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, 2012, 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));
118 atomic_set (&peer->ksnp_refcount, 1); /* 1 ref for caller */
119 peer->ksnp_closing = 0;
120 peer->ksnp_accepting = 0;
121 peer->ksnp_proto = NULL;
122 peer->ksnp_last_alive = 0;
123 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
125 INIT_LIST_HEAD (&peer->ksnp_conns);
126 INIT_LIST_HEAD (&peer->ksnp_routes);
127 INIT_LIST_HEAD (&peer->ksnp_tx_queue);
128 INIT_LIST_HEAD (&peer->ksnp_zc_req_list);
129 spin_lock_init(&peer->ksnp_lock);
131 spin_lock_bh(&net->ksnn_lock);
133 if (net->ksnn_shutdown) {
134 spin_unlock_bh(&net->ksnn_lock);
136 LIBCFS_FREE(peer, sizeof(*peer));
137 CERROR("Can't create peer: network shutdown\n");
143 spin_unlock_bh(&net->ksnn_lock);
150 ksocknal_destroy_peer (ksock_peer_t *peer)
152 ksock_net_t *net = peer->ksnp_ni->ni_data;
154 CDEBUG (D_NET, "peer %s %p deleted\n",
155 libcfs_id2str(peer->ksnp_id), peer);
157 LASSERT (atomic_read (&peer->ksnp_refcount) == 0);
158 LASSERT (peer->ksnp_accepting == 0);
159 LASSERT (list_empty (&peer->ksnp_conns));
160 LASSERT (list_empty (&peer->ksnp_routes));
161 LASSERT (list_empty (&peer->ksnp_tx_queue));
162 LASSERT (list_empty (&peer->ksnp_zc_req_list));
164 LIBCFS_FREE (peer, sizeof (*peer));
166 /* NB a peer's connections and routes keep a reference on their peer
167 * until they are destroyed, so we can be assured that _all_ state to
168 * do with this peer has been cleaned up when its refcount drops to
170 spin_lock_bh(&net->ksnn_lock);
172 spin_unlock_bh(&net->ksnn_lock);
176 ksocknal_find_peer_locked (lnet_ni_t *ni, lnet_process_id_t id)
178 struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
179 struct list_head *tmp;
182 list_for_each (tmp, peer_list) {
184 peer = list_entry (tmp, ksock_peer_t, ksnp_list);
186 LASSERT (!peer->ksnp_closing);
188 if (peer->ksnp_ni != ni)
191 if (peer->ksnp_id.nid != id.nid ||
192 peer->ksnp_id.pid != id.pid)
195 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
196 peer, libcfs_id2str(id),
197 atomic_read(&peer->ksnp_refcount));
204 ksocknal_find_peer (lnet_ni_t *ni, lnet_process_id_t id)
208 read_lock(&ksocknal_data.ksnd_global_lock);
209 peer = ksocknal_find_peer_locked(ni, id);
210 if (peer != NULL) /* +1 ref for caller? */
211 ksocknal_peer_addref(peer);
212 read_unlock(&ksocknal_data.ksnd_global_lock);
218 ksocknal_unlink_peer_locked (ksock_peer_t *peer)
222 ksock_interface_t *iface;
224 for (i = 0; i < peer->ksnp_n_passive_ips; i++) {
225 LASSERT (i < LNET_MAX_INTERFACES);
226 ip = peer->ksnp_passive_ips[i];
228 iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
229 /* All IPs in peer->ksnp_passive_ips[] come from the
230 * interface list, therefore the call must succeed. */
231 LASSERT (iface != NULL);
233 CDEBUG(D_NET, "peer=%p iface=%p ksni_nroutes=%d\n",
234 peer, iface, iface->ksni_nroutes);
235 iface->ksni_npeers--;
238 LASSERT (list_empty(&peer->ksnp_conns));
239 LASSERT (list_empty(&peer->ksnp_routes));
240 LASSERT (!peer->ksnp_closing);
241 peer->ksnp_closing = 1;
242 list_del (&peer->ksnp_list);
243 /* lose peerlist's ref */
244 ksocknal_peer_decref(peer);
248 ksocknal_get_peer_info (lnet_ni_t *ni, int index,
249 lnet_process_id_t *id, __u32 *myip, __u32 *peer_ip,
250 int *port, int *conn_count, int *share_count)
253 struct list_head *ptmp;
254 ksock_route_t *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++) {
264 list_for_each (ptmp, &ksocknal_data.ksnd_peers[i]) {
265 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
267 if (peer->ksnp_ni != ni)
270 if (peer->ksnp_n_passive_ips == 0 &&
271 list_empty(&peer->ksnp_routes)) {
285 for (j = 0; j < peer->ksnp_n_passive_ips; j++) {
290 *myip = peer->ksnp_passive_ips[j];
299 list_for_each (rtmp, &peer->ksnp_routes) {
303 route = list_entry(rtmp, ksock_route_t,
307 *myip = route->ksnr_myipaddr;
308 *peer_ip = route->ksnr_ipaddr;
309 *port = route->ksnr_port;
310 *conn_count = route->ksnr_conn_count;
311 *share_count = route->ksnr_share_count;
318 read_unlock(&ksocknal_data.ksnd_global_lock);
323 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
325 ksock_peer_t *peer = route->ksnr_peer;
326 int type = conn->ksnc_type;
327 ksock_interface_t *iface;
329 conn->ksnc_route = route;
330 ksocknal_route_addref(route);
332 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
333 if (route->ksnr_myipaddr == 0) {
334 /* route wasn't bound locally yet (the initial route) */
335 CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
336 libcfs_id2str(peer->ksnp_id),
338 &conn->ksnc_myipaddr);
340 CDEBUG(D_NET, "Rebinding %s %pI4h from "
342 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)
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 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
583 hi = ksocknal_data.ksnd_peer_hash_size - 1;
586 for (i = lo; i <= hi; i++) {
587 list_for_each_safe (ptmp, pnxt,
588 &ksocknal_data.ksnd_peers[i]) {
589 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
591 if (peer->ksnp_ni != ni)
594 if (!((id.nid == LNET_NID_ANY || peer->ksnp_id.nid == id.nid) &&
595 (id.pid == LNET_PID_ANY || peer->ksnp_id.pid == id.pid)))
598 ksocknal_peer_addref(peer); /* a ref for me... */
600 ksocknal_del_peer_locked (peer, ip);
602 if (peer->ksnp_closing &&
603 !list_empty(&peer->ksnp_tx_queue)) {
604 LASSERT (list_empty(&peer->ksnp_conns));
605 LASSERT (list_empty(&peer->ksnp_routes));
607 list_splice_init(&peer->ksnp_tx_queue,
611 ksocknal_peer_decref(peer); /* ...till here */
613 rc = 0; /* matched! */
617 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
619 ksocknal_txlist_done(ni, &zombies, 1);
624 static ksock_conn_t *
625 ksocknal_get_conn_by_idx (lnet_ni_t *ni, int index)
628 struct list_head *ptmp;
630 struct list_head *ctmp;
633 read_lock(&ksocknal_data.ksnd_global_lock);
635 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
636 list_for_each (ptmp, &ksocknal_data.ksnd_peers[i]) {
637 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
639 LASSERT (!peer->ksnp_closing);
641 if (peer->ksnp_ni != ni)
644 list_for_each (ctmp, &peer->ksnp_conns) {
648 conn = list_entry (ctmp, ksock_conn_t,
650 ksocknal_conn_addref(conn);
651 read_unlock(&ksocknal_data.ksnd_global_lock);
657 read_unlock(&ksocknal_data.ksnd_global_lock);
661 static ksock_sched_t *
662 ksocknal_choose_scheduler_locked(unsigned int cpt)
664 struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
665 ksock_sched_t *sched;
668 LASSERT(info->ksi_nthreads > 0);
670 sched = &info->ksi_scheds[0];
672 * NB: it's safe so far, but info->ksi_nthreads could be changed
673 * at runtime when we have dynamic LNet configuration, then we
674 * need to take care of this.
