2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2016, Intel Corporation.
6 * This file is part of Lustre, https://wiki.hpdd.intel.com/
8 * Portals is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Portals is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Portals; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_LNET
24 #include <lnet/lib-lnet.h>
26 #define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
27 #define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4)
28 #define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */
29 #define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4)
30 #define LNET_NRB_SMALL_PAGES 1
31 #define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */
32 #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
33 #define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \
36 static char *forwarding = "";
37 module_param(forwarding, charp, 0444);
38 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
40 static int tiny_router_buffers;
41 module_param(tiny_router_buffers, int, 0444);
42 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
43 static int small_router_buffers;
44 module_param(small_router_buffers, int, 0444);
45 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
46 static int large_router_buffers;
47 module_param(large_router_buffers, int, 0444);
48 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
49 static int peer_buffer_credits;
50 module_param(peer_buffer_credits, int, 0444);
51 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
53 static int auto_down = 1;
54 module_param(auto_down, int, 0444);
55 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
58 lnet_peer_buffer_credits(struct lnet_net *net)
60 /* NI option overrides LNet default */
61 if (net->net_tunables.lct_peer_rtr_credits > 0)
62 return net->net_tunables.lct_peer_rtr_credits;
63 if (peer_buffer_credits > 0)
64 return peer_buffer_credits;
66 /* As an approximation, allow this peer the same number of router
67 * buffers as it is allowed outstanding sends */
68 return net->net_tunables.lct_peer_tx_credits;
72 static int lnet_router_checker(void *);
74 static int check_routers_before_use;
75 module_param(check_routers_before_use, int, 0444);
76 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
78 int avoid_asym_router_failure = 1;
79 module_param(avoid_asym_router_failure, int, 0644);
80 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
82 static int dead_router_check_interval = 60;
83 module_param(dead_router_check_interval, int, 0644);
84 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
86 static int live_router_check_interval = 60;
87 module_param(live_router_check_interval, int, 0644);
88 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
90 static int router_ping_timeout = 50;
91 module_param(router_ping_timeout, int, 0644);
92 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
95 lnet_peers_start_down(void)
97 return check_routers_before_use;
101 lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive,
104 if (cfs_time_before(when, lp->lpni_timestamp)) { /* out of date information */
105 CDEBUG(D_NET, "Out of date\n");
110 * This function can be called with different cpt locks being
111 * held. lpni_alive_count modification needs to be properly protected.
112 * Significant reads to lpni_alive_count are also protected with
115 spin_lock(&lp->lpni_lock);
117 lp->lpni_timestamp = when; /* update timestamp */
118 lp->lpni_ping_deadline = 0; /* disable ping timeout */
120 if (lp->lpni_alive_count != 0 && /* got old news */
121 (!lp->lpni_alive) == (!alive)) { /* new date for old news */
122 spin_unlock(&lp->lpni_lock);
123 CDEBUG(D_NET, "Old news\n");
127 /* Flag that notification is outstanding */
129 lp->lpni_alive_count++;
130 lp->lpni_alive = (alive) ? 1 : 0;
132 lp->lpni_notifylnd = notifylnd;
134 lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
136 spin_unlock(&lp->lpni_lock);
138 CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
142 * This function will always be called with lp->lpni_cpt lock held.
145 lnet_ni_notify_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
150 /* Notify only in 1 thread at any time to ensure ordered notification.
151 * NB individual events can be missed; the only guarantee is that you
152 * always get the most recent news */
154 spin_lock(&lp->lpni_lock);
156 if (lp->lpni_notifying || ni == NULL) {
157 spin_unlock(&lp->lpni_lock);
161 lp->lpni_notifying = 1;
164 * lp->lpni_notify needs to be protected because it can be set in
165 * lnet_notify_locked().
