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(struct lnet_ni *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;
208 rtr = 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++;
245 struct lnet_remotenet *
246 lnet_find_rnet_locked(__u32 net)
248 struct lnet_remotenet *rnet;
249 struct list_head *tmp;
250 struct list_head *rn_list;
252 LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
254 rn_list = lnet_net2rnethash(net);
255 list_for_each(tmp, rn_list) {
256 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
258 if (rnet->lrn_net == net)
264 static void lnet_shuffle_seed(void)
269 struct timespec64 ts;
270 struct lnet_ni *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(struct lnet_remotenet *rnet, struct lnet_route *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 struct lnet_remotenet *rnet;
326 struct lnet_remotenet *rnet2;
327 struct lnet_route *route;
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_state == LNET_STATE_RUNNING);
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 struct lnet_route *route2;
398 route2 = list_entry(e, struct lnet_route, lr_list);
399 if (route2->lr_gateway == route->lr_gateway) {
404 /* our lookups must be true */
405 LASSERT(route2->lr_gateway->lpni_nid != gateway);
409 lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */
410 lnet_add_route_to_rnet(rnet2, route);
412 ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL);
413 lnet_net_unlock(LNET_LOCK_EX);
415 /* XXX Assume alive */
416 if (ni->ni_net->net_lnd->lnd_notify != NULL)
417 (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
419 lnet_net_lock(LNET_LOCK_EX);
422 /* -1 for notify or !add_route */
423 lnet_peer_ni_decref_locked(route->lr_gateway);
424 lnet_net_unlock(LNET_LOCK_EX);
430 LIBCFS_FREE(route, sizeof(*route));
434 LIBCFS_FREE(rnet, sizeof(*rnet));
436 /* indicate to startup the router checker if configured */
437 wake_up(&the_lnet.ln_rc_waitq);
443 lnet_check_routes(void)
445 struct lnet_remotenet *rnet;
446 struct lnet_route *route;
447 struct lnet_route *route2;
448 struct list_head *e1;
449 struct list_head *e2;
451 struct list_head *rn_list;
454 cpt = lnet_net_lock_current();
456 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
457 rn_list = &the_lnet.ln_remote_nets_hash[i];
458 list_for_each(e1, rn_list) {
459 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
462 list_for_each(e2, &rnet->lrn_routes) {
467 route = list_entry(e2, struct lnet_route,
470 if (route2 == NULL) {
475 if (route->lr_gateway->lpni_net ==
476 route2->lr_gateway->lpni_net)
479 nid1 = route->lr_gateway->lpni_nid;
480 nid2 = route2->lr_gateway->lpni_nid;
483 lnet_net_unlock(cpt);
485 CERROR("Routes to %s via %s and %s not "
488 libcfs_nid2str(nid1),
489 libcfs_nid2str(nid2));
495 lnet_net_unlock(cpt);
500 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
502 struct lnet_peer_ni *gateway;
503 struct lnet_remotenet *rnet;
504 struct lnet_route *route;
505 struct list_head *e1;
506 struct list_head *e2;
508 struct list_head *rn_list;
511 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
512 libcfs_net2str(net), libcfs_nid2str(gw_nid));
514 /* NB Caller may specify either all routes via the given gateway
515 * or a specific route entry actual NIDs) */
517 lnet_net_lock(LNET_LOCK_EX);
518 if (net == LNET_NIDNET(LNET_NID_ANY))
519 rn_list = &the_lnet.ln_remote_nets_hash[0];
521 rn_list = lnet_net2rnethash(net);
524 list_for_each(e1, rn_list) {
525 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
527 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
528 net == rnet->lrn_net))
531 list_for_each(e2, &rnet->lrn_routes) {
532 route = list_entry(e2, struct lnet_route, lr_list);
534 gateway = route->lr_gateway;
535 if (!(gw_nid == LNET_NID_ANY ||
536 gw_nid == gateway->lpni_nid))
539 list_del(&route->lr_list);
540 list_del(&route->lr_gwlist);
541 the_lnet.ln_remote_nets_version++;
543 if (list_empty(&rnet->lrn_routes))
544 list_del(&rnet->lrn_list);
