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_conn_count++;
347 /* Successful connection => further attempts can
348 * proceed immediately */
349 route->ksnr_timeout = cfs_time_current();
350 route->ksnr_retry_interval = SOCKNAL_MIN_RECONNECT_INTERVAL;
354 ksocknal_add_route_locked (ksock_peer_t *peer, ksock_route_t *route)
356 struct list_head *tmp;
359 ksock_route_t *route2;
361 LASSERT (route->ksnr_peer == NULL);
362 LASSERT (!route->ksnr_connecting);
363 LASSERT (route->ksnr_connected == 0);
365 /* LASSERT(unique) */
366 list_for_each(tmp, &peer->ksnp_routes) {
367 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
369 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
370 CERROR ("Duplicate route "LPX64" %u.%u.%u.%u\n",
371 peer->ksnp_nid, HIPQUAD(route->ksnr_ipaddr));
376 route->ksnr_peer = peer;
377 atomic_inc (&peer->ksnp_refcount);
378 /* peer's routelist takes over my ref on 'route' */
379 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
381 list_for_each(tmp, &peer->ksnp_conns) {
382 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
383 type = conn->ksnc_type;
385 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
388 ksocknal_associate_route_conn_locked(route, conn);
389 /* keep going (typed routes) */
394 ksocknal_del_route_locked (ksock_route_t *route)
396 ksock_peer_t *peer = route->ksnr_peer;
397 ksock_interface_t *iface;
399 struct list_head *ctmp;
400 struct list_head *cnxt;
402 LASSERT (!route->ksnr_deleted);
404 /* Close associated conns */
405 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
406 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
408 if (conn->ksnc_route != route)
411 ksocknal_close_conn_locked (conn, 0);
414 if (route->ksnr_myipaddr != 0) {
415 iface = ksocknal_ip2iface(route->ksnr_myipaddr);
417 iface->ksni_nroutes--;
420 route->ksnr_deleted = 1;
421 list_del (&route->ksnr_list);
422 ksocknal_put_route (route); /* drop peer's ref */
424 if (list_empty (&peer->ksnp_routes) &&
425 list_empty (&peer->ksnp_conns)) {
426 /* I've just removed the last autoconnect route of a peer
427 * with no active connections */
428 ksocknal_unlink_peer_locked (peer);
433 ksocknal_add_peer (ptl_nid_t nid, __u32 ipaddr, int port)
436 struct list_head *tmp;
439 ksock_route_t *route;
440 ksock_route_t *route2;
442 if (nid == PTL_NID_ANY)
445 /* Have a brand new peer ready... */
446 peer = ksocknal_create_peer (nid);
450 route = ksocknal_create_route (ipaddr, port);
452 ksocknal_put_peer (peer);
456 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
458 peer2 = ksocknal_find_peer_locked (nid);
460 ksocknal_put_peer (peer);
463 /* peer table takes my ref on peer */
464 list_add_tail (&peer->ksnp_list,
465 ksocknal_nid2peerlist (nid));
469 list_for_each (tmp, &peer->ksnp_routes) {
470 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
472 if (route2->ksnr_ipaddr == ipaddr)
477 if (route2 == NULL) {
478 ksocknal_add_route_locked(peer, route);
479 route->ksnr_share_count++;
481 ksocknal_put_route(route);
482 route2->ksnr_share_count++;
485 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
491 ksocknal_del_peer_locked (ksock_peer_t *peer, __u32 ip, int single_share)
494 ksock_route_t *route;
495 struct list_head *tmp;
496 struct list_head *nxt;
499 LASSERT (!peer->ksnp_closing);
501 /* Extra ref prevents peer disappearing until I'm done with it */
502 atomic_inc(&peer->ksnp_refcount);
504 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
505 route = list_entry(tmp, ksock_route_t, ksnr_list);
507 if (single_share && route->ksnr_share_count == 0)
511 if (!(ip == 0 || route->ksnr_ipaddr == ip))
515 route->ksnr_share_count = 0;
516 else if (route->ksnr_share_count > 0)
517 route->ksnr_share_count--;
519 if (route->ksnr_share_count == 0) {
520 /* This deletes associated conns too */
521 ksocknal_del_route_locked (route);
529 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
530 route = list_entry(tmp, ksock_route_t, ksnr_list);
531 nshared += route->ksnr_share_count;
535 /* remove everything else if there are no explicit entries
538 list_for_each_safe (tmp, nxt, &peer->ksnp_routes) {
539 route = list_entry(tmp, ksock_route_t, ksnr_list);
541 /* we should only be removing auto-entries */
542 LASSERT(route->ksnr_share_count == 0);
543 ksocknal_del_route_locked (route);
546 list_for_each_safe (tmp, nxt, &peer->ksnp_conns) {
547 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
549 ksocknal_close_conn_locked(conn, 0);
553 ksocknal_put_peer(peer);
554 /* NB peer unlinks itself when last conn/route is removed */
558 ksocknal_del_peer (ptl_nid_t nid, __u32 ip, int single_share)
561 struct list_head *ptmp;
562 struct list_head *pnxt;
569 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
571 if (nid != PTL_NID_ANY)
572 lo = hi = ksocknal_nid2peerlist(nid) - ksocknal_data.ksnd_peers;
575 hi = ksocknal_data.ksnd_peer_hash_size - 1;
578 for (i = lo; i <= hi; i++) {
579 list_for_each_safe (ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
580 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
582 if (!(nid == PTL_NID_ANY || peer->ksnp_nid == nid))
585 ksocknal_del_peer_locked (peer, ip, single_share);
586 rc = 0; /* matched! */
593 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
599 ksocknal_get_conn_by_idx (int index)
602 struct list_head *ptmp;
604 struct list_head *ctmp;
607 read_lock (&ksocknal_data.ksnd_global_lock);
609 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
610 list_for_each (ptmp, &ksocknal_data.ksnd_peers[i]) {
611 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
613 LASSERT (!peer->ksnp_closing);
615 list_for_each (ctmp, &peer->ksnp_conns) {
619 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
620 atomic_inc (&conn->ksnc_refcount);
621 read_unlock (&ksocknal_data.ksnd_global_lock);
627 read_unlock (&ksocknal_data.ksnd_global_lock);
632 ksocknal_choose_scheduler_locked (unsigned int irq)
634 ksock_sched_t *sched;
635 ksock_irqinfo_t *info;
638 LASSERT (irq < NR_IRQS);
639 info = &ksocknal_data.ksnd_irqinfo[irq];
641 if (irq != 0 && /* hardware NIC */
642 info->ksni_valid) { /* already set up */
643 return (&ksocknal_data.ksnd_schedulers[info->ksni_sched]);
646 /* software NIC (irq == 0) || not associated with a scheduler yet.
