4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/socklnd/socklnd.c
34 * Author: Zach Brown <zab@zabbo.net>
35 * Author: Peter J. Braam <braam@clusterfs.com>
36 * Author: Phil Schwan <phil@clusterfs.com>
37 * Author: Eric Barton <eric@bartonsoftware.com>
40 #include <linux/pci.h>
43 static struct lnet_lnd the_ksocklnd;
44 struct ksock_nal_data ksocknal_data;
46 static struct ksock_interface *
47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
49 struct ksock_net *net = ni->ni_data;
51 struct ksock_interface *iface;
53 for (i = 0; i < net->ksnn_ninterfaces; i++) {
54 LASSERT(i < LNET_INTERFACES_NUM);
55 iface = &net->ksnn_interfaces[i];
57 if (iface->ksni_ipaddr == ip)
64 static struct ksock_route *
65 ksocknal_create_route(__u32 ipaddr, int port)
67 struct ksock_route *route;
69 LIBCFS_ALLOC (route, sizeof (*route));
73 atomic_set (&route->ksnr_refcount, 1);
74 route->ksnr_peer = NULL;
75 route->ksnr_retry_interval = 0; /* OK to connect at any time */
76 route->ksnr_ipaddr = ipaddr;
77 route->ksnr_port = port;
78 route->ksnr_scheduled = 0;
79 route->ksnr_connecting = 0;
80 route->ksnr_connected = 0;
81 route->ksnr_deleted = 0;
82 route->ksnr_conn_count = 0;
83 route->ksnr_share_count = 0;
89 ksocknal_destroy_route(struct ksock_route *route)
91 LASSERT (atomic_read(&route->ksnr_refcount) == 0);
93 if (route->ksnr_peer != NULL)
94 ksocknal_peer_decref(route->ksnr_peer);
96 LIBCFS_FREE (route, sizeof (*route));
100 ksocknal_create_peer(struct ksock_peer_ni **peerp, struct lnet_ni *ni,
101 struct lnet_process_id id)
103 int cpt = lnet_cpt_of_nid(id.nid, ni);
104 struct ksock_net *net = ni->ni_data;
105 struct ksock_peer_ni *peer_ni;
107 LASSERT(id.nid != LNET_NID_ANY);
108 LASSERT(id.pid != LNET_PID_ANY);
109 LASSERT(!in_interrupt());
111 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
115 peer_ni->ksnp_ni = ni;
116 peer_ni->ksnp_id = id;
117 atomic_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
118 peer_ni->ksnp_closing = 0;
119 peer_ni->ksnp_accepting = 0;
120 peer_ni->ksnp_proto = NULL;
121 peer_ni->ksnp_last_alive = 0;
122 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
124 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
125 INIT_LIST_HEAD(&peer_ni->ksnp_routes);
126 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
127 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
128 spin_lock_init(&peer_ni->ksnp_lock);
130 spin_lock_bh(&net->ksnn_lock);
132 if (net->ksnn_shutdown) {
133 spin_unlock_bh(&net->ksnn_lock);
135 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
136 CERROR("Can't create peer_ni: network shutdown\n");
142 spin_unlock_bh(&net->ksnn_lock);
149 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
151 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
153 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
154 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
156 LASSERT(atomic_read(&peer_ni->ksnp_refcount) == 0);
157 LASSERT(peer_ni->ksnp_accepting == 0);
158 LASSERT(list_empty(&peer_ni->ksnp_conns));
159 LASSERT(list_empty(&peer_ni->ksnp_routes));
160 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
161 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
163 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
165 /* NB a peer_ni's connections and routes keep a reference on their peer_ni
166 * until they are destroyed, so we can be assured that _all_ state to
167 * do with this peer_ni has been cleaned up when its refcount drops to
169 spin_lock_bh(&net->ksnn_lock);
171 spin_unlock_bh(&net->ksnn_lock);
174 struct ksock_peer_ni *
175 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
177 struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
178 struct list_head *tmp;
179 struct ksock_peer_ni *peer_ni;
181 list_for_each(tmp, peer_list) {
182 peer_ni = list_entry(tmp, struct ksock_peer_ni, ksnp_list);
184 LASSERT(!peer_ni->ksnp_closing);
186 if (peer_ni->ksnp_ni != ni)
189 if (peer_ni->ksnp_id.nid != id.nid ||
190 peer_ni->ksnp_id.pid != id.pid)
193 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
194 peer_ni, libcfs_id2str(id),
195 atomic_read(&peer_ni->ksnp_refcount));
201 struct ksock_peer_ni *
202 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
204 struct ksock_peer_ni *peer_ni;
206 read_lock(&ksocknal_data.ksnd_global_lock);
207 peer_ni = ksocknal_find_peer_locked(ni, id);
208 if (peer_ni != NULL) /* +1 ref for caller? */
209 ksocknal_peer_addref(peer_ni);
210 read_unlock(&ksocknal_data.ksnd_global_lock);
216 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
220 struct ksock_interface *iface;
222 for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
223 LASSERT(i < LNET_INTERFACES_NUM);
224 ip = peer_ni->ksnp_passive_ips[i];
226 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
228 * All IPs in peer_ni->ksnp_passive_ips[] come from the
229 * interface list, therefore the call must succeed.
231 LASSERT(iface != NULL);
233 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
234 peer_ni, iface, iface->ksni_nroutes);
235 iface->ksni_npeers--;
238 LASSERT(list_empty(&peer_ni->ksnp_conns));
239 LASSERT(list_empty(&peer_ni->ksnp_routes));
240 LASSERT(!peer_ni->ksnp_closing);
241 peer_ni->ksnp_closing = 1;
242 list_del(&peer_ni->ksnp_list);
243 /* lose peerlist's ref */
244 ksocknal_peer_decref(peer_ni);
248 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
249 struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
250 int *port, int *conn_count, int *share_count)
252 struct ksock_peer_ni *peer_ni;
253 struct list_head *ptmp;
254 struct ksock_route *route;
255 struct list_head *rtmp;
260 read_lock(&ksocknal_data.ksnd_global_lock);
262 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
263 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
264 peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
266 if (peer_ni->ksnp_ni != ni)
269 if (peer_ni->ksnp_n_passive_ips == 0 &&
270 list_empty(&peer_ni->ksnp_routes)) {
274 *id = peer_ni->ksnp_id;
284 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
288 *id = peer_ni->ksnp_id;
289 *myip = peer_ni->ksnp_passive_ips[j];
298 list_for_each(rtmp, &peer_ni->ksnp_routes) {
302 route = list_entry(rtmp, struct ksock_route,
305 *id = peer_ni->ksnp_id;
306 *myip = route->ksnr_myipaddr;
307 *peer_ip = route->ksnr_ipaddr;
308 *port = route->ksnr_port;
309 *conn_count = route->ksnr_conn_count;
310 *share_count = route->ksnr_share_count;
317 read_unlock(&ksocknal_data.ksnd_global_lock);
322 ksocknal_associate_route_conn_locked(struct ksock_route *route, struct ksock_conn *conn)
324 struct ksock_peer_ni *peer_ni = route->ksnr_peer;
325 int type = conn->ksnc_type;
326 struct ksock_interface *iface;
328 conn->ksnc_route = route;
329 ksocknal_route_addref(route);
331 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
332 if (route->ksnr_myipaddr == 0) {
333 /* route wasn't bound locally yet (the initial route) */
334 CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
335 libcfs_id2str(peer_ni->ksnp_id),
337 &conn->ksnc_myipaddr);
339 CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
340 "to %pI4h\n", libcfs_id2str(peer_ni->ksnp_id),
342 &route->ksnr_myipaddr,
343 &conn->ksnc_myipaddr);
345 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
346 route->ksnr_myipaddr);
348 iface->ksni_nroutes--;
350 route->ksnr_myipaddr = conn->ksnc_myipaddr;
351 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
352 route->ksnr_myipaddr);
354 iface->ksni_nroutes++;
357 route->ksnr_connected |= (1<<type);
358 route->ksnr_conn_count++;
360 /* Successful connection => further attempts can
361 * proceed immediately */
362 route->ksnr_retry_interval = 0;
366 ksocknal_add_route_locked(struct ksock_peer_ni *peer_ni, struct ksock_route *route)
368 struct list_head *tmp;
369 struct ksock_conn *conn;
370 struct ksock_route *route2;
372 LASSERT(!peer_ni->ksnp_closing);
373 LASSERT(route->ksnr_peer == NULL);
374 LASSERT(!route->ksnr_scheduled);
375 LASSERT(!route->ksnr_connecting);
376 LASSERT(route->ksnr_connected == 0);
378 /* LASSERT(unique) */
379 list_for_each(tmp, &peer_ni->ksnp_routes) {
380 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
382 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
383 CERROR("Duplicate route %s %pI4h\n",
384 libcfs_id2str(peer_ni->ksnp_id),
385 &route->ksnr_ipaddr);
390 route->ksnr_peer = peer_ni;
391 ksocknal_peer_addref(peer_ni);
392 /* peer_ni's routelist takes over my ref on 'route' */
393 list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
395 list_for_each(tmp, &peer_ni->ksnp_conns) {
