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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/socklnd/socklnd.c
38 * Author: Zach Brown <zab@zabbo.net>
39 * Author: Peter J. Braam <braam@clusterfs.com>
40 * Author: Phil Schwan <phil@clusterfs.com>
41 * Author: Eric Barton <eric@bartonsoftware.com>
46 static lnd_t the_ksocklnd;
47 ksock_nal_data_t ksocknal_data;
49 static ksock_interface_t *
50 ksocknal_ip2iface(lnet_ni_t *ni, __u32 ip)
52 ksock_net_t *net = ni->ni_data;
54 ksock_interface_t *iface;
56 for (i = 0; i < net->ksnn_ninterfaces; i++) {
57 LASSERT(i < LNET_MAX_INTERFACES);
58 iface = &net->ksnn_interfaces[i];
60 if (iface->ksni_ipaddr == ip)
67 static ksock_route_t *
68 ksocknal_create_route (__u32 ipaddr, int port)
72 LIBCFS_ALLOC (route, sizeof (*route));
76 atomic_set (&route->ksnr_refcount, 1);
77 route->ksnr_peer = NULL;
78 route->ksnr_retry_interval = 0; /* OK to connect at any time */
79 route->ksnr_ipaddr = ipaddr;
80 route->ksnr_port = port;
81 route->ksnr_scheduled = 0;
82 route->ksnr_connecting = 0;
83 route->ksnr_connected = 0;
84 route->ksnr_deleted = 0;
85 route->ksnr_conn_count = 0;
86 route->ksnr_share_count = 0;
92 ksocknal_destroy_route (ksock_route_t *route)
94 LASSERT (atomic_read(&route->ksnr_refcount) == 0);
96 if (route->ksnr_peer != NULL)
97 ksocknal_peer_decref(route->ksnr_peer);
99 LIBCFS_FREE (route, sizeof (*route));
103 ksocknal_create_peer (ksock_peer_t **peerp, lnet_ni_t *ni, lnet_process_id_t id)
105 ksock_net_t *net = ni->ni_data;
108 LASSERT (id.nid != LNET_NID_ANY);
109 LASSERT (id.pid != LNET_PID_ANY);
110 LASSERT (!in_interrupt());
112 LIBCFS_ALLOC (peer, sizeof (*peer));
116 memset (peer, 0, sizeof (*peer)); /* NULL pointers/clear flags etc */
120 atomic_set (&peer->ksnp_refcount, 1); /* 1 ref for caller */
121 peer->ksnp_closing = 0;
122 peer->ksnp_accepting = 0;
123 peer->ksnp_proto = NULL;
124 peer->ksnp_last_alive = 0;
125 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
127 INIT_LIST_HEAD(&peer->ksnp_conns);
128 INIT_LIST_HEAD(&peer->ksnp_routes);
129 INIT_LIST_HEAD(&peer->ksnp_tx_queue);
130 INIT_LIST_HEAD(&peer->ksnp_zc_req_list);
131 spin_lock_init(&peer->ksnp_lock);
133 spin_lock_bh(&net->ksnn_lock);
135 if (net->ksnn_shutdown) {
136 spin_unlock_bh(&net->ksnn_lock);
138 LIBCFS_FREE(peer, sizeof(*peer));
139 CERROR("Can't create peer: network shutdown\n");
145 spin_unlock_bh(&net->ksnn_lock);
152 ksocknal_destroy_peer (ksock_peer_t *peer)
154 ksock_net_t *net = peer->ksnp_ni->ni_data;
156 CDEBUG (D_NET, "peer %s %p deleted\n",
157 libcfs_id2str(peer->ksnp_id), peer);
159 LASSERT(atomic_read(&peer->ksnp_refcount) == 0);
160 LASSERT(peer->ksnp_accepting == 0);
161 LASSERT(list_empty(&peer->ksnp_conns));
162 LASSERT(list_empty(&peer->ksnp_routes));
163 LASSERT(list_empty(&peer->ksnp_tx_queue));
164 LASSERT(list_empty(&peer->ksnp_zc_req_list));
166 LIBCFS_FREE(peer, sizeof(*peer));
168 /* NB a peer's connections and routes keep a reference on their peer
169 * until they are destroyed, so we can be assured that _all_ state to
170 * do with this peer has been cleaned up when its refcount drops to
172 spin_lock_bh(&net->ksnn_lock);
174 spin_unlock_bh(&net->ksnn_lock);
178 ksocknal_find_peer_locked (lnet_ni_t *ni, lnet_process_id_t id)
180 struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
181 struct list_head *tmp;
184 list_for_each(tmp, peer_list) {
186 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
188 LASSERT(!peer->ksnp_closing);
190 if (peer->ksnp_ni != ni)
193 if (peer->ksnp_id.nid != id.nid ||
194 peer->ksnp_id.pid != id.pid)
197 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
198 peer, libcfs_id2str(id),
199 atomic_read(&peer->ksnp_refcount));
206 ksocknal_find_peer (lnet_ni_t *ni, lnet_process_id_t id)
210 read_lock(&ksocknal_data.ksnd_global_lock);
211 peer = ksocknal_find_peer_locked(ni, id);
212 if (peer != NULL) /* +1 ref for caller? */
213 ksocknal_peer_addref(peer);
214 read_unlock(&ksocknal_data.ksnd_global_lock);
220 ksocknal_unlink_peer_locked (ksock_peer_t *peer)
224 ksock_interface_t *iface;
226 for (i = 0; i < peer->ksnp_n_passive_ips; i++) {
227 LASSERT (i < LNET_MAX_INTERFACES);
228 ip = peer->ksnp_passive_ips[i];
230 iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
231 /* All IPs in peer->ksnp_passive_ips[] come from the
232 * interface list, therefore the call must succeed. */
233 LASSERT (iface != NULL);
235 CDEBUG(D_NET, "peer=%p iface=%p ksni_nroutes=%d\n",
236 peer, iface, iface->ksni_nroutes);
237 iface->ksni_npeers--;
240 LASSERT(list_empty(&peer->ksnp_conns));
241 LASSERT(list_empty(&peer->ksnp_routes));
242 LASSERT(!peer->ksnp_closing);
243 peer->ksnp_closing = 1;
244 list_del(&peer->ksnp_list);
245 /* lose peerlist's ref */
246 ksocknal_peer_decref(peer);
250 ksocknal_get_peer_info (lnet_ni_t *ni, int index,
251 lnet_process_id_t *id, __u32 *myip, __u32 *peer_ip,
252 int *port, int *conn_count, int *share_count)
255 struct list_head *ptmp;
256 ksock_route_t *route;
257 struct list_head *rtmp;
262 read_lock(&ksocknal_data.ksnd_global_lock);
264 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
265 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
266 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
268 if (peer->ksnp_ni != ni)
271 if (peer->ksnp_n_passive_ips == 0 &&
272 list_empty(&peer->ksnp_routes)) {
286 for (j = 0; j < peer->ksnp_n_passive_ips; j++) {
291 *myip = peer->ksnp_passive_ips[j];
300 list_for_each(rtmp, &peer->ksnp_routes) {
304 route = list_entry(rtmp, ksock_route_t,
308 *myip = route->ksnr_myipaddr;
309 *peer_ip = route->ksnr_ipaddr;
310 *port = route->ksnr_port;
311 *conn_count = route->ksnr_conn_count;
312 *share_count = route->ksnr_share_count;
319 read_unlock(&ksocknal_data.ksnd_global_lock);
324 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
326 ksock_peer_t *peer = route->ksnr_peer;
327 int type = conn->ksnc_type;
328 ksock_interface_t *iface;
330 conn->ksnc_route = route;
331 ksocknal_route_addref(route);
333 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
334 if (route->ksnr_myipaddr == 0) {
335 /* route wasn't bound locally yet (the initial route) */
336 CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
337 libcfs_id2str(peer->ksnp_id),
339 &conn->ksnc_myipaddr);
341 CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
342 "to %pI4h\n", libcfs_id2str(peer->ksnp_id),
344 &route->ksnr_myipaddr,
345 &conn->ksnc_myipaddr);
347 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
348 route->ksnr_myipaddr);
350 iface->ksni_nroutes--;
352 route->ksnr_myipaddr = conn->ksnc_myipaddr;
353 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
354 route->ksnr_myipaddr);
356 iface->ksni_nroutes++;
359 route->ksnr_connected |= (1<<type);
360 route->ksnr_conn_count++;
362 /* Successful connection => further attempts can
363 * proceed immediately */
364 route->ksnr_retry_interval = 0;
368 ksocknal_add_route_locked (ksock_peer_t *peer, ksock_route_t *route)
370 struct list_head *tmp;
372 ksock_route_t *route2;
374 LASSERT(!peer->ksnp_closing);
375 LASSERT(route->ksnr_peer == NULL);
376 LASSERT(!route->ksnr_scheduled);
377 LASSERT(!route->ksnr_connecting);
378 LASSERT(route->ksnr_connected == 0);
380 /* LASSERT(unique) */
381 list_for_each(tmp, &peer->ksnp_routes) {
382 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
384 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
385 CERROR("Duplicate route %s %pI4h\n",
386 libcfs_id2str(peer->ksnp_id),
387 &route->ksnr_ipaddr);
392 route->ksnr_peer = peer;
393 ksocknal_peer_addref(peer);
394 /* peer's routelist takes over my ref on 'route' */
395 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
397 list_for_each(tmp, &peer->ksnp_conns) {
398 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
400 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