647 * Choose the CPU with the fewest connections... */
648 sched = &ksocknal_data.ksnd_schedulers[0];
649 for (i = 1; i < ksocknal_data.ksnd_nschedulers; i++)
650 if (sched->kss_nconns >
651 ksocknal_data.ksnd_schedulers[i].kss_nconns)
652 sched = &ksocknal_data.ksnd_schedulers[i];
654 if (irq != 0) { /* Hardware NIC */
655 info->ksni_valid = 1;
656 info->ksni_sched = sched - ksocknal_data.ksnd_schedulers;
659 LASSERT (info->ksni_sched == sched - ksocknal_data.ksnd_schedulers);
666 ksocknal_local_ipvec (__u32 *ipaddrs)
671 read_lock (&ksocknal_data.ksnd_global_lock);
673 nip = ksocknal_data.ksnd_ninterfaces;
674 for (i = 0; i < nip; i++) {
675 LASSERT (i < SOCKNAL_MAX_INTERFACES);
677 ipaddrs[i] = ksocknal_data.ksnd_interfaces[i].ksni_ipaddr;
678 LASSERT (ipaddrs[i] != 0);
681 read_unlock (&ksocknal_data.ksnd_global_lock);
686 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
688 int best_netmatch = 0;
695 for (i = 0; i < nips; i++) {
699 this_xor = (ips[i] ^ iface->ksni_ipaddr);
700 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
703 best_netmatch < this_netmatch ||
704 (best_netmatch == this_netmatch &&
705 best_xor > this_xor)))
709 best_netmatch = this_netmatch;
718 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
720 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
722 ksock_interface_t *iface;
723 ksock_interface_t *best_iface;
734 /* CAVEAT EMPTOR: We do all our interface matching with an
735 * exclusive hold of global lock at IRQ priority. We're only
736 * expecting to be dealing with small numbers of interfaces, so the
737 * O(n**3)-ness shouldn't matter */
739 /* Also note that I'm not going to return more than n_peerips
740 * interfaces, even if I have more myself */
742 write_lock_irqsave(global_lock, flags);
744 LASSERT (n_peerips <= SOCKNAL_MAX_INTERFACES);
745 LASSERT (ksocknal_data.ksnd_ninterfaces <= SOCKNAL_MAX_INTERFACES);
747 n_ips = MIN(n_peerips, ksocknal_data.ksnd_ninterfaces);
749 for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
750 /* ^ yes really... */
752 /* If we have any new interfaces, first tick off all the
753 * peer IPs that match old interfaces, then choose new
754 * interfaces to match the remaining peer IPS.
755 * We don't forget interfaces we've stopped using; we might
756 * start using them again... */
758 if (i < peer->ksnp_n_passive_ips) {
760 ip = peer->ksnp_passive_ips[i];
761 best_iface = ksocknal_ip2iface(ip);
763 /* peer passive ips are kept up to date */
764 LASSERT(best_iface != NULL);
766 /* choose a new interface */
767 LASSERT (i == peer->ksnp_n_passive_ips);
773 for (j = 0; j < ksocknal_data.ksnd_ninterfaces; j++) {
774 iface = &ksocknal_data.ksnd_interfaces[j];
775 ip = iface->ksni_ipaddr;
777 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
778 if (peer->ksnp_passive_ips[k] == ip)
781 if (k < peer->ksnp_n_passive_ips) /* using it already */
784 k = ksocknal_match_peerip(iface, peerips, n_peerips);
785 xor = (ip ^ peerips[k]);
786 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
788 if (!(best_iface == NULL ||
789 best_netmatch < this_netmatch ||
790 (best_netmatch == this_netmatch &&
791 best_npeers > iface->ksni_npeers)))
795 best_netmatch = this_netmatch;
796 best_npeers = iface->ksni_npeers;
799 best_iface->ksni_npeers++;
800 ip = best_iface->ksni_ipaddr;
801 peer->ksnp_passive_ips[i] = ip;
802 peer->ksnp_n_passive_ips = i+1;
805 LASSERT (best_iface != NULL);
807 /* mark the best matching peer IP used */
808 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
812 /* Overwrite input peer IP addresses */
813 memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
815 write_unlock_irqrestore(global_lock, flags);
821 ksocknal_create_routes(ksock_peer_t *peer, int port,
822 __u32 *peer_ipaddrs, int npeer_ipaddrs)
824 ksock_route_t *newroute = NULL;
825 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
827 struct list_head *rtmp;
828 ksock_route_t *route;
829 ksock_interface_t *iface;
830 ksock_interface_t *best_iface;
837 /* CAVEAT EMPTOR: We do all our interface matching with an
838 * exclusive hold of global lock at IRQ priority. We're only
839 * expecting to be dealing with small numbers of interfaces, so the
840 * O(n**3)-ness here shouldn't matter */
842 write_lock_irqsave(global_lock, flags);
844 LASSERT (npeer_ipaddrs <= SOCKNAL_MAX_INTERFACES);
846 for (i = 0; i < npeer_ipaddrs; i++) {
847 if (newroute != NULL) {
848 newroute->ksnr_ipaddr = peer_ipaddrs[i];
850 write_unlock_irqrestore(global_lock, flags);
852 newroute = ksocknal_create_route(peer_ipaddrs[i], port);
853 if (newroute == NULL)
856 write_lock_irqsave(global_lock, flags);
859 /* Already got a route? */
861 list_for_each(rtmp, &peer->ksnp_routes) {
862 route = list_entry(rtmp, ksock_route_t, ksnr_list);
864 if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
876 LASSERT (ksocknal_data.ksnd_ninterfaces <= SOCKNAL_MAX_INTERFACES);
878 /* Select interface to connect from */
879 for (j = 0; j < ksocknal_data.ksnd_ninterfaces; j++) {
880 iface = &ksocknal_data.ksnd_interfaces[j];
882 /* Using this interface already? */
883 list_for_each(rtmp, &peer->ksnp_routes) {
884 route = list_entry(rtmp, ksock_route_t, ksnr_list);
886 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
894 this_netmatch = (((iface->ksni_ipaddr ^
895 newroute->ksnr_ipaddr) &
896 iface->ksni_netmask) == 0) ? 1 : 0;
898 if (!(best_iface == NULL ||
899 best_netmatch < this_netmatch ||
900 (best_netmatch == this_netmatch &&
901 best_nroutes > iface->ksni_nroutes)))
905 best_netmatch = this_netmatch;
906 best_nroutes = iface->ksni_nroutes;
909 if (best_iface == NULL)
912 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
913 best_iface->ksni_nroutes++;
915 ksocknal_add_route_locked(peer, newroute);
919 write_unlock_irqrestore(global_lock, flags);
920 if (newroute != NULL)
921 ksocknal_put_route(newroute);
925 ksocknal_create_conn (ksock_route_t *route, struct socket *sock, int type)
927 int passive = (type == SOCKNAL_CONN_NONE);
928 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
929 __u32 ipaddrs[SOCKNAL_MAX_INTERFACES];
932 struct list_head *tmp;
937 ksock_peer_t *peer = NULL;
939 ksock_sched_t *sched;
944 /* NB, sock has an associated file since (a) this connection might
945 * have been created in userland and (b) we need to refcount the