396 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
398 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
401 ksocknal_associate_route_conn_locked(route, conn);
402 /* keep going (typed routes) */
407 ksocknal_del_route_locked(struct ksock_route *route)
409 struct ksock_peer_ni *peer_ni = route->ksnr_peer;
410 struct ksock_interface *iface;
411 struct ksock_conn *conn;
412 struct list_head *ctmp;
413 struct list_head *cnxt;
415 LASSERT(!route->ksnr_deleted);
417 /* Close associated conns */
418 list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
419 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
421 if (conn->ksnc_route != route)
424 ksocknal_close_conn_locked(conn, 0);
427 if (route->ksnr_myipaddr != 0) {
428 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
429 route->ksnr_myipaddr);
431 iface->ksni_nroutes--;
434 route->ksnr_deleted = 1;
435 list_del(&route->ksnr_list);
436 ksocknal_route_decref(route); /* drop peer_ni's ref */
438 if (list_empty(&peer_ni->ksnp_routes) &&
439 list_empty(&peer_ni->ksnp_conns)) {
440 /* I've just removed the last route to a peer_ni with no active
442 ksocknal_unlink_peer_locked(peer_ni);
447 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
450 struct list_head *tmp;
451 struct ksock_peer_ni *peer_ni;
452 struct ksock_peer_ni *peer2;
453 struct ksock_route *route;
454 struct ksock_route *route2;
457 if (id.nid == LNET_NID_ANY ||
458 id.pid == LNET_PID_ANY)
461 /* Have a brand new peer_ni ready... */
462 rc = ksocknal_create_peer(&peer_ni, ni, id);
466 route = ksocknal_create_route (ipaddr, port);
468 ksocknal_peer_decref(peer_ni);
472 write_lock_bh(&ksocknal_data.ksnd_global_lock);
474 /* always called with a ref on ni, so shutdown can't have started */
475 LASSERT(((struct ksock_net *) ni->ni_data)->ksnn_shutdown == 0);
477 peer2 = ksocknal_find_peer_locked(ni, id);
479 ksocknal_peer_decref(peer_ni);
482 /* peer_ni table takes my ref on peer_ni */
483 list_add_tail(&peer_ni->ksnp_list,
484 ksocknal_nid2peerlist(id.nid));
488 list_for_each(tmp, &peer_ni->ksnp_routes) {
489 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
491 if (route2->ksnr_ipaddr == ipaddr)
496 if (route2 == NULL) {
497 ksocknal_add_route_locked(peer_ni, route);
498 route->ksnr_share_count++;
500 ksocknal_route_decref(route);
501 route2->ksnr_share_count++;
504 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
510 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
512 struct ksock_conn *conn;
513 struct ksock_route *route;
514 struct list_head *tmp;
515 struct list_head *nxt;
518 LASSERT(!peer_ni->ksnp_closing);
520 /* Extra ref prevents peer_ni disappearing until I'm done with it */
521 ksocknal_peer_addref(peer_ni);
523 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
524 route = list_entry(tmp, struct ksock_route, ksnr_list);
527 if (!(ip == 0 || route->ksnr_ipaddr == ip))
530 route->ksnr_share_count = 0;
531 /* This deletes associated conns too */
532 ksocknal_del_route_locked(route);
536 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
537 route = list_entry(tmp, struct ksock_route, ksnr_list);
538 nshared += route->ksnr_share_count;
542 /* remove everything else if there are no explicit entries
545 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
546 route = list_entry(tmp, struct ksock_route, ksnr_list);
548 /* we should only be removing auto-entries */
549 LASSERT(route->ksnr_share_count == 0);
550 ksocknal_del_route_locked(route);
553 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
554 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
556 ksocknal_close_conn_locked(conn, 0);
560 ksocknal_peer_decref(peer_ni);
561 /* NB peer_ni unlinks itself when last conn/route is removed */
565 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
567 struct list_head zombies = LIST_HEAD_INIT(zombies);
568 struct list_head *ptmp;
569 struct list_head *pnxt;
570 struct ksock_peer_ni *peer_ni;
576 write_lock_bh(&ksocknal_data.ksnd_global_lock);
578 if (id.nid != LNET_NID_ANY) {
579 hi = (int)(ksocknal_nid2peerlist(id.nid) -
580 ksocknal_data.ksnd_peers);
584 hi = ksocknal_data.ksnd_peer_hash_size - 1;
587 for (i = lo; i <= hi; i++) {
588 list_for_each_safe(ptmp, pnxt,
589 &ksocknal_data.ksnd_peers[i]) {
590 peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
592 if (peer_ni->ksnp_ni != ni)
595 if (!((id.nid == LNET_NID_ANY ||
596 peer_ni->ksnp_id.nid == id.nid) &&
597 (id.pid == LNET_PID_ANY ||
598 peer_ni->ksnp_id.pid == id.pid)))
601 ksocknal_peer_addref(peer_ni); /* a ref for me... */
603 ksocknal_del_peer_locked(peer_ni, ip);
605 if (peer_ni->ksnp_closing &&
606 !list_empty(&peer_ni->ksnp_tx_queue)) {
607 LASSERT(list_empty(&peer_ni->ksnp_conns));
608 LASSERT(list_empty(&peer_ni->ksnp_routes));
610 list_splice_init(&peer_ni->ksnp_tx_queue,
614 ksocknal_peer_decref(peer_ni); /* ...till here */
616 rc = 0; /* matched! */
620 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
622 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
627 static struct ksock_conn *
628 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
630 struct ksock_peer_ni *peer_ni;
631 struct list_head *ptmp;
632 struct ksock_conn *conn;
633 struct list_head *ctmp;
636 read_lock(&ksocknal_data.ksnd_global_lock);
638 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
639 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
640 peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
642 LASSERT(!peer_ni->ksnp_closing);
644 if (peer_ni->ksnp_ni != ni)
647 list_for_each(ctmp, &peer_ni->ksnp_conns) {
651 conn = list_entry(ctmp, struct ksock_conn,
653 ksocknal_conn_addref(conn);
654 read_unlock(&ksocknal_data. \
661 read_unlock(&ksocknal_data.ksnd_global_lock);
665 static struct ksock_sched *
666 ksocknal_choose_scheduler_locked(unsigned int cpt)
668 struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
669 struct ksock_sched *sched;
672 if (info->ksi_nthreads == 0) {
673 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
674 if (info->ksi_nthreads > 0) {
675 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
684 sched = &info->ksi_scheds[0];
686 * NB: it's safe so far, but info->ksi_nthreads could be changed
687 * at runtime when we have dynamic LNet configuration, then we
688 * need to take care of this.
690 for (i = 1; i < info->ksi_nthreads; i++) {
691 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
692 sched = &info->ksi_scheds[i];
699 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
701 struct ksock_net *net = ni->ni_data;
705 read_lock(&ksocknal_data.ksnd_global_lock);
707 nip = net->ksnn_ninterfaces;
708 LASSERT(nip <= LNET_INTERFACES_NUM);
711 * Only offer interfaces for additional connections if I have
715 read_unlock(&ksocknal_data.ksnd_global_lock);
719 for (i = 0; i < nip; i++) {
720 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
721 LASSERT(ipaddrs[i] != 0);
724 read_unlock(&ksocknal_data.ksnd_global_lock);
729 ksocknal_match_peerip(struct ksock_interface *iface, __u32 *ips, int nips)
731 int best_netmatch = 0;
738 for (i = 0; i < nips; i++) {
742 this_xor = (ips[i] ^ iface->ksni_ipaddr);
743 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
746 best_netmatch < this_netmatch ||
747 (best_netmatch == this_netmatch &&
748 best_xor > this_xor)))
752 best_netmatch = this_netmatch;
761 ksocknal_select_ips(struct ksock_peer_ni *peer_ni, __u32 *peerips, int n_peerips)
763 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
764 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
765 struct ksock_interface *iface;
766 struct ksock_interface *best_iface;
777 /* CAVEAT EMPTOR: We do all our interface matching with an
778 * exclusive hold of global lock at IRQ priority. We're only
779 * expecting to be dealing with small numbers of interfaces, so the
780 * O(n**3)-ness shouldn't matter */
782 /* Also note that I'm not going to return more than n_peerips
783 * interfaces, even if I have more myself */
785 write_lock_bh(global_lock);
787 LASSERT(n_peerips <= LNET_INTERFACES_NUM);
788 LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
790 /* Only match interfaces for additional connections
791 * if I have > 1 interface */
792 n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
793 MIN(n_peerips, net->ksnn_ninterfaces);
795 for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
796 /* ^ yes really... */
798 /* If we have any new interfaces, first tick off all the
799 * peer_ni IPs that match old interfaces, then choose new
800 * interfaces to match the remaining peer_ni IPS.