548 lnet_rtr_decref_locked(gateway);
549 lnet_peer_ni_decref_locked(gateway);
551 lnet_net_unlock(LNET_LOCK_EX);
553 LIBCFS_FREE(route, sizeof(*route));
556 LIBCFS_FREE(rnet, sizeof(*rnet));
559 lnet_net_lock(LNET_LOCK_EX);
564 if (net == LNET_NIDNET(LNET_NID_ANY) &&
565 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
566 rn_list = &the_lnet.ln_remote_nets_hash[idx];
569 lnet_net_unlock(LNET_LOCK_EX);
575 lnet_destroy_routes (void)
577 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
580 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
582 int i, rc = -ENOENT, j;
584 if (the_lnet.ln_rtrpools == NULL)
587 for (i = 0; i < LNET_NRBPOOLS; i++) {
588 struct lnet_rtrbufpool *rbp;
590 lnet_net_lock(LNET_LOCK_EX);
591 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
595 pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
596 pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
597 pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
598 pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
602 lnet_net_unlock(LNET_LOCK_EX);
605 lnet_net_lock(LNET_LOCK_EX);
606 pool_cfg->pl_routing = the_lnet.ln_routing;
607 lnet_net_unlock(LNET_LOCK_EX);
613 lnet_get_route(int idx, __u32 *net, __u32 *hops,
614 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
616 struct list_head *e1;
617 struct list_head *e2;
618 struct lnet_remotenet *rnet;
619 struct lnet_route *route;
622 struct list_head *rn_list;
624 cpt = lnet_net_lock_current();
626 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
627 rn_list = &the_lnet.ln_remote_nets_hash[i];
628 list_for_each(e1, rn_list) {
629 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
631 list_for_each(e2, &rnet->lrn_routes) {
632 route = list_entry(e2, struct lnet_route,
636 *net = rnet->lrn_net;
637 *hops = route->lr_hops;
638 *priority = route->lr_priority;
639 *gateway = route->lr_gateway->lpni_nid;
640 *alive = lnet_is_route_alive(route);
641 lnet_net_unlock(cpt);
648 lnet_net_unlock(cpt);
653 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
655 struct lnet_ni_status *stat;
659 __swab32s(&pbuf->pb_info.pi_magic);
660 __swab32s(&pbuf->pb_info.pi_features);
661 __swab32s(&pbuf->pb_info.pi_pid);
662 __swab32s(&pbuf->pb_info.pi_nnis);
663 nnis = pbuf->pb_info.pi_nnis;
664 if (nnis > pbuf->pb_nnis)
665 nnis = pbuf->pb_nnis;
666 for (i = 0; i < nnis; i++) {
667 stat = &pbuf->pb_info.pi_ni[i];
668 __swab64s(&stat->ns_nid);
669 __swab32s(&stat->ns_status);
675 * parse router-checker pinginfo, record number of down NIs for remote
676 * networks on that router.
679 lnet_parse_rc_info(struct lnet_rc_data *rcd)
681 struct lnet_ping_buffer *pbuf = rcd->rcd_pingbuffer;
682 struct lnet_peer_ni *gw = rcd->rcd_gateway;
683 struct lnet_route *rte;
686 if (!gw->lpni_alive || !pbuf)
690 * Protect gw->lpni_ping_feats. This can be set from
691 * lnet_notify_locked with different locks being held
693 spin_lock(&gw->lpni_lock);
695 if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
696 lnet_swap_pinginfo(pbuf);
698 /* NB always racing with network! */
699 if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
700 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
701 libcfs_nid2str(gw->lpni_nid), pbuf->pb_info.pi_magic);
702 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
706 gw->lpni_ping_feats = pbuf->pb_info.pi_features;
708 /* Without NI status info there's nothing more to do. */
709 if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
712 /* Determine the number of NIs for which there is data. */
713 nnis = pbuf->pb_info.pi_nnis;
714 if (pbuf->pb_nnis < nnis) {
715 if (rcd->rcd_nnis < nnis)
716 rcd->rcd_nnis = nnis;
717 nnis = pbuf->pb_nnis;
720 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
725 /* If routing disabled then the route is down. */
726 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
731 for (i = 0; i < nnis; i++) {
732 struct lnet_ni_status *stat = &pbuf->pb_info.pi_ni[i];
733 lnet_nid_t nid = stat->ns_nid;