946 * socket so that we don't close it while I/O is being done on
947 * it, and sock->file has that pre-cooked... */
948 LASSERT (KSN_SOCK2FILE(sock) != NULL);
949 LASSERT (cfs_file_count(KSN_SOCK2FILE(sock)) > 0);
950 LASSERT (route == NULL || !passive);
952 rc = ksocknal_lib_setup_sock (sock);
956 irq = ksocknal_lib_sock_irq (sock);
958 PORTAL_ALLOC(conn, sizeof(*conn));
962 memset (conn, 0, sizeof (*conn));
963 conn->ksnc_peer = NULL;
964 conn->ksnc_route = NULL;
965 conn->ksnc_sock = sock;
966 conn->ksnc_type = type;
967 ksocknal_lib_save_callback(sock, conn);
968 atomic_set (&conn->ksnc_refcount, 1); /* 1 ref for me */
970 conn->ksnc_rx_ready = 0;
971 conn->ksnc_rx_scheduled = 0;
972 ksocknal_new_packet (conn, 0);
974 CFS_INIT_LIST_HEAD (&conn->ksnc_tx_queue);
975 conn->ksnc_tx_ready = 0;
976 conn->ksnc_tx_scheduled = 0;
977 atomic_set (&conn->ksnc_tx_nob, 0);
979 /* stash conn's local and remote addrs */
980 rc = ksocknal_lib_get_conn_addrs (conn);
985 /* Active connection sends HELLO eagerly */
986 rc = ksocknal_local_ipvec(ipaddrs);
991 rc = ksocknal_send_hello (conn, ipaddrs, nipaddrs);
996 /* Find out/confirm peer's NID and connection type and get the
997 * vector of interfaces she's willing to let me connect to */
998 nid = (route == NULL) ? PTL_NID_ANY : route->ksnr_peer->ksnp_nid;
999 rc = ksocknal_recv_hello (conn, &nid, &incarnation, ipaddrs);
1003 LASSERT (nid != PTL_NID_ANY);
1005 if (route != NULL) {
1006 peer = route->ksnr_peer;
1007 atomic_inc(&peer->ksnp_refcount);
1009 peer = ksocknal_create_peer(nid);
1015 write_lock_irqsave(global_lock, flags);
1017 peer2 = ksocknal_find_peer_locked(nid);
1018 if (peer2 == NULL) {
1019 /* NB this puts an "empty" peer in the peer
1020 * table (which takes my ref) */
1021 list_add_tail(&peer->ksnp_list,
1022 ksocknal_nid2peerlist(nid));
1024 ksocknal_put_peer(peer);
1028 atomic_inc(&peer->ksnp_refcount);
1030 write_unlock_irqrestore(global_lock, flags);
1034 ksocknal_create_routes(peer, conn->ksnc_port,
1038 rc = ksocknal_select_ips(peer, ipaddrs, nipaddrs);
1040 rc = ksocknal_send_hello (conn, ipaddrs, rc);
1045 write_lock_irqsave (global_lock, flags);
1047 if (peer->ksnp_closing ||
1048 (route != NULL && route->ksnr_deleted)) {
1049 /* route/peer got closed under me */
1054 /* Refuse to duplicate an existing connection (both sides might
1055 * autoconnect at once), unless this is a loopback connection */
1056 if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1057 list_for_each(tmp, &peer->ksnp_conns) {
1058 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1060 if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1061 conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1062 conn2->ksnc_type != conn->ksnc_type ||
1063 conn2->ksnc_incarnation != incarnation)
1066 CWARN("Not creating duplicate connection to "
1067 "%u.%u.%u.%u type %d\n",
1068 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_type);
1074 /* If the connection created by this route didn't bind to the IP
1075 * address the route connected to, the connection/route matching
1076 * code below probably isn't going to work. */
1077 if (route != NULL &&
1078 route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1079 CERROR("Route "LPX64" %u.%u.%u.%u connected to %u.%u.%u.%u\n",
1081 HIPQUAD(route->ksnr_ipaddr),
1082 HIPQUAD(conn->ksnc_ipaddr));
1085 /* Search for a route corresponding to the new connection and
1086 * create an association. This allows incoming connections created
1087 * by routes in my peer to match my own route entries so I don't
1088 * continually create duplicate routes. */
1089 list_for_each (tmp, &peer->ksnp_routes) {
1090 route = list_entry(tmp, ksock_route_t, ksnr_list);
1092 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1095 ksocknal_associate_route_conn_locked(route, conn);
1099 /* Give conn a ref on sock->file since we're going to return success */
1100 cfs_get_file(KSN_SOCK2FILE(sock));
1102 conn->ksnc_peer = peer; /* conn takes my ref on peer */
1103 conn->ksnc_incarnation = incarnation;
1104 peer->ksnp_last_alive = cfs_time_current();
1105 peer->ksnp_error = 0;
1107 sched = ksocknal_choose_scheduler_locked (irq);
1108 sched->kss_nconns++;
1109 conn->ksnc_scheduler = sched;
1111 /* Set the deadline for the outgoing HELLO to drain */
1112 conn->ksnc_tx_bufnob = SOCK_WMEM_QUEUED(sock);
1113 conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
1114 mb(); /* order with adding to peer's conn list */
1116 list_add (&conn->ksnc_list, &peer->ksnp_conns);
1117 atomic_inc (&conn->ksnc_refcount);
1119 /* NB my callbacks block while I hold ksnd_global_lock */
1120 ksocknal_lib_set_callback(sock, conn);
1122 /* Take all the packets blocking for a connection.
1123 * NB, it might be nicer to share these blocked packets among any
1124 * other connections that are becoming established. */
1125 while (!list_empty (&peer->ksnp_tx_queue)) {
1126 tx = list_entry (peer->ksnp_tx_queue.next,
1127 ksock_tx_t, tx_list);
1129 list_del (&tx->tx_list);
1130 ksocknal_queue_tx_locked (tx, conn);
1133 rc = ksocknal_close_stale_conns_locked(peer, incarnation);
1135 CERROR ("Closed %d stale conns to nid "LPX64" ip %d.%d.%d.%d\n",
1136 rc, conn->ksnc_peer->ksnp_nid,
1137 HIPQUAD(conn->ksnc_ipaddr));
1139 write_unlock_irqrestore (global_lock, flags);
1141 ksocknal_lib_bind_irq (irq);
1143 /* Call the callbacks right now to get things going. */
1144 if (ksocknal_getconnsock(conn) == 0) {
1145 ksocknal_lib_act_callback(sock, conn);
1146 ksocknal_putconnsock(conn);
1149 CWARN("New conn nid:"LPX64" %u.%u.%u.%u -> %u.%u.%u.%u/%d"
1150 " incarnation:"LPX64" sched[%d]/%d\n",
1151 nid, HIPQUAD(conn->ksnc_myipaddr),
1152 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port, incarnation,
1153 (int)(conn->ksnc_scheduler - ksocknal_data.ksnd_schedulers), irq);
1155 ksocknal_put_conn (conn);
1159 if (!peer->ksnp_closing &&
1160 list_empty (&peer->ksnp_conns) &&
1161 list_empty (&peer->ksnp_routes))
1162 ksocknal_unlink_peer_locked(peer);
1163 write_unlock_irqrestore(global_lock, flags);
1166 ksocknal_put_peer (peer);
1169 PORTAL_FREE (conn, sizeof(*conn));
1176 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1178 /* This just does the immmediate housekeeping, and queues the
1179 * connection for the reaper to terminate.