801 * We don't forget interfaces we've stopped using; we might
802 * start using them again... */
804 if (i < peer_ni->ksnp_n_passive_ips) {
806 ip = peer_ni->ksnp_passive_ips[i];
807 best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
809 /* peer_ni passive ips are kept up to date */
810 LASSERT(best_iface != NULL);
812 /* choose a new interface */
813 LASSERT (i == peer_ni->ksnp_n_passive_ips);
819 for (j = 0; j < net->ksnn_ninterfaces; j++) {
820 iface = &net->ksnn_interfaces[j];
821 ip = iface->ksni_ipaddr;
823 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
824 if (peer_ni->ksnp_passive_ips[k] == ip)
827 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
830 k = ksocknal_match_peerip(iface, peerips, n_peerips);
831 xor = (ip ^ peerips[k]);
832 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
834 if (!(best_iface == NULL ||
835 best_netmatch < this_netmatch ||
836 (best_netmatch == this_netmatch &&
837 best_npeers > iface->ksni_npeers)))
841 best_netmatch = this_netmatch;
842 best_npeers = iface->ksni_npeers;
845 LASSERT(best_iface != NULL);
847 best_iface->ksni_npeers++;
848 ip = best_iface->ksni_ipaddr;
849 peer_ni->ksnp_passive_ips[i] = ip;
850 peer_ni->ksnp_n_passive_ips = i+1;
853 /* mark the best matching peer_ni IP used */
854 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
858 /* Overwrite input peer_ni IP addresses */
859 memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
861 write_unlock_bh(global_lock);
867 ksocknal_create_routes(struct ksock_peer_ni *peer_ni, int port,
868 __u32 *peer_ipaddrs, int npeer_ipaddrs)
870 struct ksock_route *newroute = NULL;
871 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
872 struct lnet_ni *ni = peer_ni->ksnp_ni;
873 struct ksock_net *net = ni->ni_data;
874 struct list_head *rtmp;
875 struct ksock_route *route;
876 struct ksock_interface *iface;
877 struct ksock_interface *best_iface;
884 /* CAVEAT EMPTOR: We do all our interface matching with an
885 * exclusive hold of global lock at IRQ priority. We're only
886 * expecting to be dealing with small numbers of interfaces, so the
887 * O(n**3)-ness here shouldn't matter */
889 write_lock_bh(global_lock);
891 if (net->ksnn_ninterfaces < 2) {
892 /* Only create additional connections
893 * if I have > 1 interface */
894 write_unlock_bh(global_lock);
898 LASSERT(npeer_ipaddrs <= LNET_INTERFACES_NUM);
900 for (i = 0; i < npeer_ipaddrs; i++) {
901 if (newroute != NULL) {
902 newroute->ksnr_ipaddr = peer_ipaddrs[i];
904 write_unlock_bh(global_lock);
906 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
907 if (newroute == NULL)
910 write_lock_bh(global_lock);
913 if (peer_ni->ksnp_closing) {
914 /* peer_ni got closed under me */
918 /* Already got a route? */
920 list_for_each(rtmp, &peer_ni->ksnp_routes) {
921 route = list_entry(rtmp, struct ksock_route, ksnr_list);
923 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
935 LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
937 /* Select interface to connect from */
938 for (j = 0; j < net->ksnn_ninterfaces; j++) {
939 iface = &net->ksnn_interfaces[j];
941 /* Using this interface already? */
942 list_for_each(rtmp, &peer_ni->ksnp_routes) {
943 route = list_entry(rtmp, struct ksock_route,
946 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
954 this_netmatch = (((iface->ksni_ipaddr ^
955 newroute->ksnr_ipaddr) &
956 iface->ksni_netmask) == 0) ? 1 : 0;
958 if (!(best_iface == NULL ||
959 best_netmatch < this_netmatch ||
960 (best_netmatch == this_netmatch &&
961 best_nroutes > iface->ksni_nroutes)))
965 best_netmatch = this_netmatch;
966 best_nroutes = iface->ksni_nroutes;
969 if (best_iface == NULL)
972 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
973 best_iface->ksni_nroutes++;
975 ksocknal_add_route_locked(peer_ni, newroute);
979 write_unlock_bh(global_lock);
980 if (newroute != NULL)
981 ksocknal_route_decref(newroute);
985 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
987 struct ksock_connreq *cr;
992 rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
993 LASSERT(rc == 0); /* we succeeded before */
995 LIBCFS_ALLOC(cr, sizeof(*cr));
997 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
998 "%pI4h: memory exhausted\n", &peer_ip);
1004 cr->ksncr_sock = sock;
1006 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
1008 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
1009 wake_up(&ksocknal_data.ksnd_connd_waitq);
1011 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1016 ksocknal_connecting(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
1018 struct ksock_route *route;
1020 list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
1021 if (route->ksnr_ipaddr == ipaddr)
1022 return route->ksnr_connecting;
1028 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
1029 struct socket *sock, int type)
1031 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1032 struct list_head zombies = LIST_HEAD_INIT(zombies);
1033 struct lnet_process_id peerid;
1034 struct list_head *tmp;
1036 struct ksock_conn *conn;
1037 struct ksock_conn *conn2;
1038 struct ksock_peer_ni *peer_ni = NULL;
1039 struct ksock_peer_ni *peer2;
1040 struct ksock_sched *sched;
1041 struct ksock_hello_msg *hello;
1043 struct ksock_tx *tx;
1044 struct ksock_tx *txtmp;
1050 active = (route != NULL);
1052 LASSERT (active == (type != SOCKLND_CONN_NONE));
1054 LIBCFS_ALLOC(conn, sizeof(*conn));
1060 conn->ksnc_peer = NULL;
1061 conn->ksnc_route = NULL;
1062 conn->ksnc_sock = sock;
1063 /* 2 ref, 1 for conn, another extra ref prevents socket
1064 * being closed before establishment of connection */
1065 atomic_set (&conn->ksnc_sock_refcount, 2);
1066 conn->ksnc_type = type;
1067 ksocknal_lib_save_callback(sock, conn);
1068 atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1070 conn->ksnc_rx_ready = 0;
1071 conn->ksnc_rx_scheduled = 0;
1073 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1074 conn->ksnc_tx_ready = 0;
1075 conn->ksnc_tx_scheduled = 0;
1076 conn->ksnc_tx_carrier = NULL;
1077 atomic_set (&conn->ksnc_tx_nob, 0);
1079 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1080 kshm_ips[LNET_INTERFACES_NUM]));
1081 if (hello == NULL) {
1086 /* stash conn's local and remote addrs */
1087 rc = ksocknal_lib_get_conn_addrs (conn);
1091 /* Find out/confirm peer_ni's NID and connection type and get the
1092 * vector of interfaces she's willing to let me connect to.