403 ksocknal_associate_route_conn_locked(route, conn);
404 /* keep going (typed routes) */
409 ksocknal_del_route_locked (ksock_route_t *route)
411 ksock_peer_t *peer = route->ksnr_peer;
412 ksock_interface_t *iface;
414 struct list_head *ctmp;
415 struct list_head *cnxt;
417 LASSERT(!route->ksnr_deleted);
419 /* Close associated conns */
420 list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
421 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
423 if (conn->ksnc_route != route)
426 ksocknal_close_conn_locked(conn, 0);
429 if (route->ksnr_myipaddr != 0) {
430 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
431 route->ksnr_myipaddr);
433 iface->ksni_nroutes--;
436 route->ksnr_deleted = 1;
437 list_del(&route->ksnr_list);
438 ksocknal_route_decref(route); /* drop peer's ref */
440 if (list_empty(&peer->ksnp_routes) &&
441 list_empty(&peer->ksnp_conns)) {
442 /* I've just removed the last route to a peer with no active
444 ksocknal_unlink_peer_locked(peer);
449 ksocknal_add_peer(lnet_ni_t *ni, lnet_process_id_t id, __u32 ipaddr, int port)
451 struct list_head *tmp;
454 ksock_route_t *route;
455 ksock_route_t *route2;
458 if (id.nid == LNET_NID_ANY ||
459 id.pid == LNET_PID_ANY)
462 /* Have a brand new peer ready... */
463 rc = ksocknal_create_peer(&peer, ni, id);
467 route = ksocknal_create_route (ipaddr, port);
469 ksocknal_peer_decref(peer);
473 write_lock_bh(&ksocknal_data.ksnd_global_lock);
475 /* always called with a ref on ni, so shutdown can't have started */
476 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
478 peer2 = ksocknal_find_peer_locked(ni, id);
480 ksocknal_peer_decref(peer);
483 /* peer table takes my ref on peer */
484 list_add_tail(&peer->ksnp_list,
485 ksocknal_nid2peerlist(id.nid));
489 list_for_each(tmp, &peer->ksnp_routes) {
490 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
492 if (route2->ksnr_ipaddr == ipaddr)
497 if (route2 == NULL) {
498 ksocknal_add_route_locked(peer, route);
499 route->ksnr_share_count++;
501 ksocknal_route_decref(route);
502 route2->ksnr_share_count++;
505 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
511 ksocknal_del_peer_locked (ksock_peer_t *peer, __u32 ip)
514 ksock_route_t *route;
515 struct list_head *tmp;
516 struct list_head *nxt;
519 LASSERT(!peer->ksnp_closing);
521 /* Extra ref prevents peer disappearing until I'm done with it */
522 ksocknal_peer_addref(peer);
524 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
525 route = list_entry(tmp, ksock_route_t, ksnr_list);
528 if (!(ip == 0 || route->ksnr_ipaddr == ip))
531 route->ksnr_share_count = 0;
532 /* This deletes associated conns too */
533 ksocknal_del_route_locked(route);
537 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
538 route = list_entry(tmp, ksock_route_t, ksnr_list);
539 nshared += route->ksnr_share_count;
543 /* remove everything else if there are no explicit entries
546 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
547 route = list_entry(tmp, ksock_route_t, ksnr_list);
549 /* we should only be removing auto-entries */
550 LASSERT(route->ksnr_share_count == 0);
551 ksocknal_del_route_locked(route);
554 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
555 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
557 ksocknal_close_conn_locked(conn, 0);
561 ksocknal_peer_decref(peer);
562 /* NB peer unlinks itself when last conn/route is removed */
566 ksocknal_del_peer (lnet_ni_t *ni, lnet_process_id_t id, __u32 ip)
568 struct list_head zombies = LIST_HEAD_INIT(zombies);
569 struct list_head *ptmp;
570 struct list_head *pnxt;
577 write_lock_bh(&ksocknal_data.ksnd_global_lock);
579 if (id.nid != LNET_NID_ANY) {
580 hi = (int)(ksocknal_nid2peerlist(id.nid) -
581 ksocknal_data.ksnd_peers);
585 hi = ksocknal_data.ksnd_peer_hash_size - 1;
588 for (i = lo; i <= hi; i++) {
589 list_for_each_safe(ptmp, pnxt,
590 &ksocknal_data.ksnd_peers[i]) {
591 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
593 if (peer->ksnp_ni != ni)
596 if (!((id.nid == LNET_NID_ANY ||
597 peer->ksnp_id.nid == id.nid) &&
598 (id.pid == LNET_PID_ANY ||
599 peer->ksnp_id.pid == id.pid)))
602 ksocknal_peer_addref(peer); /* a ref for me... */
604 ksocknal_del_peer_locked(peer, ip);
606 if (peer->ksnp_closing &&
607 !list_empty(&peer->ksnp_tx_queue)) {
608 LASSERT(list_empty(&peer->ksnp_conns));
609 LASSERT(list_empty(&peer->ksnp_routes));
611 list_splice_init(&peer->ksnp_tx_queue,
615 ksocknal_peer_decref(peer); /* ...till here */
617 rc = 0; /* matched! */
621 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
623 ksocknal_txlist_done(ni, &zombies, 1);
628 static ksock_conn_t *
629 ksocknal_get_conn_by_idx (lnet_ni_t *ni, int index)
632 struct list_head *ptmp;
634 struct list_head *ctmp;
637 read_lock(&ksocknal_data.ksnd_global_lock);
639 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
640 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
641 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
643 LASSERT(!peer->ksnp_closing);
645 if (peer->ksnp_ni != ni)
648 list_for_each(ctmp, &peer->ksnp_conns) {
652 conn = list_entry(ctmp, ksock_conn_t,
654 ksocknal_conn_addref(conn);
655 read_unlock(&ksocknal_data. \
662 read_unlock(&ksocknal_data.ksnd_global_lock);
666 static ksock_sched_t *
667 ksocknal_choose_scheduler_locked(unsigned int cpt)
669 struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
670 ksock_sched_t *sched;
673 LASSERT(info->ksi_nthreads > 0);
675 sched = &info->ksi_scheds[0];
677 * NB: it's safe so far, but info->ksi_nthreads could be changed
678 * at runtime when we have dynamic LNet configuration, then we
679 * need to take care of this.
681 for (i = 1; i < info->ksi_nthreads; i++) {
682 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
683 sched = &info->ksi_scheds[i];
690 ksocknal_local_ipvec (lnet_ni_t *ni, __u32 *ipaddrs)
692 ksock_net_t *net = ni->ni_data;
696 read_lock(&ksocknal_data.ksnd_global_lock);
698 nip = net->ksnn_ninterfaces;
699 LASSERT (nip <= LNET_MAX_INTERFACES);
701 /* Only offer interfaces for additional connections if I have
704 read_unlock(&ksocknal_data.ksnd_global_lock);
708 for (i = 0; i < nip; i++) {
709 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
710 LASSERT (ipaddrs[i] != 0);
713 read_unlock(&ksocknal_data.ksnd_global_lock);
718 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
720 int best_netmatch = 0;
727 for (i = 0; i < nips; i++) {
731 this_xor = (ips[i] ^ iface->ksni_ipaddr);
732 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
735 best_netmatch < this_netmatch ||
736 (best_netmatch == this_netmatch &&
737 best_xor > this_xor)))
741 best_netmatch = this_netmatch;
750 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
752 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
753 ksock_net_t *net = peer->ksnp_ni->ni_data;
754 ksock_interface_t *iface;
755 ksock_interface_t *best_iface;
766 /* CAVEAT EMPTOR: We do all our interface matching with an
767 * exclusive hold of global lock at IRQ priority. We're only
768 * expecting to be dealing with small numbers of interfaces, so the
769 * O(n**3)-ness shouldn't matter */
771 /* Also note that I'm not going to return more than n_peerips
772 * interfaces, even if I have more myself */
774 write_lock_bh(global_lock);
776 LASSERT (n_peerips <= LNET_MAX_INTERFACES);
777 LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
779 /* Only match interfaces for additional connections
780 * if I have > 1 interface */
781 n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
782 MIN(n_peerips, net->ksnn_ninterfaces);
784 for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
785 /* ^ yes really... */
787 /* If we have any new interfaces, first tick off all the
788 * peer IPs that match old interfaces, then choose new
789 * interfaces to match the remaining peer IPS.