1180 * Caller holds ksnd_global_lock exclusively in irq context */
1181 ksock_peer_t *peer = conn->ksnc_peer;
1182 ksock_route_t *route;
1183 ksock_conn_t *conn2;
1184 struct list_head *tmp;
1186 LASSERT (peer->ksnp_error == 0);
1187 LASSERT (!conn->ksnc_closing);
1188 conn->ksnc_closing = 1;
1189 atomic_inc (&ksocknal_data.ksnd_nclosing_conns);
1191 /* ksnd_deathrow_conns takes over peer's ref */
1192 list_del (&conn->ksnc_list);
1194 route = conn->ksnc_route;
1195 if (route != NULL) {
1196 /* dissociate conn from route... */
1197 LASSERT (!route->ksnr_deleted);
1198 LASSERT ((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1201 list_for_each(tmp, &peer->ksnp_conns) {
1202 conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1204 if (conn2->ksnc_route == route &&
1205 conn2->ksnc_type == conn->ksnc_type)
1211 route->ksnr_connected &= ~(1 << conn->ksnc_type);
1213 conn->ksnc_route = NULL;
1215 #if 0 /* irrelevent with only eager routes */
1216 list_del (&route->ksnr_list); /* make route least favourite */
1217 list_add_tail (&route->ksnr_list, &peer->ksnp_routes);
1219 ksocknal_put_route (route); /* drop conn's ref on route */
1222 if (list_empty (&peer->ksnp_conns)) {
1223 /* No more connections to this peer */
1225 peer->ksnp_error = error; /* stash last conn close reason */
1227 if (list_empty (&peer->ksnp_routes)) {
1228 /* I've just closed last conn belonging to a
1229 * non-autoconnecting peer */
1230 ksocknal_unlink_peer_locked (peer);
1234 spin_lock (&ksocknal_data.ksnd_reaper_lock);
1236 list_add_tail (&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1237 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
1239 spin_unlock (&ksocknal_data.ksnd_reaper_lock);
1243 ksocknal_terminate_conn (ksock_conn_t *conn)
1245 /* This gets called by the reaper (guaranteed thread context) to
1246 * disengage the socket from its callbacks and close it.
1247 * ksnc_refcount will eventually hit zero, and then the reaper will
1249 unsigned long flags;
1250 ksock_peer_t *peer = conn->ksnc_peer;
1251 ksock_sched_t *sched = conn->ksnc_scheduler;
1256 LASSERT(conn->ksnc_closing);
1258 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1259 spin_lock_irqsave(&sched->kss_lock, flags);
1261 if (!conn->ksnc_tx_scheduled &&
1262 !list_empty(&conn->ksnc_tx_queue)){
1263 list_add_tail (&conn->ksnc_tx_list,
1264 &sched->kss_tx_conns);
1265 /* a closing conn is always ready to tx */
1266 conn->ksnc_tx_ready = 1;
1267 conn->ksnc_tx_scheduled = 1;
1268 /* extra ref for scheduler */
1269 atomic_inc (&conn->ksnc_refcount);
1271 cfs_waitq_signal (&sched->kss_waitq);
1274 spin_unlock_irqrestore (&sched->kss_lock, flags);
1276 /* serialise with callbacks */
1277 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1279 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1281 /* OK, so this conn may not be completely disengaged from its
1282 * scheduler yet, but it _has_ committed to terminate... */
1283 conn->ksnc_scheduler->kss_nconns--;
1285 if (peer->ksnp_error != 0) {
1286 /* peer's last conn closed in error */
1287 LASSERT (list_empty (&peer->ksnp_conns));
1289 /* convert peer's last-known-alive timestamp from jiffies */
1290 do_gettimeofday (&now);
1291 then = now.tv_sec - cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1292 peer->ksnp_last_alive));
1296 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1298 /* The socket is closed on the final put; either here, or in
1299 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1300 * when the connection was established, this will close the socket
1301 * immediately, aborting anything buffered in it. Any hung
1302 * zero-copy transmits will therefore complete in finite time. */
1303 ksocknal_putconnsock (conn);
1306 kpr_notify (&ksocknal_data.ksnd_router, peer->ksnp_nid,
1311 ksocknal_destroy_conn (ksock_conn_t *conn)
1313 /* Final coup-de-grace of the reaper */
1314 CDEBUG (D_NET, "connection %p\n", conn);
1316 LASSERT (atomic_read (&conn->ksnc_refcount) == 0);
1317 LASSERT (conn->ksnc_route == NULL);
1318 LASSERT (!conn->ksnc_tx_scheduled);
1319 LASSERT (!conn->ksnc_rx_scheduled);
1320 LASSERT (list_empty(&conn->ksnc_tx_queue));
1322 /* complete current receive if any */
1323 switch (conn->ksnc_rx_state) {
1324 case SOCKNAL_RX_BODY:
1325 CERROR("Completing partial receive from "LPX64
1326 ", ip %d.%d.%d.%d:%d, with error\n",
1327 conn->ksnc_peer->ksnp_nid,
1328 HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
1329 lib_finalize (&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_FAIL);
1331 case SOCKNAL_RX_BODY_FWD:
1332 ksocknal_fmb_callback (conn->ksnc_cookie, -ECONNABORTED);
1334 case SOCKNAL_RX_HEADER:
1335 case SOCKNAL_RX_SLOP:
1342 ksocknal_put_peer (conn->ksnc_peer);
1344 PORTAL_FREE (conn, sizeof (*conn));
1345 atomic_dec (&ksocknal_data.ksnd_nclosing_conns);
1349 ksocknal_put_conn (ksock_conn_t *conn)
1351 unsigned long flags;
1353 CDEBUG (D_OTHER, "putting conn[%p] -> "LPX64" (%d)\n",
1354 conn, conn->ksnc_peer->ksnp_nid,
1355 atomic_read (&conn->ksnc_refcount));
1357 LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
1358 if (!atomic_dec_and_test (&conn->ksnc_refcount))
1361 spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
1363 list_add (&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1364 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
1366 spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
1370 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1373 struct list_head *ctmp;
1374 struct list_head *cnxt;
1377 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