676 for (i = 1; i < info->ksi_nthreads; i++) {
677 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
678 sched = &info->ksi_scheds[i];
685 ksocknal_local_ipvec (lnet_ni_t *ni, __u32 *ipaddrs)
687 ksock_net_t *net = ni->ni_data;
691 read_lock(&ksocknal_data.ksnd_global_lock);
693 nip = net->ksnn_ninterfaces;
694 LASSERT (nip <= LNET_MAX_INTERFACES);
696 /* Only offer interfaces for additional connections if I have
699 read_unlock(&ksocknal_data.ksnd_global_lock);
703 for (i = 0; i < nip; i++) {
704 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
705 LASSERT (ipaddrs[i] != 0);
708 read_unlock(&ksocknal_data.ksnd_global_lock);
713 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
715 int best_netmatch = 0;
722 for (i = 0; i < nips; i++) {
726 this_xor = (ips[i] ^ iface->ksni_ipaddr);
727 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
730 best_netmatch < this_netmatch ||
731 (best_netmatch == this_netmatch &&
732 best_xor > this_xor)))
736 best_netmatch = this_netmatch;
745 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
747 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
748 ksock_net_t *net = peer->ksnp_ni->ni_data;
749 ksock_interface_t *iface;
750 ksock_interface_t *best_iface;
761 /* CAVEAT EMPTOR: We do all our interface matching with an
762 * exclusive hold of global lock at IRQ priority. We're only
763 * expecting to be dealing with small numbers of interfaces, so the
764 * O(n**3)-ness shouldn't matter */
766 /* Also note that I'm not going to return more than n_peerips
767 * interfaces, even if I have more myself */
769 write_lock_bh(global_lock);
771 LASSERT (n_peerips <= LNET_MAX_INTERFACES);
772 LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
774 /* Only match interfaces for additional connections
775 * if I have > 1 interface */
776 n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
777 MIN(n_peerips, net->ksnn_ninterfaces);
779 for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
780 /* ^ yes really... */
782 /* If we have any new interfaces, first tick off all the
783 * peer IPs that match old interfaces, then choose new
784 * interfaces to match the remaining peer IPS.
785 * We don't forget interfaces we've stopped using; we might
786 * start using them again... */
788 if (i < peer->ksnp_n_passive_ips) {
790 ip = peer->ksnp_passive_ips[i];
791 best_iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
794 /* choose a new interface */
795 LASSERT (i == peer->ksnp_n_passive_ips);
801 for (j = 0; j < net->ksnn_ninterfaces; j++) {
802 iface = &net->ksnn_interfaces[j];
803 ip = iface->ksni_ipaddr;
805 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
806 if (peer->ksnp_passive_ips[k] == ip)
809 if (k < peer->ksnp_n_passive_ips) /* using it already */
812 k = ksocknal_match_peerip(iface, peerips, n_peerips);
813 xor = (ip ^ peerips[k]);
814 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
816 if (!(best_iface == NULL ||
817 best_netmatch < this_netmatch ||
818 (best_netmatch == this_netmatch &&
819 best_npeers > iface->ksni_npeers)))
823 best_netmatch = this_netmatch;
824 best_npeers = iface->ksni_npeers;
827 best_iface->ksni_npeers++;
828 ip = best_iface->ksni_ipaddr;
829 peer->ksnp_passive_ips[i] = ip;
830 peer->ksnp_n_passive_ips = i+1;
833 /* mark the best matching peer IP used */
834 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
838 /* Overwrite input peer IP addresses */
839 memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
841 write_unlock_bh(global_lock);
847 ksocknal_create_routes(ksock_peer_t *peer, int port,
848 __u32 *peer_ipaddrs, int npeer_ipaddrs)
850 ksock_route_t *newroute = NULL;
851 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
852 lnet_ni_t *ni = peer->ksnp_ni;
853 ksock_net_t *net = ni->ni_data;
854 struct list_head *rtmp;
855 ksock_route_t *route;
856 ksock_interface_t *iface;
857 ksock_interface_t *best_iface;
864 /* CAVEAT EMPTOR: We do all our interface matching with an
865 * exclusive hold of global lock at IRQ priority. We're only
866 * expecting to be dealing with small numbers of interfaces, so the
867 * O(n**3)-ness here shouldn't matter */
869 write_lock_bh(global_lock);
871 if (net->ksnn_ninterfaces < 2) {
872 /* Only create additional connections
873 * if I have > 1 interface */
874 write_unlock_bh(global_lock);
878 LASSERT (npeer_ipaddrs <= LNET_MAX_INTERFACES);
880 for (i = 0; i < npeer_ipaddrs; i++) {
881 if (newroute != NULL) {
882 newroute->ksnr_ipaddr = peer_ipaddrs[i];
884 write_unlock_bh(global_lock);
886 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
887 if (newroute == NULL)
890 write_lock_bh(global_lock);
893 if (peer->ksnp_closing) {
894 /* peer got closed under me */
898 /* Already got a route? */
900 list_for_each(rtmp, &peer->ksnp_routes) {
901 route = list_entry(rtmp, ksock_route_t, ksnr_list);
903 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
915 LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
917 /* Select interface to connect from */
918 for (j = 0; j < net->ksnn_ninterfaces; j++) {
919 iface = &net->ksnn_interfaces[j];
921 /* Using this interface already? */
922 list_for_each(rtmp, &peer->ksnp_routes) {
923 route = list_entry(rtmp, ksock_route_t,
926 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
934 this_netmatch = (((iface->ksni_ipaddr ^
935 newroute->ksnr_ipaddr) &
936 iface->ksni_netmask) == 0) ? 1 : 0;
938 if (!(best_iface == NULL ||
939 best_netmatch < this_netmatch ||
940 (best_netmatch == this_netmatch &&
941 best_nroutes > iface->ksni_nroutes)))
945 best_netmatch = this_netmatch;
946 best_nroutes = iface->ksni_nroutes;
949 if (best_iface == NULL)
952 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
953 best_iface->ksni_nroutes++;
955 ksocknal_add_route_locked(peer, newroute);
959 write_unlock_bh(global_lock);
960 if (newroute != NULL)
961 ksocknal_route_decref(newroute);
965 ksocknal_accept (lnet_ni_t *ni, struct socket *sock)
972 rc = libcfs_sock_getaddr(sock, 1, &peer_ip, &peer_port);
973 LASSERT (rc == 0); /* we succeeded before */
975 LIBCFS_ALLOC(cr, sizeof(*cr));
977 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
978 "%pI4h: memory exhausted\n",
985 cr->ksncr_sock = sock;
987 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
989 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
990 wake_up(&ksocknal_data.ksnd_connd_waitq);
992 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
997 ksocknal_connecting (ksock_peer_t *peer, __u32 ipaddr)
999 ksock_route_t *route;
1001 list_for_each_entry (route, &peer->ksnp_routes, ksnr_list) {
1003 if (route->ksnr_ipaddr == ipaddr)
1004 return route->ksnr_connecting;
1010 ksocknal_create_conn (lnet_ni_t *ni, ksock_route_t *route,
1011 struct socket *sock, int type)
1013 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1014 LIST_HEAD (zombies);
1015 lnet_process_id_t peerid;
1016 struct list_head *tmp;
1019 ksock_conn_t *conn2;
1020 ksock_peer_t *peer = NULL;
1021 ksock_peer_t *peer2;
1022 ksock_sched_t *sched;
1023 ksock_hello_msg_t *hello;
1031 active = (route != NULL);
1033 LASSERT (active == (type != SOCKLND_CONN_NONE));
1035 LIBCFS_ALLOC(conn, sizeof(*conn));
1041 conn->ksnc_peer = NULL;
1042 conn->ksnc_route = NULL;
1043 conn->ksnc_sock = sock;
1044 /* 2 ref, 1 for conn, another extra ref prevents socket
1045 * being closed before establishment of connection */
1046 atomic_set (&conn->ksnc_sock_refcount, 2);
1047 conn->ksnc_type = type;
1048 ksocknal_lib_save_callback(sock, conn);
1049 atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1051 conn->ksnc_rx_ready = 0;
1052 conn->ksnc_rx_scheduled = 0;
1054 INIT_LIST_HEAD (&conn->ksnc_tx_queue);
1055 conn->ksnc_tx_ready = 0;
1056 conn->ksnc_tx_scheduled = 0;
1057 conn->ksnc_tx_carrier = NULL;
1058 atomic_set (&conn->ksnc_tx_nob, 0);
1060 LIBCFS_ALLOC(hello, offsetof(ksock_hello_msg_t,
1061 kshm_ips[LNET_MAX_INTERFACES]));
1062 if (hello == NULL) {
1067 /* stash conn's local and remote addrs */
1068 rc = ksocknal_lib_get_conn_addrs (conn);
1072 /* Find out/confirm peer's NID and connection type and get the
1073 * vector of interfaces she's willing to let me connect to.
