_nroutes)))
949 best_netmatch = this_netmatch;
950 best_nroutes = iface->ksni_nroutes;
953 if (best_iface == NULL)
956 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
957 best_iface->ksni_nroutes++;
959 ksocknal_add_route_locked(peer, newroute);
963 write_unlock_bh(global_lock);
964 if (newroute != NULL)
965 ksocknal_route_decref(newroute);
969 ksocknal_accept(lnet_ni_t *ni, struct socket *sock)
976 rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
977 LASSERT(rc == 0); /* we succeeded before */
979 LIBCFS_ALLOC(cr, sizeof(*cr));
981 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
982 "%pI4h: memory exhausted\n", &peer_ip);
988 cr->ksncr_sock = sock;
990 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
992 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
993 wake_up(&ksocknal_data.ksnd_connd_waitq);
995 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1000 ksocknal_connecting (ksock_peer_t *peer, __u32 ipaddr)
1002 ksock_route_t *route;
1004 list_for_each_entry(route, &peer->ksnp_routes, ksnr_list) {
1005 if (route->ksnr_ipaddr == ipaddr)
1006 return route->ksnr_connecting;
1012 ksocknal_create_conn(lnet_ni_t *ni, ksock_route_t *route,
1013 struct socket *sock, int type)
1015 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1016 struct list_head zombies = LIST_HEAD_INIT(zombies);
1017 lnet_process_id_t peerid;
1018 struct list_head *tmp;
1021 ksock_conn_t *conn2;
1022 ksock_peer_t *peer = NULL;
1023 ksock_peer_t *peer2;
1024 ksock_sched_t *sched;
1025 ksock_hello_msg_t *hello;
1033 active = (route != NULL);
1035 LASSERT (active == (type != SOCKLND_CONN_NONE));
1037 LIBCFS_ALLOC(conn, sizeof(*conn));
1043 conn->ksnc_peer = NULL;
1044 conn->ksnc_route = NULL;
1045 conn->ksnc_sock = sock;
1046 /* 2 ref, 1 for conn, another extra ref prevents socket
1047 * being closed before establishment of connection */
1048 atomic_set (&conn->ksnc_sock_refcount, 2);
1049 conn->ksnc_type = type;
1050 ksocknal_lib_save_callback(sock, conn);
1051 atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1053 conn->ksnc_rx_ready = 0;
1054 conn->ksnc_rx_scheduled = 0;
1056 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1057 conn->ksnc_tx_ready = 0;
1058 conn->ksnc_tx_scheduled = 0;
1059 conn->ksnc_tx_carrier = NULL;
1060 atomic_set (&conn->ksnc_tx_nob, 0);
1062 LIBCFS_ALLOC(hello, offsetof(ksock_hello_msg_t,
1063 kshm_ips[LNET_MAX_INTERFACES]));
1064 if (hello == NULL) {
1069 /* stash conn's local and remote addrs */
1070 rc = ksocknal_lib_get_conn_addrs (conn);
1074 /* Find out/confirm peer's NID and connection type and get the
1075 * vector of interfaces she's willing to let me connect to.
1076 * Passive connections use the listener timeout since the peer sends
1080 peer = route->ksnr_peer;
1081 LASSERT(ni == peer->ksnp_ni);
1083 /* Active connection sends HELLO eagerly */
1084 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1085 peerid = peer->ksnp_id;
1087 write_lock_bh(global_lock);
1088 conn->ksnc_proto = peer->ksnp_proto;
1089 write_unlock_bh(global_lock);
1091 if (conn->ksnc_proto == NULL) {
1092 conn->ksnc_proto = &ksocknal_protocol_v3x;
1093 #if SOCKNAL_VERSION_DEBUG
1094 if (*ksocknal_tunables.ksnd_protocol == 2)
1095 conn->ksnc_proto = &ksocknal_protocol_v2x;
1096 else if (*ksocknal_tunables.ksnd_protocol == 1)
1097 conn->ksnc_proto = &ksocknal_protocol_v1x;
1101 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1105 peerid.nid = LNET_NID_ANY;
1106 peerid.pid = LNET_PID_ANY;
1108 /* Passive, get protocol from peer */
1109 conn->ksnc_proto = NULL;
1112 rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1116 LASSERT (rc == 0 || active);
1117 LASSERT (conn->ksnc_proto != NULL);
1118 LASSERT (peerid.nid != LNET_NID_ANY);
1120 cpt = lnet_cpt_of_nid(peerid.nid);
1123 ksocknal_peer_addref(peer);
1124 write_lock_bh(global_lock);
1126 rc = ksocknal_create_peer(&peer, ni, peerid);
1130 write_lock_bh(global_lock);
1132 /* called with a ref on ni, so shutdown can't have started */
1133 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1135 peer2 = ksocknal_find_peer_locked(ni, peerid);
1136 if (peer2 == NULL) {
1137 /* NB this puts an "empty" peer in the peer
1138 * table (which takes my ref) */
1139 list_add_tail(&peer->ksnp_list,
1140 ksocknal_nid2peerlist(peerid.nid));
1142 ksocknal_peer_decref(peer);
1147 ksocknal_peer_addref(peer);
1148 peer->ksnp_accepting++;
1150 /* Am I already connecting to this guy? Resolve in
1151 * favour of higher NID... */
1152 if (peerid.nid < ni->ni_nid &&
1153 ksocknal_connecting(peer, conn->ksnc_ipaddr)) {
1155 warn = "connection race resolution";
1160 if (peer->ksnp_closing ||
1161 (active && route->ksnr_deleted)) {
1162 /* peer/route got closed under me */
1164 warn = "peer/route removed";
1168 if (peer->ksnp_proto == NULL) {
1169 /* Never connected before.
1170 * NB recv_hello may have returned EPROTO to signal my peer
1171 * wants a different protocol than the one I asked for.