1378 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
1381 conn->ksnc_ipaddr == ipaddr) {
1383 ksocknal_close_conn_locked (conn, why);
1391 ksocknal_close_stale_conns_locked (ksock_peer_t *peer, __u64 incarnation)
1394 struct list_head *ctmp;
1395 struct list_head *cnxt;
1398 list_for_each_safe (ctmp, cnxt, &peer->ksnp_conns) {
1399 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
1401 if (conn->ksnc_incarnation == incarnation)
1404 CWARN("Closing stale conn nid:"LPX64" ip:%08x/%d "
1405 "incarnation:"LPX64"("LPX64")\n",
1406 peer->ksnp_nid, conn->ksnc_ipaddr, conn->ksnc_port,
1407 conn->ksnc_incarnation, incarnation);
1410 ksocknal_close_conn_locked (conn, -ESTALE);
1417 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1419 ksock_peer_t *peer = conn->ksnc_peer;
1420 __u32 ipaddr = conn->ksnc_ipaddr;
1421 unsigned long flags;
1424 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1426 count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1428 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1434 ksocknal_close_matching_conns (ptl_nid_t nid, __u32 ipaddr)
1436 unsigned long flags;
1438 struct list_head *ptmp;
1439 struct list_head *pnxt;
1445 write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
1447 if (nid != PTL_NID_ANY)
1448 lo = hi = ksocknal_nid2peerlist(nid) - ksocknal_data.ksnd_peers;
1451 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1454 for (i = lo; i <= hi; i++) {
1455 list_for_each_safe (ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1457 peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
1459 if (!(nid == PTL_NID_ANY || nid == peer->ksnp_nid))
1462 count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1466 write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
1468 /* wildcards always succeed */
1469 if (nid == PTL_NID_ANY || ipaddr == 0)
1472 return (count == 0 ? -ENOENT : 0);
1476 ksocknal_notify (void *arg, ptl_nid_t gw_nid, int alive)
1478 /* The router is telling me she's been notified of a change in
1479 * gateway state.... */
1481 CDEBUG (D_NET, "gw "LPX64" %s\n", gw_nid, alive ? "up" : "down");
1484 /* If the gateway crashed, close all open connections... */
1485 ksocknal_close_matching_conns (gw_nid, 0);
1489 /* ...otherwise do nothing. We can only establish new connections
1490 * if we have autroutes, and these connect on demand. */
1494 ksocknal_push_peer (ksock_peer_t *peer)
1498 struct list_head *tmp;
1501 for (index = 0; ; index++) {
1502 read_lock (&ksocknal_data.ksnd_global_lock);
1507 list_for_each (tmp, &peer->ksnp_conns) {
1509 conn = list_entry (tmp, ksock_conn_t, ksnc_list);
1510 atomic_inc (&conn->ksnc_refcount);
1515 read_unlock (&ksocknal_data.ksnd_global_lock);
1520 ksocknal_lib_push_conn (conn);
1521 ksocknal_put_conn (conn);
1526 ksocknal_push (ptl_nid_t nid)
1529 struct list_head *tmp;
1535 if (nid != PTL_NID_ANY) {
1536 peer = ksocknal_get_peer (nid);
1540 ksocknal_push_peer (peer);
1541 ksocknal_put_peer (peer);
1546 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1547 for (j = 0; ; j++) {
1548 read_lock (&ksocknal_data.ksnd_global_lock);
1553 list_for_each (tmp, &ksocknal_data.ksnd_peers[i]) {
1555 peer = list_entry(tmp, ksock_peer_t,
1557 atomic_inc (&peer->ksnp_refcount);
1562 read_unlock (&ksocknal_data.ksnd_global_lock);
1566 ksocknal_push_peer (peer);
1567 ksocknal_put_peer (peer);
1577 ksocknal_add_interface(__u32 ipaddress, __u32 netmask)
1579 unsigned long flags;
1580 ksock_interface_t *iface;
1584 struct list_head *ptmp;
1586 struct list_head *rtmp;
1587 ksock_route_t *route;
1589 if (ipaddress == 0 ||
1593 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
1595 iface = ksocknal_ip2iface(ipaddress);
1596 if (iface != NULL) {
1597 /* silently ignore dups */
1599 } else if (ksocknal_data.ksnd_ninterfaces == SOCKNAL_MAX_INTERFACES) {
1602 iface = &ksocknal_data.ksnd_interfaces[ksocknal_data.ksnd_ninterfaces++];
1604 iface->ksni_ipaddr = ipaddress;
1605 iface->ksni_netmask = netmask;
1606 iface->ksni_nroutes = 0;
1607 iface->ksni_npeers = 0;
1609 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1610 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1611 peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1613 for (j = 0; i < peer->ksnp_n_passive_ips; j++)
1614 if (peer->ksnp_passive_ips[j] == ipaddress)
1615 iface->ksni_npeers++;
1617 list_for_each(rtmp, &peer->ksnp_routes) {
1618 route = list_entry(rtmp, ksock_route_t, ksnr_list);
1620 if (route->ksnr_myipaddr == ipaddress)
1621 iface->ksni_nroutes++;
1627 /* NB only new connections will pay attention to the new interface! */
1630 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
1636 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
1638 struct list_head *tmp;
1639 struct list_head *nxt;
1640 ksock_route_t *route;
1645 for (i = 0; i < peer->ksnp_n_passive_ips; i++)
1646 if (peer->ksnp_passive_ips[i] == ipaddr) {
1647 for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
1648 peer->ksnp_passive_ips[j-1] =
1649 peer->ksnp_passive_ips[j];
1650 peer->ksnp_n_passive_ips--;
1654 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
1655 route = list_entry (tmp, ksock_route_t, ksnr_list);
1657 if (route->ksnr_myipaddr != ipaddr)
1660 if (route->ksnr_share_count != 0) {
1661 /* Manually created; keep, but unbind */
1662 route->ksnr_myipaddr = 0;
1664 ksocknal_del_route_locked(route);
1668 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
1669 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1671 if (conn->ksnc_myipaddr == ipaddr)
1672 ksocknal_close_conn_locked (conn, 0);