1074 * Passive connections use the listener timeout since the peer sends
1078 peer = route->ksnr_peer;
1079 LASSERT(ni == peer->ksnp_ni);
1081 /* Active connection sends HELLO eagerly */
1082 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1083 peerid = peer->ksnp_id;
1085 write_lock_bh(global_lock);
1086 conn->ksnc_proto = peer->ksnp_proto;
1087 write_unlock_bh(global_lock);
1089 if (conn->ksnc_proto == NULL) {
1090 conn->ksnc_proto = &ksocknal_protocol_v3x;
1091 #if SOCKNAL_VERSION_DEBUG
1092 if (*ksocknal_tunables.ksnd_protocol == 2)
1093 conn->ksnc_proto = &ksocknal_protocol_v2x;
1094 else if (*ksocknal_tunables.ksnd_protocol == 1)
1095 conn->ksnc_proto = &ksocknal_protocol_v1x;
1099 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1103 peerid.nid = LNET_NID_ANY;
1104 peerid.pid = LNET_PID_ANY;
1106 /* Passive, get protocol from peer */
1107 conn->ksnc_proto = NULL;
1110 rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1114 LASSERT (rc == 0 || active);
1115 LASSERT (conn->ksnc_proto != NULL);
1116 LASSERT (peerid.nid != LNET_NID_ANY);
1118 cpt = lnet_cpt_of_nid(peerid.nid);
1121 ksocknal_peer_addref(peer);
1122 write_lock_bh(global_lock);
1124 rc = ksocknal_create_peer(&peer, ni, peerid);
1128 write_lock_bh(global_lock);
1130 /* called with a ref on ni, so shutdown can't have started */
1131 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1133 peer2 = ksocknal_find_peer_locked(ni, peerid);
1134 if (peer2 == NULL) {
1135 /* NB this puts an "empty" peer in the peer
1136 * table (which takes my ref) */
1137 list_add_tail(&peer->ksnp_list,
1138 ksocknal_nid2peerlist(peerid.nid));
1140 ksocknal_peer_decref(peer);
1145 ksocknal_peer_addref(peer);
1146 peer->ksnp_accepting++;
1148 /* Am I already connecting to this guy? Resolve in
1149 * favour of higher NID... */
1150 if (peerid.nid < ni->ni_nid &&
1151 ksocknal_connecting(peer, conn->ksnc_ipaddr)) {
1153 warn = "connection race resolution";
1158 if (peer->ksnp_closing ||
1159 (active && route->ksnr_deleted)) {
1160 /* peer/route got closed under me */
1162 warn = "peer/route removed";
1166 if (peer->ksnp_proto == NULL) {
1167 /* Never connected before.
1168 * NB recv_hello may have returned EPROTO to signal my peer
1169 * wants a different protocol than the one I asked for.
1171 LASSERT (list_empty(&peer->ksnp_conns));
1173 peer->ksnp_proto = conn->ksnc_proto;
1174 peer->ksnp_incarnation = incarnation;
1177 if (peer->ksnp_proto != conn->ksnc_proto ||
1178 peer->ksnp_incarnation != incarnation) {
1179 /* Peer rebooted or I've got the wrong protocol version */
1180 ksocknal_close_peer_conns_locked(peer, 0, 0);
1182 peer->ksnp_proto = NULL;
1184 warn = peer->ksnp_incarnation != incarnation ?
1186 "wrong proto version";
1196 warn = "lost conn race";
1199 warn = "retry with different protocol version";
1203 /* Refuse to duplicate an existing connection, unless this is a
1204 * loopback connection */
1205 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1206 list_for_each(tmp, &peer->ksnp_conns) {
1207 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1209 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1210 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1211 conn2->ksnc_type != conn->ksnc_type)
1214 /* Reply on a passive connection attempt so the peer
1215 * realises we're connected. */
1225 /* If the connection created by this route didn't bind to the IP
1226 * address the route connected to, the connection/route matching
1227 * code below probably isn't going to work. */
1229 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1230 CERROR("Route %s %pI4h connected to %pI4h\n",
1231 libcfs_id2str(peer->ksnp_id),
1232 &route->ksnr_ipaddr,
1233 &conn->ksnc_ipaddr);
1236 /* Search for a route corresponding to the new connection and
1237 * create an association. This allows incoming connections created
1238 * by routes in my peer to match my own route entries so I don't
1239 * continually create duplicate routes. */
1240 list_for_each (tmp, &peer->ksnp_routes) {
1241 route = list_entry(tmp, ksock_route_t, ksnr_list);
1243 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1246 ksocknal_associate_route_conn_locked(route, conn);
1250 conn->ksnc_peer = peer; /* conn takes my ref on peer */
1251 peer->ksnp_last_alive = cfs_time_current();
1252 peer->ksnp_send_keepalive = 0;
1253 peer->ksnp_error = 0;
1255 sched = ksocknal_choose_scheduler_locked(cpt);
1256 sched->kss_nconns++;
1257 conn->ksnc_scheduler = sched;
1259 conn->ksnc_tx_last_post = cfs_time_current();
1260 /* Set the deadline for the outgoing HELLO to drain */
1261 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1262 conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1263 mb(); /* order with adding to peer's conn list */
1265 list_add (&conn->ksnc_list, &peer->ksnp_conns);
1266 ksocknal_conn_addref(conn);
1268 ksocknal_new_packet(conn, 0);
1270 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1272 /* Take packets blocking for this connection. */
1273 list_for_each_entry_safe(tx, txtmp, &peer->ksnp_tx_queue, tx_list) {
1274 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) == SOCKNAL_MATCH_NO)
1277 list_del (&tx->tx_list);
1278 ksocknal_queue_tx_locked (tx, conn);
1281 write_unlock_bh(global_lock);
1283 /* We've now got a new connection. Any errors from here on are just
1284 * like "normal" comms errors and we close the connection normally.