1173 LASSERT(list_empty(&peer->ksnp_conns));
1175 peer->ksnp_proto = conn->ksnc_proto;
1176 peer->ksnp_incarnation = incarnation;
1179 if (peer->ksnp_proto != conn->ksnc_proto ||
1180 peer->ksnp_incarnation != incarnation) {
1181 /* Peer rebooted or I've got the wrong protocol version */
1182 ksocknal_close_peer_conns_locked(peer, 0, 0);
1184 peer->ksnp_proto = NULL;
1186 warn = peer->ksnp_incarnation != incarnation ?
1188 "wrong proto version";
1198 warn = "lost conn race";
1201 warn = "retry with different protocol version";
1205 /* Refuse to duplicate an existing connection, unless this is a
1206 * loopback connection */
1207 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1208 list_for_each(tmp, &peer->ksnp_conns) {
1209 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1211 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1212 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1213 conn2->ksnc_type != conn->ksnc_type)
1216 /* Reply on a passive connection attempt so the peer
1217 * realises we're connected. */
1227 /* If the connection created by this route didn't bind to the IP
1228 * address the route connected to, the connection/route matching
1229 * code below probably isn't going to work. */
1231 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1232 CERROR("Route %s %pI4h connected to %pI4h\n",
1233 libcfs_id2str(peer->ksnp_id),
1234 &route->ksnr_ipaddr,
1235 &conn->ksnc_ipaddr);
1238 /* Search for a route corresponding to the new connection and
1239 * create an association. This allows incoming connections created
1240 * by routes in my peer to match my own route entries so I don't
1241 * continually create duplicate routes. */
1242 list_for_each(tmp, &peer->ksnp_routes) {
1243 route = list_entry(tmp, ksock_route_t, ksnr_list);
1245 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1248 ksocknal_associate_route_conn_locked(route, conn);
1252 conn->ksnc_peer = peer; /* conn takes my ref on peer */
1253 peer->ksnp_last_alive = cfs_time_current();
1254 peer->ksnp_send_keepalive = 0;
1255 peer->ksnp_error = 0;
1257 sched = ksocknal_choose_scheduler_locked(cpt);
1258 sched->kss_nconns++;
1259 conn->ksnc_scheduler = sched;
1261 conn->ksnc_tx_last_post = cfs_time_current();
1262 /* Set the deadline for the outgoing HELLO to drain */
1263 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1264 conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1265 smp_mb(); /* order with adding to peer's conn list */
1267 list_add(&conn->ksnc_list, &peer->ksnp_conns);
1268 ksocknal_conn_addref(conn);
1270 ksocknal_new_packet(conn, 0);
1272 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1274 /* Take packets blocking for this connection. */
1275 list_for_each_entry_safe(tx, txtmp, &peer->ksnp_tx_queue, tx_list) {
1276 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1280 list_del(&tx->tx_list);
1281 ksocknal_queue_tx_locked(tx, conn);
1284 write_unlock_bh(global_lock);
1286 /* We've now got a new connection. Any errors from here on are just
1287 * like "normal" comms errors and we close the connection normally.
1288 * NB (a) we still have to send the reply HELLO for passive
1290 * (b) normal I/O on the conn is blocked until I setup and call the
1294 CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1295 " incarnation:%lld sched[%d:%d]\n",
1296 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1297 &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1298 conn->ksnc_port, incarnation, cpt,
1299 (int)(sched - &sched->kss_info->ksi_scheds[0]));
1302 /* additional routes after interface exchange? */
1303 ksocknal_create_routes(peer, conn->ksnc_port,
1304 hello->kshm_ips, hello->kshm_nips);
1306 hello->kshm_nips = ksocknal_select_ips(peer, hello->kshm_ips,
1308 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1311 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1312 kshm_ips[LNET_MAX_INTERFACES]));
1314 /* setup the socket AFTER I've received hello (it disables
1315 * SO_LINGER). I might call back to the acceptor who may want
1316 * to send a protocol version response and then close the
1317 * socket; this ensures the socket only tears down after the
1318 * response has been sent. */
1320 rc = ksocknal_lib_setup_sock(sock);
1322 write_lock_bh(global_lock);
1324 /* NB my callbacks block while I hold ksnd_global_lock */
1325 ksocknal_lib_set_callback(sock, conn);
1328 peer->ksnp_accepting--;
1330 write_unlock_bh(global_lock);
1333 write_lock_bh(global_lock);
1334 if (!conn->ksnc_closing) {
1335 /* could be closed by another thread */
1336 ksocknal_close_conn_locked(conn, rc);
1338 write_unlock_bh(global_lock);
1339 } else if (ksocknal_connsock_addref(conn) == 0) {
1340 /* Allow I/O to proceed. */
1341 ksocknal_read_callback(conn);
1342 ksocknal_write_callback(conn);
1343 ksocknal_connsock_decref(conn);
1346 ksocknal_connsock_decref(conn);
1347 ksocknal_conn_decref(conn);
1351 if (!peer->ksnp_closing &&
1352 list_empty(&peer->ksnp_conns) &&
1353 list_empty(&peer->ksnp_routes)) {
1354 list_add(&zombies, &peer->ksnp_tx_queue);
1355 list_del_init(&peer->ksnp_tx_queue);
1356 ksocknal_unlink_peer_locked(peer);
1359 write_unlock_bh(global_lock);
1363 CERROR("Not creating conn %s type %d: %s\n",
1364 libcfs_id2str(peerid), conn->ksnc_type, warn);
1366 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1367 libcfs_id2str(peerid), conn->ksnc_type, warn);
1372 /* Request retry by replying with CONN_NONE
1373 * ksnc_proto has been set already */
1374 conn->ksnc_type = SOCKLND_CONN_NONE;
1375 hello->kshm_nips = 0;
1376 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1379 write_lock_bh(global_lock);
1380 peer->ksnp_accepting--;
1381 write_unlock_bh(global_lock);
1384 ksocknal_txlist_done(ni, &zombies, 1);
1385 ksocknal_peer_decref(peer);
1389 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1390 kshm_ips[LNET_MAX_INTERFACES]));
1392 LIBCFS_FREE(conn, sizeof(*conn));
1400 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1402 /* This just does the immmediate housekeeping, and queues the
1403 * connection for the reaper to terminate.