1677 ksocknal_del_interface(__u32 ipaddress)
1680 unsigned long flags;
1681 struct list_head *tmp;
1682 struct list_head *nxt;
1688 write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
1690 for (i = 0; i < ksocknal_data.ksnd_ninterfaces; i++) {
1691 this_ip = ksocknal_data.ksnd_interfaces[i].ksni_ipaddr;
1693 if (!(ipaddress == 0 ||
1694 ipaddress == this_ip))
1699 for (j = i+1; j < ksocknal_data.ksnd_ninterfaces; j++)
1700 ksocknal_data.ksnd_interfaces[j-1] =
1701 ksocknal_data.ksnd_interfaces[j];
1703 ksocknal_data.ksnd_ninterfaces--;
1705 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
1706 list_for_each_safe(tmp, nxt, &ksocknal_data.ksnd_peers[j]) {
1707 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
1709 ksocknal_peer_del_interface_locked(peer, this_ip);
1714 write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
1720 ksocknal_cmd(struct portals_cfg *pcfg, void * private)
1724 switch(pcfg->pcfg_command) {
1725 case NAL_CMD_GET_INTERFACE: {
1726 ksock_interface_t *iface;
1728 read_lock (&ksocknal_data.ksnd_global_lock);
1730 if (pcfg->pcfg_count < 0 ||
1731 pcfg->pcfg_count >= ksocknal_data.ksnd_ninterfaces) {
1735 iface = &ksocknal_data.ksnd_interfaces[pcfg->pcfg_count];
1737 pcfg->pcfg_id = iface->ksni_ipaddr;
1738 pcfg->pcfg_misc = iface->ksni_netmask;
1739 pcfg->pcfg_fd = iface->ksni_npeers;
1740 pcfg->pcfg_count = iface->ksni_nroutes;
1743 read_unlock (&ksocknal_data.ksnd_global_lock);
1746 case NAL_CMD_ADD_INTERFACE: {
1747 rc = ksocknal_add_interface(pcfg->pcfg_id, /* IP address */
1748 pcfg->pcfg_misc); /* net mask */
1751 case NAL_CMD_DEL_INTERFACE: {
1752 rc = ksocknal_del_interface(pcfg->pcfg_id); /* IP address */
1755 case NAL_CMD_GET_PEER: {
1761 int share_count = 0;
1763 rc = ksocknal_get_peer_info(pcfg->pcfg_count, &nid,
1765 &conn_count, &share_count);
1766 pcfg->pcfg_nid = nid;
1767 pcfg->pcfg_size = myip;
1769 pcfg->pcfg_misc = port;
1770 pcfg->pcfg_count = conn_count;
1771 pcfg->pcfg_wait = share_count;
1774 case NAL_CMD_ADD_PEER: {
1775 rc = ksocknal_add_peer (pcfg->pcfg_nid,
1776 pcfg->pcfg_id, /* IP */
1777 pcfg->pcfg_misc); /* port */
1780 case NAL_CMD_DEL_PEER: {
1781 rc = ksocknal_del_peer (pcfg->pcfg_nid,
1782 pcfg->pcfg_id, /* IP */
1783 pcfg->pcfg_flags); /* single_share? */
1786 case NAL_CMD_GET_CONN: {
1787 ksock_conn_t *conn = ksocknal_get_conn_by_idx (pcfg->pcfg_count);
1796 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1799 pcfg->pcfg_nid = conn->ksnc_peer->ksnp_nid;
1800 pcfg->pcfg_id = conn->ksnc_ipaddr;
1801 pcfg->pcfg_misc = conn->ksnc_port;
1802 pcfg->pcfg_fd = conn->ksnc_myipaddr;
1803 pcfg->pcfg_flags = conn->ksnc_type;
1804 pcfg->pcfg_gw_nal = conn->ksnc_scheduler -
1805 ksocknal_data.ksnd_schedulers;
1806 pcfg->pcfg_count = txmem;
1807 pcfg->pcfg_size = rxmem;
1808 pcfg->pcfg_wait = nagle;
1809 ksocknal_put_conn (conn);
1813 case NAL_CMD_REGISTER_PEER_FD: {
1814 struct socket *sock = sockfd_lookup (pcfg->pcfg_fd, &rc);
1815 int type = pcfg->pcfg_misc;
1821 case SOCKNAL_CONN_NONE:
1822 case SOCKNAL_CONN_ANY:
1823 case SOCKNAL_CONN_CONTROL:
1824 case SOCKNAL_CONN_BULK_IN:
1825 case SOCKNAL_CONN_BULK_OUT:
1826 rc = ksocknal_create_conn(NULL, sock, type);
1832 cfs_put_file (KSN_SOCK2FILE(sock));
1835 case NAL_CMD_CLOSE_CONNECTION: {
1836 rc = ksocknal_close_matching_conns (pcfg->pcfg_nid,
1840 case NAL_CMD_REGISTER_MYNID: {
1841 rc = ksocknal_set_mynid (pcfg->pcfg_nid);
1844 case NAL_CMD_PUSH_CONNECTION: {
1845 rc = ksocknal_push (pcfg->pcfg_nid);
1857 ksocknal_free_fmbs (ksock_fmb_pool_t *p)
1859 int npages = p->fmp_buff_pages;
1863 LASSERT (list_empty(&p->fmp_blocked_conns));
1864 LASSERT (p->fmp_nactive_fmbs == 0);
1866 while (!list_empty(&p->fmp_idle_fmbs)) {
1868 fmb = list_entry(p->fmp_idle_fmbs.next,
1869 ksock_fmb_t, fmb_list);
1871 for (i = 0; i < npages; i++)
1872 if (fmb->fmb_kiov[i].kiov_page != NULL)
1873 cfs_free_page(fmb->fmb_kiov[i].kiov_page);
1875 list_del(&fmb->fmb_list);
1876 PORTAL_FREE(fmb, offsetof(ksock_fmb_t, fmb_kiov[npages]));
1881 ksocknal_free_buffers (void)
1883 ksocknal_free_fmbs(&ksocknal_data.ksnd_small_fmp);
1884 ksocknal_free_fmbs(&ksocknal_data.ksnd_large_fmp);
1886 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_ltxs) == 0);
1888 if (ksocknal_data.ksnd_schedulers != NULL)
1889 PORTAL_FREE (ksocknal_data.ksnd_schedulers,
1890 sizeof (ksock_sched_t) * ksocknal_data.ksnd_nschedulers);
1892 PORTAL_FREE (ksocknal_data.ksnd_peers,
1893 sizeof (struct list_head) *
1894 ksocknal_data.ksnd_peer_hash_size);
1898 ksocknal_api_shutdown (nal_t *nal)
1900 ksock_sched_t *sched;
1903 if (nal->nal_refct != 0) {
1904 /* This module got the first ref */
1905 PORTAL_MODULE_UNUSE;
1909 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
1910 atomic_read (&portal_kmemory));
1912 LASSERT(nal == &ksocknal_api);
1914 switch (ksocknal_data.ksnd_init) {
1918 case SOCKNAL_INIT_ALL:
1919 libcfs_nal_cmd_unregister(SOCKNAL);
1921 ksocknal_data.ksnd_init = SOCKNAL_INIT_LIB;
1924 case SOCKNAL_INIT_LIB:
1925 /* No more calls to ksocknal_cmd() to create new
1926 * autoroutes/connections since we're being unloaded. */
1928 /* Delete all peers */
1929 ksocknal_del_peer(PTL_NID_ANY, 0, 0);
1931 /* Wait for all peer state to clean up */
1933 while (atomic_read (&ksocknal_data.ksnd_npeers) != 0) {
1935 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1936 "waiting for %d peers to disconnect\n",
1937 atomic_read (&ksocknal_data.ksnd_npeers));
1938 set_current_state (TASK_UNINTERRUPTIBLE);
1939 schedule_timeout (cfs_time_seconds(1));
1942 /* Tell lib we've stopped calling into her. */
1943 lib_fini(&ksocknal_lib);