790 * We don't forget interfaces we've stopped using; we might
791 * start using them again... */
793 if (i < peer->ksnp_n_passive_ips) {
795 ip = peer->ksnp_passive_ips[i];
796 best_iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
798 /* peer passive ips are kept up to date */
799 LASSERT(best_iface != NULL);
801 /* choose a new interface */
802 LASSERT (i == peer->ksnp_n_passive_ips);
808 for (j = 0; j < net->ksnn_ninterfaces; j++) {
809 iface = &net->ksnn_interfaces[j];
810 ip = iface->ksni_ipaddr;
812 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
813 if (peer->ksnp_passive_ips[k] == ip)
816 if (k < peer->ksnp_n_passive_ips) /* using it already */
819 k = ksocknal_match_peerip(iface, peerips, n_peerips);
820 xor = (ip ^ peerips[k]);
821 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
823 if (!(best_iface == NULL ||
824 best_netmatch < this_netmatch ||
825 (best_netmatch == this_netmatch &&
826 best_npeers > iface->ksni_npeers)))
830 best_netmatch = this_netmatch;
831 best_npeers = iface->ksni_npeers;
834 LASSERT(best_iface != NULL);
836 best_iface->ksni_npeers++;
837 ip = best_iface->ksni_ipaddr;
838 peer->ksnp_passive_ips[i] = ip;
839 peer->ksnp_n_passive_ips = i+1;
842 /* mark the best matching peer IP used */
843 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
847 /* Overwrite input peer IP addresses */
848 memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
850 write_unlock_bh(global_lock);
856 ksocknal_create_routes(ksock_peer_t *peer, int port,
857 __u32 *peer_ipaddrs, int npeer_ipaddrs)
859 ksock_route_t *newroute = NULL;
860 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
861 lnet_ni_t *ni = peer->ksnp_ni;
862 ksock_net_t *net = ni->ni_data;
863 struct list_head *rtmp;
864 ksock_route_t *route;
865 ksock_interface_t *iface;
866 ksock_interface_t *best_iface;
873 /* CAVEAT EMPTOR: We do all our interface matching with an
874 * exclusive hold of global lock at IRQ priority. We're only
875 * expecting to be dealing with small numbers of interfaces, so the
876 * O(n**3)-ness here shouldn't matter */
878 write_lock_bh(global_lock);
880 if (net->ksnn_ninterfaces < 2) {
881 /* Only create additional connections
882 * if I have > 1 interface */
883 write_unlock_bh(global_lock);
887 LASSERT (npeer_ipaddrs <= LNET_MAX_INTERFACES);
889 for (i = 0; i < npeer_ipaddrs; i++) {
890 if (newroute != NULL) {
891 newroute->ksnr_ipaddr = peer_ipaddrs[i];
893 write_unlock_bh(global_lock);
895 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
896 if (newroute == NULL)
899 write_lock_bh(global_lock);
902 if (peer->ksnp_closing) {
903 /* peer got closed under me */
907 /* Already got a route? */
909 list_for_each(rtmp, &peer->ksnp_routes) {
910 route = list_entry(rtmp, ksock_route_t, ksnr_list);
912 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
924 LASSERT(net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
926 /* Select interface to connect from */
927 for (j = 0; j < net->ksnn_ninterfaces; j++) {
928 iface = &net->ksnn_interfaces[j];
930 /* Using this interface already? */
931 list_for_each(rtmp, &peer->ksnp_routes) {
932 route = list_entry(rtmp, ksock_route_t,
935 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
943 this_netmatch = (((iface->ksni_ipaddr ^
944 newroute->ksnr_ipaddr) &
945 iface->ksni_netmask) == 0) ? 1 : 0;
947 if (!(best_iface == NULL ||
948 best_netmatch < this_netmatch ||
949 (best_netmatch == this_netmatch &&
950 best_nroutes > iface->ksni_nroutes)))
954 best_netmatch = this_netmatch;
955 best_nroutes = iface->ksni_nroutes;
958 if (best_iface == NULL)
961 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
962 best_iface->ksni_nroutes++;
964 ksocknal_add_route_locked(peer, newroute);
968 write_unlock_bh(global_lock);
969 if (newroute != NULL)
970 ksocknal_route_decref(newroute);
974 ksocknal_accept(lnet_ni_t *ni, struct socket *sock)
981 rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
982 LASSERT(rc == 0); /* we succeeded before */
984 LIBCFS_ALLOC(cr, sizeof(*cr));
986 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
987 "%pI4h: memory exhausted\n", &peer_ip);
993 cr->ksncr_sock = sock;
995 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
997 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
998 wake_up(&ksocknal_data.ksnd_connd_waitq);
1000 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1005 ksocknal_connecting (ksock_peer_t *peer, __u32 ipaddr)
1007 ksock_route_t *route;
1009 list_for_each_entry(route, &peer->ksnp_routes, ksnr_list) {
1010 if (route->ksnr_ipaddr == ipaddr)
1011 return route->ksnr_connecting;
1017 ksocknal_create_conn(lnet_ni_t *ni, ksock_route_t *route,
1018 struct socket *sock, int type)
1020 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1021 struct list_head zombies = LIST_HEAD_INIT(zombies);
1022 lnet_process_id_t peerid;
1023 struct list_head *tmp;
1026 ksock_conn_t *conn2;
1027 ksock_peer_t *peer = NULL;
1028 ksock_peer_t *peer2;
1029 ksock_sched_t *sched;
1030 ksock_hello_msg_t *hello;
1038 active = (route != NULL);
1040 LASSERT (active == (type != SOCKLND_CONN_NONE));
1042 LIBCFS_ALLOC(conn, sizeof(*conn));
1048 memset (conn, 0, sizeof (*conn));
1050 conn->ksnc_peer = NULL;
1051 conn->ksnc_route = NULL;
1052 conn->ksnc_sock = sock;
1053 /* 2 ref, 1 for conn, another extra ref prevents socket
1054 * being closed before establishment of connection */
1055 atomic_set (&conn->ksnc_sock_refcount, 2);
1056 conn->ksnc_type = type;
1057 ksocknal_lib_save_callback(sock, conn);
1058 atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1060 conn->ksnc_rx_ready = 0;
1061 conn->ksnc_rx_scheduled = 0;
1063 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1064 conn->ksnc_tx_ready = 0;
1065 conn->ksnc_tx_scheduled = 0;
1066 conn->ksnc_tx_carrier = NULL;
1067 atomic_set (&conn->ksnc_tx_nob, 0);
1069 LIBCFS_ALLOC(hello, offsetof(ksock_hello_msg_t,
1070 kshm_ips[LNET_MAX_INTERFACES]));
1071 if (hello == NULL) {
1076 /* stash conn's local and remote addrs */
1077 rc = ksocknal_lib_get_conn_addrs (conn);
1081 /* Find out/confirm peer's NID and connection type and get the
1082 * vector of interfaces she's willing to let me connect to.