167 while (lp->lpni_notify) {
168 alive = lp->lpni_alive;
169 notifylnd = lp->lpni_notifylnd;
171 lp->lpni_notifylnd = 0;
174 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
175 spin_unlock(&lp->lpni_lock);
176 lnet_net_unlock(lp->lpni_cpt);
178 /* A new notification could happen now; I'll handle it
179 * when control returns to me */
181 (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
184 lnet_net_lock(lp->lpni_cpt);
185 spin_lock(&lp->lpni_lock);
189 lp->lpni_notifying = 0;
190 spin_unlock(&lp->lpni_lock);
194 lnet_rtr_addref_locked(struct lnet_peer_ni *lp)
196 LASSERT(atomic_read(&lp->lpni_refcount) > 0);
197 LASSERT(lp->lpni_rtr_refcount >= 0);
199 /* lnet_net_lock must be exclusively locked */
200 lp->lpni_rtr_refcount++;
201 if (lp->lpni_rtr_refcount == 1) {
202 struct list_head *pos;
204 /* a simple insertion sort */
205 list_for_each_prev(pos, &the_lnet.ln_routers) {
206 struct lnet_peer_ni *rtr =
207 list_entry(pos, struct lnet_peer_ni,
210 if (rtr->lpni_nid < lp->lpni_nid)
214 list_add(&lp->lpni_rtr_list, pos);
215 /* addref for the_lnet.ln_routers */
216 lnet_peer_ni_addref_locked(lp);
217 the_lnet.ln_routers_version++;
222 lnet_rtr_decref_locked(struct lnet_peer_ni *lp)
224 LASSERT(atomic_read(&lp->lpni_refcount) > 0);
225 LASSERT(lp->lpni_rtr_refcount > 0);
227 /* lnet_net_lock must be exclusively locked */
228 lp->lpni_rtr_refcount--;
229 if (lp->lpni_rtr_refcount == 0) {
230 LASSERT(list_empty(&lp->lpni_routes));
232 if (lp->lpni_rcd != NULL) {
233 list_add(&lp->lpni_rcd->rcd_list,
234 &the_lnet.ln_rcd_deathrow);
238 list_del(&lp->lpni_rtr_list);
239 /* decref for the_lnet.ln_routers */
240 lnet_peer_ni_decref_locked(lp);
241 the_lnet.ln_routers_version++;
246 lnet_find_rnet_locked(__u32 net)
248 lnet_remotenet_t *rnet;
249 struct list_head *tmp;
250 struct list_head *rn_list;
252 LASSERT(!the_lnet.ln_shutdown);
254 rn_list = lnet_net2rnethash(net);
255 list_for_each(tmp, rn_list) {
256 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
258 if (rnet->lrn_net == net)
264 static void lnet_shuffle_seed(void)
269 struct timespec64 ts;
270 lnet_ni_t *ni = NULL;
275 cfs_get_random_bytes(seed, sizeof(seed));
277 /* Nodes with small feet have little entropy
278 * the NID for this node gives the most entropy in the low bits */
279 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
280 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
282 if (lnd_type != LOLND)
283 seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
287 cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]);
292 /* NB expects LNET_LOCK held */
294 lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
296 unsigned int len = 0;
297 unsigned int offset = 0;
302 list_for_each(e, &rnet->lrn_routes) {
306 /* len+1 positions to add a new entry, also prevents division by 0 */
307 offset = cfs_rand() % (len + 1);
308 list_for_each(e, &rnet->lrn_routes) {
313 list_add(&route->lr_list, e);
314 list_add(&route->lr_gwlist, &route->lr_gateway->lpni_routes);
316 the_lnet.ln_remote_nets_version++;
317 lnet_rtr_addref_locked(route->lr_gateway);
321 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
322 unsigned int priority)
325 lnet_remotenet_t *rnet;
326 lnet_remotenet_t *rnet2;
329 struct lnet_peer_ni *lpni;
333 CDEBUG(D_NET, "Add route: net %s hops %d priority %u gw %s\n",
334 libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
336 if (gateway == LNET_NID_ANY ||
337 LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
338 net == LNET_NIDNET(LNET_NID_ANY) ||
339 LNET_NETTYP(net) == LOLND ||
340 LNET_NIDNET(gateway) == net ||
341 (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
344 if (lnet_islocalnet(net)) /* it's a local network */
347 /* Assume net, route, all new */
348 LIBCFS_ALLOC(route, sizeof(*route));
349 LIBCFS_ALLOC(rnet, sizeof(*rnet));
350 if (route == NULL || rnet == NULL) {
351 CERROR("Out of memory creating route %s %d %s\n",
352 libcfs_net2str(net), hops, libcfs_nid2str(gateway));
354 LIBCFS_FREE(route, sizeof(*route));
356 LIBCFS_FREE(rnet, sizeof(*rnet));
360 INIT_LIST_HEAD(&rnet->lrn_routes);
362 route->lr_hops = hops;
364 route->lr_priority = priority;
366 lnet_net_lock(LNET_LOCK_EX);
368 lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
370 lnet_net_unlock(LNET_LOCK_EX);
372 LIBCFS_FREE(route, sizeof(*route));
373 LIBCFS_FREE(rnet, sizeof(*rnet));
376 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
377 return rc; /* ignore the route entry */
378 CERROR("Error %d creating route %s %d %s\n", rc,
379 libcfs_net2str(net), hops,
380 libcfs_nid2str(gateway));
383 route->lr_gateway = lpni;
384 LASSERT(!the_lnet.ln_shutdown);
386 rnet2 = lnet_find_rnet_locked(net);
389 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
393 /* Search for a duplicate route (it's a NOOP if it is) */
395 list_for_each(e, &rnet2->lrn_routes) {
396 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
398 if (route2->lr_gateway == route->lr_gateway) {
403 /* our lookups must be true */
404 LASSERT(route2->lr_gateway->lpni_nid != gateway);