1404 * Caller holds ksnd_global_lock exclusively in irq context */
1405 ksock_peer_t *peer = conn->ksnc_peer;
1406 ksock_route_t *route;
1407 ksock_conn_t *conn2;
1408 struct list_head *tmp;
1410 LASSERT(peer->ksnp_error == 0);
1411 LASSERT(!conn->ksnc_closing);
1412 conn->ksnc_closing = 1;
1414 /* ksnd_deathrow_conns takes over peer's ref */
1415 list_del(&conn->ksnc_list);
1417 route = conn->ksnc_route;
1418 if (route != NULL) {
1419 /* dissociate conn from route... */
1420 LASSERT(!route->ksnr_deleted);
1421 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1424 list_for_each(tmp, &peer->ksnp_conns) {
1425 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1427 if (conn2->ksnc_route == route &&
1428 conn2->ksnc_type == conn->ksnc_type)
1434 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1436 conn->ksnc_route = NULL;
1438 ksocknal_route_decref(route); /* drop conn's ref on route */
1441 if (list_empty(&peer->ksnp_conns)) {
1442 /* No more connections to this peer */
1444 if (!list_empty(&peer->ksnp_tx_queue)) {
1447 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1449 /* throw them to the last connection...,
1450 * these TXs will be send to /dev/null by scheduler */
1451 list_for_each_entry(tx, &peer->ksnp_tx_queue,
1453 ksocknal_tx_prep(conn, tx);
1455 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1456 list_splice_init(&peer->ksnp_tx_queue,
1457 &conn->ksnc_tx_queue);
1458 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1461 /* renegotiate protocol version */
1462 peer->ksnp_proto = NULL;
1463 /* stash last conn close reason */
1464 peer->ksnp_error = error;
1466 if (list_empty(&peer->ksnp_routes)) {
1467 /* I've just closed last conn belonging to a
1468 * peer with no routes to it */
1469 ksocknal_unlink_peer_locked(peer);
1473 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1475 list_add_tail(&conn->ksnc_list,
1476 &ksocknal_data.ksnd_deathrow_conns);
1477 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1479 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1483 ksocknal_peer_failed (ksock_peer_t *peer)
1486 cfs_time_t last_alive = 0;
1488 /* There has been a connection failure or comms error; but I'll only
1489 * tell LNET I think the peer is dead if it's to another kernel and
1490 * there are no connections or connection attempts in existence. */
1492 read_lock(&ksocknal_data.ksnd_global_lock);
1494 if ((peer->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1495 list_empty(&peer->ksnp_conns) &&
1496 peer->ksnp_accepting == 0 &&
1497 ksocknal_find_connecting_route_locked(peer) == NULL) {
1499 last_alive = peer->ksnp_last_alive;
1502 read_unlock(&ksocknal_data.ksnd_global_lock);
1505 lnet_notify(peer->ksnp_ni, peer->ksnp_id.nid, 0,
1510 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1512 ksock_peer_t *peer = conn->ksnc_peer;
1515 struct list_head zlist = LIST_HEAD_INIT(zlist);
1517 /* NB safe to finalize TXs because closing of socket will
1518 * abort all buffered data */
1519 LASSERT(conn->ksnc_sock == NULL);
1521 spin_lock(&peer->ksnp_lock);
1523 list_for_each_entry_safe(tx, tmp, &peer->ksnp_zc_req_list, tx_zc_list) {
1524 if (tx->tx_conn != conn)
1527 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1529 tx->tx_msg.ksm_zc_cookies[0] = 0;
1530 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1531 list_del(&tx->tx_zc_list);
1532 list_add(&tx->tx_zc_list, &zlist);
1535 spin_unlock(&peer->ksnp_lock);
1537 while (!list_empty(&zlist)) {
1538 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1540 list_del(&tx->tx_zc_list);
1541 ksocknal_tx_decref(tx);
1546 ksocknal_terminate_conn(ksock_conn_t *conn)
1548 /* This gets called by the reaper (guaranteed thread context) to
1549 * disengage the socket from its callbacks and close it.