1093 * Passive connections use the listener timeout since the peer_ni sends
1097 peer_ni = route->ksnr_peer;
1098 LASSERT(ni == peer_ni->ksnp_ni);
1100 /* Active connection sends HELLO eagerly */
1101 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1102 peerid = peer_ni->ksnp_id;
1104 write_lock_bh(global_lock);
1105 conn->ksnc_proto = peer_ni->ksnp_proto;
1106 write_unlock_bh(global_lock);
1108 if (conn->ksnc_proto == NULL) {
1109 conn->ksnc_proto = &ksocknal_protocol_v3x;
1110 #if SOCKNAL_VERSION_DEBUG
1111 if (*ksocknal_tunables.ksnd_protocol == 2)
1112 conn->ksnc_proto = &ksocknal_protocol_v2x;
1113 else if (*ksocknal_tunables.ksnd_protocol == 1)
1114 conn->ksnc_proto = &ksocknal_protocol_v1x;
1118 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1122 peerid.nid = LNET_NID_ANY;
1123 peerid.pid = LNET_PID_ANY;
1125 /* Passive, get protocol from peer_ni */
1126 conn->ksnc_proto = NULL;
1129 rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1133 LASSERT (rc == 0 || active);
1134 LASSERT (conn->ksnc_proto != NULL);
1135 LASSERT (peerid.nid != LNET_NID_ANY);
1137 cpt = lnet_cpt_of_nid(peerid.nid, ni);
1140 ksocknal_peer_addref(peer_ni);
1141 write_lock_bh(global_lock);
1143 rc = ksocknal_create_peer(&peer_ni, ni, peerid);
1147 write_lock_bh(global_lock);
1149 /* called with a ref on ni, so shutdown can't have started */
1150 LASSERT(((struct ksock_net *) ni->ni_data)->ksnn_shutdown == 0);
1152 peer2 = ksocknal_find_peer_locked(ni, peerid);
1153 if (peer2 == NULL) {
1154 /* NB this puts an "empty" peer_ni in the peer_ni
1155 * table (which takes my ref) */
1156 list_add_tail(&peer_ni->ksnp_list,
1157 ksocknal_nid2peerlist(peerid.nid));
1159 ksocknal_peer_decref(peer_ni);
1164 ksocknal_peer_addref(peer_ni);
1165 peer_ni->ksnp_accepting++;
1167 /* Am I already connecting to this guy? Resolve in
1168 * favour of higher NID... */
1169 if (peerid.nid < ni->ni_nid &&
1170 ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1172 warn = "connection race resolution";
1177 if (peer_ni->ksnp_closing ||
1178 (active && route->ksnr_deleted)) {
1179 /* peer_ni/route got closed under me */
1181 warn = "peer_ni/route removed";
1185 if (peer_ni->ksnp_proto == NULL) {
1186 /* Never connected before.
1187 * NB recv_hello may have returned EPROTO to signal my peer_ni
1188 * wants a different protocol than the one I asked for.
1190 LASSERT(list_empty(&peer_ni->ksnp_conns));
1192 peer_ni->ksnp_proto = conn->ksnc_proto;
1193 peer_ni->ksnp_incarnation = incarnation;
1196 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1197 peer_ni->ksnp_incarnation != incarnation) {
1198 /* peer_ni rebooted or I've got the wrong protocol version */
1199 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1201 peer_ni->ksnp_proto = NULL;
1203 warn = peer_ni->ksnp_incarnation != incarnation ?
1204 "peer_ni rebooted" :
1205 "wrong proto version";
1215 warn = "lost conn race";
1218 warn = "retry with different protocol version";
1222 /* Refuse to duplicate an existing connection, unless this is a
1223 * loopback connection */
1224 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1225 list_for_each(tmp, &peer_ni->ksnp_conns) {
1226 conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1228 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1229 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1230 conn2->ksnc_type != conn->ksnc_type)
1233 /* Reply on a passive connection attempt so the peer_ni
1234 * realises we're connected. */
1244 /* If the connection created by this route didn't bind to the IP
1245 * address the route connected to, the connection/route matching
1246 * code below probably isn't going to work. */
1248 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1249 CERROR("Route %s %pI4h connected to %pI4h\n",
1250 libcfs_id2str(peer_ni->ksnp_id),
1251 &route->ksnr_ipaddr,
1252 &conn->ksnc_ipaddr);
1255 /* Search for a route corresponding to the new connection and
1256 * create an association. This allows incoming connections created
1257 * by routes in my peer_ni to match my own route entries so I don't
1258 * continually create duplicate routes. */
1259 list_for_each(tmp, &peer_ni->ksnp_routes) {
1260 route = list_entry(tmp, struct ksock_route, ksnr_list);
1262 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1265 ksocknal_associate_route_conn_locked(route, conn);
1269 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1270 peer_ni->ksnp_last_alive = ktime_get_seconds();
1271 peer_ni->ksnp_send_keepalive = 0;
1272 peer_ni->ksnp_error = 0;
1274 sched = ksocknal_choose_scheduler_locked(cpt);
1276 CERROR("no schedulers available. node is unhealthy\n");
1280 * The cpt might have changed if we ended up selecting a non cpt
1281 * native scheduler. So use the scheduler's cpt instead.
1283 cpt = sched->kss_info->ksi_cpt;
1284 sched->kss_nconns++;
1285 conn->ksnc_scheduler = sched;
1287 conn->ksnc_tx_last_post = ktime_get_seconds();
1288 /* Set the deadline for the outgoing HELLO to drain */
1289 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1290 conn->ksnc_tx_deadline = ktime_get_seconds() +
1291 *ksocknal_tunables.ksnd_timeout;
1292 smp_mb(); /* order with adding to peer_ni's conn list */
1294 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1295 ksocknal_conn_addref(conn);
1297 ksocknal_new_packet(conn, 0);
1299 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1301 /* Take packets blocking for this connection. */
1302 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1303 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1307 list_del(&tx->tx_list);
1308 ksocknal_queue_tx_locked(tx, conn);
1311 write_unlock_bh(global_lock);
1313 /* We've now got a new connection. Any errors from here on are just
1314 * like "normal" comms errors and we close the connection normally.
1315 * NB (a) we still have to send the reply HELLO for passive
1317 * (b) normal I/O on the conn is blocked until I setup and call the
1321 CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1322 " incarnation:%lld sched[%d:%d]\n",
1323 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1324 &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1325 conn->ksnc_port, incarnation, cpt,
1326 (int)(sched - &sched->kss_info->ksi_scheds[0]));
1329 /* additional routes after interface exchange? */
1330 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1331 hello->kshm_ips, hello->kshm_nips);
1333 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1335 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1338 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1339 kshm_ips[LNET_INTERFACES_NUM]));
1341 /* setup the socket AFTER I've received hello (it disables
1342 * SO_LINGER). I might call back to the acceptor who may want
1343 * to send a protocol version response and then close the
1344 * socket; this ensures the socket only tears down after the
1345 * response has been sent. */
1347 rc = ksocknal_lib_setup_sock(sock);
1349 write_lock_bh(global_lock);
1351 /* NB my callbacks block while I hold ksnd_global_lock */
1352 ksocknal_lib_set_callback(sock, conn);
1355 peer_ni->ksnp_accepting--;
1357 write_unlock_bh(global_lock);
1360 write_lock_bh(global_lock);
1361 if (!conn->ksnc_closing) {
1362 /* could be closed by another thread */
1363 ksocknal_close_conn_locked(conn, rc);
1365 write_unlock_bh(global_lock);
1366 } else if (ksocknal_connsock_addref(conn) == 0) {
1367 /* Allow I/O to proceed. */
1368 ksocknal_read_callback(conn);
1369 ksocknal_write_callback(conn);
1370 ksocknal_connsock_decref(conn);
1373 ksocknal_connsock_decref(conn);
1374 ksocknal_conn_decref(conn);
1378 if (!peer_ni->ksnp_closing &&
1379 list_empty(&peer_ni->ksnp_conns) &&
1380 list_empty(&peer_ni->ksnp_routes)) {
1381 list_add(&zombies, &peer_ni->ksnp_tx_queue);
1382 list_del_init(&peer_ni->ksnp_tx_queue);
1383 ksocknal_unlink_peer_locked(peer_ni);
1386 write_unlock_bh(global_lock);
1390 CERROR("Not creating conn %s type %d: %s\n",
1391 libcfs_id2str(peerid), conn->ksnc_type, warn);
1393 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1394 libcfs_id2str(peerid), conn->ksnc_type, warn);
1399 /* Request retry by replying with CONN_NONE
1400 * ksnc_proto has been set already */
1401 conn->ksnc_type = SOCKLND_CONN_NONE;
1402 hello->kshm_nips = 0;
1403 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1406 write_lock_bh(global_lock);
1407 peer_ni->ksnp_accepting--;
1408 write_unlock_bh(global_lock);
1412 * If we get here without an error code, just use -EALREADY.
1413 * Depending on how we got here, the error may be positive
1414 * or negative. Normalize the value for ksocknal_txlist_done().
1416 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1417 ksocknal_txlist_done(ni, &zombies, rc2);
1418 ksocknal_peer_decref(peer_ni);
1422 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1423 kshm_ips[LNET_INTERFACES_NUM]));
1425 LIBCFS_FREE(conn, sizeof(*conn));
1433 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1435 /* This just does the immmediate housekeeping, and queues the
1436 * connection for the reaper to terminate.