735 if (nid == LNET_NID_ANY) {
736 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
737 libcfs_nid2str(gw->lpni_nid));
738 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
742 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
745 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
750 if (stat->ns_status == LNET_NI_STATUS_UP) {
751 if (LNET_NIDNET(nid) == rte->lr_net) {
758 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
759 libcfs_nid2str(gw->lpni_nid), stat->ns_status);
760 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
764 if (up) { /* ignore downed NIs if NI for dest network is up */
768 /* if @down is zero and this route is single-hop, it means
769 * we can't find NI for target network */
770 if (down == 0 && rte->lr_hops == 1)
773 rte->lr_downis = down;
776 spin_unlock(&gw->lpni_lock);
780 lnet_router_checker_event(struct lnet_event *event)
782 struct lnet_rc_data *rcd = event->md.user_ptr;
783 struct lnet_peer_ni *lp;
785 LASSERT(rcd != NULL);
787 if (event->unlinked) {
788 LNetInvalidateMDHandle(&rcd->rcd_mdh);
792 LASSERT(event->type == LNET_EVENT_SEND ||
793 event->type == LNET_EVENT_REPLY);
795 lp = rcd->rcd_gateway;
798 /* NB: it's called with holding lnet_res_lock, we have a few
799 * places need to hold both locks at the same time, please take
800 * care of lock ordering */
801 lnet_net_lock(lp->lpni_cpt);
802 if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
803 /* ignore if no longer a router or rcd is replaced */
807 if (event->type == LNET_EVENT_SEND) {
808 lp->lpni_ping_notsent = 0;
809 if (event->status == 0)
813 /* LNET_EVENT_REPLY */
814 /* A successful REPLY means the router is up. If _any_ comms
815 * to the router fail I assume it's down (this will happen if
816 * we ping alive routers to try to detect router death before
817 * apps get burned). */
819 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
820 /* The router checker will wake up very shortly and do the
821 * actual notification.
822 * XXX If 'lp' stops being a router before then, it will still
823 * have the notification pending!!! */
825 if (avoid_asym_router_failure && event->status == 0)
826 lnet_parse_rc_info(rcd);
829 lnet_net_unlock(lp->lpni_cpt);
833 lnet_wait_known_routerstate(void)
835 struct lnet_peer_ni *rtr;
836 struct list_head *entry;
839 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
842 int cpt = lnet_net_lock_current();
845 list_for_each(entry, &the_lnet.ln_routers) {
846 rtr = list_entry(entry, struct lnet_peer_ni,
849 spin_lock(&rtr->lpni_lock);
851 if (rtr->lpni_alive_count == 0) {
853 spin_unlock(&rtr->lpni_lock);
856 spin_unlock(&rtr->lpni_lock);
859 lnet_net_unlock(cpt);
864 set_current_state(TASK_UNINTERRUPTIBLE);
865 schedule_timeout(cfs_time_seconds(1));
870 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
872 struct lnet_route *rte;
874 if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
875 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
876 if (rte->lr_net == net) {
885 lnet_update_ni_status_locked(void)
887 struct lnet_ni *ni = NULL;
891 LASSERT(the_lnet.ln_routing);
893 timeout = router_ping_timeout +
894 MAX(live_router_check_interval, dead_router_check_interval);
896 now = ktime_get_real_seconds();
897 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
898 if (ni->ni_net->net_lnd->lnd_type == LOLND)
901 if (now < ni->ni_last_alive + timeout)
905 /* re-check with lock */
906 if (now < ni->ni_last_alive + timeout) {
911 LASSERT(ni->ni_status != NULL);
913 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
914 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
915 libcfs_nid2str(ni->ni_nid), timeout);
916 /* NB: so far, this is the only place to set
917 * NI status to "down" */
918 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
925 lnet_destroy_rc_data(struct lnet_rc_data *rcd)
927 LASSERT(list_empty(&rcd->rcd_list));
928 /* detached from network */
929 LASSERT(LNetMDHandleIsInvalid(rcd->rcd_mdh));