1550 * ksnc_refcount will eventually hit zero, and then the reaper will
1552 ksock_peer_t *peer = conn->ksnc_peer;
1553 ksock_sched_t *sched = conn->ksnc_scheduler;
1556 LASSERT(conn->ksnc_closing);
1558 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1559 spin_lock_bh(&sched->kss_lock);
1561 /* a closing conn is always ready to tx */
1562 conn->ksnc_tx_ready = 1;
1564 if (!conn->ksnc_tx_scheduled &&
1565 !list_empty(&conn->ksnc_tx_queue)) {
1566 list_add_tail(&conn->ksnc_tx_list,
1567 &sched->kss_tx_conns);
1568 conn->ksnc_tx_scheduled = 1;
1569 /* extra ref for scheduler */
1570 ksocknal_conn_addref(conn);
1572 wake_up (&sched->kss_waitq);
1575 spin_unlock_bh(&sched->kss_lock);
1577 /* serialise with callbacks */
1578 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1580 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1582 /* OK, so this conn may not be completely disengaged from its
1583 * scheduler yet, but it _has_ committed to terminate... */
1584 conn->ksnc_scheduler->kss_nconns--;
1586 if (peer->ksnp_error != 0) {
1587 /* peer's last conn closed in error */
1588 LASSERT(list_empty(&peer->ksnp_conns));
1590 peer->ksnp_error = 0; /* avoid multiple notifications */
1593 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1596 ksocknal_peer_failed(peer);
1598 /* The socket is closed on the final put; either here, or in
1599 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1600 * when the connection was established, this will close the socket
1601 * immediately, aborting anything buffered in it. Any hung
1602 * zero-copy transmits will therefore complete in finite time. */
1603 ksocknal_connsock_decref(conn);
1607 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1609 /* Queue the conn for the reaper to destroy */
1611 LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1612 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1614 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1615 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1617 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1621 ksocknal_destroy_conn (ksock_conn_t *conn)
1623 cfs_time_t last_rcv;
1625 /* Final coup-de-grace of the reaper */
1626 CDEBUG (D_NET, "connection %p\n", conn);
1628 LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1629 LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1630 LASSERT (conn->ksnc_sock == NULL);
1631 LASSERT (conn->ksnc_route == NULL);
1632 LASSERT (!conn->ksnc_tx_scheduled);
1633 LASSERT (!conn->ksnc_rx_scheduled);
1634 LASSERT(list_empty(&conn->ksnc_tx_queue));
1636 /* complete current receive if any */
1637 switch (conn->ksnc_rx_state) {
1638 case SOCKNAL_RX_LNET_PAYLOAD:
1639 last_rcv = conn->ksnc_rx_deadline -
1640 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1641 CERROR("Completing partial receive from %s[%d], "
1642 "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1643 "last alive is %ld secs ago\n",
1644 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1645 &conn->ksnc_ipaddr, conn->ksnc_port,
1646 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1647 cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1649 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1650 conn->ksnc_cookie, -EIO);
1652 case SOCKNAL_RX_LNET_HEADER:
1653 if (conn->ksnc_rx_started)
1654 CERROR("Incomplete receive of lnet header from %s, "
1655 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1656 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1657 &conn->ksnc_ipaddr, conn->ksnc_port,
1658 conn->ksnc_proto->pro_version);
1660 case SOCKNAL_RX_KSM_HEADER:
1661 if (conn->ksnc_rx_started)
1662 CERROR("Incomplete receive of ksock message from %s, "
1663 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1664 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1665 &conn->ksnc_ipaddr, conn->ksnc_port,
1666 conn->ksnc_proto->pro_version);
1668 case SOCKNAL_RX_SLOP:
1669 if (conn->ksnc_rx_started)
1670 CERROR("Incomplete receive of slops from %s, "
1671 "ip %pI4h:%d, with error\n",
1672 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1673 &conn->ksnc_ipaddr, conn->ksnc_port);
1680 ksocknal_peer_decref(conn->ksnc_peer);
1682 LIBCFS_FREE (conn, sizeof (*conn));
1686 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1689 struct list_head *ctmp;
1690 struct list_head *cnxt;
1693 list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
1694 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1697 conn->ksnc_ipaddr == ipaddr) {
1699 ksocknal_close_conn_locked (conn, why);
1707 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1709 ksock_peer_t *peer = conn->ksnc_peer;
1710 __u32 ipaddr = conn->ksnc_ipaddr;
1713 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1715 count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1717 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1723 ksocknal_close_matching_conns (lnet_process_id_t id, __u32 ipaddr)
1726 struct list_head *ptmp;
1727 struct list_head *pnxt;
1733 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1735 if (id.nid != LNET_NID_ANY)
1736 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1739 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1742 for (i = lo; i <= hi; i++) {
1743 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1745 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1747 if (!((id.nid == LNET_NID_ANY || id.nid == peer->ksnp_id.nid) &&
1748 (id.pid == LNET_PID_ANY || id.pid == peer->ksnp_id.pid)))
1751 count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1755 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1757 /* wildcards always succeed */
1758 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1761 return (count == 0 ? -ENOENT : 0);
1765 ksocknal_notify (lnet_ni_t *ni, lnet_nid_t gw_nid, int alive)
1767 /* The router is telling me she's been notified of a change in
1768 * gateway state.... */
1769 lnet_process_id_t id = {0};
1772 id.pid = LNET_PID_ANY;
1774 CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1775 alive ? "up" : "down");
1778 /* If the gateway crashed, close all open connections... */
1779 ksocknal_close_matching_conns (id, 0);