1083 * Passive connections use the listener timeout since the peer sends
1087 peer = route->ksnr_peer;
1088 LASSERT(ni == peer->ksnp_ni);
1090 /* Active connection sends HELLO eagerly */
1091 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1092 peerid = peer->ksnp_id;
1094 write_lock_bh(global_lock);
1095 conn->ksnc_proto = peer->ksnp_proto;
1096 write_unlock_bh(global_lock);
1098 if (conn->ksnc_proto == NULL) {
1099 conn->ksnc_proto = &ksocknal_protocol_v3x;
1100 #if SOCKNAL_VERSION_DEBUG
1101 if (*ksocknal_tunables.ksnd_protocol == 2)
1102 conn->ksnc_proto = &ksocknal_protocol_v2x;
1103 else if (*ksocknal_tunables.ksnd_protocol == 1)
1104 conn->ksnc_proto = &ksocknal_protocol_v1x;
1108 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1112 peerid.nid = LNET_NID_ANY;
1113 peerid.pid = LNET_PID_ANY;
1115 /* Passive, get protocol from peer */
1116 conn->ksnc_proto = NULL;
1119 rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1123 LASSERT (rc == 0 || active);
1124 LASSERT (conn->ksnc_proto != NULL);
1125 LASSERT (peerid.nid != LNET_NID_ANY);
1127 cpt = lnet_cpt_of_nid(peerid.nid);
1130 ksocknal_peer_addref(peer);
1131 write_lock_bh(global_lock);
1133 rc = ksocknal_create_peer(&peer, ni, peerid);
1137 write_lock_bh(global_lock);
1139 /* called with a ref on ni, so shutdown can't have started */
1140 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1142 peer2 = ksocknal_find_peer_locked(ni, peerid);
1143 if (peer2 == NULL) {
1144 /* NB this puts an "empty" peer in the peer
1145 * table (which takes my ref) */
1146 list_add_tail(&peer->ksnp_list,
1147 ksocknal_nid2peerlist(peerid.nid));
1149 ksocknal_peer_decref(peer);
1154 ksocknal_peer_addref(peer);
1155 peer->ksnp_accepting++;
1157 /* Am I already connecting to this guy? Resolve in
1158 * favour of higher NID... */
1159 if (peerid.nid < ni->ni_nid &&
1160 ksocknal_connecting(peer, conn->ksnc_ipaddr)) {
1162 warn = "connection race resolution";
1167 if (peer->ksnp_closing ||
1168 (active && route->ksnr_deleted)) {
1169 /* peer/route got closed under me */
1171 warn = "peer/route removed";
1175 if (peer->ksnp_proto == NULL) {
1176 /* Never connected before.
1177 * NB recv_hello may have returned EPROTO to signal my peer
1178 * wants a different protocol than the one I asked for.
1180 LASSERT(list_empty(&peer->ksnp_conns));
1182 peer->ksnp_proto = conn->ksnc_proto;
1183 peer->ksnp_incarnation = incarnation;
1186 if (peer->ksnp_proto != conn->ksnc_proto ||
1187 peer->ksnp_incarnation != incarnation) {
1188 /* Peer rebooted or I've got the wrong protocol version */
1189 ksocknal_close_peer_conns_locked(peer, 0, 0);
1191 peer->ksnp_proto = NULL;
1193 warn = peer->ksnp_incarnation != incarnation ?
1195 "wrong proto version";
1205 warn = "lost conn race";
1208 warn = "retry with different protocol version";
1212 /* Refuse to duplicate an existing connection, unless this is a
1213 * loopback connection */
1214 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1215 list_for_each(tmp, &peer->ksnp_conns) {
1216 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1218 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1219 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1220 conn2->ksnc_type != conn->ksnc_type)
1223 /* Reply on a passive connection attempt so the peer
1224 * realises we're connected. */
1234 /* If the connection created by this route didn't bind to the IP
1235 * address the route connected to, the connection/route matching
1236 * code below probably isn't going to work. */
1238 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1239 CERROR("Route %s %pI4h connected to %pI4h\n",
1240 libcfs_id2str(peer->ksnp_id),
1241 &route->ksnr_ipaddr,
1242 &conn->ksnc_ipaddr);
1245 /* Search for a route corresponding to the new connection and
1246 * create an association. This allows incoming connections created
1247 * by routes in my peer to match my own route entries so I don't
1248 * continually create duplicate routes. */
1249 list_for_each(tmp, &peer->ksnp_routes) {
1250 route = list_entry(tmp, ksock_route_t, ksnr_list);
1252 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1255 ksocknal_associate_route_conn_locked(route, conn);
1259 conn->ksnc_peer = peer; /* conn takes my ref on peer */
1260 peer->ksnp_last_alive = cfs_time_current();
1261 peer->ksnp_send_keepalive = 0;
1262 peer->ksnp_error = 0;
1264 sched = ksocknal_choose_scheduler_locked(cpt);
1265 sched->kss_nconns++;
1266 conn->ksnc_scheduler = sched;
1268 conn->ksnc_tx_last_post = cfs_time_current();
1269 /* Set the deadline for the outgoing HELLO to drain */
1270 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1271 conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1272 smp_mb(); /* order with adding to peer's conn list */
1274 list_add(&conn->ksnc_list, &peer->ksnp_conns);
1275 ksocknal_conn_addref(conn);
1277 ksocknal_new_packet(conn, 0);
1279 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1281 /* Take packets blocking for this connection. */
1282 list_for_each_entry_safe(tx, txtmp, &peer->ksnp_tx_queue, tx_list) {
1283 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1287 list_del(&tx->tx_list);
1288 ksocknal_queue_tx_locked(tx, conn);
1291 write_unlock_bh(global_lock);
1293 /* We've now got a new connection. Any errors from here on are just
1294 * like "normal" comms errors and we close the connection normally.
1295 * NB (a) we still have to send the reply HELLO for passive
1297 * (b) normal I/O on the conn is blocked until I setup and call the
1301 CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1302 " incarnation:%lld sched[%d:%d]\n",
1303 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1304 &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1305 conn->ksnc_port, incarnation, cpt,
1306 (int)(sched - &sched->kss_info->ksi_scheds[0]));
1309 /* additional routes after interface exchange? */
1310 ksocknal_create_routes(peer, conn->ksnc_port,
1311 hello->kshm_ips, hello->kshm_nips);
1313 hello->kshm_nips = ksocknal_select_ips(peer, hello->kshm_ips,
1315 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1318 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1319 kshm_ips[LNET_MAX_INTERFACES]));
1321 /* setup the socket AFTER I've received hello (it disables
1322 * SO_LINGER). I might call back to the acceptor who may want
1323 * to send a protocol version response and then close the
1324 * socket; this ensures the socket only tears down after the
1325 * response has been sent. */
1327 rc = ksocknal_lib_setup_sock(sock);
1329 write_lock_bh(global_lock);
1331 /* NB my callbacks block while I hold ksnd_global_lock */
1332 ksocknal_lib_set_callback(sock, conn);
1335 peer->ksnp_accepting--;
1337 write_unlock_bh(global_lock);
1340 write_lock_bh(global_lock);
1341 if (!conn->ksnc_closing) {
1342 /* could be closed by another thread */
1343 ksocknal_close_conn_locked(conn, rc);
1345 write_unlock_bh(global_lock);
1346 } else if (ksocknal_connsock_addref(conn) == 0) {
1347 /* Allow I/O to proceed. */
1348 ksocknal_read_callback(conn);
1349 ksocknal_write_callback(conn);
1350 ksocknal_connsock_decref(conn);
1353 ksocknal_connsock_decref(conn);
1354 ksocknal_conn_decref(conn);
1358 if (!peer->ksnp_closing &&
1359 list_empty(&peer->ksnp_conns) &&
1360 list_empty(&peer->ksnp_routes)) {
1361 list_add(&zombies, &peer->ksnp_tx_queue);
1362 list_del_init(&peer->ksnp_tx_queue);
1363 ksocknal_unlink_peer_locked(peer);
1366 write_unlock_bh(global_lock);
1370 CERROR("Not creating conn %s type %d: %s\n",
1371 libcfs_id2str(peerid), conn->ksnc_type, warn);
1373 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1374 libcfs_id2str(peerid), conn->ksnc_type, warn);
1379 /* Request retry by replying with CONN_NONE
1380 * ksnc_proto has been set already */
1381 conn->ksnc_type = SOCKLND_CONN_NONE;
1382 hello->kshm_nips = 0;
1383 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1386 write_lock_bh(global_lock);
1387 peer->ksnp_accepting--;
1388 write_unlock_bh(global_lock);
1391 ksocknal_txlist_done(ni, &zombies, 1);
1392 ksocknal_peer_decref(peer);
1396 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1397 kshm_ips[LNET_MAX_INTERFACES]));
1399 LIBCFS_FREE(conn, sizeof(*conn));
1407 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1409 /* This just does the immmediate housekeeping, and queues the
1410 * connection for the reaper to terminate.
