2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2013, Intel Corporation.
6 * This file is part of Portals
7 * http://sourceforge.net/projects/sandiaportals/
9 * Portals is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Portals is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Portals; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LNET
25 #include <lnet/lib-lnet.h>
27 #if defined(__KERNEL__) && defined(LNET_ROUTER)
29 #define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
30 #define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4)
31 #define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */
32 #define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4)
33 #define LNET_NRB_SMALL_PAGES 1
34 #define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */
35 #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
36 #define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \
39 static char *forwarding = "";
40 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
41 "Explicitly enable/disable forwarding between networks");
43 static int tiny_router_buffers;
44 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
45 "# of 0 payload messages to buffer in the router");
46 static int small_router_buffers;
47 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
48 "# of small (1 page) messages to buffer in the router");
49 static int large_router_buffers;
50 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
51 "# of large messages to buffer in the router");
52 static int peer_buffer_credits = 0;
53 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
54 "# router buffer credits per peer");
56 static int auto_down = 1;
57 CFS_MODULE_PARM(auto_down, "i", int, 0444,
58 "Automatically mark peers down on comms error");
61 lnet_peer_buffer_credits(lnet_ni_t *ni)
63 /* NI option overrides LNet default */
64 if (ni->ni_peerrtrcredits > 0)
65 return ni->ni_peerrtrcredits;
66 if (peer_buffer_credits > 0)
67 return peer_buffer_credits;
69 /* As an approximation, allow this peer the same number of router
70 * buffers as it is allowed outstanding sends */
71 return ni->ni_peertxcredits;
75 static int lnet_router_checker(void *);
79 lnet_peer_buffer_credits(lnet_ni_t *ni)
86 static int check_routers_before_use = 0;
87 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
88 "Assume routers are down and ping them before use");
90 static int avoid_asym_router_failure = 1;
91 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
92 "Avoid asymmetrical router failures (0 to disable)");
94 static int dead_router_check_interval = 60;
95 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
96 "Seconds between dead router health checks (<= 0 to disable)");
98 static int live_router_check_interval = 60;
99 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
100 "Seconds between live router health checks (<= 0 to disable)");
102 static int router_ping_timeout = 50;
103 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
104 "Seconds to wait for the reply to a router health query");
107 lnet_peers_start_down(void)
109 return check_routers_before_use;
113 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
115 if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
116 CDEBUG(D_NET, "Out of date\n");
120 lp->lp_timestamp = when; /* update timestamp */
121 lp->lp_ping_deadline = 0; /* disable ping timeout */
123 if (lp->lp_alive_count != 0 && /* got old news */
124 (!lp->lp_alive) == (!alive)) { /* new date for old news */
125 CDEBUG(D_NET, "Old news\n");
129 /* Flag that notification is outstanding */
131 lp->lp_alive_count++;
132 lp->lp_alive = !(!alive); /* 1 bit! */
134 lp->lp_notifylnd |= notifylnd;
136 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
138 CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
142 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
147 /* Notify only in 1 thread at any time to ensure ordered notification.