1783 /* ...otherwise do nothing. We can only establish new connections
1784 * if we have autroutes, and these connect on demand. */
1788 ksocknal_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1791 cfs_time_t last_alive = 0;
1792 cfs_time_t now = cfs_time_current();
1793 ksock_peer_t *peer = NULL;
1794 rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1795 lnet_process_id_t id = {
1797 .pid = LNET_PID_LUSTRE,
1802 peer = ksocknal_find_peer_locked(ni, id);
1804 struct list_head *tmp;
1808 list_for_each(tmp, &peer->ksnp_conns) {
1809 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1810 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1812 if (bufnob < conn->ksnc_tx_bufnob) {
1813 /* something got ACKed */
1814 conn->ksnc_tx_deadline =
1815 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1816 peer->ksnp_last_alive = now;
1817 conn->ksnc_tx_bufnob = bufnob;
1821 last_alive = peer->ksnp_last_alive;
1822 if (ksocknal_find_connectable_route_locked(peer) == NULL)
1828 if (last_alive != 0)
1831 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago, connect %d\n",
1832 libcfs_nid2str(nid), peer,
1833 last_alive ? cfs_duration_sec(now - last_alive) : -1,
1839 ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1841 write_lock_bh(glock);
1843 peer = ksocknal_find_peer_locked(ni, id);
1845 ksocknal_launch_all_connections_locked(peer);
1847 write_unlock_bh(glock);
1852 ksocknal_push_peer (ksock_peer_t *peer)
1856 struct list_head *tmp;
1859 for (index = 0; ; index++) {
1860 read_lock(&ksocknal_data.ksnd_global_lock);
1865 list_for_each(tmp, &peer->ksnp_conns) {
1867 conn = list_entry(tmp, ksock_conn_t,
1869 ksocknal_conn_addref(conn);
1874 read_unlock(&ksocknal_data.ksnd_global_lock);
1879 ksocknal_lib_push_conn (conn);
1880 ksocknal_conn_decref(conn);
1885 ksocknal_push (lnet_ni_t *ni, lnet_process_id_t id)
1887 struct list_head *start;
1888 struct list_head *end;
1889 struct list_head *tmp;
1891 unsigned int hsize = ksocknal_data.ksnd_peer_hash_size;
1893 if (id.nid == LNET_NID_ANY) {
1894 start = &ksocknal_data.ksnd_peers[0];
1895 end = &ksocknal_data.ksnd_peers[hsize - 1];
1897 start = end = ksocknal_nid2peerlist(id.nid);
1900 for (tmp = start; tmp <= end; tmp++) {
1901 int peer_off; /* searching offset in peer hash table */
1903 for (peer_off = 0; ; peer_off++) {
1907 read_lock(&ksocknal_data.ksnd_global_lock);
1908 list_for_each_entry(peer, tmp, ksnp_list) {
1909 if (!((id.nid == LNET_NID_ANY ||
1910 id.nid == peer->ksnp_id.nid) &&
1911 (id.pid == LNET_PID_ANY ||
1912 id.pid == peer->ksnp_id.pid)))
1915 if (i++ == peer_off) {
1916 ksocknal_peer_addref(peer);
1920 read_unlock(&ksocknal_data.ksnd_global_lock);
1922 if (i == 0) /* no match */
1926 ksocknal_push_peer(peer);
1927 ksocknal_peer_decref(peer);
1934 ksocknal_add_interface(lnet_ni_t *ni, __u32 ipaddress, __u32 netmask)
1936 ksock_net_t *net = ni->ni_data;
1937 ksock_interface_t *iface;
1941 struct list_head *ptmp;
1943 struct list_head *rtmp;
1944 ksock_route_t *route;
1946 if (ipaddress == 0 ||
1950 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1952 iface = ksocknal_ip2iface(ni, ipaddress);
1953 if (iface != NULL) {
1954 /* silently ignore dups */
1956 } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1959 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1961 iface->ksni_ipaddr = ipaddress;
1962 iface->ksni_netmask = netmask;
1963 iface->ksni_nroutes = 0;
1964 iface->ksni_npeers = 0;
1966 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1967 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1968 peer = list_entry(ptmp, ksock_peer_t,
1971 for (j = 0; j < peer->ksnp_n_passive_ips; j++)
1972 if (peer->ksnp_passive_ips[j] == ipaddress)
1973 iface->ksni_npeers++;
1975 list_for_each(rtmp, &peer->ksnp_routes) {
1976 route = list_entry(rtmp,
1980 if (route->ksnr_myipaddr == ipaddress)
1981 iface->ksni_nroutes++;
1987 /* NB only new connections will pay attention to the new interface! */
1990 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1996 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
1998 struct list_head *tmp;
1999 struct list_head *nxt;
2000 ksock_route_t *route;
2005 for (i = 0; i < peer->ksnp_n_passive_ips; i++)
2006 if (peer->ksnp_passive_ips[i] == ipaddr) {
2007 for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
2008 peer->ksnp_passive_ips[j-1] =
2009 peer->ksnp_passive_ips[j];
2010 peer->ksnp_n_passive_ips--;
2014 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
2015 route = list_entry(tmp, ksock_route_t, ksnr_list);
2017 if (route->ksnr_myipaddr != ipaddr)
2020 if (route->ksnr_share_count != 0) {
2021 /* Manually created; keep, but unbind */
2022 route->ksnr_myipaddr = 0;
2024 ksocknal_del_route_locked(route);
2028 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
2029 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2031 if (conn->ksnc_myipaddr == ipaddr)
2032 ksocknal_close_conn_locked (conn, 0);
2037 ksocknal_del_interface(lnet_ni_t *ni, __u32 ipaddress)
2039 ksock_net_t *net = ni->ni_data;
2041 struct list_head *tmp;
2042 struct list_head *nxt;
2048 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2050 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2051 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2053 if (!(ipaddress == 0 ||
2054 ipaddress == this_ip))
2059 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2060 net->ksnn_interfaces[j-1] =
2061 net->ksnn_interfaces[j];
2063 net->ksnn_ninterfaces--;
2065 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2066 list_for_each_safe(tmp, nxt,
2067 &ksocknal_data.ksnd_peers[j]) {
2068 peer = list_entry(tmp, ksock_peer_t,
2071 if (peer->ksnp_ni != ni)
2074 ksocknal_peer_del_interface_locked(peer, this_ip);
2079 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2085 ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
2087 lnet_process_id_t id = {0};
2088 struct libcfs_ioctl_data *data = arg;
2092 case IOC_LIBCFS_GET_INTERFACE: {
2093 ksock_net_t *net = ni->ni_data;
2094 ksock_interface_t *iface;
2096 read_lock(&ksocknal_data.ksnd_global_lock);
2098 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2102 iface = &net->ksnn_interfaces[data->ioc_count];