408 lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */
409 lnet_add_route_to_rnet(rnet2, route);
411 ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL);
412 lnet_net_unlock(LNET_LOCK_EX);
414 /* XXX Assume alive */
415 if (ni->ni_net->net_lnd->lnd_notify != NULL)
416 (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
418 lnet_net_lock(LNET_LOCK_EX);
421 /* -1 for notify or !add_route */
422 lnet_peer_ni_decref_locked(route->lr_gateway);
423 lnet_net_unlock(LNET_LOCK_EX);
429 LIBCFS_FREE(route, sizeof(*route));
433 LIBCFS_FREE(rnet, sizeof(*rnet));
435 /* indicate to startup the router checker if configured */
436 wake_up(&the_lnet.ln_rc_waitq);
442 lnet_check_routes(void)
444 lnet_remotenet_t *rnet;
446 lnet_route_t *route2;
447 struct list_head *e1;
448 struct list_head *e2;
450 struct list_head *rn_list;
453 cpt = lnet_net_lock_current();
455 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
456 rn_list = &the_lnet.ln_remote_nets_hash[i];
457 list_for_each(e1, rn_list) {
458 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
461 list_for_each(e2, &rnet->lrn_routes) {
466 route = list_entry(e2, lnet_route_t,
469 if (route2 == NULL) {
474 if (route->lr_gateway->lpni_net ==
475 route2->lr_gateway->lpni_net)
478 nid1 = route->lr_gateway->lpni_nid;
479 nid2 = route2->lr_gateway->lpni_nid;
482 lnet_net_unlock(cpt);
484 CERROR("Routes to %s via %s and %s not "
487 libcfs_nid2str(nid1),
488 libcfs_nid2str(nid2));
494 lnet_net_unlock(cpt);
499 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
501 struct lnet_peer_ni *gateway;
502 lnet_remotenet_t *rnet;
504 struct list_head *e1;
505 struct list_head *e2;
507 struct list_head *rn_list;
510 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
511 libcfs_net2str(net), libcfs_nid2str(gw_nid));
513 /* NB Caller may specify either all routes via the given gateway
514 * or a specific route entry actual NIDs) */
516 lnet_net_lock(LNET_LOCK_EX);
517 if (net == LNET_NIDNET(LNET_NID_ANY))
518 rn_list = &the_lnet.ln_remote_nets_hash[0];
520 rn_list = lnet_net2rnethash(net);
523 list_for_each(e1, rn_list) {
524 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
526 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
527 net == rnet->lrn_net))
530 list_for_each(e2, &rnet->lrn_routes) {
531 route = list_entry(e2, lnet_route_t, lr_list);
533 gateway = route->lr_gateway;
534 if (!(gw_nid == LNET_NID_ANY ||
535 gw_nid == gateway->lpni_nid))
538 list_del(&route->lr_list);
539 list_del(&route->lr_gwlist);
540 the_lnet.ln_remote_nets_version++;
542 if (list_empty(&rnet->lrn_routes))
543 list_del(&rnet->lrn_list);
547 lnet_rtr_decref_locked(gateway);
548 lnet_peer_ni_decref_locked(gateway);
550 lnet_net_unlock(LNET_LOCK_EX);
552 LIBCFS_FREE(route, sizeof(*route));
555 LIBCFS_FREE(rnet, sizeof(*rnet));
558 lnet_net_lock(LNET_LOCK_EX);
563 if (net == LNET_NIDNET(LNET_NID_ANY) &&
564 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
565 rn_list = &the_lnet.ln_remote_nets_hash[idx];
568 lnet_net_unlock(LNET_LOCK_EX);
574 lnet_destroy_routes (void)
576 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
579 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
581 int i, rc = -ENOENT, j;
583 if (the_lnet.ln_rtrpools == NULL)
586 for (i = 0; i < LNET_NRBPOOLS; i++) {
587 lnet_rtrbufpool_t *rbp;
589 lnet_net_lock(LNET_LOCK_EX);
590 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
594 pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
595 pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
596 pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
597 pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
601 lnet_net_unlock(LNET_LOCK_EX);
604 lnet_net_lock(LNET_LOCK_EX);
605 pool_cfg->pl_routing = the_lnet.ln_routing;
606 lnet_net_unlock(LNET_LOCK_EX);
612 lnet_get_route(int idx, __u32 *net, __u32 *hops,
613 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
615 struct list_head *e1;
616 struct list_head *e2;
617 lnet_remotenet_t *rnet;
621 struct list_head *rn_list;
623 cpt = lnet_net_lock_current();
625 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
626 rn_list = &the_lnet.ln_remote_nets_hash[i];
627 list_for_each(e1, rn_list) {
628 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
630 list_for_each(e2, &rnet->lrn_routes) {
631 route = list_entry(e2, lnet_route_t,
635 *net = rnet->lrn_net;
636 *hops = route->lr_hops;
637 *priority = route->lr_priority;
638 *gateway = route->lr_gateway->lpni_nid;
639 *alive = lnet_is_route_alive(route);
640 lnet_net_unlock(cpt);
647 lnet_net_unlock(cpt);
652 lnet_swap_pinginfo(struct lnet_ping_info *info)
655 struct lnet_ni_status *stat;
657 __swab32s(&info->pi_magic);
658 __swab32s(&info->pi_features);
659 __swab32s(&info->pi_pid);
660 __swab32s(&info->pi_nnis);
661 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
662 stat = &info->pi_ni[i];
663 __swab64s(&stat->ns_nid);
664 __swab32s(&stat->ns_status);
670 * parse router-checker pinginfo, record number of down NIs for remote
671 * networks on that router.
