1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
5 * Author: Zach Brown <zab@zabbo.net>
6 * Author: Peter J. Braam <braam@clusterfs.com>
7 * Author: Phil Schwan <phil@clusterfs.com>
8 * Author: Eric Barton <eric@bartonsoftware.com>
10 * This file is part of Portals, http://www.sf.net/projects/sandiaportals/
12 * Portals is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Portals is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Portals; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 ksock_nal_data_t ksocknal_data;
30 ptl_handle_ni_t ksocknal_ni;
31 ksock_tunables_t ksocknal_tunables;
33 kpr_nal_interface_t ksocknal_router_interface = {
35 kprni_arg: &ksocknal_data,
36 kprni_fwd: ksocknal_fwd_packet,
37 kprni_notify: ksocknal_notify,
41 ksocknal_set_mynid(ptl_nid_t nid)
43 lib_ni_t *ni = &ksocknal_lib.libnal_ni;
45 /* FIXME: we have to do this because we call lib_init() at module
46 * insertion time, which is before we have 'mynid' available. lib_init
47 * sets the NAL's nid, which it uses to tell other nodes where packets
48 * are coming from. This is not a very graceful solution to this
51 CDEBUG(D_IOCTL, "setting mynid to "LPX64" (old nid="LPX64")\n",
59 ksocknal_ip2iface(__u32 ip)
62 ksock_interface_t *iface;
64 for (i = 0; i < ksocknal_data.ksnd_ninterfaces; i++) {
65 LASSERT(i < SOCKNAL_MAX_INTERFACES);
66 iface = &ksocknal_data.ksnd_interfaces[i];
68 if (iface->ksni_ipaddr == ip)
76 ksocknal_create_route (__u32 ipaddr, int port)
80 PORTAL_ALLOC (route, sizeof (*route));
84 atomic_set (&route->ksnr_refcount, 1);
85 route->ksnr_peer = NULL;
86 route->ksnr_timeout = cfs_time_current();
87 route->ksnr_retry_interval = SOCKNAL_MIN_RECONNECT_INTERVAL;
88 route->ksnr_ipaddr = ipaddr;
89 route->ksnr_port = port;
90 route->ksnr_connecting = 0;
91 route->ksnr_connected = 0;
92 route->ksnr_deleted = 0;
93 route->ksnr_conn_count = 0;
94 route->ksnr_share_count = 0;
100 ksocknal_destroy_route (ksock_route_t *route)
102 if (route->ksnr_peer != NULL)
103 ksocknal_put_peer (route->ksnr_peer);
105 PORTAL_FREE (route, sizeof (*route));
109 ksocknal_put_route (ksock_route_t *route)
111 CDEBUG (D_OTHER, "putting route[%p] (%d)\n",
112 route, atomic_read (&route->ksnr_refcount));
114 LASSERT (atomic_read (&route->ksnr_refcount) > 0);
115 if (!atomic_dec_and_test (&route->ksnr_refcount))
118 ksocknal_destroy_route (route);
122 ksocknal_create_peer (ptl_nid_t nid)
126 LASSERT (nid != PTL_NID_ANY);
128 PORTAL_ALLOC (peer, sizeof (*peer));
132 memset (peer, 0, sizeof (*peer)); /* NULL pointers/clear flags etc */
134 peer->ksnp_nid = nid;
135 atomic_set (&peer->ksnp_refcount, 1); /* 1 ref for caller */
136 peer->ksnp_closing = 0;
137 CFS_INIT_LIST_HEAD (&peer->ksnp_conns);
138 CFS_INIT_LIST_HEAD (&peer->ksnp_routes);
139 CFS_INIT_LIST_HEAD (&peer->ksnp_tx_queue);
141 atomic_inc (&ksocknal_data.ksnd_npeers);
146 ksocknal_destroy_peer (ksock_peer_t *peer)
148 CDEBUG (D_NET, "peer "LPX64" %p deleted\n", peer->ksnp_nid, peer);
150 LASSERT (atomic_read (&peer->ksnp_refcount) == 0);
151 LASSERT (list_empty (&peer->ksnp_conns));
152 LASSERT (list_empty (&peer->ksnp_routes));
153 LASSERT (list_empty (&peer->ksnp_tx_queue));
155 PORTAL_FREE (peer, sizeof (*peer));
157 /* NB a peer's connections and autoconnect routes keep a reference
158 * on their peer until they are destroyed, so we can be assured
159 * that _all_ state to do with this peer has been cleaned up when
160 * its refcount drops to zero. */
161 atomic_dec (&ksocknal_data.ksnd_npeers);
165 ksocknal_put_peer (ksock_peer_t *peer)
167 CDEBUG (D_OTHER, "putting peer[%p] -> "LPX64" (%d)\n",
168 peer, peer->ksnp_nid,
169 atomic_read (&peer->ksnp_refcount));
171 LASSERT (atomic_read (&peer->ksnp_refcount) > 0);
172 if (!atomic_dec_and_test (&peer->ksnp_refcount))
175 ksocknal_destroy_peer (peer);
179 ksocknal_find_peer_locked (ptl_nid_t nid)
181 struct list_head *peer_list = ksocknal_nid2peerlist (nid);
182 struct list_head *tmp;
185 list_for_each (tmp, peer_list) {
187 peer = list_entry (tmp, ksock_peer_t, ksnp_list);
189 LASSERT (!peer->ksnp_closing);
191 if (peer->ksnp_nid != nid)
194 CDEBUG(D_NET, "got peer [%p] -> "LPX64" (%d)\n",
195 peer, nid, atomic_read (&peer->ksnp_refcount));
202 ksocknal_get_peer (ptl_nid_t nid)
206 read_lock (&ksocknal_data.ksnd_global_lock);
207 peer = ksocknal_find_peer_locked (nid);
208 if (peer != NULL) /* +1 ref for caller? */
209 atomic_inc (&peer->ksnp_refcount);
210 read_unlock (&ksocknal_data.ksnd_global_lock);
216 ksocknal_unlink_peer_locked (ksock_peer_t *peer)
221 for (i = 0; i < peer->ksnp_n_passive_ips; i++) {
222 LASSERT (i < SOCKNAL_MAX_INTERFACES);
223 ip = peer->ksnp_passive_ips[i];
225 ksocknal_ip2iface(ip)->ksni_npeers--;
228 LASSERT (list_empty(&peer->ksnp_conns));
229 LASSERT (list_empty(&peer->ksnp_routes));
230 LASSERT (!peer->ksnp_closing);
231 peer->ksnp_closing = 1;
232 list_del (&peer->ksnp_list);
233 /* lose peerlist's ref */
234 ksocknal_put_peer (peer);
238 ksocknal_get_peer_info (int index, ptl_nid_t *nid,
239 __u32 *myip, __u32 *peer_ip, int *port,
240 int *conn_count, int *share_count)
243 struct list_head *ptmp;
244 ksock_route_t *route;
245 struct list_head *rtmp;
250 read_lock (&ksocknal_data.ksnd_global_lock);
252 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
254 list_for_each (ptmp, &ksocknal_data.ksnd_peers[i]) {
255 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
257 if (peer->ksnp_n_passive_ips == 0 &&
258 list_empty(&peer->ksnp_routes)) {
262 *nid = peer->ksnp_nid;
272 for (j = 0; j < peer->ksnp_n_passive_ips; j++) {
276 *nid = peer->ksnp_nid;
277 *myip = peer->ksnp_passive_ips[j];
286 list_for_each (rtmp, &peer->ksnp_routes) {
290 route = list_entry(rtmp, ksock_route_t,
293 *nid = peer->ksnp_nid;
294 *myip = route->ksnr_myipaddr;
295 *peer_ip = route->ksnr_ipaddr;
296 *port = route->ksnr_port;
297 *conn_count = route->ksnr_conn_count;
298 *share_count = route->ksnr_share_count;
305 read_unlock (&ksocknal_data.ksnd_global_lock);
310 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
312 ksock_peer_t *peer = route->ksnr_peer;
313 int type = conn->ksnc_type;
314 ksock_interface_t *iface;
316 conn->ksnc_route = route;
317 atomic_inc (&route->ksnr_refcount);
319 if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
320 if (route->ksnr_myipaddr == 0) {
321 /* route wasn't bound locally yet (the initial route) */
322 CWARN("Binding "LPX64" %u.%u.%u.%u to %u.%u.%u.%u\n",
324 HIPQUAD(route->ksnr_ipaddr),
325 HIPQUAD(conn->ksnc_myipaddr));
327 CWARN("Rebinding "LPX64" %u.%u.%u.%u from "