1285 * NB (a) we still have to send the reply HELLO for passive
1287 * (b) normal I/O on the conn is blocked until I setup and call the
1291 CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1292 " incarnation:%lld sched[%d:%d]\n",
1293 libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1294 &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1295 conn->ksnc_port, incarnation, cpt,
1296 (int)(sched - &sched->kss_info->ksi_scheds[0]));
1299 /* additional routes after interface exchange? */
1300 ksocknal_create_routes(peer, conn->ksnc_port,
1301 hello->kshm_ips, hello->kshm_nips);
1303 hello->kshm_nips = ksocknal_select_ips(peer, hello->kshm_ips,
1305 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1308 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1309 kshm_ips[LNET_MAX_INTERFACES]));
1311 /* setup the socket AFTER I've received hello (it disables
1312 * SO_LINGER). I might call back to the acceptor who may want
1313 * to send a protocol version response and then close the
1314 * socket; this ensures the socket only tears down after the
1315 * response has been sent. */
1317 rc = ksocknal_lib_setup_sock(sock);
1319 write_lock_bh(global_lock);
1321 /* NB my callbacks block while I hold ksnd_global_lock */
1322 ksocknal_lib_set_callback(sock, conn);
1325 peer->ksnp_accepting--;
1327 write_unlock_bh(global_lock);
1330 write_lock_bh(global_lock);
1331 if (!conn->ksnc_closing) {
1332 /* could be closed by another thread */
1333 ksocknal_close_conn_locked(conn, rc);
1335 write_unlock_bh(global_lock);
1336 } else if (ksocknal_connsock_addref(conn) == 0) {
1337 /* Allow I/O to proceed. */
1338 ksocknal_read_callback(conn);
1339 ksocknal_write_callback(conn);
1340 ksocknal_connsock_decref(conn);
1343 ksocknal_connsock_decref(conn);
1344 ksocknal_conn_decref(conn);
1348 if (!peer->ksnp_closing &&
1349 list_empty (&peer->ksnp_conns) &&
1350 list_empty (&peer->ksnp_routes)) {
1351 list_add(&zombies, &peer->ksnp_tx_queue);
1352 list_del_init(&peer->ksnp_tx_queue);
1353 ksocknal_unlink_peer_locked(peer);
1356 write_unlock_bh(global_lock);
1360 CERROR("Not creating conn %s type %d: %s\n",
1361 libcfs_id2str(peerid), conn->ksnc_type, warn);
1363 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1364 libcfs_id2str(peerid), conn->ksnc_type, warn);
1369 /* Request retry by replying with CONN_NONE
1370 * ksnc_proto has been set already */
1371 conn->ksnc_type = SOCKLND_CONN_NONE;
1372 hello->kshm_nips = 0;
1373 ksocknal_send_hello(ni, conn, peerid.nid, hello);
1376 write_lock_bh(global_lock);
1377 peer->ksnp_accepting--;
1378 write_unlock_bh(global_lock);
1381 ksocknal_txlist_done(ni, &zombies, 1);
1382 ksocknal_peer_decref(peer);
1386 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1387 kshm_ips[LNET_MAX_INTERFACES]));
1389 LIBCFS_FREE (conn, sizeof(*conn));
1392 libcfs_sock_release(sock);
1397 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1399 /* This just does the immmediate housekeeping, and queues the
1400 * connection for the reaper to terminate.
1401 * Caller holds ksnd_global_lock exclusively in irq context */
1402 ksock_peer_t *peer = conn->ksnc_peer;
1403 ksock_route_t *route;
1404 ksock_conn_t *conn2;
1405 struct list_head *tmp;
1407 LASSERT (peer->ksnp_error == 0);
1408 LASSERT (!conn->ksnc_closing);
1409 conn->ksnc_closing = 1;
1411 /* ksnd_deathrow_conns takes over peer's ref */
1412 list_del (&conn->ksnc_list);
1414 route = conn->ksnc_route;
1415 if (route != NULL) {
1416 /* dissociate conn from route... */
1417 LASSERT (!route->ksnr_deleted);
1418 LASSERT ((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1421 list_for_each(tmp, &peer->ksnp_conns) {
1422 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1424 if (conn2->ksnc_route == route &&
1425 conn2->ksnc_type == conn->ksnc_type)
1431 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1433 conn->ksnc_route = NULL;
1435 #if 0 /* irrelevant with only eager routes */
1436 /* make route least favourite */
1437 list_del (&route->ksnr_list);
1438 list_add_tail (&route->ksnr_list, &peer->ksnp_routes);
1440 ksocknal_route_decref(route); /* drop conn's ref on route */
1443 if (list_empty (&peer->ksnp_conns)) {
1444 /* No more connections to this peer */
1446 if (!list_empty(&peer->ksnp_tx_queue)) {
1449 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1451 /* throw them to the last connection...,
1452 * these TXs will be send to /dev/null by scheduler */
1453 list_for_each_entry(tx, &peer->ksnp_tx_queue,
1455 ksocknal_tx_prep(conn, tx);
1457 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1458 list_splice_init(&peer->ksnp_tx_queue,
1459 &conn->ksnc_tx_queue);
1460 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1463 peer->ksnp_proto = NULL; /* renegotiate protocol version */
1464 peer->ksnp_error = error; /* stash last conn close reason */
1466 if (list_empty (&peer->ksnp_routes)) {
1467 /* I've just closed last conn belonging to a
1468 * peer with no routes to it */
1469 ksocknal_unlink_peer_locked (peer);
1473 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1475 list_add_tail(&conn->ksnc_list,
1476 &ksocknal_data.ksnd_deathrow_conns);
1477 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1479 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1483 ksocknal_peer_failed (ksock_peer_t *peer)
1486 unsigned long last_alive = 0;
1488 /* There has been a connection failure or comms error; but I'll only
1489 * tell LNET I think the peer is dead if it's to another kernel and
1490 * there are no connections or connection attempts in existence. */
1492 read_lock(&ksocknal_data.ksnd_global_lock);
1494 if ((peer->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1495 list_empty(&peer->ksnp_conns) &&
1496 peer->ksnp_accepting == 0 &&
1497 ksocknal_find_connecting_route_locked(peer) == NULL) {
1499 last_alive = peer->ksnp_last_alive;
1502 read_unlock(&ksocknal_data.ksnd_global_lock);
1505 lnet_notify (peer->ksnp_ni, peer->ksnp_id.nid, 0,
1510 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1512 ksock_peer_t *peer = conn->ksnc_peer;
1517 /* NB safe to finalize TXs because closing of socket will
1518 * abort all buffered data */
1519 LASSERT (conn->ksnc_sock == NULL);
1521 spin_lock(&peer->ksnp_lock);
1523 list_for_each_entry_safe(tx, tmp, &peer->ksnp_zc_req_list, tx_zc_list) {
1524 if (tx->tx_conn != conn)
1527 LASSERT (tx->tx_msg.ksm_zc_cookies[0] != 0);
1529 tx->tx_msg.ksm_zc_cookies[0] = 0;
1530 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1531 list_del(&tx->tx_zc_list);
1532 list_add(&tx->tx_zc_list, &zlist);
1535 spin_unlock(&peer->ksnp_lock);
1537 while (!list_empty(&zlist)) {
1538 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1540 list_del(&tx->tx_zc_list);
1541 ksocknal_tx_decref(tx);
1546 ksocknal_terminate_conn (ksock_conn_t *conn)
1548 /* This gets called by the reaper (guaranteed thread context) to
1549 * disengage the socket from its callbacks and close it.