674 lnet_parse_rc_info(lnet_rc_data_t *rcd)
676 struct lnet_ping_info *info = rcd->rcd_pinginfo;
677 struct lnet_peer_ni *gw = rcd->rcd_gateway;
684 * Protect gw->lpni_ping_feats. This can be set from
685 * lnet_notify_locked with different locks being held
687 spin_lock(&gw->lpni_lock);
689 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
690 lnet_swap_pinginfo(info);
692 /* NB always racing with network! */
693 if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
694 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
695 libcfs_nid2str(gw->lpni_nid), info->pi_magic);
696 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
697 spin_unlock(&gw->lpni_lock);
701 gw->lpni_ping_feats = info->pi_features;
702 if ((gw->lpni_ping_feats & LNET_PING_FEAT_MASK) == 0) {
703 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
704 libcfs_nid2str(gw->lpni_nid), gw->lpni_ping_feats);
705 spin_unlock(&gw->lpni_lock);
706 return; /* nothing I can understand */
709 if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) {
710 spin_unlock(&gw->lpni_lock);
711 return; /* can't carry NI status info */
714 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
719 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
724 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
725 struct lnet_ni_status *stat = &info->pi_ni[i];
726 lnet_nid_t nid = stat->ns_nid;
728 if (nid == LNET_NID_ANY) {
729 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
730 libcfs_nid2str(gw->lpni_nid));
731 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
732 spin_unlock(&gw->lpni_lock);
736 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
739 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
744 if (stat->ns_status == LNET_NI_STATUS_UP) {
745 if (LNET_NIDNET(nid) == rte->lr_net) {
752 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
753 libcfs_nid2str(gw->lpni_nid), stat->ns_status);
754 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
755 spin_unlock(&gw->lpni_lock);
759 if (up) { /* ignore downed NIs if NI for dest network is up */
763 /* if @down is zero and this route is single-hop, it means
764 * we can't find NI for target network */
765 if (down == 0 && rte->lr_hops == 1)
768 rte->lr_downis = down;
771 spin_unlock(&gw->lpni_lock);
775 lnet_router_checker_event(lnet_event_t *event)
777 lnet_rc_data_t *rcd = event->md.user_ptr;
778 struct lnet_peer_ni *lp;
780 LASSERT(rcd != NULL);
782 if (event->unlinked) {
783 LNetInvalidateHandle(&rcd->rcd_mdh);
787 LASSERT(event->type == LNET_EVENT_SEND ||
788 event->type == LNET_EVENT_REPLY);
790 lp = rcd->rcd_gateway;
793 /* NB: it's called with holding lnet_res_lock, we have a few
794 * places need to hold both locks at the same time, please take
795 * care of lock ordering */
796 lnet_net_lock(lp->lpni_cpt);
797 if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
798 /* ignore if no longer a router or rcd is replaced */
802 if (event->type == LNET_EVENT_SEND) {
803 lp->lpni_ping_notsent = 0;
804 if (event->status == 0)
808 /* LNET_EVENT_REPLY */
809 /* A successful REPLY means the router is up. If _any_ comms
810 * to the router fail I assume it's down (this will happen if
811 * we ping alive routers to try to detect router death before
812 * apps get burned). */
814 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
815 /* The router checker will wake up very shortly and do the
816 * actual notification.
817 * XXX If 'lp' stops being a router before then, it will still
818 * have the notification pending!!! */
820 if (avoid_asym_router_failure && event->status == 0)
821 lnet_parse_rc_info(rcd);
824 lnet_net_unlock(lp->lpni_cpt);
828 lnet_wait_known_routerstate(void)
830 struct lnet_peer_ni *rtr;
831 struct list_head *entry;
834 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
837 int cpt = lnet_net_lock_current();
840 list_for_each(entry, &the_lnet.ln_routers) {
841 rtr = list_entry(entry, struct lnet_peer_ni,
844 spin_lock(&rtr->lpni_lock);
846 if (rtr->lpni_alive_count == 0) {
848 spin_unlock(&rtr->lpni_lock);
851 spin_unlock(&rtr->lpni_lock);
854 lnet_net_unlock(cpt);
859 set_current_state(TASK_UNINTERRUPTIBLE);
860 schedule_timeout(cfs_time_seconds(1));
865 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
869 if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
870 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
871 if (rte->lr_net == net) {
880 lnet_update_ni_status_locked(void)
882 lnet_ni_t *ni = NULL;
886 LASSERT(the_lnet.ln_routing);
888 timeout = router_ping_timeout +
889 MAX(live_router_check_interval, dead_router_check_interval);
891 now = ktime_get_real_seconds();
892 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
893 if (ni->ni_net->net_lnd->lnd_type == LOLND)
896 if (now < ni->ni_last_alive + timeout)
900 /* re-check with lock */
901 if (now < ni->ni_last_alive + timeout) {
906 LASSERT(ni->ni_status != NULL);
908 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
909 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
910 libcfs_nid2str(ni->ni_nid), timeout);
911 /* NB: so far, this is the only place to set
912 * NI status to "down" */
913 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
920 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
922 LASSERT(list_empty(&rcd->rcd_list));
923 /* detached from network */
924 LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
926 if (rcd->rcd_gateway != NULL) {
927 int cpt = rcd->rcd_gateway->lpni_cpt;
930 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
931 lnet_net_unlock(cpt);
934 if (rcd->rcd_pinginfo != NULL)