148 * NB individual events can be missed; the only guarantee is that you
149 * always get the most recent news */
151 if (lp->lp_notifying)
154 lp->lp_notifying = 1;
156 while (lp->lp_notify) {
157 alive = lp->lp_alive;
158 notifylnd = lp->lp_notifylnd;
160 lp->lp_notifylnd = 0;
163 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
164 lnet_net_unlock(lp->lp_cpt);
166 /* A new notification could happen now; I'll handle it
167 * when control returns to me */
169 (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
171 lnet_net_lock(lp->lp_cpt);
175 lp->lp_notifying = 0;
180 lnet_rtr_addref_locked(lnet_peer_t *lp)
182 LASSERT(lp->lp_refcount > 0);
183 LASSERT(lp->lp_rtr_refcount >= 0);
185 /* lnet_net_lock must be exclusively locked */
186 lp->lp_rtr_refcount++;
187 if (lp->lp_rtr_refcount == 1) {
188 struct list_head *pos;
190 /* a simple insertion sort */
191 list_for_each_prev(pos, &the_lnet.ln_routers) {
192 lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
195 if (rtr->lp_nid < lp->lp_nid)
199 list_add(&lp->lp_rtr_list, pos);
200 /* addref for the_lnet.ln_routers */
201 lnet_peer_addref_locked(lp);
202 the_lnet.ln_routers_version++;
207 lnet_rtr_decref_locked(lnet_peer_t *lp)
209 LASSERT(lp->lp_refcount > 0);
210 LASSERT(lp->lp_rtr_refcount > 0);
212 /* lnet_net_lock must be exclusively locked */
213 lp->lp_rtr_refcount--;
214 if (lp->lp_rtr_refcount == 0) {
215 LASSERT(list_empty(&lp->lp_routes));
217 if (lp->lp_rcd != NULL) {
218 list_add(&lp->lp_rcd->rcd_list,
219 &the_lnet.ln_rcd_deathrow);
223 list_del(&lp->lp_rtr_list);
224 /* decref for the_lnet.ln_routers */
225 lnet_peer_decref_locked(lp);
226 the_lnet.ln_routers_version++;
231 lnet_find_net_locked (__u32 net)
233 lnet_remotenet_t *rnet;
234 struct list_head *tmp;
235 struct list_head *rn_list;
237 LASSERT(!the_lnet.ln_shutdown);
239 rn_list = lnet_net2rnethash(net);
240 list_for_each(tmp, rn_list) {
241 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
243 if (rnet->lrn_net == net)
249 static void lnet_shuffle_seed(void)
251 static int seeded = 0;
252 int lnd_type, seed[2];
255 struct list_head *tmp;
260 cfs_get_random_bytes(seed, sizeof(seed));
262 /* Nodes with small feet have little entropy
263 * the NID for this node gives the most entropy in the low bits */
264 list_for_each(tmp, &the_lnet.ln_nis) {
265 ni = list_entry(tmp, lnet_ni_t, ni_list);
266 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
268 if (lnd_type != LOLND)
269 seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
272 do_gettimeofday(&tv);
273 cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
278 /* NB expects LNET_LOCK held */
280 lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
282 unsigned int len = 0;
283 unsigned int offset = 0;
288 list_for_each(e, &rnet->lrn_routes) {
292 /* len+1 positions to add a new entry, also prevents division by 0 */
293 offset = cfs_rand() % (len + 1);
294 list_for_each(e, &rnet->lrn_routes) {
299 list_add(&route->lr_list, e);
300 list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
302 the_lnet.ln_remote_nets_version++;
303 lnet_rtr_addref_locked(route->lr_gateway);
307 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
308 unsigned int priority)
311 lnet_remotenet_t *rnet;
312 lnet_remotenet_t *rnet2;
318 CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
319 libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
321 if (gateway == LNET_NID_ANY ||
322 LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
323 net == LNET_NIDNET(LNET_NID_ANY) ||
324 LNET_NETTYP(net) == LOLND ||
325 LNET_NIDNET(gateway) == net ||
326 hops < 1 || hops > 255)
329 if (lnet_islocalnet(net)) /* it's a local network */
330 return 0; /* ignore the route entry */
332 /* Assume net, route, all new */
333 LIBCFS_ALLOC(route, sizeof(*route));
334 LIBCFS_ALLOC(rnet, sizeof(*rnet));
335 if (route == NULL || rnet == NULL) {
336 CERROR("Out of memory creating route %s %d %s\n",
337 libcfs_net2str(net), hops, libcfs_nid2str(gateway));
339 LIBCFS_FREE(route, sizeof(*route));
341 LIBCFS_FREE(rnet, sizeof(*rnet));
345 INIT_LIST_HEAD(&rnet->lrn_routes);
347 route->lr_hops = hops;
349 route->lr_priority = priority;
351 lnet_net_lock(LNET_LOCK_EX);