328 "%u.%u.%u.%u to %u.%u.%u.%u\n",
330 HIPQUAD(route->ksnr_ipaddr),
331 HIPQUAD(route->ksnr_myipaddr),
332 HIPQUAD(conn->ksnc_myipaddr));
334 iface = ksocknal_ip2iface(route->ksnr_myipaddr);
336 iface->ksni_nroutes--;
338 route->ksnr_myipaddr = conn->ksnc_myipaddr;
339 iface = ksocknal_ip2iface(route->ksnr_myipaddr);
341 iface->ksni_nroutes++;
344 route->ksnr_connected |= (1<<type);
345 route->ksnr_connecting &= ~(1<<type);
346 route->ksnr_conn_count++;
348 /* Successful connection => further attempts can
349 * proceed immediately */
350 route->ksnr_timeout = cfs_time_current();
351 route->ksnr_retry_interval = SOCKNAL_MIN_RECONNECT_INTERVAL;
355 ksocknal_add_route_locked (ksock_peer_t *peer, ksock_route_t *route)
357 struct list_head *tmp;
360 ksock_route_t *route2;
362 LASSERT (route->ksnr_peer == NULL);
363 LASSERT (route->ksnr_connecting == 0);
364 LASSERT (route->ksnr_connected == 0);
366 /* LASSERT(unique) */
367 list_for_each(tmp, &peer->ksnp_routes) {
368 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
370 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
371 CERROR ("Duplicate route "LPX64" %u.%u.%u.%u\n",
372 peer->ksnp_nid, HIPQUAD(route->ksnr_ipaddr));
377 route->ksnr_peer = peer;
378 atomic_inc (&peer->ksnp_refcount);
379 /* peer's routelist takes over my ref on 'route' */
380 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
382 list_for_each(tmp, &peer->ksnp_conns) {
383 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
384 type = conn->ksnc_type;
386 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
389 ksocknal_associate_route_conn_locked(route, conn);
390 /* keep going (typed routes) */
395 ksocknal_del_route_locked (ksock_route_t *route)
397 ksock_peer_t *peer = route->ksnr_peer;
398 ksock_interface_t *iface;
400 struct list_head *ctmp;
401 struct list_head *cnxt;
403 LASSERT (!route->ksnr_deleted);
405 /* Close associated conns */
406 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
407 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
409 if (conn->ksnc_route != route)
412 ksocknal_close_conn_locked (conn, 0);
415 if (route->ksnr_myipaddr != 0) {
416 iface = ksocknal_ip2iface(route->ksnr_myipaddr);
418 iface->ksni_nroutes--;
421 route->ksnr_deleted = 1;
422 list_del (&route->ksnr_list);
423 ksocknal_put_route (route); /* drop peer's ref */
425 if (list_empty (&peer->ksnp_routes) &&
426 list_empty (&peer->ksnp_conns)) {
427 /* I've just removed the last autoconnect route of a peer
428 * with no active connections */
429 ksocknal_unlink_peer_locked (peer);
434 ksocknal_add_peer (ptl_nid_t nid, __u32 ipaddr, int port)
437 struct list_head *tmp;
440 ksock_route_t *route;
441 ksock_route_t *route2;
443 if (nid == PTL_NID_ANY)
446 /* Have a brand new peer ready... */
447 peer = ksocknal_create_peer (nid);
451 route = ksocknal_create_route (ipaddr, port);
453 ksocknal_put_peer (peer);
457 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
459 peer2 = ksocknal_find_peer_locked (nid);
461 ksocknal_put_peer (peer);
464 /* peer table takes my ref on peer */
465 list_add_tail (&peer->ksnp_list,
466 ksocknal_nid2peerlist (nid));
470 list_for_each (tmp, &peer->ksnp_routes) {
471 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
473 if (route2->ksnr_ipaddr == ipaddr)
478 if (route2 == NULL) {
479 ksocknal_add_route_locked(peer, route);
480 route->ksnr_share_count++;
482 ksocknal_put_route(route);
483 route2->ksnr_share_count++;
486 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
492 ksocknal_del_peer_locked (ksock_peer_t *peer, __u32 ip, int single_share)
495 ksock_route_t *route;
496 struct list_head *tmp;
497 struct list_head *nxt;
500 LASSERT (!peer->ksnp_closing);
502 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
503 route = list_entry(tmp, ksock_route_t, ksnr_list);
505 if (single_share && route->ksnr_share_count == 0)
509 if (!(ip == 0 || route->ksnr_ipaddr == ip))
513 route->ksnr_share_count = 0;
514 else if (route->ksnr_share_count > 0)
515 route->ksnr_share_count--;
517 if (route->ksnr_share_count == 0) {
518 /* This deletes associated conns too */
519 ksocknal_del_route_locked (route);
527 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
528 route = list_entry(tmp, ksock_route_t, ksnr_list);
529 nshared += route->ksnr_share_count;
533 /* remove everything else if there are no explicit entries
536 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
537 route = list_entry(tmp, ksock_route_t, ksnr_list);
539 /* we should only be removing auto-entries */
540 LASSERT(route->ksnr_share_count == 0);
541 ksocknal_del_route_locked (route);
544 list_for_each_safe (tmp, nxt, &peer->ksnp_conns) {
545 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
547 ksocknal_close_conn_locked(conn, 0);
551 /* NB peer unlinks itself when last conn/route is removed */
555 ksocknal_del_peer (ptl_nid_t nid, __u32 ip, int single_share)
558 struct list_head *ptmp;
559 struct list_head *pnxt;
566 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
568 if (nid != PTL_NID_ANY)
569 lo = hi = ksocknal_nid2peerlist(nid) - ksocknal_data.ksnd_peers;
572 hi = ksocknal_data.ksnd_peer_hash_size - 1;
575 for (i = lo; i <= hi; i++) {
576 list_for_each_safe (ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
577 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
579 if (!(nid == PTL_NID_ANY || peer->ksnp_nid == nid))
582 ksocknal_del_peer_locked (peer, ip, single_share);
583 rc = 0; /* matched! */
590 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
596 ksocknal_get_conn_by_idx (int index)
599 struct list_head *ptmp;
601 struct list_head *ctmp;
604 read_lock (&ksocknal_data.ksnd_global_lock);
606 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
607 list_for_each (ptmp, &ksocknal_data.ksnd_peers[i]) {
608 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
610 LASSERT (!peer->ksnp_closing);
612 list_for_each (ctmp, &peer->ksnp_conns) {
616 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
617 atomic_inc (&conn->ksnc_refcount);
618 read_unlock (&ksocknal_data.ksnd_global_lock);
624 read_unlock (&ksocknal_data.ksnd_global_lock);
629 ksocknal_choose_scheduler_locked (unsigned int irq)
631 ksock_sched_t *sched;
632 ksock_irqinfo_t *info;
635 LASSERT (irq < NR_IRQS);
636 info = &ksocknal_data.ksnd_irqinfo[irq];
638 if (irq != 0 && /* hardware NIC */
639 info->ksni_valid) { /* already set up */
640 return (&ksocknal_data.ksnd_schedulers[info->ksni_sched]);
643 /* software NIC (irq == 0) || not associated with a scheduler yet.