1550 * ksnc_refcount will eventually hit zero, and then the reaper will
1552 ksock_peer_t *peer = conn->ksnc_peer;
1553 ksock_sched_t *sched = conn->ksnc_scheduler;
1556 LASSERT(conn->ksnc_closing);
1558 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1559 spin_lock_bh(&sched->kss_lock);
1561 /* a closing conn is always ready to tx */
1562 conn->ksnc_tx_ready = 1;
1564 if (!conn->ksnc_tx_scheduled &&
1565 !list_empty(&conn->ksnc_tx_queue)) {
1566 list_add_tail (&conn->ksnc_tx_list,
1567 &sched->kss_tx_conns);
1568 conn->ksnc_tx_scheduled = 1;
1569 /* extra ref for scheduler */
1570 ksocknal_conn_addref(conn);
1572 wake_up (&sched->kss_waitq);
1575 spin_unlock_bh(&sched->kss_lock);
1577 /* serialise with callbacks */
1578 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1580 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1582 /* OK, so this conn may not be completely disengaged from its
1583 * scheduler yet, but it _has_ committed to terminate... */
1584 conn->ksnc_scheduler->kss_nconns--;
1586 if (peer->ksnp_error != 0) {
1587 /* peer's last conn closed in error */
1588 LASSERT (list_empty (&peer->ksnp_conns));
1590 peer->ksnp_error = 0; /* avoid multiple notifications */
1593 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1596 ksocknal_peer_failed(peer);
1598 /* The socket is closed on the final put; either here, or in
1599 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1600 * when the connection was established, this will close the socket
1601 * immediately, aborting anything buffered in it. Any hung
1602 * zero-copy transmits will therefore complete in finite time. */
1603 ksocknal_connsock_decref(conn);
1607 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1609 /* Queue the conn for the reaper to destroy */
1611 LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1612 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1614 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1615 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1617 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1621 ksocknal_destroy_conn (ksock_conn_t *conn)
1623 unsigned long last_rcv;
1625 /* Final coup-de-grace of the reaper */
1626 CDEBUG (D_NET, "connection %p\n", conn);
1628 LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1629 LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1630 LASSERT (conn->ksnc_sock == NULL);
1631 LASSERT (conn->ksnc_route == NULL);
1632 LASSERT (!conn->ksnc_tx_scheduled);
1633 LASSERT (!conn->ksnc_rx_scheduled);
1634 LASSERT (list_empty(&conn->ksnc_tx_queue));
1636 /* complete current receive if any */
1637 switch (conn->ksnc_rx_state) {
1638 case SOCKNAL_RX_LNET_PAYLOAD:
1639 last_rcv = conn->ksnc_rx_deadline -
1640 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1641 CERROR("Completing partial receive from %s[%d]"
1642 ", ip %pI4h:%d, with error, wanted: %d, left: %d, "
1643 "last alive is %ld secs ago\n",
1644 libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1645 &conn->ksnc_ipaddr, conn->ksnc_port,
1646 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1647 cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1649 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1650 conn->ksnc_cookie, -EIO);
1652 case SOCKNAL_RX_LNET_HEADER:
1653 if (conn->ksnc_rx_started)
1654 CERROR("Incomplete receive of lnet header from %s"
1655 ", ip %pI4h:%d, with error, protocol: %d.x.\n",
1656 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1657 &conn->ksnc_ipaddr, conn->ksnc_port,
1658 conn->ksnc_proto->pro_version);
1660 case SOCKNAL_RX_KSM_HEADER:
1661 if (conn->ksnc_rx_started)
1662 CERROR("Incomplete receive of ksock message from %s"
1663 ", ip %pI4h:%d, with error, protocol: %d.x.\n",
1664 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1665 &conn->ksnc_ipaddr, conn->ksnc_port,
1666 conn->ksnc_proto->pro_version);
1668 case SOCKNAL_RX_SLOP:
1669 if (conn->ksnc_rx_started)
1670 CERROR("Incomplete receive of slops from %s"
1671 ", ip %pI4h:%d, with error\n",
1672 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1673 &conn->ksnc_ipaddr, conn->ksnc_port);
1680 ksocknal_peer_decref(conn->ksnc_peer);
1682 LIBCFS_FREE (conn, sizeof (*conn));
1686 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1689 struct list_head *ctmp;
1690 struct list_head *cnxt;
1693 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
1694 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
1697 conn->ksnc_ipaddr == ipaddr) {
1699 ksocknal_close_conn_locked (conn, why);
1707 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1709 ksock_peer_t *peer = conn->ksnc_peer;
1710 __u32 ipaddr = conn->ksnc_ipaddr;
1713 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1715 count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1717 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1723 ksocknal_close_matching_conns (lnet_process_id_t id, __u32 ipaddr)
1726 struct list_head *ptmp;
1727 struct list_head *pnxt;
1733 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1735 if (id.nid != LNET_NID_ANY)
1736 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1739 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1742 for (i = lo; i <= hi; i++) {
1743 list_for_each_safe (ptmp, pnxt,
1744 &ksocknal_data.ksnd_peers[i]) {
1746 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
1748 if (!((id.nid == LNET_NID_ANY || id.nid == peer->ksnp_id.nid) &&
1749 (id.pid == LNET_PID_ANY || id.pid == peer->ksnp_id.pid)))
1752 count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1756 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1758 /* wildcards always succeed */
1759 if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1769 ksocknal_notify (lnet_ni_t *ni, lnet_nid_t gw_nid, int alive)
1771 /* The router is telling me she's been notified of a change in
1772 * gateway state.... */
1773 lnet_process_id_t id = {0};
1776 id.pid = LNET_PID_ANY;
1778 CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1779 alive ? "up" : "down");
1782 /* If the gateway crashed, close all open connections... */
1783 ksocknal_close_matching_conns (id, 0);
1787 /* ...otherwise do nothing. We can only establish new connections
1788 * if we have autroutes, and these connect on demand. */
1792 ksocknal_query (lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1795 unsigned long last_alive = 0;
1796 unsigned long now = cfs_time_current();
1797 ksock_peer_t *peer = NULL;
1798 rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1799 lnet_process_id_t id = {.nid = nid, .pid = LUSTRE_SRV_LNET_PID};
1803 peer = ksocknal_find_peer_locked(ni, id);
1805 struct list_head *tmp;
1809 list_for_each (tmp, &peer->ksnp_conns) {
1810 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1811 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1813 if (bufnob < conn->ksnc_tx_bufnob) {
1814 /* something got ACKed */
1815 conn->ksnc_tx_deadline =
1816 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1817 peer->ksnp_last_alive = now;
1818 conn->ksnc_tx_bufnob = bufnob;
1822 last_alive = peer->ksnp_last_alive;
1823 if (ksocknal_find_connectable_route_locked(peer) == NULL)
1829 if (last_alive != 0)
1832 CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago, connect %d\n",
1833 libcfs_nid2str(nid), peer,
1834 last_alive ? cfs_duration_sec(now - last_alive) : -1,
1840 ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1842 write_lock_bh(glock);
1844 peer = ksocknal_find_peer_locked(ni, id);
1846 ksocknal_launch_all_connections_locked(peer);
1848 write_unlock_bh(glock);
1853 ksocknal_push_peer (ksock_peer_t *peer)