1437 * Caller holds ksnd_global_lock exclusively in irq context */
1438 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1439 struct ksock_route *route;
1440 struct ksock_conn *conn2;
1441 struct list_head *tmp;
1443 LASSERT(peer_ni->ksnp_error == 0);
1444 LASSERT(!conn->ksnc_closing);
1445 conn->ksnc_closing = 1;
1447 /* ksnd_deathrow_conns takes over peer_ni's ref */
1448 list_del(&conn->ksnc_list);
1450 route = conn->ksnc_route;
1451 if (route != NULL) {
1452 /* dissociate conn from route... */
1453 LASSERT(!route->ksnr_deleted);
1454 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1457 list_for_each(tmp, &peer_ni->ksnp_conns) {
1458 conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1460 if (conn2->ksnc_route == route &&
1461 conn2->ksnc_type == conn->ksnc_type)
1467 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1469 conn->ksnc_route = NULL;
1471 ksocknal_route_decref(route); /* drop conn's ref on route */
1474 if (list_empty(&peer_ni->ksnp_conns)) {
1475 /* No more connections to this peer_ni */
1477 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1478 struct ksock_tx *tx;
1480 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1482 /* throw them to the last connection...,
1483 * these TXs will be send to /dev/null by scheduler */
1484 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1486 ksocknal_tx_prep(conn, tx);
1488 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1489 list_splice_init(&peer_ni->ksnp_tx_queue,
1490 &conn->ksnc_tx_queue);
1491 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1494 /* renegotiate protocol version */
1495 peer_ni->ksnp_proto = NULL;
1496 /* stash last conn close reason */
1497 peer_ni->ksnp_error = error;
1499 if (list_empty(&peer_ni->ksnp_routes)) {
1500 /* I've just closed last conn belonging to a
1501 * peer_ni with no routes to it */
1502 ksocknal_unlink_peer_locked(peer_ni);
1506 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1508 list_add_tail(&conn->ksnc_list,
1509 &ksocknal_data.ksnd_deathrow_conns);
1510 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1512 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1516 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1519 time64_t last_alive = 0;
1521 /* There has been a connection failure or comms error; but I'll only
1522 * tell LNET I think the peer_ni is dead if it's to another kernel and
1523 * there are no connections or connection attempts in existence. */
1525 read_lock(&ksocknal_data.ksnd_global_lock);
1527 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1528 list_empty(&peer_ni->ksnp_conns) &&
1529 peer_ni->ksnp_accepting == 0 &&
1530 ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1532 last_alive = peer_ni->ksnp_last_alive;
1535 read_unlock(&ksocknal_data.ksnd_global_lock);
1538 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid, 0,
1539 cfs_time_seconds(last_alive)); /* to jiffies */
1543 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1545 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1546 struct ksock_tx *tx;
1547 struct ksock_tx *tmp;
1548 struct list_head zlist = LIST_HEAD_INIT(zlist);
1550 /* NB safe to finalize TXs because closing of socket will
1551 * abort all buffered data */
1552 LASSERT(conn->ksnc_sock == NULL);
1554 spin_lock(&peer_ni->ksnp_lock);
1556 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1557 if (tx->tx_conn != conn)
1560 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1562 tx->tx_msg.ksm_zc_cookies[0] = 0;
1563 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1564 list_del(&tx->tx_zc_list);
1565 list_add(&tx->tx_zc_list, &zlist);
1568 spin_unlock(&peer_ni->ksnp_lock);
1570 while (!list_empty(&zlist)) {
1571 tx = list_entry(zlist.next, struct ksock_tx, tx_zc_list);
1573 list_del(&tx->tx_zc_list);
1574 ksocknal_tx_decref(tx);
1579 ksocknal_terminate_conn(struct ksock_conn *conn)
1581 /* This gets called by the reaper (guaranteed thread context) to
1582 * disengage the socket from its callbacks and close it.
1583 * ksnc_refcount will eventually hit zero, and then the reaper will
1585 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1586 struct ksock_sched *sched = conn->ksnc_scheduler;
1589 LASSERT(conn->ksnc_closing);
1591 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1592 spin_lock_bh(&sched->kss_lock);
1594 /* a closing conn is always ready to tx */
1595 conn->ksnc_tx_ready = 1;
1597 if (!conn->ksnc_tx_scheduled &&
1598 !list_empty(&conn->ksnc_tx_queue)) {
1599 list_add_tail(&conn->ksnc_tx_list,
1600 &sched->kss_tx_conns);
1601 conn->ksnc_tx_scheduled = 1;
1602 /* extra ref for scheduler */
1603 ksocknal_conn_addref(conn);
1605 wake_up (&sched->kss_waitq);
1608 spin_unlock_bh(&sched->kss_lock);
1610 /* serialise with callbacks */
1611 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1613 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1615 /* OK, so this conn may not be completely disengaged from its
1616 * scheduler yet, but it _has_ committed to terminate... */
1617 conn->ksnc_scheduler->kss_nconns--;
1619 if (peer_ni->ksnp_error != 0) {
1620 /* peer_ni's last conn closed in error */
1621 LASSERT(list_empty(&peer_ni->ksnp_conns));
1623 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1626 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1629 ksocknal_peer_failed(peer_ni);
1631 /* The socket is closed on the final put; either here, or in
1632 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1633 * when the connection was established, this will close the socket
1634 * immediately, aborting anything buffered in it. Any hung
1635 * zero-copy transmits will therefore complete in finite time. */
1636 ksocknal_connsock_decref(conn);
1640 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1642 /* Queue the conn for the reaper to destroy */
1643 LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1644 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1646 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1647 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1649 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1653 ksocknal_destroy_conn(struct ksock_conn *conn)
1657 /* Final coup-de-grace of the reaper */
1658 CDEBUG (D_NET, "connection %p\n", conn);
1660 LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1661 LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1662 LASSERT (conn->ksnc_sock == NULL);
1663 LASSERT (conn->ksnc_route == NULL);
1664 LASSERT (!conn->ksnc_tx_scheduled);
1665 LASSERT (!conn->ksnc_rx_scheduled);
1666 LASSERT(list_empty(&conn->ksnc_tx_queue));
1668 /* complete current receive if any */
1669 switch (conn->ksnc_rx_state) {
1670 case SOCKNAL_RX_LNET_PAYLOAD:
1671 last_rcv = conn->ksnc_rx_deadline -
1672 *ksocknal_tunables.ksnd_timeout;
1673 CERROR("Completing partial receive from %s[%d], "
1674 "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1675 "last alive is %lld secs ago\n",
1676 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1677 &conn->ksnc_ipaddr, conn->ksnc_port,
1678 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1679 ktime_get_seconds() - last_rcv);
1680 lnet_finalize(conn->ksnc_cookie, -EIO);
1682 case SOCKNAL_RX_LNET_HEADER:
1683 if (conn->ksnc_rx_started)
1684 CERROR("Incomplete receive of lnet header from %s, "
1685 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1686 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1687 &conn->ksnc_ipaddr, conn->ksnc_port,
1688 conn->ksnc_proto->pro_version);
1690 case SOCKNAL_RX_KSM_HEADER:
1691 if (conn->ksnc_rx_started)
1692 CERROR("Incomplete receive of ksock message from %s, "
1693 "ip %pI4h:%d, with error, protocol: %d.x.\n",
1694 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1695 &conn->ksnc_ipaddr, conn->ksnc_port,
1696 conn->ksnc_proto->pro_version);
1698 case SOCKNAL_RX_SLOP:
1699 if (conn->ksnc_rx_started)
1700 CERROR("Incomplete receive of slops from %s, "
1701 "ip %pI4h:%d, with error\n",
1702 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1703 &conn->ksnc_ipaddr, conn->ksnc_port);
1710 ksocknal_peer_decref(conn->ksnc_peer);
1712 LIBCFS_FREE (conn, sizeof (*conn));
1716 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr, int why)
1718 struct ksock_conn *conn;
1719 struct list_head *ctmp;
1720 struct list_head *cnxt;
1723 list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1724 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
1727 conn->ksnc_ipaddr == ipaddr) {
1729 ksocknal_close_conn_locked (conn, why);
1737 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1739 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1740 u32 ipaddr = conn->ksnc_ipaddr;
1743 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1745 count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1747 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1753 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1755 struct ksock_peer_ni *peer_ni;