644 * Choose the CPU with the fewest connections... */
645 sched = &ksocknal_data.ksnd_schedulers[0];
646 for (i = 1; i < ksocknal_data.ksnd_nschedulers; i++)
647 if (sched->kss_nconns >
648 ksocknal_data.ksnd_schedulers[i].kss_nconns)
649 sched = &ksocknal_data.ksnd_schedulers[i];
651 if (irq != 0) { /* Hardware NIC */
652 info->ksni_valid = 1;
653 info->ksni_sched = sched - ksocknal_data.ksnd_schedulers;
656 LASSERT (info->ksni_sched == sched - ksocknal_data.ksnd_schedulers);
663 ksocknal_local_ipvec (__u32 *ipaddrs)
668 read_lock (&ksocknal_data.ksnd_global_lock);
670 nip = ksocknal_data.ksnd_ninterfaces;
671 for (i = 0; i < nip; i++) {
672 LASSERT (i < SOCKNAL_MAX_INTERFACES);
674 ipaddrs[i] = ksocknal_data.ksnd_interfaces[i].ksni_ipaddr;
675 LASSERT (ipaddrs[i] != 0);
678 read_unlock (&ksocknal_data.ksnd_global_lock);
683 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
685 int best_netmatch = 0;
692 for (i = 0; i < nips; i++) {
696 this_xor = (ips[i] ^ iface->ksni_ipaddr);
697 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
700 best_netmatch < this_netmatch ||
701 (best_netmatch == this_netmatch &&
702 best_xor > this_xor)))
706 best_netmatch = this_netmatch;
715 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
717 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
719 ksock_interface_t *iface;
720 ksock_interface_t *best_iface;
731 /* CAVEAT EMPTOR: We do all our interface matching with an
732 * exclusive hold of global lock at IRQ priority. We're only
733 * expecting to be dealing with small numbers of interfaces, so the
734 * O(n**3)-ness shouldn't matter */
736 /* Also note that I'm not going to return more than n_peerips
737 * interfaces, even if I have more myself */
739 write_lock_irqsave(global_lock, flags);
741 LASSERT (n_peerips <= SOCKNAL_MAX_INTERFACES);
742 LASSERT (ksocknal_data.ksnd_ninterfaces <= SOCKNAL_MAX_INTERFACES);
744 n_ips = MIN(n_peerips, ksocknal_data.ksnd_ninterfaces);
746 for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
747 /* ^ yes really... */
749 /* If we have any new interfaces, first tick off all the
750 * peer IPs that match old interfaces, then choose new
751 * interfaces to match the remaining peer IPS.
752 * We don't forget interfaces we've stopped using; we might
753 * start using them again... */
755 if (i < peer->ksnp_n_passive_ips) {
757 ip = peer->ksnp_passive_ips[i];
758 best_iface = ksocknal_ip2iface(ip);
760 /* peer passive ips are kept up to date */
761 LASSERT(best_iface != NULL);
763 /* choose a new interface */
764 LASSERT (i == peer->ksnp_n_passive_ips);
770 for (j = 0; j < ksocknal_data.ksnd_ninterfaces; j++) {
771 iface = &ksocknal_data.ksnd_interfaces[j];
772 ip = iface->ksni_ipaddr;
774 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
775 if (peer->ksnp_passive_ips[k] == ip)
778 if (k < peer->ksnp_n_passive_ips) /* using it already */
781 k = ksocknal_match_peerip(iface, peerips, n_peerips);
782 xor = (ip ^ peerips[k]);
783 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
785 if (!(best_iface == NULL ||
786 best_netmatch < this_netmatch ||
787 (best_netmatch == this_netmatch &&
788 best_npeers > iface->ksni_npeers)))
792 best_netmatch = this_netmatch;
793 best_npeers = iface->ksni_npeers;
796 best_iface->ksni_npeers++;
797 ip = best_iface->ksni_ipaddr;
798 peer->ksnp_passive_ips[i] = ip;
799 peer->ksnp_n_passive_ips = i+1;
802 LASSERT (best_iface != NULL);
804 /* mark the best matching peer IP used */
805 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
809 /* Overwrite input peer IP addresses */
810 memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
812 write_unlock_irqrestore(global_lock, flags);
818 ksocknal_create_routes(ksock_peer_t *peer, int port,
819 __u32 *peer_ipaddrs, int npeer_ipaddrs)
821 ksock_route_t *newroute = NULL;
822 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
824 struct list_head *rtmp;
825 ksock_route_t *route;
826 ksock_interface_t *iface;
827 ksock_interface_t *best_iface;
834 /* CAVEAT EMPTOR: We do all our interface matching with an
835 * exclusive hold of global lock at IRQ priority. We're only
836 * expecting to be dealing with small numbers of interfaces, so the
837 * O(n**3)-ness here shouldn't matter */
839 write_lock_irqsave(global_lock, flags);
841 LASSERT (npeer_ipaddrs <= SOCKNAL_MAX_INTERFACES);
843 for (i = 0; i < npeer_ipaddrs; i++) {
844 if (newroute != NULL) {
845 newroute->ksnr_ipaddr = peer_ipaddrs[i];
847 write_unlock_irqrestore(global_lock, flags);
849 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
850 if (newroute == NULL)
853 write_lock_irqsave(global_lock, flags);
856 /* Already got a route? */
858 list_for_each(rtmp, &peer->ksnp_routes) {
859 route = list_entry(rtmp, ksock_route_t, ksnr_list);
861 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
873 LASSERT (ksocknal_data.ksnd_ninterfaces <= SOCKNAL_MAX_INTERFACES);
875 /* Select interface to connect from */
876 for (j = 0; j < ksocknal_data.ksnd_ninterfaces; j++) {
877 iface = &ksocknal_data.ksnd_interfaces[j];
879 /* Using this interface already? */
880 list_for_each(rtmp, &peer->ksnp_routes) {
881 route = list_entry(rtmp, ksock_route_t, ksnr_list);
883 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
891 this_netmatch = (((iface->ksni_ipaddr ^
892 newroute->ksnr_ipaddr) &
893 iface->ksni_netmask) == 0) ? 1 : 0;
895 if (!(best_iface == NULL ||
896 best_netmatch < this_netmatch ||
897 (best_netmatch == this_netmatch &&
898 best_nroutes > iface->ksni_nroutes)))
902 best_netmatch = this_netmatch;
903 best_nroutes = iface->ksni_nroutes;
906 if (best_iface == NULL)
909 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
910 best_iface->ksni_nroutes++;
912 ksocknal_add_route_locked(peer, newroute);
916 write_unlock_irqrestore(global_lock, flags);
917 if (newroute != NULL)
918 ksocknal_put_route(newroute);
922 ksocknal_create_conn (ksock_route_t *route, struct socket *sock, int type)
924 int passive = (type == SOCKNAL_CONN_NONE);
925 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
926 __u32 ipaddrs[SOCKNAL_MAX_INTERFACES];
929 struct list_head *tmp;
934 ksock_peer_t *peer = NULL;
936 ksock_sched_t *sched;
941 /* NB, sock has an associated file since (a) this connection might
942 * have been created in userland and (b) we need to refcount the
943 * socket so that we don't close it while I/O is being done on
944 * it, and sock->file has that pre-cooked... */
945 LASSERT (KSN_SOCK2FILE(sock) != NULL);
946 LASSERT (cfs_file_count(KSN_SOCK2FILE(sock)) > 0);
947 LASSERT (route == NULL || !passive);
949 rc = ksocknal_lib_setup_sock (sock);
953 irq = ksocknal_lib_sock_irq (sock);
955 PORTAL_ALLOC(conn, sizeof(*conn));
959 memset (conn, 0, sizeof (*conn));
960 conn->ksnc_peer = NULL;
961 conn->ksnc_route = NULL;
962 conn->ksnc_sock = sock;
963 conn->ksnc_type = type;
964 ksocknal_lib_save_callback(sock, conn);
965 atomic_set (&conn->ksnc_refcount, 1); /* 1 ref for me */
967 conn->ksnc_rx_ready = 0;
968 conn->ksnc_rx_scheduled = 0;
969 ksocknal_new_packet (conn, 0);
971 CFS_INIT_LIST_HEAD (&conn->ksnc_tx_queue);
972 conn->ksnc_tx_ready = 0;
973 conn->ksnc_tx_scheduled = 0;
974 atomic_set (&conn->ksnc_tx_nob, 0);
976 /* stash conn's local and remote addrs */
977 rc = ksocknal_lib_get_conn_addrs (conn);