1411 * Caller holds ksnd_global_lock exclusively in irq context */
1412 ksock_peer_t *peer = conn->ksnc_peer;
1413 ksock_route_t *route;
1414 ksock_conn_t *conn2;
1415 struct list_head *tmp;
1417 LASSERT(peer->ksnp_error == 0);
1418 LASSERT(!conn->ksnc_closing);
1419 conn->ksnc_closing = 1;
1421 /* ksnd_deathrow_conns takes over peer's ref */
1422 list_del(&conn->ksnc_list);
1424 route = conn->ksnc_route;
1425 if (route != NULL) {
1426 /* dissociate conn from route... */
1427 LASSERT(!route->ksnr_deleted);
1428 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1431 list_for_each(tmp, &peer->ksnp_conns) {
1432 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1434 if (conn2->ksnc_route == route &&
1435 conn2->ksnc_type == conn->ksnc_type)
1441 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1443 conn->ksnc_route = NULL;
1445 #if 0 /* irrelevent with only eager routes */
1446 /* make route least favourite */
1447 list_del(&route->ksnr_list);
1448 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1450 ksocknal_route_decref(route); /* drop conn's ref on route */
1453 if (list_empty(&peer->ksnp_conns)) {
1454 /* No more connections to this peer */
1456 if (!list_empty(&peer->ksnp_tx_queue)) {
1459 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1461 /* throw them to the last connection...,
1462 * these TXs will be send to /dev/null by scheduler */
1463 list_for_each_entry(tx, &peer->ksnp_tx_queue,
1465 ksocknal_tx_prep(conn, tx);
1467 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1468 list_splice_init(&peer->ksnp_tx_queue,
1469 &conn->ksnc_tx_queue);
1470 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1473 /* renegotiate protocol version */
1474 peer->ksnp_proto = NULL;
1475 /* stash last conn close reason */
1476 peer->ksnp_error = error;
1478 if (list_empty(&peer->ksnp_routes)) {
1479 /* I've just closed last conn belonging to a
1480 * peer with no routes to it */
1481 ksocknal_unlink_peer_locked(peer);
1485 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1487 list_add_tail(&conn->ksnc_list,
1488 &ksocknal_data.ksnd_deathrow_conns);
1489 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1491 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1495 ksocknal_peer_failed (ksock_peer_t *peer)
1498 cfs_time_t last_alive = 0;
1500 /* There has been a connection failure or comms error; but I'll only
1501 * tell LNET I think the peer is dead if it's to another kernel and
1502 * there are no connections or connection attempts in existance. */
1504 read_lock(&ksocknal_data.ksnd_global_lock);
1506 if ((peer->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1507 list_empty(&peer->ksnp_conns) &&
1508 peer->ksnp_accepting == 0 &&
1509 ksocknal_find_connecting_route_locked(peer) == NULL) {
1511 last_alive = peer->ksnp_last_alive;
1514 read_unlock(&ksocknal_data.ksnd_global_lock);
1517 lnet_notify(peer->ksnp_ni, peer->ksnp_id.nid, 0,
1522 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1524 ksock_peer_t *peer = conn->ksnc_peer;
1527 struct list_head zlist = LIST_HEAD_INIT(zlist);
1529 /* NB safe to finalize TXs because closing of socket will
1530 * abort all buffered data */
1531 LASSERT(conn->ksnc_sock == NULL);
1533 spin_lock(&peer->ksnp_lock);
1535 list_for_each_entry_safe(tx, tmp, &peer->ksnp_zc_req_list, tx_zc_list) {
1536 if (tx->tx_conn != conn)
1539 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1541 tx->tx_msg.ksm_zc_cookies[0] = 0;
1542 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1543 list_del(&tx->tx_zc_list);
1544 list_add(&tx->tx_zc_list, &zlist);
1547 spin_unlock(&peer->ksnp_lock);
1549 while (!list_empty(&zlist)) {
1550 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1552 list_del(&tx->tx_zc_list);
1553 ksocknal_tx_decref(tx);
1558 ksocknal_terminate_conn(ksock_conn_t *conn)
1560 /* This gets called by the reaper (guaranteed thread context) to
1561 * disengage the socket from its callbacks and close it.
1562 * ksnc_refcount will eventually hit zero, and then the reaper will
1564 ksock_peer_t *peer = conn->ksnc_peer;
1565 ksock_sched_t *sched = conn->ksnc_scheduler;
1568 LASSERT(conn->ksnc_closing);
1570 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1571 spin_lock_bh(&sched->kss_lock);
1573 /* a closing conn is always ready to tx */
1574 conn->ksnc_tx_ready = 1;
1576 if (!conn->ksnc_tx_scheduled &&
1577 !list_empty(&conn->ksnc_tx_queue)) {
1578 list_add_tail(&conn->ksnc_tx_list,
1579 &sched->kss_tx_conns);
1580 conn->ksnc_tx_scheduled = 1;
1581 /* extra ref for scheduler */
1582 ksocknal_conn_addref(conn);
1584 wake_up (&sched->kss_waitq);
1587 spin_unlock_bh(&sched->kss_lock);
1589 /* serialise with callbacks */
1590 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1592 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1594 /* OK, so this conn may not be completely disengaged from its
1595 * scheduler yet, but it _has_ committed to terminate... */
1596 conn->ksnc_scheduler->kss_nconns--;
1598 if (peer->ksnp_error != 0) {
1599 /* peer's last conn closed in error */
1600 LASSERT(list_empty(&peer->ksnp_conns));
1602 peer->ksnp_error = 0; /* avoid multiple notifications */
1605 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1608 ksocknal_peer_failed(peer);
1610 /* The socket is closed on the final put; either here, or in
1611 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1612 * when the connection was established, this will close the socket
1613 * immediately, aborting anything buffered in it. Any hung
1614 * zero-copy transmits will therefore complete in finite time. */
1615 ksocknal_connsock_decref(conn);
1619 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1621 /* Queue the conn for the reaper to destroy */
1623 LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1624 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1626 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1627 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1629 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1633 ksocknal_destroy_conn (ksock_conn_t *conn)
1635 cfs_time_t last_rcv;
1637 /* Final coup-de-grace of the reaper */
1638 CDEBUG (D_NET, "connection %p\n", conn);
1640 LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1641 LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1642 LASSERT (conn->ksnc_sock == NULL);
1643 LASSERT (conn->ksnc_route == NULL);
1644 LASSERT (!conn->ksnc_tx_scheduled);
1645 LASSERT (!conn->ksnc_rx_scheduled);
1646 LASSERT(list_empty(&conn->ksnc_tx_queue));
1648 /* complete current receive if any */
1649 switch (conn->ksnc_rx_state) {
1650 case SOCKNAL_RX_LNET_PAYLOAD:
1651 last_rcv = conn->ksnc_rx_deadline -
1652 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1653 CERROR("Completing partial receive from %s[%d], "
1654 "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1655 "last alive is %ld secs ago\n",
1656 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1657 &conn->ksnc_ipaddr, conn->ksnc_port,
1658 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1659 cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1661 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1662 conn->ksnc_cookie, -EIO);
1664 case SOCKNAL_RX_LNET_HEADER:
1665 if (conn->ksnc_rx_started)
1666 CERROR("Incomplete receive of lnet header from %s, "
1667 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1668 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1669 &conn->ksnc_ipaddr, conn->ksnc_port,
1670 conn->ksnc_proto->pro_version);
1672 case SOCKNAL_RX_KSM_HEADER:
1673 if (conn->ksnc_rx_started)
1674 CERROR("Incomplete receive of ksock message from %s, "
1675 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1676 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1677 &conn->ksnc_ipaddr, conn->ksnc_port,
1678 conn->ksnc_proto->pro_version);
1680 case SOCKNAL_RX_SLOP:
1681 if (conn->ksnc_rx_started)
1682 CERROR("Incomplete receive of slops from %s, "
1683 "ip %pI4h:%d, with error\n",
1684 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1685 &conn->ksnc_ipaddr, conn->ksnc_port);
1692 ksocknal_peer_decref(conn->ksnc_peer);
1694 LIBCFS_FREE (conn, sizeof (*conn));
1698 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1701 struct list_head *ctmp;
1702 struct list_head *cnxt;
1705 list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
1706 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1709 conn->ksnc_ipaddr == ipaddr) {
1711 ksocknal_close_conn_locked (conn, why);
1719 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1721 ksock_peer_t *peer = conn->ksnc_peer;