1756 struct list_head *ptmp;
1757 struct list_head *pnxt;
1763 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1765 if (id.nid != LNET_NID_ANY)
1766 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1769 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1772 for (i = lo; i <= hi; i++) {
1773 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1775 peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
1777 if (!((id.nid == LNET_NID_ANY || id.nid == peer_ni->ksnp_id.nid) &&
1778 (id.pid == LNET_PID_ANY || id.pid == peer_ni->ksnp_id.pid)))
1781 count += ksocknal_close_peer_conns_locked (peer_ni, ipaddr, 0);
1785 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1787 /* wildcards always succeed */
1788 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1791 return (count == 0 ? -ENOENT : 0);
1795 ksocknal_notify(struct lnet_ni *ni, lnet_nid_t gw_nid, int alive)
1797 /* The router is telling me she's been notified of a change in
1800 struct lnet_process_id id = {
1802 .pid = LNET_PID_ANY,
1805 CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1806 alive ? "up" : "down");
1809 /* If the gateway crashed, close all open connections... */
1810 ksocknal_close_matching_conns (id, 0);
1814 /* ...otherwise do nothing. We can only establish new connections
1815 * if we have autroutes, and these connect on demand. */
1819 ksocknal_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1822 time64_t last_alive = 0;
1823 time64_t now = ktime_get_seconds();
1824 struct ksock_peer_ni *peer_ni = NULL;
1825 rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1826 struct lnet_process_id id = {
1828 .pid = LNET_PID_LUSTRE,
1833 peer_ni = ksocknal_find_peer_locked(ni, id);
1834 if (peer_ni != NULL) {
1835 struct list_head *tmp;
1836 struct ksock_conn *conn;
1839 list_for_each(tmp, &peer_ni->ksnp_conns) {
1840 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
1841 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1843 if (bufnob < conn->ksnc_tx_bufnob) {
1844 /* something got ACKed */
1845 conn->ksnc_tx_deadline = ktime_get_seconds() +
1846 *ksocknal_tunables.ksnd_timeout;
1847 peer_ni->ksnp_last_alive = now;
1848 conn->ksnc_tx_bufnob = bufnob;
1852 last_alive = peer_ni->ksnp_last_alive;
1853 if (ksocknal_find_connectable_route_locked(peer_ni) == NULL)
1859 if (last_alive != 0)
1862 CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago, connect %d\n",
1863 libcfs_nid2str(nid), peer_ni,
1864 last_alive ? now - last_alive : -1,
1870 ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1872 write_lock_bh(glock);
1874 peer_ni = ksocknal_find_peer_locked(ni, id);
1875 if (peer_ni != NULL)
1876 ksocknal_launch_all_connections_locked(peer_ni);
1878 write_unlock_bh(glock);
1883 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1887 struct list_head *tmp;
1888 struct ksock_conn *conn;
1890 for (index = 0; ; index++) {
1891 read_lock(&ksocknal_data.ksnd_global_lock);
1896 list_for_each(tmp, &peer_ni->ksnp_conns) {
1898 conn = list_entry(tmp, struct ksock_conn,
1900 ksocknal_conn_addref(conn);
1905 read_unlock(&ksocknal_data.ksnd_global_lock);
1910 ksocknal_lib_push_conn (conn);
1911 ksocknal_conn_decref(conn);
1916 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1918 struct list_head *start;
1919 struct list_head *end;
1920 struct list_head *tmp;
1922 unsigned int hsize = ksocknal_data.ksnd_peer_hash_size;
1924 if (id.nid == LNET_NID_ANY) {
1925 start = &ksocknal_data.ksnd_peers[0];
1926 end = &ksocknal_data.ksnd_peers[hsize - 1];
1928 start = end = ksocknal_nid2peerlist(id.nid);
1931 for (tmp = start; tmp <= end; tmp++) {
1932 int peer_off; /* searching offset in peer_ni hash table */
1934 for (peer_off = 0; ; peer_off++) {
1935 struct ksock_peer_ni *peer_ni;
1938 read_lock(&ksocknal_data.ksnd_global_lock);
1939 list_for_each_entry(peer_ni, tmp, ksnp_list) {
1940 if (!((id.nid == LNET_NID_ANY ||
1941 id.nid == peer_ni->ksnp_id.nid) &&
1942 (id.pid == LNET_PID_ANY ||
1943 id.pid == peer_ni->ksnp_id.pid)))
1946 if (i++ == peer_off) {
1947 ksocknal_peer_addref(peer_ni);
1951 read_unlock(&ksocknal_data.ksnd_global_lock);
1953 if (i == 0) /* no match */
1957 ksocknal_push_peer(peer_ni);
1958 ksocknal_peer_decref(peer_ni);
1965 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1967 struct ksock_net *net = ni->ni_data;
1968 struct ksock_interface *iface;
1972 struct list_head *ptmp;
1973 struct ksock_peer_ni *peer_ni;
1974 struct list_head *rtmp;
1975 struct ksock_route *route;
1977 if (ipaddress == 0 ||
1981 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1983 iface = ksocknal_ip2iface(ni, ipaddress);
1984 if (iface != NULL) {
1985 /* silently ignore dups */
1987 } else if (net->ksnn_ninterfaces == LNET_INTERFACES_NUM) {
1990 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1992 iface->ksni_ipaddr = ipaddress;
1993 iface->ksni_netmask = netmask;
1994 iface->ksni_nroutes = 0;
1995 iface->ksni_npeers = 0;
1997 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1998 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1999 peer_ni = list_entry(ptmp, struct ksock_peer_ni,
2002 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
2003 if (peer_ni->ksnp_passive_ips[j] == ipaddress)
2004 iface->ksni_npeers++;
2006 list_for_each(rtmp, &peer_ni->ksnp_routes) {
2007 route = list_entry(rtmp,
2011 if (route->ksnr_myipaddr == ipaddress)
2012 iface->ksni_nroutes++;
2018 /* NB only new connections will pay attention to the new interface! */
2021 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2027 ksocknal_peer_del_interface_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
2029 struct list_head *tmp;
2030 struct list_head *nxt;
2031 struct ksock_route *route;
2032 struct ksock_conn *conn;
2036 for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
2037 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
2038 for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
2039 peer_ni->ksnp_passive_ips[j-1] =
2040 peer_ni->ksnp_passive_ips[j];
2041 peer_ni->ksnp_n_passive_ips--;
2045 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
2046 route = list_entry(tmp, struct ksock_route, ksnr_list);
2048 if (route->ksnr_myipaddr != ipaddr)
2051 if (route->ksnr_share_count != 0) {
2052 /* Manually created; keep, but unbind */
2053 route->ksnr_myipaddr = 0;
2055 ksocknal_del_route_locked(route);
2059 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
2060 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
2062 if (conn->ksnc_myipaddr == ipaddr)
2063 ksocknal_close_conn_locked (conn, 0);
2068 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
2070 struct ksock_net *net = ni->ni_data;
2072 struct list_head *tmp;
2073 struct list_head *nxt;
2074 struct ksock_peer_ni *peer_ni;
2079 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2081 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2082 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2084 if (!(ipaddress == 0 ||
2085 ipaddress == this_ip))
2090 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2091 net->ksnn_interfaces[j-1] =
2092 net->ksnn_interfaces[j];
2094 net->ksnn_ninterfaces--;
2096 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2097 list_for_each_safe(tmp, nxt,
2098 &ksocknal_data.ksnd_peers[j]) {
2099 peer_ni = list_entry(tmp, struct ksock_peer_ni,
2102 if (peer_ni->ksnp_ni != ni)
2105 ksocknal_peer_del_interface_locked(peer_ni, this_ip);
2110 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2116 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2118 struct lnet_process_id id = {0};
2119 struct libcfs_ioctl_data *data = arg;
2123 case IOC_LIBCFS_GET_INTERFACE: {
2124 struct ksock_net *net = ni->ni_data;
2125 struct ksock_interface *iface;
2127 read_lock(&ksocknal_data.ksnd_global_lock);
2129 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2133 iface = &net->ksnn_interfaces[data->ioc_count];
2135 data->ioc_u32[0] = iface->ksni_ipaddr;
2136 data->ioc_u32[1] = iface->ksni_netmask;
2137 data->ioc_u32[2] = iface->ksni_npeers;
2138 data->ioc_u32[3] = iface->ksni_nroutes;
2141 read_unlock(&ksocknal_data.ksnd_global_lock);
2145 case IOC_LIBCFS_ADD_INTERFACE:
2146 return ksocknal_add_interface(ni,
2147 data->ioc_u32[0], /* IP address */
2148 data->ioc_u32[1]); /* net mask */
2150 case IOC_LIBCFS_DEL_INTERFACE:
2151 return ksocknal_del_interface(ni,
2152 data->ioc_u32[0]); /* IP address */
2154 case IOC_LIBCFS_GET_PEER: {
2159 int share_count = 0;
2161 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2162 &id, &myip, &ip, &port,
2163 &conn_count, &share_count);
2167 data->ioc_nid = id.nid;
2168 data->ioc_count = share_count;
2169 data->ioc_u32[0] = ip;
2170 data->ioc_u32[1] = port;
2171 data->ioc_u32[2] = myip;
2172 data->ioc_u32[3] = conn_count;
2173 data->ioc_u32[4] = id.pid;
2177 case IOC_LIBCFS_ADD_PEER:
2178 id.nid = data->ioc_nid;
2179 id.pid = LNET_PID_LUSTRE;
2180 return ksocknal_add_peer (ni, id,