982 /* Active connection sends HELLO eagerly */
983 rc = ksocknal_local_ipvec(ipaddrs);
988 rc = ksocknal_send_hello (conn, ipaddrs, nipaddrs);
993 /* Find out/confirm peer's NID and connection type and get the
994 * vector of interfaces she's willing to let me connect to */
995 nid = (route == NULL) ? PTL_NID_ANY : route->ksnr_peer->ksnp_nid;
996 rc = ksocknal_recv_hello (conn, &nid, &incarnation, ipaddrs);
1000 LASSERT (nid != PTL_NID_ANY);
1002 if (route != NULL) {
1003 peer = route->ksnr_peer;
1004 atomic_inc(&peer->ksnp_refcount);
1006 peer = ksocknal_create_peer(nid);
1012 write_lock_irqsave(global_lock, flags);
1014 peer2 = ksocknal_find_peer_locked(nid);
1015 if (peer2 == NULL) {
1016 /* NB this puts an "empty" peer in the peer
1017 * table (which takes my ref) */
1018 list_add_tail(&peer->ksnp_list,
1019 ksocknal_nid2peerlist(nid));
1021 ksocknal_put_peer(peer);
1025 atomic_inc(&peer->ksnp_refcount);
1027 write_unlock_irqrestore(global_lock, flags);
1031 ksocknal_create_routes(peer, conn->ksnc_port,
1035 rc = ksocknal_select_ips(peer, ipaddrs, nipaddrs);
1037 rc = ksocknal_send_hello (conn, ipaddrs, rc);
1042 write_lock_irqsave (global_lock, flags);
1044 if (peer->ksnp_closing ||
1045 (route != NULL && route->ksnr_deleted)) {
1046 /* route/peer got closed under me */
1051 /* Refuse to duplicate an existing connection (both sides might
1052 * autoconnect at once), unless this is a loopback connection */
1053 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1054 list_for_each(tmp, &peer->ksnp_conns) {
1055 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1057 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1058 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1059 conn2->ksnc_type != conn->ksnc_type ||
1060 conn2->ksnc_incarnation != incarnation)
1063 CWARN("Not creating duplicate connection to "
1064 "%u.%u.%u.%u type %d\n",
1065 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_type);
1071 /* If the connection created by this route didn't bind to the IP
1072 * address the route connected to, the connection/route matching
1073 * code below probably isn't going to work. */
1074 if (route != NULL &&
1075 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1076 CERROR("Route "LPX64" %u.%u.%u.%u connected to %u.%u.%u.%u\n",
1078 HIPQUAD(route->ksnr_ipaddr),
1079 HIPQUAD(conn->ksnc_ipaddr));
1082 /* Search for a route corresponding to the new connection and
1083 * create an association. This allows incoming connections created
1084 * by routes in my peer to match my own route entries so I don't
1085 * continually create duplicate routes. */
1086 list_for_each (tmp, &peer->ksnp_routes) {
1087 route = list_entry(tmp, ksock_route_t, ksnr_list);
1089 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1092 ksocknal_associate_route_conn_locked(route, conn);
1096 /* Give conn a ref on sock->file since we're going to return success */
1097 cfs_get_file(KSN_SOCK2FILE(sock));
1099 conn->ksnc_peer = peer; /* conn takes my ref on peer */
1100 conn->ksnc_incarnation = incarnation;
1101 peer->ksnp_last_alive = cfs_time_current();
1102 peer->ksnp_error = 0;
1104 sched = ksocknal_choose_scheduler_locked (irq);
1105 sched->kss_nconns++;
1106 conn->ksnc_scheduler = sched;
1108 /* Set the deadline for the outgoing HELLO to drain */
1109 conn->ksnc_tx_bufnob = SOCK_WMEM_QUEUED(sock);
1110 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
1111 mb(); /* order with adding to peer's conn list */
1113 list_add (&conn->ksnc_list, &peer->ksnp_conns);
1114 atomic_inc (&conn->ksnc_refcount);
1116 /* NB my callbacks block while I hold ksnd_global_lock */
1117 ksocknal_lib_set_callback(sock, conn);
1119 /* Take all the packets blocking for a connection.
1120 * NB, it might be nicer to share these blocked packets among any
1121 * other connections that are becoming established. */
1122 while (!list_empty (&peer->ksnp_tx_queue)) {
1123 tx = list_entry (peer->ksnp_tx_queue.next,
1124 ksock_tx_t, tx_list);
1126 list_del (&tx->tx_list);
1127 ksocknal_queue_tx_locked (tx, conn);
1130 rc = ksocknal_close_stale_conns_locked(peer, incarnation);
1132 CERROR ("Closed %d stale conns to nid "LPX64" ip %d.%d.%d.%d\n",
1133 rc, conn->ksnc_peer->ksnp_nid,
1134 HIPQUAD(conn->ksnc_ipaddr));
1136 write_unlock_irqrestore (global_lock, flags);
1138 ksocknal_lib_bind_irq (irq);
1140 /* Call the callbacks right now to get things going. */
1141 if (ksocknal_getconnsock(conn) == 0) {
1142 ksocknal_lib_act_callback(sock, conn);
1143 ksocknal_putconnsock(conn);
1146 CWARN("New conn nid:"LPX64" %u.%u.%u.%u -> %u.%u.%u.%u/%d"
1147 " incarnation:"LPX64" sched[%d]/%d\n",
1148 nid, HIPQUAD(conn->ksnc_myipaddr),
1149 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port, incarnation,
1150 (int)(conn->ksnc_scheduler - ksocknal_data.ksnd_schedulers), irq);
1152 ksocknal_put_conn (conn);
1156 if (!peer->ksnp_closing &&
1157 list_empty (&peer->ksnp_conns) &&
1158 list_empty (&peer->ksnp_routes))
1159 ksocknal_unlink_peer_locked(peer);
1160 write_unlock_irqrestore(global_lock, flags);
1163 ksocknal_put_peer (peer);
1166 PORTAL_FREE (conn, sizeof(*conn));
1173 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1175 /* This just does the immmediate housekeeping, and queues the
1176 * connection for the reaper to terminate.
1177 * Caller holds ksnd_global_lock exclusively in irq context */
1178 ksock_peer_t *peer = conn->ksnc_peer;
1179 ksock_route_t *route;
1180 ksock_conn_t *conn2;
1181 struct list_head *tmp;
1183 LASSERT (peer->ksnp_error == 0);
1184 LASSERT (!conn->ksnc_closing);
1185 conn->ksnc_closing = 1;
1186 atomic_inc (&ksocknal_data.ksnd_nclosing_conns);
1188 /* ksnd_deathrow_conns takes over peer's ref */
1189 list_del (&conn->ksnc_list);
1191 route = conn->ksnc_route;
1192 if (route != NULL) {
1193 /* dissociate conn from route... */
1194 LASSERT (!route->ksnr_deleted);
1195 LASSERT ((route->ksnr_connecting & (1 << conn->ksnc_type)) == 0);
1196 LASSERT ((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1199 list_for_each(tmp, &peer->ksnp_conns) {
1200 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1202 if (conn2->ksnc_route == route &&
1203 conn2->ksnc_type == conn->ksnc_type)
1209 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1211 conn->ksnc_route = NULL;
1213 #if 0 /* irrelevent with only eager routes */
1214 list_del (&route->ksnr_list); /* make route least favourite */
1215 list_add_tail (&route->ksnr_list, &peer->ksnp_routes);
1217 ksocknal_put_route (route); /* drop conn's ref on route */
1220 if (list_empty (&peer->ksnp_conns)) {
1221 /* No more connections to this peer */
1223 peer->ksnp_error = error; /* stash last conn close reason */
1225 if (list_empty (&peer->ksnp_routes)) {
1226 /* I've just closed last conn belonging to a
1227 * non-autoconnecting peer */
1228 ksocknal_unlink_peer_locked (peer);
1232 spin_lock (&ksocknal_data.ksnd_reaper_lock);
1234 list_add_tail (&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1235 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
1237 spin_unlock (&ksocknal_data.ksnd_reaper_lock);
1241 ksocknal_terminate_conn (ksock_conn_t *conn)
1243 /* This gets called by the reaper (guaranteed thread context) to
1244 * disengage the socket from its callbacks and close it.