2104 data->ioc_u32[0] = iface->ksni_ipaddr;
2105 data->ioc_u32[1] = iface->ksni_netmask;
2106 data->ioc_u32[2] = iface->ksni_npeers;
2107 data->ioc_u32[3] = iface->ksni_nroutes;
2110 read_unlock(&ksocknal_data.ksnd_global_lock);
2114 case IOC_LIBCFS_ADD_INTERFACE:
2115 return ksocknal_add_interface(ni,
2116 data->ioc_u32[0], /* IP address */
2117 data->ioc_u32[1]); /* net mask */
2119 case IOC_LIBCFS_DEL_INTERFACE:
2120 return ksocknal_del_interface(ni,
2121 data->ioc_u32[0]); /* IP address */
2123 case IOC_LIBCFS_GET_PEER: {
2128 int share_count = 0;
2130 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2131 &id, &myip, &ip, &port,
2132 &conn_count, &share_count);
2136 data->ioc_nid = id.nid;
2137 data->ioc_count = share_count;
2138 data->ioc_u32[0] = ip;
2139 data->ioc_u32[1] = port;
2140 data->ioc_u32[2] = myip;
2141 data->ioc_u32[3] = conn_count;
2142 data->ioc_u32[4] = id.pid;
2146 case IOC_LIBCFS_ADD_PEER:
2147 id.nid = data->ioc_nid;
2148 id.pid = LNET_PID_LUSTRE;
2149 return ksocknal_add_peer (ni, id,
2150 data->ioc_u32[0], /* IP */
2151 data->ioc_u32[1]); /* port */
2153 case IOC_LIBCFS_DEL_PEER:
2154 id.nid = data->ioc_nid;
2155 id.pid = LNET_PID_ANY;
2156 return ksocknal_del_peer (ni, id,
2157 data->ioc_u32[0]); /* IP */
2159 case IOC_LIBCFS_GET_CONN: {
2163 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2168 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2170 data->ioc_count = txmem;
2171 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
2172 data->ioc_flags = nagle;
2173 data->ioc_u32[0] = conn->ksnc_ipaddr;
2174 data->ioc_u32[1] = conn->ksnc_port;
2175 data->ioc_u32[2] = conn->ksnc_myipaddr;
2176 data->ioc_u32[3] = conn->ksnc_type;
2177 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2178 data->ioc_u32[5] = rxmem;
2179 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2180 ksocknal_conn_decref(conn);
2184 case IOC_LIBCFS_CLOSE_CONNECTION:
2185 id.nid = data->ioc_nid;
2186 id.pid = LNET_PID_ANY;
2187 return ksocknal_close_matching_conns (id,
2190 case IOC_LIBCFS_REGISTER_MYNID:
2191 /* Ignore if this is a noop */
2192 if (data->ioc_nid == ni->ni_nid)
2195 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2196 libcfs_nid2str(data->ioc_nid),
2197 libcfs_nid2str(ni->ni_nid));
2200 case IOC_LIBCFS_PUSH_CONNECTION:
2201 id.nid = data->ioc_nid;
2202 id.pid = LNET_PID_ANY;
2203 return ksocknal_push(ni, id);
2212 ksocknal_free_buffers (void)
2214 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2216 if (ksocknal_data.ksnd_sched_info != NULL) {
2217 struct ksock_sched_info *info;
2220 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2221 if (info->ksi_scheds != NULL) {
2222 LIBCFS_FREE(info->ksi_scheds,
2223 info->ksi_nthreads_max *
2224 sizeof(info->ksi_scheds[0]));
2227 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2230 LIBCFS_FREE (ksocknal_data.ksnd_peers,
2231 sizeof(struct list_head) *
2232 ksocknal_data.ksnd_peer_hash_size);
2234 spin_lock(&ksocknal_data.ksnd_tx_lock);
2236 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2237 struct list_head zlist;
2240 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2241 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2242 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2244 while (!list_empty(&zlist)) {
2245 tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2246 list_del(&tx->tx_list);
2247 LIBCFS_FREE(tx, tx->tx_desc_size);
2250 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2255 ksocknal_base_shutdown(void)
2257 struct ksock_sched_info *info;
2258 ksock_sched_t *sched;
2262 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2263 atomic_read (&libcfs_kmemory));
2264 LASSERT (ksocknal_data.ksnd_nnets == 0);
2266 switch (ksocknal_data.ksnd_init) {
2270 case SOCKNAL_INIT_ALL:
2271 case SOCKNAL_INIT_DATA:
2272 LASSERT (ksocknal_data.ksnd_peers != NULL);
2273 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2274 LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2277 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2278 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2279 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2280 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2281 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2283 if (ksocknal_data.ksnd_sched_info != NULL) {
2284 cfs_percpt_for_each(info, i,
2285 ksocknal_data.ksnd_sched_info) {
2286 if (info->ksi_scheds == NULL)
2289 for (j = 0; j < info->ksi_nthreads_max; j++) {
2291 sched = &info->ksi_scheds[j];
2292 LASSERT(list_empty(&sched->\
2294 LASSERT(list_empty(&sched->\
2296 LASSERT(list_empty(&sched-> \
2297 kss_zombie_noop_txs));
2298 LASSERT(sched->kss_nconns == 0);
2303 /* flag threads to terminate; wake and wait for them to die */
2304 ksocknal_data.ksnd_shuttingdown = 1;
2305 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2306 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2308 if (ksocknal_data.ksnd_sched_info != NULL) {
2309 cfs_percpt_for_each(info, i,
2310 ksocknal_data.ksnd_sched_info) {
2311 if (info->ksi_scheds == NULL)
2314 for (j = 0; j < info->ksi_nthreads_max; j++) {
2315 sched = &info->ksi_scheds[j];
2316 wake_up_all(&sched->kss_waitq);
2322 read_lock(&ksocknal_data.ksnd_global_lock);
2323 while (ksocknal_data.ksnd_nthreads != 0) {
2326 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2327 "waiting for %d threads to terminate\n",
2328 ksocknal_data.ksnd_nthreads);
2329 read_unlock(&ksocknal_data.ksnd_global_lock);
2330 set_current_state(TASK_UNINTERRUPTIBLE);
2331 schedule_timeout(cfs_time_seconds(1));
2332 read_lock(&ksocknal_data.ksnd_global_lock);
2334 read_unlock(&ksocknal_data.ksnd_global_lock);
2336 ksocknal_free_buffers();
2338 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2342 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2343 atomic_read (&libcfs_kmemory));
2345 module_put(THIS_MODULE);
2348 static __u64 ksocknal_new_incarnation(void)
2352 /* The incarnation number is the time this module loaded and it
2353 * identifies this particular instance of the socknal. Hopefully
2354 * we won't be able to reboot more frequently than 1MHz for the
2355 * forseeable future :) */
2357 do_gettimeofday(&tv);