935 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
937 LIBCFS_FREE(rcd, sizeof(*rcd));
940 static lnet_rc_data_t *
941 lnet_create_rc_data_locked(struct lnet_peer_ni *gateway)
943 lnet_rc_data_t *rcd = NULL;
944 struct lnet_ping_info *pi;
948 lnet_net_unlock(gateway->lpni_cpt);
950 LIBCFS_ALLOC(rcd, sizeof(*rcd));
954 LNetInvalidateHandle(&rcd->rcd_mdh);
955 INIT_LIST_HEAD(&rcd->rcd_list);
957 LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
961 for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
962 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
963 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
965 rcd->rcd_pinginfo = pi;
967 LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
968 rc = LNetMDBind((lnet_md_t){.start = pi,
970 .length = LNET_PINGINFO_SIZE,
971 .threshold = LNET_MD_THRESH_INF,
972 .options = LNET_MD_TRUNCATE,
973 .eq_handle = the_lnet.ln_rc_eqh},
977 CERROR("Can't bind MD: %d\n", rc);
982 lnet_net_lock(gateway->lpni_cpt);
983 /* router table changed or someone has created rcd for this gateway */
984 if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) {
985 lnet_net_unlock(gateway->lpni_cpt);
989 lnet_peer_ni_addref_locked(gateway);
990 rcd->rcd_gateway = gateway;
991 gateway->lpni_rcd = rcd;
992 gateway->lpni_ping_notsent = 0;
998 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
999 rc = LNetMDUnlink(rcd->rcd_mdh);
1002 lnet_destroy_rc_data(rcd);
1005 lnet_net_lock(gateway->lpni_cpt);
1006 return gateway->lpni_rcd;
1010 lnet_router_check_interval (struct lnet_peer_ni *rtr)
1014 secs = rtr->lpni_alive ? live_router_check_interval :
1015 dead_router_check_interval;
1023 lnet_ping_router_locked (struct lnet_peer_ni *rtr)
1025 lnet_rc_data_t *rcd = NULL;
1026 cfs_time_t now = cfs_time_current();
1030 lnet_peer_ni_addref_locked(rtr);
1032 if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
1033 cfs_time_after(now, rtr->lpni_ping_deadline))
1034 lnet_notify_locked(rtr, 1, 0, now);
1036 /* Run any outstanding notifications */
1037 ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
1038 lnet_ni_notify_locked(ni, rtr);
1040 if (!lnet_isrouter(rtr) ||
1041 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1042 /* router table changed or router checker is shutting down */
1043 lnet_peer_ni_decref_locked(rtr);
1047 rcd = rtr->lpni_rcd != NULL ?
1048 rtr->lpni_rcd : lnet_create_rc_data_locked(rtr);
1053 secs = lnet_router_check_interval(rtr);
1056 "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1057 "alive_count %d lpni_ping_timestamp %lu\n",
1058 libcfs_nid2str(rtr->lpni_nid), secs,
1059 rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1060 rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1062 if (secs != 0 && !rtr->lpni_ping_notsent &&
1063 cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp,
1064 cfs_time_seconds(secs)))) {
1066 lnet_process_id_t id;
1067 lnet_handle_md_t mdh;
1069 id.nid = rtr->lpni_nid;
1070 id.pid = LNET_PID_LUSTRE;
1071 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1073 rtr->lpni_ping_notsent = 1;
1074 rtr->lpni_ping_timestamp = now;
1078 if (rtr->lpni_ping_deadline == 0) {
1079 rtr->lpni_ping_deadline =
1080 cfs_time_shift(router_ping_timeout);
1083 lnet_net_unlock(rtr->lpni_cpt);
1085 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1086 LNET_PROTO_PING_MATCHBITS, 0);
1088 lnet_net_lock(rtr->lpni_cpt);
1090 rtr->lpni_ping_notsent = 0; /* no event pending */
1093 lnet_peer_ni_decref_locked(rtr);
1098 lnet_router_checker_start(void)
1102 struct task_struct *task;
1104 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1106 if (check_routers_before_use &&
1107 dead_router_check_interval <= 0) {
1108 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1109 " set if 'check_routers_before_use' is set"
1114 sema_init(&the_lnet.ln_rc_signal, 0);
1116 rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1118 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1122 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1123 task = kthread_run(lnet_router_checker, NULL, "router_checker");
1126 CERROR("Can't start router checker thread: %d\n", rc);
1127 /* block until event callback signals exit */
1128 down(&the_lnet.ln_rc_signal);
1129 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1131 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1135 if (check_routers_before_use) {
1136 /* Note that a helpful side-effect of pinging all known routers
1137 * at startup is that it makes them drop stale connections they
1138 * may have to a previous instance of me. */
1139 lnet_wait_known_routerstate();
1146 lnet_router_checker_stop (void)
1150 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1153 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1154 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1155 /* wakeup the RC thread if it's sleeping */
1156 wake_up(&the_lnet.ln_rc_waitq);
1158 /* block until event callback signals exit */
1159 down(&the_lnet.ln_rc_signal);
1160 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1162 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1168 lnet_prune_rc_data(int wait_unlink)
1170 lnet_rc_data_t *rcd;
1171 lnet_rc_data_t *tmp;
1172 struct lnet_peer_ni *lp;
1173 struct list_head head;
1176 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1177 list_empty(&the_lnet.ln_rcd_deathrow) &&
1178 list_empty(&the_lnet.ln_rcd_zombie)))
1181 INIT_LIST_HEAD(&head);
1183 lnet_net_lock(LNET_LOCK_EX);
1185 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1186 /* router checker is stopping, prune all */