1945 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
1948 case SOCKNAL_INIT_DATA:
1949 LASSERT (atomic_read (&ksocknal_data.ksnd_npeers) == 0);
1950 LASSERT (ksocknal_data.ksnd_peers != NULL);
1951 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1952 LASSERT (list_empty (&ksocknal_data.ksnd_peers[i]));
1954 LASSERT (list_empty (&ksocknal_data.ksnd_enomem_conns));
1955 LASSERT (list_empty (&ksocknal_data.ksnd_zombie_conns));
1956 LASSERT (list_empty (&ksocknal_data.ksnd_autoconnectd_routes));
1957 LASSERT (list_empty (&ksocknal_data.ksnd_small_fmp.fmp_blocked_conns));
1958 LASSERT (list_empty (&ksocknal_data.ksnd_large_fmp.fmp_blocked_conns));
1960 if (ksocknal_data.ksnd_schedulers != NULL)
1961 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
1962 ksock_sched_t *kss =
1963 &ksocknal_data.ksnd_schedulers[i];
1965 LASSERT (list_empty (&kss->kss_tx_conns));
1966 LASSERT (list_empty (&kss->kss_rx_conns));
1967 LASSERT (kss->kss_nconns == 0);
1970 /* stop router calling me */
1971 kpr_shutdown (&ksocknal_data.ksnd_router);
1973 /* flag threads to terminate; wake and wait for them to die */
1974 ksocknal_data.ksnd_shuttingdown = 1;
1975 cfs_waitq_broadcast (&ksocknal_data.ksnd_autoconnectd_waitq);
1976 cfs_waitq_broadcast (&ksocknal_data.ksnd_reaper_waitq);
1978 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
1979 sched = &ksocknal_data.ksnd_schedulers[i];
1980 cfs_waitq_broadcast(&sched->kss_waitq);
1984 read_lock(&ksocknal_data.ksnd_global_lock);
1985 while (ksocknal_data.ksnd_nthreads != 0) {
1987 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1988 "waiting for %d threads to terminate\n",
1989 ksocknal_data.ksnd_nthreads);
1990 read_unlock(&ksocknal_data.ksnd_global_lock);
1991 set_current_state (TASK_UNINTERRUPTIBLE);
1992 schedule_timeout (cfs_time_seconds(1));
1993 read_lock(&ksocknal_data.ksnd_global_lock);
1995 read_unlock(&ksocknal_data.ksnd_global_lock);
1997 kpr_deregister (&ksocknal_data.ksnd_router);
1999 ksocknal_free_buffers();
2001 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2004 case SOCKNAL_INIT_NOTHING:
2008 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2009 atomic_read (&portal_kmemory));
2011 printk(KERN_INFO "Lustre: Routing socket NAL unloaded (final mem %d)\n",
2012 atomic_read(&portal_kmemory));
2017 ksocknal_init_incarnation (void)
2021 /* The incarnation number is the time this module loaded and it
2022 * identifies this particular instance of the socknal. Hopefully
2023 * we won't be able to reboot more frequently than 1MHz for the
2024 * forseeable future :) */
2026 do_gettimeofday(&tv);
2028 ksocknal_data.ksnd_incarnation =
2029 (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2033 ksocknal_api_startup (nal_t *nal, ptl_pid_t requested_pid,
2034 ptl_ni_limits_t *requested_limits,
2035 ptl_ni_limits_t *actual_limits)
2037 ptl_process_id_t process_id;
2038 int pkmem = atomic_read(&portal_kmemory);
2043 LASSERT (nal == &ksocknal_api);
2045 if (nal->nal_refct != 0) {
2046 if (actual_limits != NULL)
2047 *actual_limits = ksocknal_lib.libnal_ni.ni_actual_limits;
2048 /* This module got the first ref */
2053 LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2055 memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2057 ksocknal_init_incarnation();
2059 ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2060 PORTAL_ALLOC (ksocknal_data.ksnd_peers,
2061 sizeof (struct list_head) * ksocknal_data.ksnd_peer_hash_size);
2062 if (ksocknal_data.ksnd_peers == NULL)
2065 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2066 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2068 rwlock_init(&ksocknal_data.ksnd_global_lock);
2070 spin_lock_init(&ksocknal_data.ksnd_small_fmp.fmp_lock);
2071 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_small_fmp.fmp_idle_fmbs);
2072 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_small_fmp.fmp_blocked_conns);
2073 ksocknal_data.ksnd_small_fmp.fmp_buff_pages = SOCKNAL_SMALL_FWD_PAGES;
2075 spin_lock_init(&ksocknal_data.ksnd_large_fmp.fmp_lock);
2076 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_large_fmp.fmp_idle_fmbs);
2077 CFS_INIT_LIST_HEAD(&ksocknal_data.ksnd_large_fmp.fmp_blocked_conns);
2078 ksocknal_data.ksnd_large_fmp.fmp_buff_pages = SOCKNAL_LARGE_FWD_PAGES;
2080 spin_lock_init (&ksocknal_data.ksnd_reaper_lock);
2081 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_enomem_conns);
2082 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_zombie_conns);
2083 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_deathrow_conns);
2084 cfs_waitq_init(&ksocknal_data.ksnd_reaper_waitq);
2086 spin_lock_init (&ksocknal_data.ksnd_autoconnectd_lock);
2087 CFS_INIT_LIST_HEAD (&ksocknal_data.ksnd_autoconnectd_routes);
2088 cfs_waitq_init(&ksocknal_data.ksnd_autoconnectd_waitq);
2090 /* NB memset above zeros whole of ksocknal_data, including
2091 * ksocknal_data.ksnd_irqinfo[all].ksni_valid */
2093 /* flag lists/ptrs/locks initialised */
2094 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2096 ksocknal_data.ksnd_nschedulers = ksocknal_nsched();
2097 PORTAL_ALLOC(ksocknal_data.ksnd_schedulers,
2098 sizeof(ksock_sched_t) * ksocknal_data.ksnd_nschedulers);
2099 if (ksocknal_data.ksnd_schedulers == NULL) {
2100 ksocknal_api_shutdown (nal);
2104 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
2105 ksock_sched_t *kss = &ksocknal_data.ksnd_schedulers[i];
2107 spin_lock_init (&kss->kss_lock);
2108 CFS_INIT_LIST_HEAD (&kss->kss_rx_conns);
2109 CFS_INIT_LIST_HEAD (&kss->kss_tx_conns);
2111 CFS_INIT_LIST_HEAD (&kss->kss_zctxdone_list);