2359 return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2363 ksocknal_base_startup(void)
2365 struct ksock_sched_info *info;
2369 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2370 LASSERT (ksocknal_data.ksnd_nnets == 0);
2372 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2374 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2375 LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2376 sizeof(struct list_head) *
2377 ksocknal_data.ksnd_peer_hash_size);
2378 if (ksocknal_data.ksnd_peers == NULL)
2381 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2382 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2384 rwlock_init(&ksocknal_data.ksnd_global_lock);
2385 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2387 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2388 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2389 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2390 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2391 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2393 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2394 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2395 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2396 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2398 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2399 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2401 /* NB memset above zeros whole of ksocknal_data */
2403 /* flag lists/ptrs/locks initialised */
2404 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2405 try_module_get(THIS_MODULE);
2407 ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2409 if (ksocknal_data.ksnd_sched_info == NULL)
2412 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2413 ksock_sched_t *sched;
2416 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2417 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2418 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2420 /* max to half of CPUs, assume another half should be
2421 * reserved for upper layer modules */
2422 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2425 info->ksi_nthreads_max = nthrs;
2428 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2429 info->ksi_nthreads_max * sizeof(*sched));
2430 if (info->ksi_scheds == NULL)
2433 for (; nthrs > 0; nthrs--) {
2434 sched = &info->ksi_scheds[nthrs - 1];
2436 sched->kss_info = info;
2437 spin_lock_init(&sched->kss_lock);
2438 INIT_LIST_HEAD(&sched->kss_rx_conns);
2439 INIT_LIST_HEAD(&sched->kss_tx_conns);
2440 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2441 init_waitqueue_head(&sched->kss_waitq);
2445 ksocknal_data.ksnd_connd_starting = 0;
2446 ksocknal_data.ksnd_connd_failed_stamp = 0;
2447 ksocknal_data.ksnd_connd_starting_stamp = cfs_time_current_sec();
2448 /* must have at least 2 connds to remain responsive to accepts while
2450 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2451 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2453 if (*ksocknal_tunables.ksnd_nconnds_max <
2454 *ksocknal_tunables.ksnd_nconnds) {
2455 ksocknal_tunables.ksnd_nconnds_max =
2456 ksocknal_tunables.ksnd_nconnds;
2459 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2461 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2462 ksocknal_data.ksnd_connd_starting++;
2463 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2466 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2467 rc = ksocknal_thread_start(ksocknal_connd,
2468 (void *)((uintptr_t)i), name);
2470 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2471 ksocknal_data.ksnd_connd_starting--;
2472 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2473 CERROR("Can't spawn socknal connd: %d\n", rc);
2478 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2480 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2484 /* flag everything initialised */
2485 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2490 ksocknal_base_shutdown();
2495 ksocknal_debug_peerhash (lnet_ni_t *ni)
2497 ksock_peer_t *peer = NULL;
2498 struct list_head *tmp;
2501 read_lock(&ksocknal_data.ksnd_global_lock);
2503 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2504 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2505 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
2507 if (peer->ksnp_ni == ni) break;
2514 ksock_route_t *route;
2517 CWARN ("Active peer on shutdown: %s, ref %d, scnt %d, "
2518 "closing %d, accepting %d, err %d, zcookie %llu, "
2519 "txq %d, zc_req %d\n", libcfs_id2str(peer->ksnp_id),
2520 atomic_read(&peer->ksnp_refcount),
2521 peer->ksnp_sharecount, peer->ksnp_closing,
2522 peer->ksnp_accepting, peer->ksnp_error,
2523 peer->ksnp_zc_next_cookie,
2524 !list_empty(&peer->ksnp_tx_queue),
2525 !list_empty(&peer->ksnp_zc_req_list));
2527 list_for_each(tmp, &peer->ksnp_routes) {
2528 route = list_entry(tmp, ksock_route_t, ksnr_list);
2529 CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2530 "del %d\n", atomic_read(&route->ksnr_refcount),
2531 route->ksnr_scheduled, route->ksnr_connecting,
2532 route->ksnr_connected, route->ksnr_deleted);
2535 list_for_each(tmp, &peer->ksnp_conns) {
2536 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2537 CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2538 atomic_read(&conn->ksnc_conn_refcount),
2539 atomic_read(&conn->ksnc_sock_refcount),
2540 conn->ksnc_type, conn->ksnc_closing);
2544 read_unlock(&ksocknal_data.ksnd_global_lock);
2549 ksocknal_shutdown (lnet_ni_t *ni)
2551 ksock_net_t *net = ni->ni_data;
2553 lnet_process_id_t anyid = {0};
2555 anyid.nid = LNET_NID_ANY;
2556 anyid.pid = LNET_PID_ANY;
2558 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2559 LASSERT(ksocknal_data.ksnd_nnets > 0);
2561 spin_lock_bh(&net->ksnn_lock);
2562 net->ksnn_shutdown = 1; /* prevent new peers */
2563 spin_unlock_bh(&net->ksnn_lock);
2565 /* Delete all peers */
2566 ksocknal_del_peer(ni, anyid, 0);
2568 /* Wait for all peer state to clean up */
2570 spin_lock_bh(&net->ksnn_lock);
2571 while (net->ksnn_npeers != 0) {
2572 spin_unlock_bh(&net->ksnn_lock);
2575 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2576 "waiting for %d peers to disconnect\n",