1245 * ksnc_refcount will eventually hit zero, and then the reaper will
1247 unsigned long flags;
1248 ksock_peer_t *peer = conn->ksnc_peer;
1249 ksock_sched_t *sched = conn->ksnc_scheduler;
1254 LASSERT(conn->ksnc_closing);
1256 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1257 spin_lock_irqsave(&sched->kss_lock, flags);
1259 if (!conn->ksnc_tx_scheduled &&
1260 !list_empty(&conn->ksnc_tx_queue)){
1261 list_add_tail (&conn->ksnc_tx_list,
1262 &sched->kss_tx_conns);
1263 /* a closing conn is always ready to tx */
1264 conn->ksnc_tx_ready = 1;
1265 conn->ksnc_tx_scheduled = 1;
1266 /* extra ref for scheduler */
1267 atomic_inc (&conn->ksnc_refcount);
1269 cfs_waitq_signal (&sched->kss_waitq);
1272 spin_unlock_irqrestore (&sched->kss_lock, flags);
1274 /* serialise with callbacks */
1275 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1277 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1279 /* OK, so this conn may not be completely disengaged from its
1280 * scheduler yet, but it _has_ committed to terminate... */
1281 conn->ksnc_scheduler->kss_nconns--;
1283 if (peer->ksnp_error != 0) {
1284 /* peer's last conn closed in error */
1285 LASSERT (list_empty (&peer->ksnp_conns));
1287 /* convert peer's last-known-alive timestamp from jiffies */
1288 do_gettimeofday (&now);
1289 then = now.tv_sec - cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1290 peer->ksnp_last_alive));
1294 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1296 /* The socket is closed on the final put; either here, or in
1297 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1298 * when the connection was established, this will close the socket
1299 * immediately, aborting anything buffered in it. Any hung
1300 * zero-copy transmits will therefore complete in finite time. */
1301 ksocknal_putconnsock (conn);
1304 kpr_notify (&ksocknal_data.ksnd_router, peer->ksnp_nid,
1309 ksocknal_destroy_conn (ksock_conn_t *conn)
1311 /* Final coup-de-grace of the reaper */
1312 CDEBUG (D_NET, "connection %p\n", conn);
1314 LASSERT (atomic_read (&conn->ksnc_refcount) == 0);
1315 LASSERT (conn->ksnc_route == NULL);
1316 LASSERT (!conn->ksnc_tx_scheduled);
1317 LASSERT (!conn->ksnc_rx_scheduled);
1318 LASSERT (list_empty(&conn->ksnc_tx_queue));
1320 /* complete current receive if any */
1321 switch (conn->ksnc_rx_state) {
1322 case SOCKNAL_RX_BODY:
1323 CERROR("Completing partial receive from "LPX64
1324 ", ip %d.%d.%d.%d:%d, with error\n",
1325 conn->ksnc_peer->ksnp_nid,
1326 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1327 lib_finalize (&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_FAIL);
1329 case SOCKNAL_RX_BODY_FWD:
1330 ksocknal_fmb_callback (conn->ksnc_cookie, -ECONNABORTED);
1332 case SOCKNAL_RX_HEADER:
1333 case SOCKNAL_RX_SLOP:
1340 ksocknal_put_peer (conn->ksnc_peer);
1342 PORTAL_FREE (conn, sizeof (*conn));
1343 atomic_dec (&ksocknal_data.ksnd_nclosing_conns);
1347 ksocknal_put_conn (ksock_conn_t *conn)
1349 unsigned long flags;
1351 CDEBUG (D_OTHER, "putting conn[%p] -> "LPX64" (%d)\n",
1352 conn, conn->ksnc_peer->ksnp_nid,
1353 atomic_read (&conn->ksnc_refcount));
1355 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1356 if (!atomic_dec_and_test (&conn->ksnc_refcount))
1359 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
1361 list_add (&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1362 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
1364 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
1368 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1371 struct list_head *ctmp;
1372 struct list_head *cnxt;
1375 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
1376 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
1379 conn->ksnc_ipaddr == ipaddr) {
1381 ksocknal_close_conn_locked (conn, why);
1389 ksocknal_close_stale_conns_locked (ksock_peer_t *peer, __u64 incarnation)
1392 struct list_head *ctmp;
1393 struct list_head *cnxt;
1396 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
1397 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
1399 if (conn->ksnc_incarnation == incarnation)
1402 CWARN("Closing stale conn nid:"LPX64" ip:%08x/%d "
1403 "incarnation:"LPX64"("LPX64")\n",
1404 peer->ksnp_nid, conn->ksnc_ipaddr, conn->ksnc_port,
1405 conn->ksnc_incarnation, incarnation);
1408 ksocknal_close_conn_locked (conn, -ESTALE);
1415 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1417 ksock_peer_t *peer = conn->ksnc_peer;
1418 __u32 ipaddr = conn->ksnc_ipaddr;
1419 unsigned long flags;
1422 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1424 count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1426 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1432 ksocknal_close_matching_conns (ptl_nid_t nid, __u32 ipaddr)
1434 unsigned long flags;
1436 struct list_head *ptmp;
1437 struct list_head *pnxt;
1443 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1445 if (nid != PTL_NID_ANY)
1446 lo = hi = ksocknal_nid2peerlist(nid) - ksocknal_data.ksnd_peers;
1449 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1452 for (i = lo; i <= hi; i++) {
1453 list_for_each_safe (ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1455 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
1457 if (!(nid == PTL_NID_ANY || nid == peer->ksnp_nid))
1460 count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1464 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1466 /* wildcards always succeed */
1467 if (nid == PTL_NID_ANY || ipaddr == 0)
1470 return (count == 0 ? -ENOENT : 0);
1474 ksocknal_notify (void *arg, ptl_nid_t gw_nid, int alive)
1476 /* The router is telling me she's been notified of a change in
1477 * gateway state.... */
1479 CDEBUG (D_NET, "gw "LPX64" %s\n", gw_nid, alive ? "up" : "down");
1482 /* If the gateway crashed, close all open connections... */
1483 ksocknal_close_matching_conns (gw_nid, 0);
1487 /* ...otherwise do nothing. We can only establish new connections
1488 * if we have autroutes, and these connect on demand. */
1492 ksocknal_push_peer (ksock_peer_t *peer)
1496 struct list_head *tmp;
1499 for (index = 0; ; index++) {
1500 read_lock (&ksocknal_data.ksnd_global_lock);
1505 list_for_each (tmp, &peer->ksnp_conns) {
1507 conn = list_entry (tmp, ksock_conn_t, ksnc_list);
1508 atomic_inc (&conn->ksnc_refcount);
1513 read_unlock (&ksocknal_data.ksnd_global_lock);
1518 ksocknal_lib_push_conn (conn);
1519 ksocknal_put_conn (conn);
1524 ksocknal_push (ptl_nid_t nid)
1527 struct list_head *tmp;