1187 list_for_each_entry(lp, &the_lnet.ln_routers,
1189 if (lp->lpni_rcd == NULL)
1192 LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1193 list_add(&lp->lpni_rcd->rcd_list,
1194 &the_lnet.ln_rcd_deathrow);
1195 lp->lpni_rcd = NULL;
1199 /* unlink all RCDs on deathrow list */
1200 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1202 if (!list_empty(&head)) {
1203 lnet_net_unlock(LNET_LOCK_EX);
1205 list_for_each_entry(rcd, &head, rcd_list)
1206 LNetMDUnlink(rcd->rcd_mdh);
1208 lnet_net_lock(LNET_LOCK_EX);
1211 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1213 /* release all zombie RCDs */
1214 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1215 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1217 if (LNetHandleIsInvalid(rcd->rcd_mdh))
1218 list_move(&rcd->rcd_list, &head);
1221 wait_unlink = wait_unlink &&
1222 !list_empty(&the_lnet.ln_rcd_zombie);
1224 lnet_net_unlock(LNET_LOCK_EX);
1226 while (!list_empty(&head)) {
1227 rcd = list_entry(head.next,
1228 lnet_rc_data_t, rcd_list);
1229 list_del_init(&rcd->rcd_list);
1230 lnet_destroy_rc_data(rcd);
1237 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1238 "Waiting for rc buffers to unlink\n");
1239 set_current_state(TASK_UNINTERRUPTIBLE);
1240 schedule_timeout(cfs_time_seconds(1) / 4);
1242 lnet_net_lock(LNET_LOCK_EX);
1245 lnet_net_unlock(LNET_LOCK_EX);
1249 * This function is called to check if the RC should block indefinitely.
1250 * It's called from lnet_router_checker() as well as being passed to
1251 * wait_event_interruptible() to avoid the lost wake_up problem.
1253 * When it's called from wait_event_interruptible() it is necessary to
1254 * also not sleep if the rc state is not running to avoid a deadlock
1255 * when the system is shutting down
1258 lnet_router_checker_active(void)
1260 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1263 /* Router Checker thread needs to run when routing is enabled in
1264 * order to call lnet_update_ni_status_locked() */
1265 if (the_lnet.ln_routing)
1268 return !list_empty(&the_lnet.ln_routers) &&
1269 (live_router_check_interval > 0 ||
1270 dead_router_check_interval > 0);
1274 lnet_router_checker(void *arg)
1276 struct lnet_peer_ni *rtr;
1277 struct list_head *entry;
1279 cfs_block_allsigs();
1281 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1286 cpt = lnet_net_lock_current();
1288 version = the_lnet.ln_routers_version;
1290 list_for_each(entry, &the_lnet.ln_routers) {
1291 rtr = list_entry(entry, struct lnet_peer_ni,
1294 cpt2 = rtr->lpni_cpt;
1296 lnet_net_unlock(cpt);
1299 /* the routers list has changed */
1300 if (version != the_lnet.ln_routers_version)
1304 lnet_ping_router_locked(rtr);
1306 /* NB dropped lock */
1307 if (version != the_lnet.ln_routers_version) {
1308 /* the routers list has changed */
1313 if (the_lnet.ln_routing)
1314 lnet_update_ni_status_locked();
1316 lnet_net_unlock(cpt);
1318 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1320 /* Call schedule_timeout() here always adds 1 to load average
1321 * because kernel counts # active tasks as nr_running
1322 * + nr_uninterruptible. */
1323 /* if there are any routes then wakeup every second. If
1324 * there are no routes then sleep indefinitely until woken
1325 * up by a user adding a route */
1326 if (!lnet_router_checker_active())
1327 wait_event_interruptible(the_lnet.ln_rc_waitq,
1328 lnet_router_checker_active());
1330 wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1332 cfs_time_seconds(1));
1335 lnet_prune_rc_data(1); /* wait for UNLINK */
1337 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1338 up(&the_lnet.ln_rc_signal);
1339 /* The unlink event callback will signal final completion */
1344 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1346 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1348 while (--npages >= 0)
1349 __free_page(rb->rb_kiov[npages].kiov_page);
1351 LIBCFS_FREE(rb, sz);
1354 static lnet_rtrbuf_t *
1355 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1357 int npages = rbp->rbp_npages;
1358 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1363 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1369 for (i = 0; i < npages; i++) {
1370 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1371 GFP_KERNEL | __GFP_ZERO);
1374 __free_page(rb->rb_kiov[i].kiov_page);
1376 LIBCFS_FREE(rb, sz);
1380 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1381 rb->rb_kiov[i].kiov_offset = 0;
1382 rb->rb_kiov[i].kiov_page = page;
1389 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1391 int npages = rbp->rbp_npages;
1393 struct list_head tmp;
1395 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1398 INIT_LIST_HEAD(&tmp);
1401 list_splice_init(&rbp->rbp_msgs, &tmp);
1402 lnet_drop_routed_msgs_locked(&tmp, cpt);
1403 list_splice_init(&rbp->rbp_bufs, &tmp);
1404 rbp->rbp_req_nbuffers = 0;
1405 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1406 rbp->rbp_mincredits = 0;
1407 lnet_net_unlock(cpt);
1409 /* Free buffers on the free list. */
1410 while (!list_empty(&tmp)) {
1411 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1412 list_del(&rb->rb_list);
1413 lnet_destroy_rtrbuf(rb, npages);
1418 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1420 struct list_head rb_list;
1423 int num_buffers = 0;
1425 int npages = rbp->rbp_npages;
1428 /* If we are called for less buffers than already in the pool, we
1429 * just lower the req_nbuffers number and excess buffers will be
1430 * thrown away as they are returned to the free list. Credits
1431 * then get adjusted as well.