1722 __u32 ipaddr = conn->ksnc_ipaddr;
1725 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1727 count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1729 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1735 ksocknal_close_matching_conns (lnet_process_id_t id, __u32 ipaddr)
1738 struct list_head *ptmp;
1739 struct list_head *pnxt;
1745 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1747 if (id.nid != LNET_NID_ANY)
1748 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1751 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1754 for (i = lo; i <= hi; i++) {
1755 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1757 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1759 if (!((id.nid == LNET_NID_ANY || id.nid == peer->ksnp_id.nid) &&
1760 (id.pid == LNET_PID_ANY || id.pid == peer->ksnp_id.pid)))
1763 count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1767 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1769 /* wildcards always succeed */
1770 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1773 return (count == 0 ? -ENOENT : 0);
1777 ksocknal_notify (lnet_ni_t *ni, lnet_nid_t gw_nid, int alive)
1779 /* The router is telling me she's been notified of a change in
1780 * gateway state.... */
1781 lnet_process_id_t id = {0};
1784 id.pid = LNET_PID_ANY;
1786 CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1787 alive ? "up" : "down");
1790 /* If the gateway crashed, close all open connections... */
1791 ksocknal_close_matching_conns (id, 0);
1795 /* ...otherwise do nothing. We can only establish new connections
1796 * if we have autroutes, and these connect on demand. */
1800 ksocknal_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1803 cfs_time_t last_alive = 0;
1804 cfs_time_t now = cfs_time_current();
1805 ksock_peer_t *peer = NULL;
1806 rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1807 lnet_process_id_t id = {
1809 .pid = LNET_PID_LUSTRE,
1814 peer = ksocknal_find_peer_locked(ni, id);
1816 struct list_head *tmp;
1820 list_for_each(tmp, &peer->ksnp_conns) {
1821 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1822 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1824 if (bufnob < conn->ksnc_tx_bufnob) {
1825 /* something got ACKed */
1826 conn->ksnc_tx_deadline =
1827 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1828 peer->ksnp_last_alive = now;
1829 conn->ksnc_tx_bufnob = bufnob;
1833 last_alive = peer->ksnp_last_alive;
1834 if (ksocknal_find_connectable_route_locked(peer) == NULL)
1840 if (last_alive != 0)
1843 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago, connect %d\n",
1844 libcfs_nid2str(nid), peer,
1845 last_alive ? cfs_duration_sec(now - last_alive) : -1,
1851 ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1853 write_lock_bh(glock);
1855 peer = ksocknal_find_peer_locked(ni, id);
1857 ksocknal_launch_all_connections_locked(peer);
1859 write_unlock_bh(glock);
1864 ksocknal_push_peer (ksock_peer_t *peer)
1868 struct list_head *tmp;
1871 for (index = 0; ; index++) {
1872 read_lock(&ksocknal_data.ksnd_global_lock);
1877 list_for_each(tmp, &peer->ksnp_conns) {
1879 conn = list_entry(tmp, ksock_conn_t,
1881 ksocknal_conn_addref(conn);
1886 read_unlock(&ksocknal_data.ksnd_global_lock);
1891 ksocknal_lib_push_conn (conn);
1892 ksocknal_conn_decref(conn);
1897 ksocknal_push (lnet_ni_t *ni, lnet_process_id_t id)
1900 struct list_head *tmp;
1906 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1907 for (j = 0; ; j++) {
1908 read_lock(&ksocknal_data.ksnd_global_lock);
1913 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
1914 peer = list_entry(tmp, ksock_peer_t,
1917 if (!((id.nid == LNET_NID_ANY ||
1918 id.nid == peer->ksnp_id.nid) &&
1919 (id.pid == LNET_PID_ANY ||
1920 id.pid == peer->ksnp_id.pid))) {
1926 ksocknal_peer_addref(peer);
1931 read_unlock(&ksocknal_data.ksnd_global_lock);
1935 ksocknal_push_peer (peer);
1936 ksocknal_peer_decref(peer);
1946 ksocknal_add_interface(lnet_ni_t *ni, __u32 ipaddress, __u32 netmask)
1948 ksock_net_t *net = ni->ni_data;
1949 ksock_interface_t *iface;
1953 struct list_head *ptmp;
1955 struct list_head *rtmp;
1956 ksock_route_t *route;
1958 if (ipaddress == 0 ||
1962 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1964 iface = ksocknal_ip2iface(ni, ipaddress);
1965 if (iface != NULL) {
1966 /* silently ignore dups */
1968 } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1971 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1973 iface->ksni_ipaddr = ipaddress;
1974 iface->ksni_netmask = netmask;
1975 iface->ksni_nroutes = 0;
1976 iface->ksni_npeers = 0;
1978 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1979 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1980 peer = list_entry(ptmp, ksock_peer_t,
1983 for (j = 0; j < peer->ksnp_n_passive_ips; j++)
1984 if (peer->ksnp_passive_ips[j] == ipaddress)
1985 iface->ksni_npeers++;
1987 list_for_each(rtmp, &peer->ksnp_routes) {
1988 route = list_entry(rtmp,
1992 if (route->ksnr_myipaddr == ipaddress)
1993 iface->ksni_nroutes++;
1999 /* NB only new connections will pay attention to the new interface! */
2002 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2008 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
2010 struct list_head *tmp;
2011 struct list_head *nxt;
2012 ksock_route_t *route;
2017 for (i = 0; i < peer->ksnp_n_passive_ips; i++)
2018 if (peer->ksnp_passive_ips[i] == ipaddr) {
2019 for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
2020 peer->ksnp_passive_ips[j-1] =
2021 peer->ksnp_passive_ips[j];
2022 peer->ksnp_n_passive_ips--;
2026 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
2027 route = list_entry(tmp, ksock_route_t, ksnr_list);
2029 if (route->ksnr_myipaddr != ipaddr)
2032 if (route->ksnr_share_count != 0) {
2033 /* Manually created; keep, but unbind */
2034 route->ksnr_myipaddr = 0;
2036 ksocknal_del_route_locked(route);
2040 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
2041 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2043 if (conn->ksnc_myipaddr == ipaddr)
2044 ksocknal_close_conn_locked (conn, 0);
2049 ksocknal_del_interface(lnet_ni_t *ni, __u32 ipaddress)
2051 ksock_net_t *net = ni->ni_data;
2053 struct list_head *tmp;
2054 struct list_head *nxt;
2060 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2062 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2063 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2065 if (!(ipaddress == 0 ||
2066 ipaddress == this_ip))
2071 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2072 net->ksnn_interfaces[j-1] =
2073 net->ksnn_interfaces[j];
2075 net->ksnn_ninterfaces--;
2077 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2078 list_for_each_safe(tmp, nxt,
2079 &ksocknal_data.ksnd_peers[j]) {
2080 peer = list_entry(tmp, ksock_peer_t,
2083 if (peer->ksnp_ni != ni)
2086 ksocknal_peer_del_interface_locked(peer, this_ip);
2091 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2097 ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
2099 lnet_process_id_t id = {0};
2100 struct libcfs_ioctl_data *data = arg;
2104 case IOC_LIBCFS_GET_INTERFACE: {
2105 ksock_net_t *net = ni->ni_data;
2106 ksock_interface_t *iface;
2108 read_lock(&ksocknal_data.ksnd_global_lock);
2110 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2114 iface = &net->ksnn_interfaces[data->ioc_count];
2116 data->ioc_u32[0] = iface->ksni_ipaddr;
2117 data->ioc_u32[1] = iface->ksni_netmask;
2118 data->ioc_u32[2] = iface->ksni_npeers;
2119 data->ioc_u32[3] = iface->ksni_nroutes;
2122 read_unlock(&ksocknal_data.ksnd_global_lock);
2126 case IOC_LIBCFS_ADD_INTERFACE:
2127 return ksocknal_add_interface(ni,
2128 data->ioc_u32[0], /* IP address */
2129 data->ioc_u32[1]); /* net mask */
2131 case IOC_LIBCFS_DEL_INTERFACE:
2132 return ksocknal_del_interface(ni,
2133 data->ioc_u32[0]); /* IP address */
2135 case IOC_LIBCFS_GET_PEER: {
2140 int share_count = 0;
2142 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2143 &id, &myip, &ip, &port,
2144 &conn_count, &share_count);
2148 data->ioc_nid = id.nid;
2149 data->ioc_count = share_count;
2150 data->ioc_u32[0] = ip;
2151 data->ioc_u32[1] = port;
2152 data->ioc_u32[2] = myip;
2153 data->ioc_u32[3] = conn_count;
2154 data->ioc_u32[4] = id.pid;
2158 case IOC_LIBCFS_ADD_PEER:
2159 id.nid = data->ioc_nid;
2160 id.pid = LNET_PID_LUSTRE;
2161 return ksocknal_add_peer (ni, id,
2162 data->ioc_u32[0], /* IP */
2163 data->ioc_u32[1]); /* port */
2165 case IOC_LIBCFS_DEL_PEER:
2166 id.nid = data->ioc_nid;
2167 id.pid = LNET_PID_ANY;
2168 return ksocknal_del_peer (ni, id,
2169 data->ioc_u32[0]); /* IP */