1533 if (nid != PTL_NID_ANY) {
1534 peer = ksocknal_get_peer (nid);
1538 ksocknal_push_peer (peer);
1539 ksocknal_put_peer (peer);
1544 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1545 for (j = 0; ; j++) {
1546 read_lock (&ksocknal_data.ksnd_global_lock);
1551 list_for_each (tmp, &ksocknal_data.ksnd_peers[i]) {
1553 peer = list_entry(tmp, ksock_peer_t,
1555 atomic_inc (&peer->ksnp_refcount);
1560 read_unlock (&ksocknal_data.ksnd_global_lock);
1564 ksocknal_push_peer (peer);
1565 ksocknal_put_peer (peer);
1575 ksocknal_add_interface(__u32 ipaddress, __u32 netmask)
1577 unsigned long flags;
1578 ksock_interface_t *iface;
1582 struct list_head *ptmp;
1584 struct list_head *rtmp;
1585 ksock_route_t *route;
1587 if (ipaddress == 0 ||
1591 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
1593 iface = ksocknal_ip2iface(ipaddress);
1594 if (iface != NULL) {
1595 /* silently ignore dups */
1597 } else if (ksocknal_data.ksnd_ninterfaces == SOCKNAL_MAX_INTERFACES) {
1600 iface = &ksocknal_data.ksnd_interfaces[ksocknal_data.ksnd_ninterfaces++];
1602 iface->ksni_ipaddr = ipaddress;
1603 iface->ksni_netmask = netmask;
1604 iface->ksni_nroutes = 0;
1605 iface->ksni_npeers = 0;
1607 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1608 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1609 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1611 for (j = 0; i < peer->ksnp_n_passive_ips; j++)
1612 if (peer->ksnp_passive_ips[j] == ipaddress)
1613 iface->ksni_npeers++;
1615 list_for_each(rtmp, &peer->ksnp_routes) {
1616 route = list_entry(rtmp, ksock_route_t, ksnr_list);
1618 if (route->ksnr_myipaddr == ipaddress)
1619 iface->ksni_nroutes++;
1625 /* NB only new connections will pay attention to the new interface! */
1628 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
1634 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
1636 struct list_head *tmp;
1637 struct list_head *nxt;
1638 ksock_route_t *route;
1643 for (i = 0; i < peer->ksnp_n_passive_ips; i++)
1644 if (peer->ksnp_passive_ips[i] == ipaddr) {
1645 for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
1646 peer->ksnp_passive_ips[j-1] =
1647 peer->ksnp_passive_ips[j];
1648 peer->ksnp_n_passive_ips--;
1652 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
1653 route = list_entry (tmp, ksock_route_t, ksnr_list);
1655 if (route->ksnr_myipaddr != ipaddr)
1658 if (route->ksnr_share_count != 0) {
1659 /* Manually created; keep, but unbind */
1660 route->ksnr_myipaddr = 0;
1662 ksocknal_del_route_locked(route);
1666 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
1667 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1669 if (conn->ksnc_myipaddr == ipaddr)
1670 ksocknal_close_conn_locked (conn, 0);
1675 ksocknal_del_interface(__u32 ipaddress)
1678 unsigned long flags;
1679 struct list_head *tmp;
1680 struct list_head *nxt;
1686 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
1688 for (i = 0; i < ksocknal_data.ksnd_ninterfaces; i++) {
1689 this_ip = ksocknal_data.ksnd_interfaces[i].ksni_ipaddr;
1691 if (!(ipaddress == 0 ||
1692 ipaddress == this_ip))
1697 for (j = i+1; j < ksocknal_data.ksnd_ninterfaces; j++)
1698 ksocknal_data.ksnd_interfaces[j-1] =
1699 ksocknal_data.ksnd_interfaces[j];
1701 ksocknal_data.ksnd_ninterfaces--;
1703 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
1704 list_for_each_safe(tmp, nxt, &ksocknal_data.ksnd_peers[j]) {
1705 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
1707 ksocknal_peer_del_interface_locked(peer, this_ip);
1712 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
1718 ksocknal_cmd(struct portals_cfg *pcfg, void * private)
1722 switch(pcfg->pcfg_command) {
1723 case NAL_CMD_GET_INTERFACE: {
1724 ksock_interface_t *iface;
1726 read_lock (&ksocknal_data.ksnd_global_lock);
1728 if (pcfg->pcfg_count < 0 ||
1729 pcfg->pcfg_count >= ksocknal_data.ksnd_ninterfaces) {
1733 iface = &ksocknal_data.ksnd_interfaces[pcfg->pcfg_count];
1735 pcfg->pcfg_id = iface->ksni_ipaddr;
1736 pcfg->pcfg_misc = iface->ksni_netmask;
1737 pcfg->pcfg_fd = iface->ksni_npeers;
1738 pcfg->pcfg_count = iface->ksni_nroutes;
1741 read_unlock (&ksocknal_data.ksnd_global_lock);
1744 case NAL_CMD_ADD_INTERFACE: {
1745 rc = ksocknal_add_interface(pcfg->pcfg_id, /* IP address */
1746 pcfg->pcfg_misc); /* net mask */
1749 case NAL_CMD_DEL_INTERFACE: {
1750 rc = ksocknal_del_interface(pcfg->pcfg_id); /* IP address */
1753 case NAL_CMD_GET_PEER: {
1759 int share_count = 0;
1761 rc = ksocknal_get_peer_info(pcfg->pcfg_count, &nid,
1763 &conn_count, &share_count);
1764 pcfg->pcfg_nid = nid;
1765 pcfg->pcfg_size = myip;
1767 pcfg->pcfg_misc = port;
1768 pcfg->pcfg_count = conn_count;
1769 pcfg->pcfg_wait = share_count;
1772 case NAL_CMD_ADD_PEER: {
1773 rc = ksocknal_add_peer (pcfg->pcfg_nid,
1774 pcfg->pcfg_id, /* IP */
1775 pcfg->pcfg_misc); /* port */
1778 case NAL_CMD_DEL_PEER: {
1779 rc = ksocknal_del_peer (pcfg->pcfg_nid,
1780 pcfg->pcfg_id, /* IP */
1781 pcfg->pcfg_flags); /* single_share? */
1784 case NAL_CMD_GET_CONN: {
1785 ksock_conn_t *conn = ksocknal_get_conn_by_idx (pcfg->pcfg_count);
1794 ksocknal_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1797 pcfg->pcfg_nid = conn->ksnc_peer->ksnp_nid;
1798 pcfg->pcfg_id = conn->ksnc_ipaddr;
1799 pcfg->pcfg_misc = conn->ksnc_port;
1800 pcfg->pcfg_fd = conn->ksnc_myipaddr;
1801 pcfg->pcfg_flags = conn->ksnc_type;
1802 pcfg->pcfg_gw_nal = conn->ksnc_scheduler -
1803 ksocknal_data.ksnd_schedulers;
1804 pcfg->pcfg_count = txmem;
1805 pcfg->pcfg_size = rxmem;
1806 pcfg->pcfg_wait = nagle;
1807 ksocknal_put_conn (conn);
1811 case NAL_CMD_REGISTER_PEER_FD: {
1812 struct socket *sock = sockfd_lookup (pcfg->pcfg_fd, &rc);
1813 int type = pcfg->pcfg_misc;
1819 case SOCKNAL_CONN_NONE:
1820 case SOCKNAL_CONN_ANY:
1821 case SOCKNAL_CONN_CONTROL:
1822 case SOCKNAL_CONN_BULK_IN:
1823 case SOCKNAL_CONN_BULK_OUT:
1824 rc = ksocknal_create_conn(NULL, sock, type);
1830 cfs_put_file (KSN_SOCK2FILE(sock));
1833 case NAL_CMD_CLOSE_CONNECTION: {
1834 rc = ksocknal_close_matching_conns (pcfg->pcfg_nid,
1838 case NAL_CMD_REGISTER_MYNID: {
1839 rc = ksocknal_set_mynid (pcfg->pcfg_nid);
1842 case NAL_CMD_PUSH_CONNECTION: {
1843 rc = ksocknal_push (pcfg->pcfg_nid);
1855 ksocknal_free_fmbs (ksock_fmb_pool_t *p)
1857 int npages = p->fmp_buff_pages;
1861 LASSERT (list_empty(&p->fmp_blocked_conns));
1862 LASSERT (p->fmp_nactive_fmbs == 0);
1864 while (!list_empty(&p->fmp_idle_fmbs)) {
1866 fmb = list_entry(p->fmp_idle_fmbs.next,
1867 ksock_fmb_t, fmb_list);