931 if (rcd->rcd_gateway != NULL) {
932 int cpt = rcd->rcd_gateway->lpni_cpt;
935 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
936 lnet_net_unlock(cpt);
939 if (rcd->rcd_pingbuffer != NULL)
940 lnet_ping_buffer_decref(rcd->rcd_pingbuffer);
942 LIBCFS_FREE(rcd, sizeof(*rcd));
945 static lnet_rc_data_t *
946 lnet_update_rc_data_locked(struct lnet_peer_ni *gateway)
948 struct lnet_handle_md mdh;
949 struct lnet_rc_data *rcd;
950 struct lnet_ping_buffer *pbuf = NULL;
951 int nnis = LNET_INTERFACES_MIN;
955 rcd = gateway->lpni_rcd;
957 nnis = rcd->rcd_nnis;
959 LNetInvalidateMDHandle(&rcd->rcd_mdh);
960 pbuf = rcd->rcd_pingbuffer;
961 rcd->rcd_pingbuffer = NULL;
963 LNetInvalidateMDHandle(&mdh);
966 lnet_net_unlock(gateway->lpni_cpt);
970 lnet_ping_buffer_decref(pbuf);
972 LIBCFS_ALLOC(rcd, sizeof(*rcd));
976 LNetInvalidateMDHandle(&rcd->rcd_mdh);
977 INIT_LIST_HEAD(&rcd->rcd_list);
978 rcd->rcd_nnis = nnis;
981 pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
985 for (i = 0; i < nnis; i++) {
986 pbuf->pb_info.pi_ni[i].ns_nid = LNET_NID_ANY;
987 pbuf->pb_info.pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
989 rcd->rcd_pingbuffer = pbuf;
991 LASSERT(!LNetEQHandleIsInvalid(the_lnet.ln_rc_eqh));
992 rc = LNetMDBind((struct lnet_md){.start = &pbuf->pb_info,
994 .length = LNET_PING_INFO_SIZE(nnis),
995 .threshold = LNET_MD_THRESH_INF,
996 .options = LNET_MD_TRUNCATE,
997 .eq_handle = the_lnet.ln_rc_eqh},
1001 CERROR("Can't bind MD: %d\n", rc);
1002 goto out_ping_buffer_decref;
1006 lnet_net_lock(gateway->lpni_cpt);
1007 /* Check if this is still a router. */
1008 if (!lnet_isrouter(gateway))
1010 /* Check if someone else installed router data. */
1011 if (gateway->lpni_rcd && gateway->lpni_rcd != rcd)
1014 /* Install and/or update the router data. */
1015 if (!gateway->lpni_rcd) {
1016 lnet_peer_ni_addref_locked(gateway);
1017 rcd->rcd_gateway = gateway;
1018 gateway->lpni_rcd = rcd;
1020 gateway->lpni_ping_notsent = 0;
1025 lnet_net_unlock(gateway->lpni_cpt);
1026 rc = LNetMDUnlink(mdh);
1028 out_ping_buffer_decref:
1029 lnet_ping_buffer_decref(pbuf);
1031 if (rcd && rcd != gateway->lpni_rcd)
1032 lnet_destroy_rc_data(rcd);
1033 lnet_net_lock(gateway->lpni_cpt);
1034 return gateway->lpni_rcd;
1038 lnet_router_check_interval(struct lnet_peer_ni *rtr)
1042 secs = rtr->lpni_alive ? live_router_check_interval :
1043 dead_router_check_interval;
1051 lnet_ping_router_locked(struct lnet_peer_ni *rtr)
1053 struct lnet_rc_data *rcd = NULL;
1054 cfs_time_t now = cfs_time_current();
1058 lnet_peer_ni_addref_locked(rtr);
1060 if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
1061 cfs_time_after(now, rtr->lpni_ping_deadline))
1062 lnet_notify_locked(rtr, 1, 0, now);
1064 /* Run any outstanding notifications */
1065 ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
1066 lnet_ni_notify_locked(ni, rtr);
1068 if (!lnet_isrouter(rtr) ||
1069 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1070 /* router table changed or router checker is shutting down */
1071 lnet_peer_ni_decref_locked(rtr);
1075 rcd = rtr->lpni_rcd;
1076 if (!rcd || rcd->rcd_nnis > rcd->rcd_pingbuffer->pb_nnis)
1077 rcd = lnet_update_rc_data_locked(rtr);
1081 secs = lnet_router_check_interval(rtr);
1084 "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1085 "alive_count %d lpni_ping_timestamp %lu\n",
1086 libcfs_nid2str(rtr->lpni_nid), secs,
1087 rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1088 rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1090 if (secs != 0 && !rtr->lpni_ping_notsent &&
1091 cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp,
1092 cfs_time_seconds(secs)))) {
1094 struct lnet_process_id id;
1095 struct lnet_handle_md mdh;
1097 id.nid = rtr->lpni_nid;
1098 id.pid = LNET_PID_LUSTRE;
1099 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1101 rtr->lpni_ping_notsent = 1;
1102 rtr->lpni_ping_timestamp = now;
1106 if (rtr->lpni_ping_deadline == 0) {
1107 rtr->lpni_ping_deadline =