1857 struct list_head *tmp;
1860 for (index = 0; ; index++) {
1861 read_lock(&ksocknal_data.ksnd_global_lock);
1866 list_for_each (tmp, &peer->ksnp_conns) {
1868 conn = list_entry (tmp, ksock_conn_t,
1870 ksocknal_conn_addref(conn);
1875 read_unlock(&ksocknal_data.ksnd_global_lock);
1880 ksocknal_lib_push_conn (conn);
1881 ksocknal_conn_decref(conn);
1886 ksocknal_push (lnet_ni_t *ni, lnet_process_id_t id)
1889 struct list_head *tmp;
1895 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1896 for (j = 0; ; j++) {
1897 read_lock(&ksocknal_data.ksnd_global_lock);
1902 list_for_each (tmp, &ksocknal_data.ksnd_peers[i]) {
1903 peer = list_entry(tmp, ksock_peer_t,
1906 if (!((id.nid == LNET_NID_ANY ||
1907 id.nid == peer->ksnp_id.nid) &&
1908 (id.pid == LNET_PID_ANY ||
1909 id.pid == peer->ksnp_id.pid))) {
1915 ksocknal_peer_addref(peer);
1920 read_unlock(&ksocknal_data.ksnd_global_lock);
1924 ksocknal_push_peer (peer);
1925 ksocknal_peer_decref(peer);
1935 ksocknal_add_interface(lnet_ni_t *ni, __u32 ipaddress, __u32 netmask)
1937 ksock_net_t *net = ni->ni_data;
1938 ksock_interface_t *iface;
1942 struct list_head *ptmp;
1944 struct list_head *rtmp;
1945 ksock_route_t *route;
1947 if (ipaddress == 0 ||
1951 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1953 iface = ksocknal_ip2iface(ni, ipaddress);
1954 if (iface != NULL) {
1955 /* silently ignore dups */
1957 } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1960 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1962 iface->ksni_ipaddr = ipaddress;
1963 iface->ksni_netmask = netmask;
1964 iface->ksni_nroutes = 0;
1965 iface->ksni_npeers = 0;
1967 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1968 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1969 peer = list_entry(ptmp, ksock_peer_t,
1972 for (j = 0; j < peer->ksnp_n_passive_ips; j++)
1973 if (peer->ksnp_passive_ips[j] == ipaddress)
1974 iface->ksni_npeers++;
1976 list_for_each(rtmp, &peer->ksnp_routes) {
1977 route = list_entry(rtmp,
1981 if (route->ksnr_myipaddr == ipaddress)
1982 iface->ksni_nroutes++;
1988 /* NB only new connections will pay attention to the new interface! */
1991 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1997 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
1999 struct list_head *tmp;
2000 struct list_head *nxt;
2001 ksock_route_t *route;
2006 for (i = 0; i < peer->ksnp_n_passive_ips; i++)
2007 if (peer->ksnp_passive_ips[i] == ipaddr) {
2008 for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
2009 peer->ksnp_passive_ips[j-1] =
2010 peer->ksnp_passive_ips[j];
2011 peer->ksnp_n_passive_ips--;
2015 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
2016 route = list_entry (tmp, ksock_route_t, ksnr_list);
2018 if (route->ksnr_myipaddr != ipaddr)
2021 if (route->ksnr_share_count != 0) {
2022 /* Manually created; keep, but unbind */
2023 route->ksnr_myipaddr = 0;
2025 ksocknal_del_route_locked(route);
2029 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
2030 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2032 if (conn->ksnc_myipaddr == ipaddr)
2033 ksocknal_close_conn_locked (conn, 0);
2038 ksocknal_del_interface(lnet_ni_t *ni, __u32 ipaddress)
2040 ksock_net_t *net = ni->ni_data;
2042 struct list_head *tmp;
2043 struct list_head *nxt;
2049 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2051 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2052 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2054 if (!(ipaddress == 0 ||
2055 ipaddress == this_ip))
2060 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2061 net->ksnn_interfaces[j-1] =
2062 net->ksnn_interfaces[j];
2064 net->ksnn_ninterfaces--;
2066 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2067 list_for_each_safe(tmp, nxt,
2068 &ksocknal_data.ksnd_peers[j]) {
2069 peer = list_entry(tmp, ksock_peer_t,
2072 if (peer->ksnp_ni != ni)
2075 ksocknal_peer_del_interface_locked(peer, this_ip);
2080 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2086 ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
2088 lnet_process_id_t id = {0};
2089 struct libcfs_ioctl_data *data = arg;
2093 case IOC_LIBCFS_GET_INTERFACE: {
2094 ksock_net_t *net = ni->ni_data;
2095 ksock_interface_t *iface;
2097 read_lock(&ksocknal_data.ksnd_global_lock);
2099 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2103 iface = &net->ksnn_interfaces[data->ioc_count];
2105 data->ioc_u32[0] = iface->ksni_ipaddr;
2106 data->ioc_u32[1] = iface->ksni_netmask;
2107 data->ioc_u32[2] = iface->ksni_npeers;
2108 data->ioc_u32[3] = iface->ksni_nroutes;
2111 read_unlock(&ksocknal_data.ksnd_global_lock);
2115 case IOC_LIBCFS_ADD_INTERFACE:
2116 return ksocknal_add_interface(ni,
2117 data->ioc_u32[0], /* IP address */
2118 data->ioc_u32[1]); /* net mask */
2120 case IOC_LIBCFS_DEL_INTERFACE:
2121 return ksocknal_del_interface(ni,
2122 data->ioc_u32[0]); /* IP address */
2124 case IOC_LIBCFS_GET_PEER: {
2129 int share_count = 0;
2131 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2132 &id, &myip, &ip, &port,
2133 &conn_count, &share_count);
2137 data->ioc_nid = id.nid;
2138 data->ioc_count = share_count;
2139 data->ioc_u32[0] = ip;
2140 data->ioc_u32[1] = port;
2141 data->ioc_u32[2] = myip;
2142 data->ioc_u32[3] = conn_count;
2143 data->ioc_u32[4] = id.pid;
2147 case IOC_LIBCFS_ADD_PEER:
2148 id.nid = data->ioc_nid;
2149 id.pid = LUSTRE_SRV_LNET_PID;
2150 return ksocknal_add_peer (ni, id,
2151 data->ioc_u32[0], /* IP */
2152 data->ioc_u32[1]); /* port */
2154 case IOC_LIBCFS_DEL_PEER:
2155 id.nid = data->ioc_nid;
2156 id.pid = LNET_PID_ANY;
2157 return ksocknal_del_peer (ni, id,
2158 data->ioc_u32[0]); /* IP */
2160 case IOC_LIBCFS_GET_CONN: {
2164 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2169 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2171 data->ioc_count = txmem;
2172 data->ioc_nid = conn->ksnc_peer->ksnp_id.nid;
2173 data->ioc_flags = nagle;
2174 data->ioc_u32[0] = conn->ksnc_ipaddr;
2175 data->ioc_u32[1] = conn->ksnc_port;
2176 data->ioc_u32[2] = conn->ksnc_myipaddr;
2177 data->ioc_u32[3] = conn->ksnc_type;
2178 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2179 data->ioc_u32[5] = rxmem;
2180 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2181 ksocknal_conn_decref(conn);
2185 case IOC_LIBCFS_CLOSE_CONNECTION:
2186 id.nid = data->ioc_nid;
2187 id.pid = LNET_PID_ANY;
2188 return ksocknal_close_matching_conns (id,
2191 case IOC_LIBCFS_REGISTER_MYNID:
2192 /* Ignore if this is a noop */
2193 if (data->ioc_nid == ni->ni_nid)
2196 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2197 libcfs_nid2str(data->ioc_nid),
2198 libcfs_nid2str(ni->ni_nid));
2201 case IOC_LIBCFS_PUSH_CONNECTION:
2202 id.nid = data->ioc_nid;
2203 id.pid = LNET_PID_ANY;
2204 return ksocknal_push(ni, id);
2213 ksocknal_free_buffers (void)
2215 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2217 if (ksocknal_data.ksnd_sched_info != NULL) {
2218 struct ksock_sched_info *info;
2221 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2222 if (info->ksi_scheds != NULL) {
2223 LIBCFS_FREE(info->ksi_scheds,
2224 info->ksi_nthreads_max *
2225 sizeof(info->ksi_scheds[0]));
2228 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2231 LIBCFS_FREE (ksocknal_data.ksnd_peers,
2232 sizeof (struct list_head) *
2233 ksocknal_data.ksnd_peer_hash_size);
2235 spin_lock(&ksocknal_data.ksnd_tx_lock);
2237 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2238 struct list_head zlist;
2241 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2242 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2243 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2245 while (!list_empty(&zlist)) {