1869 for (i = 0; i < npages; i++)
1870 if (fmb->fmb_kiov[i].kiov_page != NULL)
1871 cfs_free_page(fmb->fmb_kiov[i].kiov_page);
1873 list_del(&fmb->fmb_list);
1874 PORTAL_FREE(fmb, offsetof(ksock_fmb_t, fmb_kiov[npages]));
1879 ksocknal_free_buffers (void)
1881 ksocknal_free_fmbs(&ksocknal_data.ksnd_small_fmp);
1882 ksocknal_free_fmbs(&ksocknal_data.ksnd_large_fmp);
1884 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_ltxs) == 0);
1886 if (ksocknal_data.ksnd_schedulers != NULL)
1887 PORTAL_FREE (ksocknal_data.ksnd_schedulers,
1888 sizeof (ksock_sched_t) * ksocknal_data.ksnd_nschedulers);
1890 PORTAL_FREE (ksocknal_data.ksnd_peers,
1891 sizeof (struct list_head) *
1892 ksocknal_data.ksnd_peer_hash_size);
1896 ksocknal_api_shutdown (nal_t *nal)
1898 ksock_sched_t *sched;
1901 if (nal->nal_refct != 0) {
1902 /* This module got the first ref */
1903 PORTAL_MODULE_UNUSE;
1907 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
1908 atomic_read (&portal_kmemory));
1910 LASSERT(nal == &ksocknal_api);
1912 switch (ksocknal_data.ksnd_init) {
1916 case SOCKNAL_INIT_ALL:
1917 libcfs_nal_cmd_unregister(SOCKNAL);
1919 ksocknal_data.ksnd_init = SOCKNAL_INIT_LIB;
1922 case SOCKNAL_INIT_LIB:
1923 /* No more calls to ksocknal_cmd() to create new
1924 * autoroutes/connections since we're being unloaded. */
1926 /* Delete all peers */
1927 ksocknal_del_peer(PTL_NID_ANY, 0, 0);
1929 /* Wait for all peer state to clean up */
1931 while (atomic_read (&ksocknal_data.ksnd_npeers) != 0) {
1933 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1934 "waiting for %d peers to disconnect\n",
1935 atomic_read (&ksocknal_data.ksnd_npeers));
1936 set_current_state (TASK_UNINTERRUPTIBLE);
1937 schedule_timeout (cfs_time_seconds(1));
1940 /* Tell lib we've stopped calling into her. */
1941 lib_fini(&ksocknal_lib);
1943 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
1946 case SOCKNAL_INIT_DATA:
1947 LASSERT (atomic_read (&ksocknal_data.ksnd_npeers) == 0);
1948 LASSERT (ksocknal_data.ksnd_peers != NULL);
1949 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1950 LASSERT (list_empty (&ksocknal_data.ksnd_peers[i]));
1952 LASSERT (list_empty (&ksocknal_data.ksnd_enomem_conns));
1953 LASSERT (list_empty (&ksocknal_data.ksnd_zombie_conns));
1954 LASSERT (list_empty (&ksocknal_data.ksnd_autoconnectd_routes));
1955 LASSERT (list_empty (&ksocknal_data.ksnd_small_fmp.fmp_blocked_conns));
1956 LASSERT (list_empty (&ksocknal_data.ksnd_large_fmp.fmp_blocked_conns));
1958 if (ksocknal_data.ksnd_schedulers != NULL)
1959 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
1960 ksock_sched_t *kss =
1961 &ksocknal_data.ksnd_schedulers[i];
1963 LASSERT (list_empty (&kss->kss_tx_conns));
1964 LASSERT (list_empty (&kss->kss_rx_conns));
1965 LASSERT (kss->kss_nconns == 0);
1968 /* stop router calling me */
1969 kpr_shutdown (&ksocknal_data.ksnd_router);
1971 /* flag threads to terminate; wake and wait for them to die */
1972 ksocknal_data.ksnd_shuttingdown = 1;
1973 cfs_waitq_broadcast (&ksocknal_data.ksnd_autoconnectd_waitq);
1974 cfs_waitq_broadcast (&ksocknal_data.ksnd_reaper_waitq);
1976 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
1977 sched = &ksocknal_data.ksnd_schedulers[i];
1978 cfs_waitq_broadcast(&sched->kss_waitq);
1982 read_lock(&ksocknal_data.ksnd_global_lock);
1983 while (ksocknal_data.ksnd_nthreads != 0) {
1985 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1986 "waiting for %d threads to terminate\n",
1987 ksocknal_data.ksnd_nthreads);
1988 read_unlock(&ksocknal_data.ksnd_global_lock);
1989 set_current_state (TASK_UNINTERRUPTIBLE);
1990 schedule_timeout (cfs_time_seconds(1));
1991 read_lock(&ksocknal_data.ksnd_global_lock);
1993 read_unlock(&ksocknal_data.ksnd_global_lock);
1995 kpr_deregister (&ksocknal_data.ksnd_router);
1997 ksocknal_free_buffers();
1999 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2002 case SOCKNAL_INIT_NOTHING:
2006 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2007 atomic_read (&portal_kmemory));
2009 printk(KERN_INFO "Lustre: Routing socket NAL unloaded (final mem %d)\n",
2010 atomic_read(&portal_kmemory));
2015 ksocknal_init_incarnation (void)
2019 /* The incarnation number is the time this module loaded and it
2020 * identifies this particular instance of the socknal. Hopefully
2021 * we won't be able to reboot more frequently than 1MHz for the
2022 * forseeable future :) */
2024 do_gettimeofday(&tv);
2026 ksocknal_data.ksnd_incarnation =
2027 (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2031 ksocknal_api_startup (nal_t *nal, ptl_pid_t requested_pid,
2032 ptl_ni_limits_t *requested_limits,
2033 ptl_ni_limits_t *actual_limits)
2035 ptl_process_id_t process_id;
2036 int pkmem = atomic_read(&portal_kmemory);
2041 LASSERT (nal == &ksocknal_api);
2043 if (nal->nal_refct != 0) {
2044 if (actual_limits != NULL)
2045 *actual_limits = ksocknal_lib.libnal_ni.ni_actual_limits;
2046 /* This module got the first ref */
2051 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2053 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2055 ksocknal_init_incarnation();
2057 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2058 PORTAL_ALLOC (ksocknal_data.ksnd_peers,
2059 sizeof (struct list_head) * ksocknal_data.ksnd_peer_hash_size);
2060 if (ksocknal_data.ksnd_peers == NULL)
2063 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2064 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2066 rwlock_init(&ksocknal_data.ksnd_global_lock);
2068 spin_lock_init(&ksocknal_data.ksnd_small_fmp.fmp_lock);
2069 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_small_fmp.fmp_idle_fmbs);
2070 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_small_fmp.fmp_blocked_conns);
2071 ksocknal_data.ksnd_small_fmp.fmp_buff_pages = SOCKNAL_SMALL_FWD_PAGES;
2073 spin_lock_init(&ksocknal_data.ksnd_large_fmp.fmp_lock);
2074 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_large_fmp.fmp_idle_fmbs);
2075 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_large_fmp.fmp_blocked_conns);
2076 ksocknal_data.ksnd_large_fmp.fmp_buff_pages = SOCKNAL_LARGE_FWD_PAGES;
2078 spin_lock_init (&ksocknal_data.ksnd_reaper_lock);
2079 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_enomem_conns);
2080 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_zombie_conns);
2081 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_deathrow_conns);
2082 cfs_waitq_init(&ksocknal_data.ksnd_reaper_waitq);
2084 spin_lock_init (&ksocknal_data.ksnd_autoconnectd_lock);