1108 cfs_time_shift(router_ping_timeout);
1111 lnet_net_unlock(rtr->lpni_cpt);
1113 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1114 LNET_PROTO_PING_MATCHBITS, 0);
1116 lnet_net_lock(rtr->lpni_cpt);
1118 rtr->lpni_ping_notsent = 0; /* no event pending */
1121 lnet_peer_ni_decref_locked(rtr);
1126 lnet_router_checker_start(void)
1130 struct task_struct *task;
1132 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1134 if (check_routers_before_use &&
1135 dead_router_check_interval <= 0) {
1136 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1137 " set if 'check_routers_before_use' is set"
1142 sema_init(&the_lnet.ln_rc_signal, 0);
1144 rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1146 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1150 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1151 task = kthread_run(lnet_router_checker, NULL, "router_checker");
1154 CERROR("Can't start router checker thread: %d\n", rc);
1155 /* block until event callback signals exit */
1156 down(&the_lnet.ln_rc_signal);
1157 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1159 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1163 if (check_routers_before_use) {
1164 /* Note that a helpful side-effect of pinging all known routers
1165 * at startup is that it makes them drop stale connections they
1166 * may have to a previous instance of me. */
1167 lnet_wait_known_routerstate();
1174 lnet_router_checker_stop (void)
1178 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1181 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1182 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1183 /* wakeup the RC thread if it's sleeping */
1184 wake_up(&the_lnet.ln_rc_waitq);
1186 /* block until event callback signals exit */
1187 down(&the_lnet.ln_rc_signal);
1188 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1190 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1196 lnet_prune_rc_data(int wait_unlink)
1198 struct lnet_rc_data *rcd;
1199 struct lnet_rc_data *tmp;
1200 struct lnet_peer_ni *lp;
1201 struct list_head head;
1204 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1205 list_empty(&the_lnet.ln_rcd_deathrow) &&
1206 list_empty(&the_lnet.ln_rcd_zombie)))
1209 INIT_LIST_HEAD(&head);
1211 lnet_net_lock(LNET_LOCK_EX);
1213 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1214 /* router checker is stopping, prune all */
1215 list_for_each_entry(lp, &the_lnet.ln_routers,
1217 if (lp->lpni_rcd == NULL)
1220 LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1221 list_add(&lp->lpni_rcd->rcd_list,
1222 &the_lnet.ln_rcd_deathrow);
1223 lp->lpni_rcd = NULL;
1227 /* unlink all RCDs on deathrow list */
1228 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1230 if (!list_empty(&head)) {
1231 lnet_net_unlock(LNET_LOCK_EX);
1233 list_for_each_entry(rcd, &head, rcd_list)
1234 LNetMDUnlink(rcd->rcd_mdh);
1236 lnet_net_lock(LNET_LOCK_EX);
1239 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1241 /* release all zombie RCDs */
1242 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1243 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1245 if (LNetMDHandleIsInvalid(rcd->rcd_mdh))
1246 list_move(&rcd->rcd_list, &head);
1249 wait_unlink = wait_unlink &&
1250 !list_empty(&the_lnet.ln_rcd_zombie);
1252 lnet_net_unlock(LNET_LOCK_EX);
1254 while (!list_empty(&head)) {
1255 rcd = list_entry(head.next,
1256 struct lnet_rc_data, rcd_list);
1257 list_del_init(&rcd->rcd_list);
1258 lnet_destroy_rc_data(rcd);
1265 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1266 "Waiting for rc buffers to unlink\n");
1267 set_current_state(TASK_UNINTERRUPTIBLE);
1268 schedule_timeout(cfs_time_seconds(1) / 4);
1270 lnet_net_lock(LNET_LOCK_EX);
1273 lnet_net_unlock(LNET_LOCK_EX);
1277 * This function is called to check if the RC should block indefinitely.
1278 * It's called from lnet_router_checker() as well as being passed to
1279 * wait_event_interruptible() to avoid the lost wake_up problem.