2246 tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2247 list_del(&tx->tx_list);
2248 LIBCFS_FREE(tx, tx->tx_desc_size);
2251 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2256 ksocknal_base_shutdown(void)
2258 struct ksock_sched_info *info;
2259 ksock_sched_t *sched;
2263 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2264 atomic_read (&libcfs_kmemory));
2265 LASSERT (ksocknal_data.ksnd_nnets == 0);
2267 switch (ksocknal_data.ksnd_init) {
2271 case SOCKNAL_INIT_ALL:
2272 case SOCKNAL_INIT_DATA:
2273 LASSERT (ksocknal_data.ksnd_peers != NULL);
2274 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2275 LASSERT (list_empty (&ksocknal_data.ksnd_peers[i]));
2278 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2279 LASSERT (list_empty (&ksocknal_data.ksnd_enomem_conns));
2280 LASSERT (list_empty (&ksocknal_data.ksnd_zombie_conns));
2281 LASSERT (list_empty (&ksocknal_data.ksnd_connd_connreqs));
2282 LASSERT (list_empty (&ksocknal_data.ksnd_connd_routes));
2284 if (ksocknal_data.ksnd_sched_info != NULL) {
2285 cfs_percpt_for_each(info, i,
2286 ksocknal_data.ksnd_sched_info) {
2287 if (info->ksi_scheds == NULL)
2290 for (j = 0; j < info->ksi_nthreads_max; j++) {
2292 sched = &info->ksi_scheds[j];
2294 &sched->kss_tx_conns));
2296 &sched->kss_rx_conns));
2298 &sched->kss_zombie_noop_txs));
2299 LASSERT(sched->kss_nconns == 0);
2304 /* flag threads to terminate; wake and wait for them to die */
2305 ksocknal_data.ksnd_shuttingdown = 1;
2306 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2307 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2309 if (ksocknal_data.ksnd_sched_info != NULL) {
2310 cfs_percpt_for_each(info, i,
2311 ksocknal_data.ksnd_sched_info) {
2312 if (info->ksi_scheds == NULL)
2315 for (j = 0; j < info->ksi_nthreads_max; j++) {
2316 sched = &info->ksi_scheds[j];
2317 wake_up_all(&sched->kss_waitq);
2323 read_lock(&ksocknal_data.ksnd_global_lock);
2324 while (ksocknal_data.ksnd_nthreads != 0) {
2326 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2327 "waiting for %d threads to terminate\n",
2328 ksocknal_data.ksnd_nthreads);
2329 read_unlock(&ksocknal_data.ksnd_global_lock);
2330 set_current_state(TASK_UNINTERRUPTIBLE);
2331 schedule_timeout(cfs_time_seconds(1));
2332 read_lock(&ksocknal_data.ksnd_global_lock);
2334 read_unlock(&ksocknal_data.ksnd_global_lock);
2336 ksocknal_free_buffers();
2338 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2342 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2343 atomic_read (&libcfs_kmemory));
2345 module_put(THIS_MODULE);
2349 ksocknal_new_incarnation (void)
2353 /* The incarnation number is the time this module loaded and it
2354 * identifies this particular instance of the socknal. Hopefully
2355 * we won't be able to reboot more frequently than 1MHz for the
2356 * foreseeable future :) */
2358 do_gettimeofday(&tv);
2360 return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2364 ksocknal_base_startup(void)
2366 struct ksock_sched_info *info;
2370 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2371 LASSERT (ksocknal_data.ksnd_nnets == 0);
2373 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2375 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2376 LIBCFS_ALLOC (ksocknal_data.ksnd_peers,
2377 sizeof (struct list_head) *
2378 ksocknal_data.ksnd_peer_hash_size);
2379 if (ksocknal_data.ksnd_peers == NULL)
2382 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2383 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2385 rwlock_init(&ksocknal_data.ksnd_global_lock);
2386 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2388 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2389 INIT_LIST_HEAD (&ksocknal_data.ksnd_enomem_conns);
2390 INIT_LIST_HEAD (&ksocknal_data.ksnd_zombie_conns);
2391 INIT_LIST_HEAD (&ksocknal_data.ksnd_deathrow_conns);
2392 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2394 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2395 INIT_LIST_HEAD (&ksocknal_data.ksnd_connd_connreqs);
2396 INIT_LIST_HEAD (&ksocknal_data.ksnd_connd_routes);
2397 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2399 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2400 INIT_LIST_HEAD (&ksocknal_data.ksnd_idle_noop_txs);
2402 /* NB memset above zeros whole of ksocknal_data */
2404 /* flag lists/ptrs/locks initialised */
2405 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2406 try_module_get(THIS_MODULE);
2408 ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2410 if (ksocknal_data.ksnd_sched_info == NULL)
2413 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2414 ksock_sched_t *sched;
2417 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2418 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2419 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2421 /* max to half of CPUs, assume another half should be
2422 * reserved for upper layer modules */
2423 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2426 info->ksi_nthreads_max = nthrs;
2429 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2430 info->ksi_nthreads_max * sizeof(*sched));
2431 if (info->ksi_scheds == NULL)
2434 for (; nthrs > 0; nthrs--) {
2435 sched = &info->ksi_scheds[nthrs - 1];
2437 sched->kss_info = info;
2438 spin_lock_init(&sched->kss_lock);
2439 INIT_LIST_HEAD(&sched->kss_rx_conns);
2440 INIT_LIST_HEAD(&sched->kss_tx_conns);
2441 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2442 init_waitqueue_head(&sched->kss_waitq);
2446 ksocknal_data.ksnd_connd_starting = 0;
2447 ksocknal_data.ksnd_connd_failed_stamp = 0;
2448 ksocknal_data.ksnd_connd_starting_stamp = get_seconds();
2449 /* must have at least 2 connds to remain responsive to accepts while
2451 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2452 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2454 if (*ksocknal_tunables.ksnd_nconnds_max <
2455 *ksocknal_tunables.ksnd_nconnds) {
2456 ksocknal_tunables.ksnd_nconnds_max =
2457 ksocknal_tunables.ksnd_nconnds;
2460 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2462 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2463 ksocknal_data.ksnd_connd_starting++;
2464 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2467 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2468 rc = ksocknal_thread_start(ksocknal_connd,
2469 (void *)((ulong_ptr_t)i), name);
2471 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2472 ksocknal_data.ksnd_connd_starting--;
2473 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2474 CERROR("Can't spawn socknal connd: %d\n", rc);
2479 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2481 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2485 /* flag everything initialised */
2486 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2491 ksocknal_base_shutdown();
2496 ksocknal_debug_peerhash (lnet_ni_t *ni)
2498 ksock_peer_t *peer = NULL;
2499 struct list_head *tmp;
2502 read_lock(&ksocknal_data.ksnd_global_lock);
2504 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2505 list_for_each (tmp, &ksocknal_data.ksnd_peers[i]) {
2506 peer = list_entry (tmp, ksock_peer_t, ksnp_list);
2508 if (peer->ksnp_ni == ni)
2516 ksock_route_t *route;
2519 CWARN ("Active peer on shutdown: %s, ref %d, scnt %d, "
2520 "closing %d, accepting %d, err %d, zcookie %llu, "
2521 "txq %d, zc_req %d\n", libcfs_id2str(peer->ksnp_id),
2522 atomic_read(&peer->ksnp_refcount),
2523 peer->ksnp_sharecount, peer->ksnp_closing,
2524 peer->ksnp_accepting, peer->ksnp_error,