2085 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_autoconnectd_routes);
2086 cfs_waitq_init(&ksocknal_data.ksnd_autoconnectd_waitq);
2088 /* NB memset above zeros whole of ksocknal_data, including
2089 * ksocknal_data.ksnd_irqinfo[all].ksni_valid */
2091 /* flag lists/ptrs/locks initialised */
2092 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2094 ksocknal_data.ksnd_nschedulers = ksocknal_nsched();
2095 PORTAL_ALLOC(ksocknal_data.ksnd_schedulers,
2096 sizeof(ksock_sched_t) * ksocknal_data.ksnd_nschedulers);
2097 if (ksocknal_data.ksnd_schedulers == NULL) {
2098 ksocknal_api_shutdown (nal);
2102 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
2103 ksock_sched_t *kss = &ksocknal_data.ksnd_schedulers[i];
2105 spin_lock_init (&kss->kss_lock);
2106 CFS_INIT_LIST_HEAD (&kss->kss_rx_conns);
2107 CFS_INIT_LIST_HEAD (&kss->kss_tx_conns);
2109 CFS_INIT_LIST_HEAD (&kss->kss_zctxdone_list);
2111 cfs_waitq_init (&kss->kss_waitq);
2114 /* NB we have to wait to be told our true NID... */
2115 process_id.pid = requested_pid;
2118 rc = lib_init(&ksocknal_lib, nal, process_id,
2119 requested_limits, actual_limits);
2121 CERROR("lib_init failed: error %d\n", rc);
2122 ksocknal_api_shutdown (nal);
2126 ksocknal_data.ksnd_init = SOCKNAL_INIT_LIB; // flag lib_init() called
2128 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
2129 rc = ksocknal_thread_start (ksocknal_scheduler,
2130 &ksocknal_data.ksnd_schedulers[i]);
2132 CERROR("Can't spawn socknal scheduler[%d]: %d\n",
2134 ksocknal_api_shutdown (nal);
2139 for (i = 0; i < SOCKNAL_N_AUTOCONNECTD; i++) {
2140 rc = ksocknal_thread_start (ksocknal_autoconnectd, (void *)((long)i));
2142 CERROR("Can't spawn socknal autoconnectd: %d\n", rc);
2143 ksocknal_api_shutdown (nal);
2148 rc = ksocknal_thread_start (ksocknal_reaper, NULL);
2150 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2151 ksocknal_api_shutdown (nal);
2155 rc = kpr_register(&ksocknal_data.ksnd_router,
2156 &ksocknal_router_interface);
2158 CDEBUG(D_NET, "Can't initialise routing interface "
2159 "(rc = %d): not routing\n", rc);
2161 /* Only allocate forwarding buffers if there's a router */
2163 for (i = 0; i < (SOCKNAL_SMALL_FWD_NMSGS +
2164 SOCKNAL_LARGE_FWD_NMSGS); i++) {
2166 ksock_fmb_pool_t *pool;
2169 if (i < SOCKNAL_SMALL_FWD_NMSGS)
2170 pool = &ksocknal_data.ksnd_small_fmp;
2172 pool = &ksocknal_data.ksnd_large_fmp;
2174 PORTAL_ALLOC(fmb, offsetof(ksock_fmb_t,
2175 fmb_kiov[pool->fmp_buff_pages]));
2177 ksocknal_api_shutdown(nal);
2181 fmb->fmb_pool = pool;
2183 for (j = 0; j < pool->fmp_buff_pages; j++) {
2184 fmb->fmb_kiov[j].kiov_page = cfs_alloc_page(CFS_ALLOC_STD);
2186 if (fmb->fmb_kiov[j].kiov_page == NULL) {
2187 ksocknal_api_shutdown (nal);
2191 LASSERT(cfs_page_address(fmb->fmb_kiov[j].kiov_page) != NULL);
2194 list_add(&fmb->fmb_list, &pool->fmp_idle_fmbs);
2198 rc = libcfs_nal_cmd_register(SOCKNAL, &ksocknal_cmd, NULL);
2200 CERROR ("Can't initialise command interface (rc = %d)\n", rc);
2201 ksocknal_api_shutdown (nal);
2205 /* flag everything initialised */
2206 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2208 printk(KERN_INFO "Lustre: Routing socket NAL loaded "
2209 "(Routing %s, initial mem %d, incarnation "LPX64")\n",
2210 kpr_routing (&ksocknal_data.ksnd_router) ?
2211 "enabled" : "disabled", pkmem, ksocknal_data.ksnd_incarnation);
2217 ksocknal_module_fini (void)
2219 #ifdef CONFIG_SYSCTL
2220 if (ksocknal_tunables.ksnd_sysctl != NULL)
2221 unregister_sysctl_table (ksocknal_tunables.ksnd_sysctl);
2223 PtlNIFini(ksocknal_ni);
2225 ptl_unregister_nal(SOCKNAL);
2228 extern cfs_sysctl_table_t ksocknal_top_ctl_table[];
2231 ksocknal_module_init (void)
2235 /* packet descriptor must fit in a router descriptor's scratchpad */
2236 LASSERT(sizeof (ksock_tx_t) <= sizeof (kprfd_scratch_t));
2237 /* the following must be sizeof(int) for proc_dointvec() */
2238 LASSERT(sizeof (ksocknal_tunables.ksnd_io_timeout) == sizeof (int));
2239 LASSERT(sizeof (ksocknal_tunables.ksnd_eager_ack) == sizeof (int));
2240 LASSERT(sizeof (ksocknal_tunables.ksnd_typed_conns) == sizeof (int));
2241 LASSERT(sizeof (ksocknal_tunables.ksnd_min_bulk) == sizeof (int));
2242 LASSERT(sizeof (ksocknal_tunables.ksnd_buffer_size) == sizeof (int));
2243 LASSERT(sizeof (ksocknal_tunables.ksnd_nagle) == sizeof (int));
2244 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_idle) == sizeof (int));
2245 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_count) == sizeof (int));
2246 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_intvl) == sizeof (int));
2248 LASSERT(sizeof (ksocknal_tunables.ksnd_irq_affinity) == sizeof (int));
2251 LASSERT(sizeof (ksocknal_tunables.ksnd_zc_min_frag) == sizeof (int));
2253 /* check ksnr_connected/connecting field large enough */
2254 LASSERT(SOCKNAL_CONN_NTYPES <= 4);
2256 ksocknal_api.nal_ni_init = ksocknal_api_startup;
2257 ksocknal_api.nal_ni_fini = ksocknal_api_shutdown;
2259 /* Initialise dynamic tunables to defaults once only */
2260 ksocknal_tunables.ksnd_io_timeout = SOCKNAL_IO_TIMEOUT;
2261 ksocknal_tunables.ksnd_eager_ack = SOCKNAL_EAGER_ACK;
2262 ksocknal_tunables.ksnd_typed_conns = SOCKNAL_TYPED_CONNS;
2263 ksocknal_tunables.ksnd_min_bulk = SOCKNAL_MIN_BULK;
2264 ksocknal_tunables.ksnd_buffer_size = SOCKNAL_BUFFER_SIZE;
2265 ksocknal_tunables.ksnd_nagle = SOCKNAL_NAGLE;
2266 ksocknal_tunables.ksnd_keepalive_idle = SOCKNAL_KEEPALIVE_IDLE;
2267 ksocknal_tunables.ksnd_keepalive_count = SOCKNAL_KEEPALIVE_COUNT;
2268 ksocknal_tunables.ksnd_keepalive_intvl = SOCKNAL_KEEPALIVE_INTVL;
2270 ksocknal_tunables.ksnd_irq_affinity = SOCKNAL_IRQ_AFFINITY;
2273 ksocknal_tunables.ksnd_zc_min_frag = SOCKNAL_ZC_MIN_FRAG;
2276 rc = ptl_register_nal(SOCKNAL, &ksocknal_api);
2278 CERROR("Can't register SOCKNAL: %d\n", rc);
2279 return (-ENOMEM); /* or something... */
2282 /* Pure gateways want the NAL started up at module load time... */
2283 rc = PtlNIInit(SOCKNAL, LUSTRE_SRV_PTL_PID, NULL, NULL, &ksocknal_ni);
2284 if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
2285 ptl_unregister_nal(SOCKNAL);
2289 #ifdef CONFIG_SYSCTL
2290 /* Press on regardless even if registering sysctl doesn't work */
2291 ksocknal_tunables.ksnd_sysctl =
2292 register_sysctl_table (ksocknal_top_ctl_table, 0);
2297 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2298 MODULE_DESCRIPTION("Kernel TCP Socket NAL v1.0.0");
2299 MODULE_LICENSE("GPL");
2301 cfs_module(ksocknal, "1.0.0", ksocknal_module_init, ksocknal_module_fini);