2113 cfs_waitq_init (&kss->kss_waitq);
2116 /* NB we have to wait to be told our true NID... */
2117 process_id.pid = requested_pid;
2120 rc = lib_init(&ksocknal_lib, nal, process_id,
2121 requested_limits, actual_limits);
2123 CERROR("lib_init failed: error %d\n", rc);
2124 ksocknal_api_shutdown (nal);
2128 ksocknal_data.ksnd_init = SOCKNAL_INIT_LIB; // flag lib_init() called
2130 for (i = 0; i < ksocknal_data.ksnd_nschedulers; i++) {
2131 rc = ksocknal_thread_start (ksocknal_scheduler,
2132 &ksocknal_data.ksnd_schedulers[i]);
2134 CERROR("Can't spawn socknal scheduler[%d]: %d\n",
2136 ksocknal_api_shutdown (nal);
2141 for (i = 0; i < SOCKNAL_N_AUTOCONNECTD; i++) {
2142 rc = ksocknal_thread_start (ksocknal_autoconnectd, (void *)((long)i));
2144 CERROR("Can't spawn socknal autoconnectd: %d\n", rc);
2145 ksocknal_api_shutdown (nal);
2150 rc = ksocknal_thread_start (ksocknal_reaper, NULL);
2152 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2153 ksocknal_api_shutdown (nal);
2157 rc = kpr_register(&ksocknal_data.ksnd_router,
2158 &ksocknal_router_interface);
2160 CDEBUG(D_NET, "Can't initialise routing interface "
2161 "(rc = %d): not routing\n", rc);
2163 /* Only allocate forwarding buffers if there's a router */
2165 for (i = 0; i < (SOCKNAL_SMALL_FWD_NMSGS +
2166 SOCKNAL_LARGE_FWD_NMSGS); i++) {
2168 ksock_fmb_pool_t *pool;
2171 if (i < SOCKNAL_SMALL_FWD_NMSGS)
2172 pool = &ksocknal_data.ksnd_small_fmp;
2174 pool = &ksocknal_data.ksnd_large_fmp;
2176 PORTAL_ALLOC(fmb, offsetof(ksock_fmb_t,
2177 fmb_kiov[pool->fmp_buff_pages]));
2179 ksocknal_api_shutdown(nal);
2183 fmb->fmb_pool = pool;
2185 for (j = 0; j < pool->fmp_buff_pages; j++) {
2186 fmb->fmb_kiov[j].kiov_page = cfs_alloc_page(CFS_ALLOC_STD);
2188 if (fmb->fmb_kiov[j].kiov_page == NULL) {
2189 ksocknal_api_shutdown (nal);
2193 LASSERT(cfs_page_address(fmb->fmb_kiov[j].kiov_page) != NULL);
2196 list_add(&fmb->fmb_list, &pool->fmp_idle_fmbs);
2200 rc = libcfs_nal_cmd_register(SOCKNAL, &ksocknal_cmd, NULL);
2202 CERROR ("Can't initialise command interface (rc = %d)\n", rc);
2203 ksocknal_api_shutdown (nal);
2207 /* flag everything initialised */
2208 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2210 printk(KERN_INFO "Lustre: Routing socket NAL loaded "
2211 "(Routing %s, initial mem %d, incarnation "LPX64")\n",
2212 kpr_routing (&ksocknal_data.ksnd_router) ?
2213 "enabled" : "disabled", pkmem, ksocknal_data.ksnd_incarnation);
2219 ksocknal_module_fini (void)
2221 #ifdef CONFIG_SYSCTL
2222 if (ksocknal_tunables.ksnd_sysctl != NULL)
2223 unregister_sysctl_table (ksocknal_tunables.ksnd_sysctl);
2225 PtlNIFini(ksocknal_ni);
2227 ptl_unregister_nal(SOCKNAL);
2230 extern cfs_sysctl_table_t ksocknal_top_ctl_table[];
2233 ksocknal_module_init (void)
2237 /* packet descriptor must fit in a router descriptor's scratchpad */
2238 LASSERT(sizeof (ksock_tx_t) <= sizeof (kprfd_scratch_t));
2239 /* the following must be sizeof(int) for proc_dointvec() */
2240 LASSERT(sizeof (ksocknal_tunables.ksnd_io_timeout) == sizeof (int));
2241 LASSERT(sizeof (ksocknal_tunables.ksnd_eager_ack) == sizeof (int));
2242 LASSERT(sizeof (ksocknal_tunables.ksnd_typed_conns) == sizeof (int));
2243 LASSERT(sizeof (ksocknal_tunables.ksnd_min_bulk) == sizeof (int));
2244 LASSERT(sizeof (ksocknal_tunables.ksnd_buffer_size) == sizeof (int));
2245 LASSERT(sizeof (ksocknal_tunables.ksnd_nagle) == sizeof (int));
2246 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_idle) == sizeof (int));
2247 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_count) == sizeof (int));
2248 LASSERT(sizeof (ksocknal_tunables.ksnd_keepalive_intvl) == sizeof (int));
2250 LASSERT(sizeof (ksocknal_tunables.ksnd_irq_affinity) == sizeof (int));
2253 LASSERT(sizeof (ksocknal_tunables.ksnd_zc_min_frag) == sizeof (int));
2255 /* check ksnr_connected/connecting field large enough */
2256 LASSERT(SOCKNAL_CONN_NTYPES <= 4);
2258 ksocknal_api.nal_ni_init = ksocknal_api_startup;
2259 ksocknal_api.nal_ni_fini = ksocknal_api_shutdown;
2261 /* Initialise dynamic tunables to defaults once only */
2262 ksocknal_tunables.ksnd_io_timeout = SOCKNAL_IO_TIMEOUT;
2263 ksocknal_tunables.ksnd_eager_ack = SOCKNAL_EAGER_ACK;
2264 ksocknal_tunables.ksnd_typed_conns = SOCKNAL_TYPED_CONNS;
2265 ksocknal_tunables.ksnd_min_bulk = SOCKNAL_MIN_BULK;
2266 ksocknal_tunables.ksnd_buffer_size = SOCKNAL_BUFFER_SIZE;
2267 ksocknal_tunables.ksnd_nagle = SOCKNAL_NAGLE;
2268 ksocknal_tunables.ksnd_keepalive_idle = SOCKNAL_KEEPALIVE_IDLE;
2269 ksocknal_tunables.ksnd_keepalive_count = SOCKNAL_KEEPALIVE_COUNT;
2270 ksocknal_tunables.ksnd_keepalive_intvl = SOCKNAL_KEEPALIVE_INTVL;
2272 ksocknal_tunables.ksnd_irq_affinity = SOCKNAL_IRQ_AFFINITY;
2275 ksocknal_tunables.ksnd_zc_min_frag = SOCKNAL_ZC_MIN_FRAG;
2278 rc = ptl_register_nal(SOCKNAL, &ksocknal_api);
2280 CERROR("Can't register SOCKNAL: %d\n", rc);
2281 return (-ENOMEM); /* or something... */
2284 /* Pure gateways want the NAL started up at module load time... */
2285 rc = PtlNIInit(SOCKNAL, LUSTRE_SRV_PTL_PID, NULL, NULL, &ksocknal_ni);
2286 if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
2287 ptl_unregister_nal(SOCKNAL);
2291 #ifdef CONFIG_SYSCTL
2292 /* Press on regardless even if registering sysctl doesn't work */
2293 ksocknal_tunables.ksnd_sysctl =
2294 register_sysctl_table (ksocknal_top_ctl_table, 0);
2299 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2300 MODULE_DESCRIPTION("Kernel TCP Socket NAL v1.0.0");
2301 MODULE_LICENSE("GPL");
2303 cfs_module(ksocknal, "1.0.0", ksocknal_module_init, ksocknal_module_fini);