2181 data->ioc_u32[0], /* IP */
2182 data->ioc_u32[1]); /* port */
2184 case IOC_LIBCFS_DEL_PEER:
2185 id.nid = data->ioc_nid;
2186 id.pid = LNET_PID_ANY;
2187 return ksocknal_del_peer (ni, id,
2188 data->ioc_u32[0]); /* IP */
2190 case IOC_LIBCFS_GET_CONN: {
2194 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
2199 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2201 data->ioc_count = txmem;
2202 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
2203 data->ioc_flags = nagle;
2204 data->ioc_u32[0] = conn->ksnc_ipaddr;
2205 data->ioc_u32[1] = conn->ksnc_port;
2206 data->ioc_u32[2] = conn->ksnc_myipaddr;
2207 data->ioc_u32[3] = conn->ksnc_type;
2208 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2209 data->ioc_u32[5] = rxmem;
2210 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2211 ksocknal_conn_decref(conn);
2215 case IOC_LIBCFS_CLOSE_CONNECTION:
2216 id.nid = data->ioc_nid;
2217 id.pid = LNET_PID_ANY;
2218 return ksocknal_close_matching_conns (id,
2221 case IOC_LIBCFS_REGISTER_MYNID:
2222 /* Ignore if this is a noop */
2223 if (data->ioc_nid == ni->ni_nid)
2226 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2227 libcfs_nid2str(data->ioc_nid),
2228 libcfs_nid2str(ni->ni_nid));
2231 case IOC_LIBCFS_PUSH_CONNECTION:
2232 id.nid = data->ioc_nid;
2233 id.pid = LNET_PID_ANY;
2234 return ksocknal_push(ni, id);
2243 ksocknal_free_buffers (void)
2245 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2247 if (ksocknal_data.ksnd_sched_info != NULL) {
2248 struct ksock_sched_info *info;
2251 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2252 if (info->ksi_scheds != NULL) {
2253 LIBCFS_FREE(info->ksi_scheds,
2254 info->ksi_nthreads_max *
2255 sizeof(info->ksi_scheds[0]));
2258 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2261 LIBCFS_FREE (ksocknal_data.ksnd_peers,
2262 sizeof(struct list_head) *
2263 ksocknal_data.ksnd_peer_hash_size);
2265 spin_lock(&ksocknal_data.ksnd_tx_lock);
2267 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2268 struct list_head zlist;
2269 struct ksock_tx *tx;
2271 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2272 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2273 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2275 while (!list_empty(&zlist)) {
2276 tx = list_entry(zlist.next, struct ksock_tx, tx_list);
2277 list_del(&tx->tx_list);
2278 LIBCFS_FREE(tx, tx->tx_desc_size);
2281 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2286 ksocknal_base_shutdown(void)
2288 struct ksock_sched_info *info;
2289 struct ksock_sched *sched;
2293 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2294 atomic_read (&libcfs_kmemory));
2295 LASSERT (ksocknal_data.ksnd_nnets == 0);
2297 switch (ksocknal_data.ksnd_init) {
2301 case SOCKNAL_INIT_ALL:
2302 case SOCKNAL_INIT_DATA:
2303 LASSERT (ksocknal_data.ksnd_peers != NULL);
2304 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2305 LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2308 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2309 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2310 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2311 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2312 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2314 if (ksocknal_data.ksnd_sched_info != NULL) {
2315 cfs_percpt_for_each(info, i,
2316 ksocknal_data.ksnd_sched_info) {
2317 if (info->ksi_scheds == NULL)
2320 for (j = 0; j < info->ksi_nthreads_max; j++) {
2322 sched = &info->ksi_scheds[j];
2323 LASSERT(list_empty(&sched->\
2325 LASSERT(list_empty(&sched->\
2327 LASSERT(list_empty(&sched-> \
2328 kss_zombie_noop_txs));
2329 LASSERT(sched->kss_nconns == 0);
2334 /* flag threads to terminate; wake and wait for them to die */
2335 ksocknal_data.ksnd_shuttingdown = 1;
2336 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2337 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2339 if (ksocknal_data.ksnd_sched_info != NULL) {
2340 cfs_percpt_for_each(info, i,
2341 ksocknal_data.ksnd_sched_info) {
2342 if (info->ksi_scheds == NULL)
2345 for (j = 0; j < info->ksi_nthreads_max; j++) {
2346 sched = &info->ksi_scheds[j];
2347 wake_up_all(&sched->kss_waitq);
2353 read_lock(&ksocknal_data.ksnd_global_lock);
2354 while (ksocknal_data.ksnd_nthreads != 0) {
2357 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2358 "waiting for %d threads to terminate\n",
2359 ksocknal_data.ksnd_nthreads);
2360 read_unlock(&ksocknal_data.ksnd_global_lock);
2361 set_current_state(TASK_UNINTERRUPTIBLE);
2362 schedule_timeout(cfs_time_seconds(1));
2363 read_lock(&ksocknal_data.ksnd_global_lock);
2365 read_unlock(&ksocknal_data.ksnd_global_lock);
2367 ksocknal_free_buffers();
2369 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2373 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2374 atomic_read (&libcfs_kmemory));
2376 module_put(THIS_MODULE);
2380 ksocknal_base_startup(void)
2382 struct ksock_sched_info *info;
2386 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2387 LASSERT (ksocknal_data.ksnd_nnets == 0);
2389 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2391 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2392 LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2393 sizeof(struct list_head) *
2394 ksocknal_data.ksnd_peer_hash_size);
2395 if (ksocknal_data.ksnd_peers == NULL)
2398 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2399 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2401 rwlock_init(&ksocknal_data.ksnd_global_lock);
2402 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2404 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2405 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2406 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2407 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2408 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2410 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2411 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2412 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2413 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2415 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2416 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2418 /* NB memset above zeros whole of ksocknal_data */
2420 /* flag lists/ptrs/locks initialised */
2421 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2422 try_module_get(THIS_MODULE);
2424 ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2426 if (ksocknal_data.ksnd_sched_info == NULL)
2429 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2430 struct ksock_sched *sched;
2433 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2434 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2435 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2437 /* max to half of CPUs, assume another half should be
2438 * reserved for upper layer modules */
2439 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2442 info->ksi_nthreads_max = nthrs;
2446 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2447 info->ksi_nthreads_max *
2449 if (info->ksi_scheds == NULL)
2452 for (; nthrs > 0; nthrs--) {
2453 sched = &info->ksi_scheds[nthrs - 1];
2455 sched->kss_info = info;
2456 spin_lock_init(&sched->kss_lock);
2457 INIT_LIST_HEAD(&sched->kss_rx_conns);
2458 INIT_LIST_HEAD(&sched->kss_tx_conns);
2459 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2460 init_waitqueue_head(&sched->kss_waitq);
2465 ksocknal_data.ksnd_connd_starting = 0;
2466 ksocknal_data.ksnd_connd_failed_stamp = 0;
2467 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2468 /* must have at least 2 connds to remain responsive to accepts while
2470 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2471 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2473 if (*ksocknal_tunables.ksnd_nconnds_max <
2474 *ksocknal_tunables.ksnd_nconnds) {
2475 ksocknal_tunables.ksnd_nconnds_max =
2476 ksocknal_tunables.ksnd_nconnds;
2479 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2481 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2482 ksocknal_data.ksnd_connd_starting++;
2483 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2486 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2487 rc = ksocknal_thread_start(ksocknal_connd,
2488 (void *)((uintptr_t)i), name);
2490 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2491 ksocknal_data.ksnd_connd_starting--;
2492 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2493 CERROR("Can't spawn socknal connd: %d\n", rc);
2498 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2500 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2504 /* flag everything initialised */
2505 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2510 ksocknal_base_shutdown();
2515 ksocknal_debug_peerhash(struct lnet_ni *ni)
2517 struct ksock_peer_ni *peer_ni = NULL;
2518 struct list_head *tmp;
2521 read_lock(&ksocknal_data.ksnd_global_lock);