2525 peer->ksnp_zc_next_cookie,
2526 !list_empty(&peer->ksnp_tx_queue),
2527 !list_empty(&peer->ksnp_zc_req_list));
2529 list_for_each (tmp, &peer->ksnp_routes) {
2530 route = list_entry(tmp, ksock_route_t, ksnr_list);
2531 CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2532 "del %d\n", atomic_read(&route->ksnr_refcount),
2533 route->ksnr_scheduled, route->ksnr_connecting,
2534 route->ksnr_connected, route->ksnr_deleted);
2537 list_for_each (tmp, &peer->ksnp_conns) {
2538 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2539 CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2540 atomic_read(&conn->ksnc_conn_refcount),
2541 atomic_read(&conn->ksnc_sock_refcount),
2542 conn->ksnc_type, conn->ksnc_closing);
2546 read_unlock(&ksocknal_data.ksnd_global_lock);
2551 ksocknal_shutdown (lnet_ni_t *ni)
2553 ksock_net_t *net = ni->ni_data;
2555 lnet_process_id_t anyid = {0};
2557 anyid.nid = LNET_NID_ANY;
2558 anyid.pid = LNET_PID_ANY;
2560 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2561 LASSERT(ksocknal_data.ksnd_nnets > 0);
2563 spin_lock_bh(&net->ksnn_lock);
2564 net->ksnn_shutdown = 1; /* prevent new peers */
2565 spin_unlock_bh(&net->ksnn_lock);
2567 /* Delete all peers */
2568 ksocknal_del_peer(ni, anyid, 0);
2570 /* Wait for all peer state to clean up */
2572 spin_lock_bh(&net->ksnn_lock);
2573 while (net->ksnn_npeers != 0) {
2574 spin_unlock_bh(&net->ksnn_lock);
2577 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2578 "waiting for %d peers to disconnect\n",
2580 set_current_state(TASK_UNINTERRUPTIBLE);
2581 schedule_timeout(cfs_time_seconds(1));
2583 ksocknal_debug_peerhash(ni);
2585 spin_lock_bh(&net->ksnn_lock);
2587 spin_unlock_bh(&net->ksnn_lock);
2589 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2590 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2591 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2594 list_del(&net->ksnn_list);
2595 LIBCFS_FREE(net, sizeof(*net));
2597 ksocknal_data.ksnd_nnets--;
2598 if (ksocknal_data.ksnd_nnets == 0)
2599 ksocknal_base_shutdown();
2603 ksocknal_enumerate_interfaces(ksock_net_t *net)
2611 n = libcfs_ipif_enumerate(&names);
2613 CERROR("Can't enumerate interfaces: %d\n", n);
2617 for (i = j = 0; i < n; i++) {
2622 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2625 rc = libcfs_ipif_query(names[i], &up, &ip, &mask);
2627 CWARN("Can't get interface %s info: %d\n",
2633 CWARN("Ignoring interface %s (down)\n",
2638 if (j == LNET_MAX_INTERFACES) {
2639 CWARN("Ignoring interface %s (too many interfaces)\n",
2644 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2645 net->ksnn_interfaces[j].ksni_netmask = mask;
2646 strncpy(&net->ksnn_interfaces[j].ksni_name[0],
2647 names[i], IFNAMSIZ);
2651 libcfs_ipif_free_enumeration(names, n);
2654 CERROR("Can't find any usable interfaces\n");
2660 ksocknal_search_new_ipif(ksock_net_t *net)
2665 for (i = 0; i < net->ksnn_ninterfaces; i++) {
2666 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2667 char *colon = strchr(ifnam, ':');
2672 if (colon != NULL) /* ignore alias device */
2675 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2677 for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2679 &tmp->ksnn_interfaces[j].ksni_name[0];
2680 char *colon2 = strchr(ifnam2, ':');
2685 found = strcmp(ifnam, ifnam2) == 0;
2702 ksocknal_start_schedulers(struct ksock_sched_info *info)
2708 if (info->ksi_nthreads == 0) {
2709 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2710 nthrs = info->ksi_nthreads_max;
2712 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2714 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2715 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2717 nthrs = min(nthrs, info->ksi_nthreads_max);
2719 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2720 /* increase two threads if there is new interface */
2721 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2724 for (i = 0; i < nthrs; i++) {
2727 ksock_sched_t *sched;
2728 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2729 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2730 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2731 info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2733 rc = ksocknal_thread_start(ksocknal_scheduler,
2738 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2739 info->ksi_cpt, info->ksi_nthreads + i, rc);
2743 info->ksi_nthreads += i;
2748 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2750 int newif = ksocknal_search_new_ipif(net);
2754 LASSERT(ncpts > 0 && ncpts <= cfs_cpt_number(lnet_cpt_table()));
2756 for (i = 0; i < ncpts; i++) {
2757 struct ksock_sched_info *info;
2758 int cpt = (cpts == NULL) ? i : cpts[i];
2760 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2761 info = ksocknal_data.ksnd_sched_info[cpt];
2763 if (!newif && info->ksi_nthreads > 0)
2766 rc = ksocknal_start_schedulers(info);
2774 ksocknal_startup (lnet_ni_t *ni)
2780 LASSERT (ni->ni_lnd == &the_ksocklnd);
2782 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2783 rc = ksocknal_base_startup();
2788 LIBCFS_ALLOC(net, sizeof(*net));
2792 spin_lock_init(&net->ksnn_lock);
2793 net->ksnn_incarnation = ksocknal_new_incarnation();
2795 ni->ni_peertimeout = *ksocknal_tunables.ksnd_peertimeout;
2796 ni->ni_maxtxcredits = *ksocknal_tunables.ksnd_credits;
2797 ni->ni_peertxcredits = *ksocknal_tunables.ksnd_peertxcredits;
2798 ni->ni_peerrtrcredits = *ksocknal_tunables.ksnd_peerrtrcredits;
2800 if (ni->ni_interfaces[0] == NULL) {
2801 rc = ksocknal_enumerate_interfaces(net);
2805 net->ksnn_ninterfaces = 1;
2807 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2810 if (ni->ni_interfaces[i] == NULL)
2813 rc = libcfs_ipif_query(
2814 ni->ni_interfaces[i], &up,
2815 &net->ksnn_interfaces[i].ksni_ipaddr,
2816 &net->ksnn_interfaces[i].ksni_netmask);
2819 CERROR("Can't get interface %s info: %d\n",
2820 ni->ni_interfaces[i], rc);
2825 CERROR("Interface %s is down\n",
2826 ni->ni_interfaces[i]);
2830 strncpy(&net->ksnn_interfaces[i].ksni_name[0],
2831 ni->ni_interfaces[i], IFNAMSIZ);
2833 net->ksnn_ninterfaces = i;
2836 /* call it before add it to ksocknal_data.ksnd_nets */
2837 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2841 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2842 net->ksnn_interfaces[0].ksni_ipaddr);
2843 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2845 ksocknal_data.ksnd_nnets++;
2850 LIBCFS_FREE(net, sizeof(*net));
2852 if (ksocknal_data.ksnd_nnets == 0)
2853 ksocknal_base_shutdown();
2860 ksocknal_module_fini (void)
2862 lnet_unregister_lnd(&the_ksocklnd);
2866 ksocknal_module_init (void)
2870 /* check ksnr_connected/connecting field large enough */
2871 CLASSERT (SOCKLND_CONN_NTYPES <= 4);
2872 CLASSERT (SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2874 /* initialize the_ksocklnd */
2875 the_ksocklnd.lnd_type = SOCKLND;
2876 the_ksocklnd.lnd_startup = ksocknal_startup;
2877 the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2878 the_ksocklnd.lnd_ctl = ksocknal_ctl;
2879 the_ksocklnd.lnd_send = ksocknal_send;
2880 the_ksocklnd.lnd_recv = ksocknal_recv;
2881 the_ksocklnd.lnd_notify = ksocknal_notify;
2882 the_ksocklnd.lnd_query = ksocknal_query;
2883 the_ksocklnd.lnd_accept = ksocknal_accept;
2885 rc = ksocknal_tunables_init();
2889 lnet_register_lnd(&the_ksocklnd);
2894 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2895 MODULE_DESCRIPTION("Kernel TCP Socket LND v3.0.0");
2896 MODULE_LICENSE("GPL");
2897 MODULE_VERSION("3.0.0");
2899 module_init(ksocknal_module_init);
2900 module_exit(ksocknal_module_fini);