2578 set_current_state(TASK_UNINTERRUPTIBLE);
2579 schedule_timeout(cfs_time_seconds(1));
2581 ksocknal_debug_peerhash(ni);
2583 spin_lock_bh(&net->ksnn_lock);
2585 spin_unlock_bh(&net->ksnn_lock);
2587 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2588 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2589 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2592 list_del(&net->ksnn_list);
2593 LIBCFS_FREE(net, sizeof(*net));
2595 ksocknal_data.ksnd_nnets--;
2596 if (ksocknal_data.ksnd_nnets == 0)
2597 ksocknal_base_shutdown();
2601 ksocknal_enumerate_interfaces(ksock_net_t *net)
2609 n = lnet_ipif_enumerate(&names);
2611 CERROR("Can't enumerate interfaces: %d\n", n);
2615 for (i = j = 0; i < n; i++) {
2620 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2623 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2625 CWARN("Can't get interface %s info: %d\n",
2631 CWARN("Ignoring interface %s (down)\n",
2636 if (j == LNET_MAX_INTERFACES) {
2637 CWARN("Ignoring interface %s (too many interfaces)\n",
2642 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2643 net->ksnn_interfaces[j].ksni_netmask = mask;
2644 strlcpy(net->ksnn_interfaces[j].ksni_name,
2645 names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2649 lnet_ipif_free_enumeration(names, n);
2652 CERROR("Can't find any usable interfaces\n");
2658 ksocknal_search_new_ipif(ksock_net_t *net)
2663 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2664 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2665 char *colon = strchr(ifnam, ':');
2670 if (colon != NULL) /* ignore alias device */
2673 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2675 for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2676 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2678 char *colon2 = strchr(ifnam2, ':');
2683 found = strcmp(ifnam, ifnam2) == 0;
2700 ksocknal_start_schedulers(struct ksock_sched_info *info)
2706 if (info->ksi_nthreads == 0) {
2707 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2708 nthrs = info->ksi_nthreads_max;
2710 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2712 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2713 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2715 nthrs = min(nthrs, info->ksi_nthreads_max);
2717 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2718 /* increase two threads if there is new interface */
2719 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2722 for (i = 0; i < nthrs; i++) {
2725 ksock_sched_t *sched;
2726 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2727 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2728 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2729 info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2731 rc = ksocknal_thread_start(ksocknal_scheduler,
2736 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2737 info->ksi_cpt, info->ksi_nthreads + i, rc);
2741 info->ksi_nthreads += i;
2746 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2748 int newif = ksocknal_search_new_ipif(net);
2752 LASSERT(ncpts > 0 && ncpts <= cfs_cpt_number(lnet_cpt_table()));
2754 for (i = 0; i < ncpts; i++) {
2755 struct ksock_sched_info *info;
2756 int cpt = (cpts == NULL) ? i : cpts[i];
2758 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2759 info = ksocknal_data.ksnd_sched_info[cpt];
2761 if (!newif && info->ksi_nthreads > 0)
2764 rc = ksocknal_start_schedulers(info);
2772 ksocknal_startup (lnet_ni_t *ni)
2778 LASSERT (ni->ni_lnd == &the_ksocklnd);
2780 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2781 rc = ksocknal_base_startup();
2786 LIBCFS_ALLOC(net, sizeof(*net));
2790 spin_lock_init(&net->ksnn_lock);
2791 net->ksnn_incarnation = ksocknal_new_incarnation();
2793 ni->ni_peertimeout = *ksocknal_tunables.ksnd_peertimeout;
2794 ni->ni_maxtxcredits = *ksocknal_tunables.ksnd_credits;
2795 ni->ni_peertxcredits = *ksocknal_tunables.ksnd_peertxcredits;
2796 ni->ni_peerrtrcredits = *ksocknal_tunables.ksnd_peerrtrcredits;
2798 if (ni->ni_interfaces[0] == NULL) {
2799 rc = ksocknal_enumerate_interfaces(net);
2803 net->ksnn_ninterfaces = 1;
2805 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2808 if (ni->ni_interfaces[i] == NULL)
2811 rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2812 &net->ksnn_interfaces[i].ksni_ipaddr,
2813 &net->ksnn_interfaces[i].ksni_netmask);
2816 CERROR("Can't get interface %s info: %d\n",
2817 ni->ni_interfaces[i], rc);
2822 CERROR("Interface %s is down\n",
2823 ni->ni_interfaces[i]);
2827 strlcpy(net->ksnn_interfaces[i].ksni_name,
2828 ni->ni_interfaces[i],
2829 sizeof(net->ksnn_interfaces[i].ksni_name));
2831 net->ksnn_ninterfaces = i;
2834 /* call it before add it to ksocknal_data.ksnd_nets */
2835 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2839 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2840 net->ksnn_interfaces[0].ksni_ipaddr);
2841 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2843 ksocknal_data.ksnd_nnets++;
2848 LIBCFS_FREE(net, sizeof(*net));
2850 if (ksocknal_data.ksnd_nnets == 0)
2851 ksocknal_base_shutdown();
2857 static void __exit ksocklnd_exit(void)
2859 lnet_unregister_lnd(&the_ksocklnd);
2862 static int __init ksocklnd_init(void)
2866 /* check ksnr_connected/connecting field large enough */
2867 CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2868 CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2870 /* initialize the_ksocklnd */
2871 the_ksocklnd.lnd_type = SOCKLND;
2872 the_ksocklnd.lnd_startup = ksocknal_startup;
2873 the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2874 the_ksocklnd.lnd_ctl = ksocknal_ctl;
2875 the_ksocklnd.lnd_send = ksocknal_send;
2876 the_ksocklnd.lnd_recv = ksocknal_recv;
2877 the_ksocklnd.lnd_notify = ksocknal_notify;
2878 the_ksocklnd.lnd_query = ksocknal_query;
2879 the_ksocklnd.lnd_accept = ksocknal_accept;
2881 rc = ksocknal_tunables_init();
2885 lnet_register_lnd(&the_ksocklnd);
2890 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2891 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2892 MODULE_VERSION("2.8.0");
2893 MODULE_LICENSE("GPL");
2895 module_init(ksocklnd_init);
2896 module_exit(ksocklnd_exit);