2171 case IOC_LIBCFS_GET_CONN: {
2175 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2180 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2182 data->ioc_count = txmem;
2183 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
2184 data->ioc_flags = nagle;
2185 data->ioc_u32[0] = conn->ksnc_ipaddr;
2186 data->ioc_u32[1] = conn->ksnc_port;
2187 data->ioc_u32[2] = conn->ksnc_myipaddr;
2188 data->ioc_u32[3] = conn->ksnc_type;
2189 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2190 data->ioc_u32[5] = rxmem;
2191 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2192 ksocknal_conn_decref(conn);
2196 case IOC_LIBCFS_CLOSE_CONNECTION:
2197 id.nid = data->ioc_nid;
2198 id.pid = LNET_PID_ANY;
2199 return ksocknal_close_matching_conns (id,
2202 case IOC_LIBCFS_REGISTER_MYNID:
2203 /* Ignore if this is a noop */
2204 if (data->ioc_nid == ni->ni_nid)
2207 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2208 libcfs_nid2str(data->ioc_nid),
2209 libcfs_nid2str(ni->ni_nid));
2212 case IOC_LIBCFS_PUSH_CONNECTION:
2213 id.nid = data->ioc_nid;
2214 id.pid = LNET_PID_ANY;
2215 return ksocknal_push(ni, id);
2224 ksocknal_free_buffers (void)
2226 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2228 if (ksocknal_data.ksnd_sched_info != NULL) {
2229 struct ksock_sched_info *info;
2232 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2233 if (info->ksi_scheds != NULL) {
2234 LIBCFS_FREE(info->ksi_scheds,
2235 info->ksi_nthreads_max *
2236 sizeof(info->ksi_scheds[0]));
2239 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2242 LIBCFS_FREE (ksocknal_data.ksnd_peers,
2243 sizeof(struct list_head) *
2244 ksocknal_data.ksnd_peer_hash_size);
2246 spin_lock(&ksocknal_data.ksnd_tx_lock);
2248 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2249 struct list_head zlist;
2252 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2253 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2254 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2256 while (!list_empty(&zlist)) {
2257 tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2258 list_del(&tx->tx_list);
2259 LIBCFS_FREE(tx, tx->tx_desc_size);
2262 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2267 ksocknal_base_shutdown(void)
2269 struct ksock_sched_info *info;
2270 ksock_sched_t *sched;
2274 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2275 atomic_read (&libcfs_kmemory));
2276 LASSERT (ksocknal_data.ksnd_nnets == 0);
2278 switch (ksocknal_data.ksnd_init) {
2282 case SOCKNAL_INIT_ALL:
2283 case SOCKNAL_INIT_DATA:
2284 LASSERT (ksocknal_data.ksnd_peers != NULL);
2285 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2286 LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2289 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2290 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2291 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2292 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2293 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2295 if (ksocknal_data.ksnd_sched_info != NULL) {
2296 cfs_percpt_for_each(info, i,
2297 ksocknal_data.ksnd_sched_info) {
2298 if (info->ksi_scheds == NULL)
2301 for (j = 0; j < info->ksi_nthreads_max; j++) {
2303 sched = &info->ksi_scheds[j];
2304 LASSERT(list_empty(&sched->\
2306 LASSERT(list_empty(&sched->\
2308 LASSERT(list_empty(&sched-> \
2309 kss_zombie_noop_txs));
2310 LASSERT(sched->kss_nconns == 0);
2315 /* flag threads to terminate; wake and wait for them to die */
2316 ksocknal_data.ksnd_shuttingdown = 1;
2317 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2318 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2320 if (ksocknal_data.ksnd_sched_info != NULL) {
2321 cfs_percpt_for_each(info, i,
2322 ksocknal_data.ksnd_sched_info) {
2323 if (info->ksi_scheds == NULL)
2326 for (j = 0; j < info->ksi_nthreads_max; j++) {
2327 sched = &info->ksi_scheds[j];
2328 wake_up_all(&sched->kss_waitq);
2334 read_lock(&ksocknal_data.ksnd_global_lock);
2335 while (ksocknal_data.ksnd_nthreads != 0) {
2337 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2338 "waiting for %d threads to terminate\n",
2339 ksocknal_data.ksnd_nthreads);
2340 read_unlock(&ksocknal_data.ksnd_global_lock);
2341 cfs_pause(cfs_time_seconds(1));
2342 read_lock(&ksocknal_data.ksnd_global_lock);
2344 read_unlock(&ksocknal_data.ksnd_global_lock);
2346 ksocknal_free_buffers();
2348 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2352 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2353 atomic_read (&libcfs_kmemory));
2355 module_put(THIS_MODULE);
2358 static __u64 ksocknal_new_incarnation(void)
2362 /* The incarnation number is the time this module loaded and it
2363 * identifies this particular instance of the socknal. Hopefully
2364 * we won't be able to reboot more frequently than 1MHz for the
2365 * forseeable future :) */
2367 do_gettimeofday(&tv);
2369 return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2373 ksocknal_base_startup(void)
2375 struct ksock_sched_info *info;
2379 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2380 LASSERT (ksocknal_data.ksnd_nnets == 0);
2382 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2384 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2385 LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2386 sizeof(struct list_head) *
2387 ksocknal_data.ksnd_peer_hash_size);
2388 if (ksocknal_data.ksnd_peers == NULL)
2391 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2392 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2394 rwlock_init(&ksocknal_data.ksnd_global_lock);
2395 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2397 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2398 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2399 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2400 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2401 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2403 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2404 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2405 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2406 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2408 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2409 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2411 /* NB memset above zeros whole of ksocknal_data */
2413 /* flag lists/ptrs/locks initialised */
2414 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2415 try_module_get(THIS_MODULE);
2417 ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2419 if (ksocknal_data.ksnd_sched_info == NULL)
2422 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2423 ksock_sched_t *sched;
2426 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2427 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2428 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2430 /* max to half of CPUs, assume another half should be
2431 * reserved for upper layer modules */
2432 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2435 info->ksi_nthreads_max = nthrs;
2438 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2439 info->ksi_nthreads_max * sizeof(*sched));
2440 if (info->ksi_scheds == NULL)
2443 for (; nthrs > 0; nthrs--) {
2444 sched = &info->ksi_scheds[nthrs - 1];
2446 sched->kss_info = info;
2447 spin_lock_init(&sched->kss_lock);
2448 INIT_LIST_HEAD(&sched->kss_rx_conns);
2449 INIT_LIST_HEAD(&sched->kss_tx_conns);
2450 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2451 init_waitqueue_head(&sched->kss_waitq);
2455 ksocknal_data.ksnd_connd_starting = 0;
2456 ksocknal_data.ksnd_connd_failed_stamp = 0;
2457 ksocknal_data.ksnd_connd_starting_stamp = cfs_time_current_sec();
2458 /* must have at least 2 connds to remain responsive to accepts while
2460 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2461 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2463 if (*ksocknal_tunables.ksnd_nconnds_max <
2464 *ksocknal_tunables.ksnd_nconnds) {
2465 ksocknal_tunables.ksnd_nconnds_max =
2466 ksocknal_tunables.ksnd_nconnds;
2469 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2471 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2472 ksocknal_data.ksnd_connd_starting++;
2473 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2476 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2477 rc = ksocknal_thread_start(ksocknal_connd,
2478 (void *)((ulong_ptr_t)i), name);
2480 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2481 ksocknal_data.ksnd_connd_starting--;
2482 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2483 CERROR("Can't spawn socknal connd: %d\n", rc);