1281 * When it's called from wait_event_interruptible() it is necessary to
1282 * also not sleep if the rc state is not running to avoid a deadlock
1283 * when the system is shutting down
1286 lnet_router_checker_active(void)
1288 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1291 /* Router Checker thread needs to run when routing is enabled in
1292 * order to call lnet_update_ni_status_locked() */
1293 if (the_lnet.ln_routing)
1296 return !list_empty(&the_lnet.ln_routers) &&
1297 (live_router_check_interval > 0 ||
1298 dead_router_check_interval > 0);
1302 lnet_router_checker(void *arg)
1304 struct lnet_peer_ni *rtr;
1305 struct list_head *entry;
1307 cfs_block_allsigs();
1309 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1314 cpt = lnet_net_lock_current();
1316 version = the_lnet.ln_routers_version;
1318 list_for_each(entry, &the_lnet.ln_routers) {
1319 rtr = list_entry(entry, struct lnet_peer_ni,
1322 cpt2 = rtr->lpni_cpt;
1324 lnet_net_unlock(cpt);
1327 /* the routers list has changed */
1328 if (version != the_lnet.ln_routers_version)
1332 lnet_ping_router_locked(rtr);
1334 /* NB dropped lock */
1335 if (version != the_lnet.ln_routers_version) {
1336 /* the routers list has changed */
1341 if (the_lnet.ln_routing)
1342 lnet_update_ni_status_locked();
1344 lnet_net_unlock(cpt);
1346 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1348 /* Call schedule_timeout() here always adds 1 to load average
1349 * because kernel counts # active tasks as nr_running
1350 * + nr_uninterruptible. */
1351 /* if there are any routes then wakeup every second. If
1352 * there are no routes then sleep indefinitely until woken
1353 * up by a user adding a route */
1354 if (!lnet_router_checker_active())
1355 wait_event_interruptible(the_lnet.ln_rc_waitq,
1356 lnet_router_checker_active());
1358 wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1360 cfs_time_seconds(1));
1363 lnet_prune_rc_data(1); /* wait for UNLINK */
1365 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1366 up(&the_lnet.ln_rc_signal);
1367 /* The unlink event callback will signal final completion */
1372 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1374 int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1376 while (--npages >= 0)
1377 __free_page(rb->rb_kiov[npages].kiov_page);
1379 LIBCFS_FREE(rb, sz);
1382 static struct lnet_rtrbuf *
1383 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1385 int npages = rbp->rbp_npages;
1386 int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1388 struct lnet_rtrbuf *rb;
1391 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1397 for (i = 0; i < npages; i++) {
1398 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1399 GFP_KERNEL | __GFP_ZERO);
1402 __free_page(rb->rb_kiov[i].kiov_page);
1404 LIBCFS_FREE(rb, sz);
1408 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1409 rb->rb_kiov[i].kiov_offset = 0;
1410 rb->rb_kiov[i].kiov_page = page;
1417 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1419 int npages = rbp->rbp_npages;
1420 struct lnet_rtrbuf *rb;
1421 struct list_head tmp;
1423 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1426 INIT_LIST_HEAD(&tmp);
1429 list_splice_init(&rbp->rbp_msgs, &tmp);
1430 lnet_drop_routed_msgs_locked(&tmp, cpt);
1431 list_splice_init(&rbp->rbp_bufs, &tmp);
1432 rbp->rbp_req_nbuffers = 0;
1433 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1434 rbp->rbp_mincredits = 0;
1435 lnet_net_unlock(cpt);
1437 /* Free buffers on the free list. */
1438 while (!list_empty(&tmp)) {
1439 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1440 list_del(&rb->rb_list);
1441 lnet_destroy_rtrbuf(rb, npages);
1446 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1448 struct list_head rb_list;
1449 struct lnet_rtrbuf *rb;
1451 int num_buffers = 0;
1453 int npages = rbp->rbp_npages;
1456 /* If we are called for less buffers than already in the pool, we
1457 * just lower the req_nbuffers number and excess buffers will be
1458 * thrown away as they are returned to the free list. Credits
1459 * then get adjusted as well.