1432 * If we already have enough buffers allocated to serve the
1433 * increase requested, then we can treat that the same way as we
1434 * do the decrease. */
1435 num_rb = nbufs - rbp->rbp_nbuffers;
1436 if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1437 rbp->rbp_req_nbuffers = nbufs;
1438 lnet_net_unlock(cpt);
1441 /* store the older value of rbp_req_nbuffers and then set it to
1442 * the new request to prevent lnet_return_rx_credits_locked() from
1443 * freeing buffers that we need to keep around */
1444 old_req_nbufs = rbp->rbp_req_nbuffers;
1445 rbp->rbp_req_nbuffers = nbufs;
1446 lnet_net_unlock(cpt);
1448 INIT_LIST_HEAD(&rb_list);
1450 /* allocate the buffers on a local list first. If all buffers are
1451 * allocated successfully then join this list to the rbp buffer
1452 * list. If not then free all allocated buffers. */
1453 while (num_rb-- > 0) {
1454 rb = lnet_new_rtrbuf(rbp, cpt);
1456 CERROR("Failed to allocate %d route bufs of %d pages\n",
1460 rbp->rbp_req_nbuffers = old_req_nbufs;
1461 lnet_net_unlock(cpt);
1466 list_add(&rb->rb_list, &rb_list);
1472 list_splice_tail(&rb_list, &rbp->rbp_bufs);
1473 rbp->rbp_nbuffers += num_buffers;
1474 rbp->rbp_credits += num_buffers;
1475 rbp->rbp_mincredits = rbp->rbp_credits;
1476 /* We need to schedule blocked msg using the newly
1478 while (!list_empty(&rbp->rbp_bufs) &&
1479 !list_empty(&rbp->rbp_msgs))
1480 lnet_schedule_blocked_locked(rbp);
1482 lnet_net_unlock(cpt);
1487 while (!list_empty(&rb_list)) {
1488 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1489 list_del(&rb->rb_list);
1490 lnet_destroy_rtrbuf(rb, npages);
1497 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1499 INIT_LIST_HEAD(&rbp->rbp_msgs);
1500 INIT_LIST_HEAD(&rbp->rbp_bufs);
1502 rbp->rbp_npages = npages;
1503 rbp->rbp_credits = 0;
1504 rbp->rbp_mincredits = 0;
1508 lnet_rtrpools_free(int keep_pools)
1510 lnet_rtrbufpool_t *rtrp;
1513 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1516 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1517 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1518 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1519 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1523 cfs_percpt_free(the_lnet.ln_rtrpools);
1524 the_lnet.ln_rtrpools = NULL;
1529 lnet_nrb_tiny_calculate(void)
1531 int nrbs = LNET_NRB_TINY;
1533 if (tiny_router_buffers < 0) {
1534 LCONSOLE_ERROR_MSG(0x10c,
1535 "tiny_router_buffers=%d invalid when "
1536 "routing enabled\n", tiny_router_buffers);
1540 if (tiny_router_buffers > 0)
1541 nrbs = tiny_router_buffers;
1543 nrbs /= LNET_CPT_NUMBER;
1544 return max(nrbs, LNET_NRB_TINY_MIN);
1548 lnet_nrb_small_calculate(void)
1550 int nrbs = LNET_NRB_SMALL;
1552 if (small_router_buffers < 0) {
1553 LCONSOLE_ERROR_MSG(0x10c,
1554 "small_router_buffers=%d invalid when "
1555 "routing enabled\n", small_router_buffers);
1559 if (small_router_buffers > 0)
1560 nrbs = small_router_buffers;
1562 nrbs /= LNET_CPT_NUMBER;
1563 return max(nrbs, LNET_NRB_SMALL_MIN);
1567 lnet_nrb_large_calculate(void)
1569 int nrbs = LNET_NRB_LARGE;
1571 if (large_router_buffers < 0) {
1572 LCONSOLE_ERROR_MSG(0x10c,
1573 "large_router_buffers=%d invalid when "
1574 "routing enabled\n", large_router_buffers);
1578 if (large_router_buffers > 0)
1579 nrbs = large_router_buffers;
1581 nrbs /= LNET_CPT_NUMBER;
1582 return max(nrbs, LNET_NRB_LARGE_MIN);
1586 lnet_rtrpools_alloc(int im_a_router)
1588 lnet_rtrbufpool_t *rtrp;
1595 if (!strcmp(forwarding, "")) {
1596 /* not set either way */
1599 } else if (!strcmp(forwarding, "disabled")) {
1600 /* explicitly disabled */
1602 } else if (!strcmp(forwarding, "enabled")) {
1603 /* explicitly enabled */
1605 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1606 "'enabled' or 'disabled'\n");
1610 nrb_tiny = lnet_nrb_tiny_calculate();
1614 nrb_small = lnet_nrb_small_calculate();
1618 nrb_large = lnet_nrb_large_calculate();
1622 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1624 sizeof(lnet_rtrbufpool_t));
1625 if (the_lnet.ln_rtrpools == NULL) {
1626 LCONSOLE_ERROR_MSG(0x10c,
1627 "Failed to initialize router buffe pool\n");
1631 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1632 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1633 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1638 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1639 LNET_NRB_SMALL_PAGES);