2523 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2524 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2525 peer_ni = list_entry(tmp, struct ksock_peer_ni, ksnp_list);
2527 if (peer_ni->ksnp_ni == ni) break;
2533 if (peer_ni != NULL) {
2534 struct ksock_route *route;
2535 struct ksock_conn *conn;
2537 CWARN ("Active peer_ni on shutdown: %s, ref %d, scnt %d, "
2538 "closing %d, accepting %d, err %d, zcookie %llu, "
2539 "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2540 atomic_read(&peer_ni->ksnp_refcount),
2541 peer_ni->ksnp_sharecount, peer_ni->ksnp_closing,
2542 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2543 peer_ni->ksnp_zc_next_cookie,
2544 !list_empty(&peer_ni->ksnp_tx_queue),
2545 !list_empty(&peer_ni->ksnp_zc_req_list));
2547 list_for_each(tmp, &peer_ni->ksnp_routes) {
2548 route = list_entry(tmp, struct ksock_route, ksnr_list);
2549 CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2550 "del %d\n", atomic_read(&route->ksnr_refcount),
2551 route->ksnr_scheduled, route->ksnr_connecting,
2552 route->ksnr_connected, route->ksnr_deleted);
2555 list_for_each(tmp, &peer_ni->ksnp_conns) {
2556 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
2557 CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2558 atomic_read(&conn->ksnc_conn_refcount),
2559 atomic_read(&conn->ksnc_sock_refcount),
2560 conn->ksnc_type, conn->ksnc_closing);
2564 read_unlock(&ksocknal_data.ksnd_global_lock);
2569 ksocknal_shutdown(struct lnet_ni *ni)
2571 struct ksock_net *net = ni->ni_data;
2572 struct lnet_process_id anyid = {
2573 .nid = LNET_NID_ANY,
2574 .pid = LNET_PID_ANY,
2578 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2579 LASSERT(ksocknal_data.ksnd_nnets > 0);
2581 spin_lock_bh(&net->ksnn_lock);
2582 net->ksnn_shutdown = 1; /* prevent new peers */
2583 spin_unlock_bh(&net->ksnn_lock);
2585 /* Delete all peers */
2586 ksocknal_del_peer(ni, anyid, 0);
2588 /* Wait for all peer_ni state to clean up */
2590 spin_lock_bh(&net->ksnn_lock);
2591 while (net->ksnn_npeers != 0) {
2592 spin_unlock_bh(&net->ksnn_lock);
2595 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2596 "waiting for %d peers to disconnect\n",
2598 set_current_state(TASK_UNINTERRUPTIBLE);
2599 schedule_timeout(cfs_time_seconds(1));
2601 ksocknal_debug_peerhash(ni);
2603 spin_lock_bh(&net->ksnn_lock);
2605 spin_unlock_bh(&net->ksnn_lock);
2607 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2608 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2609 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2612 list_del(&net->ksnn_list);
2613 LIBCFS_FREE(net, sizeof(*net));
2615 ksocknal_data.ksnd_nnets--;
2616 if (ksocknal_data.ksnd_nnets == 0)
2617 ksocknal_base_shutdown();
2621 ksocknal_enumerate_interfaces(struct ksock_net *net)
2629 n = lnet_ipif_enumerate(&names);
2631 CERROR("Can't enumerate interfaces: %d\n", n);
2635 for (i = j = 0; i < n; i++) {
2640 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2643 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2645 CWARN("Can't get interface %s info: %d\n",
2651 CWARN("Ignoring interface %s (down)\n",
2656 if (j == LNET_INTERFACES_NUM) {
2657 CWARN("Ignoring interface %s (too many interfaces)\n",
2662 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2663 net->ksnn_interfaces[j].ksni_netmask = mask;
2664 strlcpy(net->ksnn_interfaces[j].ksni_name,
2665 names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2669 lnet_ipif_free_enumeration(names, n);
2672 CERROR("Can't find any usable interfaces\n");
2678 ksocknal_search_new_ipif(struct ksock_net *net)
2683 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2684 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2685 char *colon = strchr(ifnam, ':');
2687 struct ksock_net *tmp;
2690 if (colon != NULL) /* ignore alias device */
2693 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2695 for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2696 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2698 char *colon2 = strchr(ifnam2, ':');
2703 found = strcmp(ifnam, ifnam2) == 0;
2720 ksocknal_start_schedulers(struct ksock_sched_info *info)
2726 if (info->ksi_nthreads == 0) {
2727 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2728 nthrs = info->ksi_nthreads_max;
2730 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2732 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2733 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2735 nthrs = min(nthrs, info->ksi_nthreads_max);
2737 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2738 /* increase two threads if there is new interface */
2739 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2742 for (i = 0; i < nthrs; i++) {
2745 struct ksock_sched *sched;
2747 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2748 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2749 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2750 info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2752 rc = ksocknal_thread_start(ksocknal_scheduler,
2757 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2758 info->ksi_cpt, info->ksi_nthreads + i, rc);
2762 info->ksi_nthreads += i;
2767 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2769 int newif = ksocknal_search_new_ipif(net);
2773 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2776 for (i = 0; i < ncpts; i++) {
2777 struct ksock_sched_info *info;
2778 int cpt = (cpts == NULL) ? i : cpts[i];
2780 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2781 info = ksocknal_data.ksnd_sched_info[cpt];
2783 if (!newif && info->ksi_nthreads > 0)
2786 rc = ksocknal_start_schedulers(info);
2794 ksocknal_startup(struct lnet_ni *ni)
2796 struct ksock_net *net;
2799 struct net_device *net_dev;
2802 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2804 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2805 rc = ksocknal_base_startup();
2810 LIBCFS_ALLOC(net, sizeof(*net));
2814 spin_lock_init(&net->ksnn_lock);
2815 net->ksnn_incarnation = ktime_get_real_ns();
2817 if (!ni->ni_net->net_tunables_set) {
2818 ni->ni_net->net_tunables.lct_peer_timeout =
2819 *ksocknal_tunables.ksnd_peertimeout;
2820 ni->ni_net->net_tunables.lct_max_tx_credits =
2821 *ksocknal_tunables.ksnd_credits;
2822 ni->ni_net->net_tunables.lct_peer_tx_credits =
2823 *ksocknal_tunables.ksnd_peertxcredits;
2824 ni->ni_net->net_tunables.lct_peer_rtr_credits =
2825 *ksocknal_tunables.ksnd_peerrtrcredits;
2826 ni->ni_net->net_tunables_set = true;
2830 if (ni->ni_interfaces[0] == NULL) {
2831 rc = ksocknal_enumerate_interfaces(net);
2835 net->ksnn_ninterfaces = 1;
2837 for (i = 0; i < LNET_INTERFACES_NUM; i++) {
2840 if (ni->ni_interfaces[i] == NULL)
2843 rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2844 &net->ksnn_interfaces[i].ksni_ipaddr,
2845 &net->ksnn_interfaces[i].ksni_netmask);
2848 CERROR("Can't get interface %s info: %d\n",
2849 ni->ni_interfaces[i], rc);
2854 CERROR("Interface %s is down\n",
2855 ni->ni_interfaces[i]);
2859 strlcpy(net->ksnn_interfaces[i].ksni_name,
2860 ni->ni_interfaces[i],
2861 sizeof(net->ksnn_interfaces[i].ksni_name));
2864 net->ksnn_ninterfaces = i;
2867 net_dev = dev_get_by_name(&init_net,
2868 net->ksnn_interfaces[0].ksni_name);
2869 if (net_dev != NULL) {
2870 node_id = dev_to_node(&net_dev->dev);
2871 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
2874 ni->ni_dev_cpt = CFS_CPT_ANY;
2877 /* call it before add it to ksocknal_data.ksnd_nets */
2878 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2882 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2883 net->ksnn_interfaces[0].ksni_ipaddr);
2884 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2886 ksocknal_data.ksnd_nnets++;
2891 LIBCFS_FREE(net, sizeof(*net));
2893 if (ksocknal_data.ksnd_nnets == 0)
2894 ksocknal_base_shutdown();
2900 static void __exit ksocklnd_exit(void)
2902 lnet_unregister_lnd(&the_ksocklnd);
2905 static int __init ksocklnd_init(void)
2909 /* check ksnr_connected/connecting field large enough */
2910 CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2911 CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2913 /* initialize the_ksocklnd */
2914 the_ksocklnd.lnd_type = SOCKLND;
2915 the_ksocklnd.lnd_startup = ksocknal_startup;
2916 the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2917 the_ksocklnd.lnd_ctl = ksocknal_ctl;
2918 the_ksocklnd.lnd_send = ksocknal_send;
2919 the_ksocklnd.lnd_recv = ksocknal_recv;
2920 the_ksocklnd.lnd_notify = ksocknal_notify;
2921 the_ksocklnd.lnd_query = ksocknal_query;
2922 the_ksocklnd.lnd_accept = ksocknal_accept;
2924 rc = ksocknal_tunables_init();
2928 lnet_register_lnd(&the_ksocklnd);
2933 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2934 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2935 MODULE_VERSION("2.8.0");
2936 MODULE_LICENSE("GPL");
2938 module_init(ksocklnd_init);
2939 module_exit(ksocklnd_exit);