1460 * If we already have enough buffers allocated to serve the
1461 * increase requested, then we can treat that the same way as we
1462 * do the decrease. */
1463 num_rb = nbufs - rbp->rbp_nbuffers;
1464 if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1465 rbp->rbp_req_nbuffers = nbufs;
1466 lnet_net_unlock(cpt);
1469 /* store the older value of rbp_req_nbuffers and then set it to
1470 * the new request to prevent lnet_return_rx_credits_locked() from
1471 * freeing buffers that we need to keep around */
1472 old_req_nbufs = rbp->rbp_req_nbuffers;
1473 rbp->rbp_req_nbuffers = nbufs;
1474 lnet_net_unlock(cpt);
1476 INIT_LIST_HEAD(&rb_list);
1478 /* allocate the buffers on a local list first. If all buffers are
1479 * allocated successfully then join this list to the rbp buffer
1480 * list. If not then free all allocated buffers. */
1481 while (num_rb-- > 0) {
1482 rb = lnet_new_rtrbuf(rbp, cpt);
1484 CERROR("Failed to allocate %d route bufs of %d pages\n",
1488 rbp->rbp_req_nbuffers = old_req_nbufs;
1489 lnet_net_unlock(cpt);
1494 list_add(&rb->rb_list, &rb_list);
1500 list_splice_tail(&rb_list, &rbp->rbp_bufs);
1501 rbp->rbp_nbuffers += num_buffers;
1502 rbp->rbp_credits += num_buffers;
1503 rbp->rbp_mincredits = rbp->rbp_credits;
1504 /* We need to schedule blocked msg using the newly
1506 while (!list_empty(&rbp->rbp_bufs) &&
1507 !list_empty(&rbp->rbp_msgs))
1508 lnet_schedule_blocked_locked(rbp);
1510 lnet_net_unlock(cpt);
1515 while (!list_empty(&rb_list)) {
1516 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1517 list_del(&rb->rb_list);
1518 lnet_destroy_rtrbuf(rb, npages);
1525 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1527 INIT_LIST_HEAD(&rbp->rbp_msgs);
1528 INIT_LIST_HEAD(&rbp->rbp_bufs);
1530 rbp->rbp_npages = npages;
1531 rbp->rbp_credits = 0;
1532 rbp->rbp_mincredits = 0;
1536 lnet_rtrpools_free(int keep_pools)
1538 struct lnet_rtrbufpool *rtrp;
1541 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1544 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1545 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1546 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1547 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1551 cfs_percpt_free(the_lnet.ln_rtrpools);
1552 the_lnet.ln_rtrpools = NULL;
1557 lnet_nrb_tiny_calculate(void)
1559 int nrbs = LNET_NRB_TINY;
1561 if (tiny_router_buffers < 0) {
1562 LCONSOLE_ERROR_MSG(0x10c,
1563 "tiny_router_buffers=%d invalid when "
1564 "routing enabled\n", tiny_router_buffers);
1568 if (tiny_router_buffers > 0)
1569 nrbs = tiny_router_buffers;
1571 nrbs /= LNET_CPT_NUMBER;
1572 return max(nrbs, LNET_NRB_TINY_MIN);
1576 lnet_nrb_small_calculate(void)
1578 int nrbs = LNET_NRB_SMALL;
1580 if (small_router_buffers < 0) {
1581 LCONSOLE_ERROR_MSG(0x10c,
1582 "small_router_buffers=%d invalid when "
1583 "routing enabled\n", small_router_buffers);
1587 if (small_router_buffers > 0)
1588 nrbs = small_router_buffers;
1590 nrbs /= LNET_CPT_NUMBER;
1591 return max(nrbs, LNET_NRB_SMALL_MIN);
1595 lnet_nrb_large_calculate(void)
1597 int nrbs = LNET_NRB_LARGE;
1599 if (large_router_buffers < 0) {
1600 LCONSOLE_ERROR_MSG(0x10c,
1601 "large_router_buffers=%d invalid when "
1602 "routing enabled\n", large_router_buffers);
1606 if (large_router_buffers > 0)
1607 nrbs = large_router_buffers;
1609 nrbs /= LNET_CPT_NUMBER;
1610 return max(nrbs, LNET_NRB_LARGE_MIN);
1614 lnet_rtrpools_alloc(int im_a_router)
1616 struct lnet_rtrbufpool *rtrp;
1623 if (!strcmp(forwarding, "")) {
1624 /* not set either way */
1627 } else if (!strcmp(forwarding, "disabled")) {
1628 /* explicitly disabled */
1630 } else if (!strcmp(forwarding, "enabled")) {
1631 /* explicitly enabled */
1633 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1634 "'enabled' or 'disabled'\n");
1638 nrb_tiny = lnet_nrb_tiny_calculate();
1642 nrb_small = lnet_nrb_small_calculate();
1646 nrb_large = lnet_nrb_large_calculate();
1650 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1652 sizeof(struct lnet_rtrbufpool));
1653 if (the_lnet.ln_rtrpools == NULL) {
1654 LCONSOLE_ERROR_MSG(0x10c,
1655 "Failed to initialize router buffe pool\n");
1659 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1660 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1661 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1666 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1667 LNET_NRB_SMALL_PAGES);