1640 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1645 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1646 LNET_NRB_LARGE_PAGES);
1647 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1653 lnet_net_lock(LNET_LOCK_EX);
1654 the_lnet.ln_routing = 1;
1655 lnet_net_unlock(LNET_LOCK_EX);
1659 lnet_rtrpools_free(0);
1664 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1669 lnet_rtrbufpool_t *rtrp;
1671 /* If the provided values for each buffer pool are different than the
1672 * configured values, we need to take action. */
1674 tiny_router_buffers = tiny;
1675 nrb = lnet_nrb_tiny_calculate();
1676 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1677 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1684 small_router_buffers = small;
1685 nrb = lnet_nrb_small_calculate();
1686 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1687 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1694 large_router_buffers = large;
1695 nrb = lnet_nrb_large_calculate();
1696 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1697 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1708 lnet_rtrpools_adjust(int tiny, int small, int large)
1710 /* this function doesn't revert the changes if adding new buffers
1711 * failed. It's up to the user space caller to revert the
1714 if (!the_lnet.ln_routing)
1717 return lnet_rtrpools_adjust_helper(tiny, small, large);
1721 lnet_rtrpools_enable(void)
1725 if (the_lnet.ln_routing)
1728 if (the_lnet.ln_rtrpools == NULL)
1729 /* If routing is turned off, and we have never
1730 * initialized the pools before, just call the
1731 * standard buffer pool allocation routine as
1732 * if we are just configuring this for the first
1734 rc = lnet_rtrpools_alloc(1);
1736 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1740 lnet_net_lock(LNET_LOCK_EX);
1741 the_lnet.ln_routing = 1;
1743 the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1744 lnet_net_unlock(LNET_LOCK_EX);
1750 lnet_rtrpools_disable(void)
1752 if (!the_lnet.ln_routing)
1755 lnet_net_lock(LNET_LOCK_EX);
1756 the_lnet.ln_routing = 0;
1757 the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1759 tiny_router_buffers = 0;
1760 small_router_buffers = 0;
1761 large_router_buffers = 0;
1762 lnet_net_unlock(LNET_LOCK_EX);
1763 lnet_rtrpools_free(1);
1767 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1769 struct lnet_peer_ni *lp = NULL;
1770 cfs_time_t now = cfs_time_current();
1771 int cpt = lnet_cpt_of_nid(nid, ni);
1773 LASSERT (!in_interrupt ());
1775 CDEBUG (D_NET, "%s notifying %s: %s\n",
1776 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1777 libcfs_nid2str(nid),
1778 alive ? "up" : "down");
1781 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1782 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1783 libcfs_nid2str(nid), alive ? "birth" : "death",
1784 libcfs_nid2str(ni->ni_nid));
1788 /* can't do predictions... */
1789 if (cfs_time_after(when, now)) {
1790 CWARN("Ignoring prediction from %s of %s %s "
1791 "%ld seconds in the future\n",
1792 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1793 libcfs_nid2str(nid), alive ? "up" : "down",
1794 cfs_duration_sec(cfs_time_sub(when, now)));
1798 if (ni != NULL && !alive && /* LND telling me she's down */
1799 !auto_down) { /* auto-down disabled */
1800 CDEBUG(D_NET, "Auto-down disabled\n");
1806 if (the_lnet.ln_shutdown) {
1807 lnet_net_unlock(cpt);
1811 lp = lnet_find_peer_ni_locked(nid);
1814 lnet_net_unlock(cpt);
1815 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1820 * It is possible for this function to be called for the same peer
1821 * but with different NIs. We want to synchronize the notification
1822 * between the different calls. So we will use the lpni_cpt to
1823 * grab the net lock.
1825 if (lp->lpni_cpt != cpt) {
1826 lnet_net_unlock(cpt);
1831 /* We can't fully trust LND on reporting exact peer last_alive
1832 * if he notifies us about dead peer. For example ksocklnd can
1833 * call us with when == _time_when_the_node_was_booted_ if
1834 * no connections were successfully established */
1835 if (ni != NULL && !alive && when < lp->lpni_last_alive)
1836 when = lp->lpni_last_alive;
1838 lnet_notify_locked(lp, ni == NULL, alive, when);
1841 lnet_ni_notify_locked(ni, lp);
1843 lnet_peer_ni_decref_locked(lp);
1845 lnet_net_unlock(cpt);
1848 EXPORT_SYMBOL(lnet_notify);