2488 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2490 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2494 /* flag everything initialised */
2495 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2500 ksocknal_base_shutdown();
2505 ksocknal_debug_peerhash (lnet_ni_t *ni)
2507 ksock_peer_t *peer = NULL;
2508 struct list_head *tmp;
2511 read_lock(&ksocknal_data.ksnd_global_lock);
2513 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2514 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2515 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
2517 if (peer->ksnp_ni == ni) break;
2524 ksock_route_t *route;
2527 CWARN ("Active peer on shutdown: %s, ref %d, scnt %d, "
2528 "closing %d, accepting %d, err %d, zcookie "LPU64", "
2529 "txq %d, zc_req %d\n", libcfs_id2str(peer->ksnp_id),
2530 atomic_read(&peer->ksnp_refcount),
2531 peer->ksnp_sharecount, peer->ksnp_closing,
2532 peer->ksnp_accepting, peer->ksnp_error,
2533 peer->ksnp_zc_next_cookie,
2534 !list_empty(&peer->ksnp_tx_queue),
2535 !list_empty(&peer->ksnp_zc_req_list));
2537 list_for_each(tmp, &peer->ksnp_routes) {
2538 route = list_entry(tmp, ksock_route_t, ksnr_list);
2539 CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2540 "del %d\n", atomic_read(&route->ksnr_refcount),
2541 route->ksnr_scheduled, route->ksnr_connecting,
2542 route->ksnr_connected, route->ksnr_deleted);
2545 list_for_each(tmp, &peer->ksnp_conns) {
2546 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2547 CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2548 atomic_read(&conn->ksnc_conn_refcount),
2549 atomic_read(&conn->ksnc_sock_refcount),
2550 conn->ksnc_type, conn->ksnc_closing);
2554 read_unlock(&ksocknal_data.ksnd_global_lock);
2559 ksocknal_shutdown (lnet_ni_t *ni)
2561 ksock_net_t *net = ni->ni_data;
2563 lnet_process_id_t anyid = {0};
2565 anyid.nid = LNET_NID_ANY;
2566 anyid.pid = LNET_PID_ANY;
2568 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2569 LASSERT(ksocknal_data.ksnd_nnets > 0);
2571 spin_lock_bh(&net->ksnn_lock);
2572 net->ksnn_shutdown = 1; /* prevent new peers */
2573 spin_unlock_bh(&net->ksnn_lock);
2575 /* Delete all peers */
2576 ksocknal_del_peer(ni, anyid, 0);
2578 /* Wait for all peer state to clean up */
2580 spin_lock_bh(&net->ksnn_lock);
2581 while (net->ksnn_npeers != 0) {
2582 spin_unlock_bh(&net->ksnn_lock);
2585 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2586 "waiting for %d peers to disconnect\n",
2588 cfs_pause(cfs_time_seconds(1));
2590 ksocknal_debug_peerhash(ni);
2592 spin_lock_bh(&net->ksnn_lock);
2594 spin_unlock_bh(&net->ksnn_lock);
2596 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2597 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2598 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2601 list_del(&net->ksnn_list);
2602 LIBCFS_FREE(net, sizeof(*net));
2604 ksocknal_data.ksnd_nnets--;
2605 if (ksocknal_data.ksnd_nnets == 0)
2606 ksocknal_base_shutdown();
2610 ksocknal_enumerate_interfaces(ksock_net_t *net)
2618 n = lnet_ipif_enumerate(&names);
2620 CERROR("Can't enumerate interfaces: %d\n", n);
2624 for (i = j = 0; i < n; i++) {
2629 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2632 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2634 CWARN("Can't get interface %s info: %d\n",
2640 CWARN("Ignoring interface %s (down)\n",
2645 if (j == LNET_MAX_INTERFACES) {
2646 CWARN("Ignoring interface %s (too many interfaces)\n",
2651 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2652 net->ksnn_interfaces[j].ksni_netmask = mask;
2653 strlcpy(net->ksnn_interfaces[j].ksni_name,
2654 names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2658 lnet_ipif_free_enumeration(names, n);
2661 CERROR("Can't find any usable interfaces\n");
2667 ksocknal_search_new_ipif(ksock_net_t *net)
2672 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2673 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2674 char *colon = strchr(ifnam, ':');
2679 if (colon != NULL) /* ignore alias device */
2682 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2684 for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2685 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2687 char *colon2 = strchr(ifnam2, ':');
2692 found = strcmp(ifnam, ifnam2) == 0;
2709 ksocknal_start_schedulers(struct ksock_sched_info *info)
2715 if (info->ksi_nthreads == 0) {
2716 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2717 nthrs = info->ksi_nthreads_max;
2719 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2721 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2722 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2724 nthrs = min(nthrs, info->ksi_nthreads_max);
2726 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2727 /* increase two threads if there is new interface */
2728 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2731 for (i = 0; i < nthrs; i++) {
2734 ksock_sched_t *sched;
2735 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2736 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2737 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2738 info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2740 rc = ksocknal_thread_start(ksocknal_scheduler,
2745 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2746 info->ksi_cpt, info->ksi_nthreads + i, rc);
2750 info->ksi_nthreads += i;
2755 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2757 int newif = ksocknal_search_new_ipif(net);
2761 LASSERT(ncpts > 0 && ncpts <= cfs_cpt_number(lnet_cpt_table()));
2763 for (i = 0; i < ncpts; i++) {
2764 struct ksock_sched_info *info;
2765 int cpt = (cpts == NULL) ? i : cpts[i];
2767 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2768 info = ksocknal_data.ksnd_sched_info[cpt];
2770 if (!newif && info->ksi_nthreads > 0)
2773 rc = ksocknal_start_schedulers(info);
2781 ksocknal_startup (lnet_ni_t *ni)
2787 LASSERT (ni->ni_lnd == &the_ksocklnd);
2789 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2790 rc = ksocknal_base_startup();
2795 LIBCFS_ALLOC(net, sizeof(*net));
2799 spin_lock_init(&net->ksnn_lock);
2800 net->ksnn_incarnation = ksocknal_new_incarnation();
2802 ni->ni_peertimeout = *ksocknal_tunables.ksnd_peertimeout;
2803 ni->ni_maxtxcredits = *ksocknal_tunables.ksnd_credits;
2804 ni->ni_peertxcredits = *ksocknal_tunables.ksnd_peertxcredits;
2805 ni->ni_peerrtrcredits = *ksocknal_tunables.ksnd_peerrtrcredits;
2807 if (ni->ni_interfaces[0] == NULL) {
2808 rc = ksocknal_enumerate_interfaces(net);
2812 net->ksnn_ninterfaces = 1;
2814 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2817 if (ni->ni_interfaces[i] == NULL)
2820 rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2821 &net->ksnn_interfaces[i].ksni_ipaddr,
2822 &net->ksnn_interfaces[i].ksni_netmask);
2825 CERROR("Can't get interface %s info: %d\n",
2826 ni->ni_interfaces[i], rc);
2831 CERROR("Interface %s is down\n",
2832 ni->ni_interfaces[i]);
2836 strlcpy(net->ksnn_interfaces[i].ksni_name,
2837 ni->ni_interfaces[i],
2838 sizeof(net->ksnn_interfaces[i].ksni_name));
2840 net->ksnn_ninterfaces = i;
2843 /* call it before add it to ksocknal_data.ksnd_nets */
2844 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2848 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2849 net->ksnn_interfaces[0].ksni_ipaddr);
2850 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2852 ksocknal_data.ksnd_nnets++;
2857 LIBCFS_FREE(net, sizeof(*net));
2859 if (ksocknal_data.ksnd_nnets == 0)
2860 ksocknal_base_shutdown();
2867 ksocknal_module_fini (void)
2869 lnet_unregister_lnd(&the_ksocklnd);
2870 ksocknal_tunables_fini();
2874 ksocknal_module_init (void)
2878 /* check ksnr_connected/connecting field large enough */
2879 CLASSERT (SOCKLND_CONN_NTYPES <= 4);
2880 CLASSERT (SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2882 /* initialize the_ksocklnd */
2883 the_ksocklnd.lnd_type = SOCKLND;
2884 the_ksocklnd.lnd_startup = ksocknal_startup;
2885 the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2886 the_ksocklnd.lnd_ctl = ksocknal_ctl;
2887 the_ksocklnd.lnd_send = ksocknal_send;
2888 the_ksocklnd.lnd_recv = ksocknal_recv;
2889 the_ksocklnd.lnd_notify = ksocknal_notify;
2890 the_ksocklnd.lnd_query = ksocknal_query;
2891 the_ksocklnd.lnd_accept = ksocknal_accept;
2893 rc = ksocknal_tunables_init();
2897 lnet_register_lnd(&the_ksocklnd);
2902 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2903 MODULE_DESCRIPTION("Kernel TCP Socket LND v3.0.0");
2904 MODULE_LICENSE("GPL");
2906 cfs_module(ksocknal, "3.0.0", ksocknal_module_init, ksocknal_module_fini);