1668 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1673 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1674 LNET_NRB_LARGE_PAGES);
1675 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1681 lnet_net_lock(LNET_LOCK_EX);
1682 the_lnet.ln_routing = 1;
1683 lnet_net_unlock(LNET_LOCK_EX);
1687 lnet_rtrpools_free(0);
1692 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1697 struct lnet_rtrbufpool *rtrp;
1699 /* If the provided values for each buffer pool are different than the
1700 * configured values, we need to take action. */
1702 tiny_router_buffers = tiny;
1703 nrb = lnet_nrb_tiny_calculate();
1704 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1705 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1712 small_router_buffers = small;
1713 nrb = lnet_nrb_small_calculate();
1714 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1715 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1722 large_router_buffers = large;
1723 nrb = lnet_nrb_large_calculate();
1724 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1725 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1736 lnet_rtrpools_adjust(int tiny, int small, int large)
1738 /* this function doesn't revert the changes if adding new buffers
1739 * failed. It's up to the user space caller to revert the
1742 if (!the_lnet.ln_routing)
1745 return lnet_rtrpools_adjust_helper(tiny, small, large);
1749 lnet_rtrpools_enable(void)
1753 if (the_lnet.ln_routing)
1756 if (the_lnet.ln_rtrpools == NULL)
1757 /* If routing is turned off, and we have never
1758 * initialized the pools before, just call the
1759 * standard buffer pool allocation routine as
1760 * if we are just configuring this for the first
1762 rc = lnet_rtrpools_alloc(1);
1764 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1768 lnet_net_lock(LNET_LOCK_EX);
1769 the_lnet.ln_routing = 1;
1771 the_lnet.ln_ping_target->pb_info.pi_features &=
1772 ~LNET_PING_FEAT_RTE_DISABLED;
1773 lnet_net_unlock(LNET_LOCK_EX);
1779 lnet_rtrpools_disable(void)
1781 if (!the_lnet.ln_routing)
1784 lnet_net_lock(LNET_LOCK_EX);
1785 the_lnet.ln_routing = 0;
1786 the_lnet.ln_ping_target->pb_info.pi_features |=
1787 LNET_PING_FEAT_RTE_DISABLED;
1789 tiny_router_buffers = 0;
1790 small_router_buffers = 0;
1791 large_router_buffers = 0;
1792 lnet_net_unlock(LNET_LOCK_EX);
1793 lnet_rtrpools_free(1);
1797 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1799 struct lnet_peer_ni *lp = NULL;
1800 cfs_time_t now = cfs_time_current();
1801 int cpt = lnet_cpt_of_nid(nid, ni);
1803 LASSERT (!in_interrupt ());
1805 CDEBUG (D_NET, "%s notifying %s: %s\n",
1806 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1807 libcfs_nid2str(nid),
1808 alive ? "up" : "down");
1811 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1812 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1813 libcfs_nid2str(nid), alive ? "birth" : "death",
1814 libcfs_nid2str(ni->ni_nid));
1818 /* can't do predictions... */
1819 if (cfs_time_after(when, now)) {
1820 CWARN("Ignoring prediction from %s of %s %s "
1821 "%ld seconds in the future\n",
1822 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1823 libcfs_nid2str(nid), alive ? "up" : "down",
1824 cfs_duration_sec(cfs_time_sub(when, now)));
1828 if (ni != NULL && !alive && /* LND telling me she's down */
1829 !auto_down) { /* auto-down disabled */
1830 CDEBUG(D_NET, "Auto-down disabled\n");
1836 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1837 lnet_net_unlock(cpt);
1841 lp = lnet_find_peer_ni_locked(nid);
1844 lnet_net_unlock(cpt);
1845 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1850 * It is possible for this function to be called for the same peer
1851 * but with different NIs. We want to synchronize the notification
1852 * between the different calls. So we will use the lpni_cpt to
1853 * grab the net lock.
1855 if (lp->lpni_cpt != cpt) {
1856 lnet_net_unlock(cpt);
1861 /* We can't fully trust LND on reporting exact peer last_alive
1862 * if he notifies us about dead peer. For example ksocklnd can
1863 * call us with when == _time_when_the_node_was_booted_ if
1864 * no connections were successfully established */
1865 if (ni != NULL && !alive && when < lp->lpni_last_alive)
1866 when = lp->lpni_last_alive;
1868 lnet_notify_locked(lp, ni == NULL, alive, when);
1871 lnet_ni_notify_locked(ni, lp);
1873 lnet_peer_ni_decref_locked(lp);
1875 lnet_net_unlock(cpt);
1878 EXPORT_SYMBOL(lnet_notify);