353 rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
355 lnet_net_unlock(LNET_LOCK_EX);
357 LIBCFS_FREE(route, sizeof(*route));
358 LIBCFS_FREE(rnet, sizeof(*rnet));
360 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
361 return 0; /* ignore the route entry */
362 CERROR("Error %d creating route %s %d %s\n", rc,
363 libcfs_net2str(net), hops,
364 libcfs_nid2str(gateway));
368 LASSERT(!the_lnet.ln_shutdown);
370 rnet2 = lnet_find_net_locked(net);
373 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
377 /* Search for a duplicate route (it's a NOOP if it is) */
379 list_for_each(e, &rnet2->lrn_routes) {
380 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
382 if (route2->lr_gateway == route->lr_gateway) {
387 /* our lookups must be true */
388 LASSERT(route2->lr_gateway->lp_nid != gateway);
392 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
393 lnet_add_route_to_rnet(rnet2, route);
395 ni = route->lr_gateway->lp_ni;
396 lnet_net_unlock(LNET_LOCK_EX);
398 /* XXX Assume alive */
399 if (ni->ni_lnd->lnd_notify != NULL)
400 (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
402 lnet_net_lock(LNET_LOCK_EX);
405 /* -1 for notify or !add_route */
406 lnet_peer_decref_locked(route->lr_gateway);
407 lnet_net_unlock(LNET_LOCK_EX);
410 LIBCFS_FREE(route, sizeof(*route));
413 LIBCFS_FREE(rnet, sizeof(*rnet));
419 lnet_check_routes(void)
421 lnet_remotenet_t *rnet;
423 lnet_route_t *route2;
424 struct list_head *e1;
425 struct list_head *e2;
427 struct list_head *rn_list;
430 cpt = lnet_net_lock_current();
432 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
433 rn_list = &the_lnet.ln_remote_nets_hash[i];
434 list_for_each(e1, rn_list) {
435 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
438 list_for_each(e2, &rnet->lrn_routes) {
443 route = list_entry(e2, lnet_route_t,
446 if (route2 == NULL) {
451 if (route->lr_gateway->lp_ni ==
452 route2->lr_gateway->lp_ni)
455 nid1 = route->lr_gateway->lp_nid;
456 nid2 = route2->lr_gateway->lp_nid;
459 lnet_net_unlock(cpt);
461 CERROR("Routes to %s via %s and %s not "
464 libcfs_nid2str(nid1),
465 libcfs_nid2str(nid2));
471 lnet_net_unlock(cpt);
476 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
478 struct lnet_peer *gateway;
479 lnet_remotenet_t *rnet;
481 struct list_head *e1;
482 struct list_head *e2;
484 struct list_head *rn_list;
487 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
488 libcfs_net2str(net), libcfs_nid2str(gw_nid));
490 /* NB Caller may specify either all routes via the given gateway
491 * or a specific route entry actual NIDs) */
493 lnet_net_lock(LNET_LOCK_EX);
494 if (net == LNET_NIDNET(LNET_NID_ANY))
495 rn_list = &the_lnet.ln_remote_nets_hash[0];
497 rn_list = lnet_net2rnethash(net);
500 list_for_each(e1, rn_list) {
501 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
503 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
504 net == rnet->lrn_net))
507 list_for_each(e2, &rnet->lrn_routes) {
508 route = list_entry(e2, lnet_route_t, lr_list);
510 gateway = route->lr_gateway;
511 if (!(gw_nid == LNET_NID_ANY ||
512 gw_nid == gateway->lp_nid))
515 list_del(&route->lr_list);
516 list_del(&route->lr_gwlist);
517 the_lnet.ln_remote_nets_version++;
519 if (list_empty(&rnet->lrn_routes))
520 list_del(&rnet->lrn_list);
524 lnet_rtr_decref_locked(gateway);
525 lnet_peer_decref_locked(gateway);
527 lnet_net_unlock(LNET_LOCK_EX);
529 LIBCFS_FREE(route, sizeof(*route));
532 LIBCFS_FREE(rnet, sizeof(*rnet));
535 lnet_net_lock(LNET_LOCK_EX);
540 if (net == LNET_NIDNET(LNET_NID_ANY) &&
541 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
542 rn_list = &the_lnet.ln_remote_nets_hash[idx];
545 lnet_net_unlock(LNET_LOCK_EX);
551 lnet_destroy_routes (void)
553 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
556 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
558 int i, rc = -ENOENT, lidx, j;
560 if (the_lnet.ln_rtrpools == NULL)
563 for (i = 0; i < LNET_NRBPOOLS; i++) {
564 lnet_rtrbufpool_t *rbp;
566 lnet_net_lock(LNET_LOCK_EX);
568 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
571 pool_cfg->pl_pools[i].pl_npages =
573 pool_cfg->pl_pools[i].pl_nbuffers =
575 pool_cfg->pl_pools[i].pl_credits =
577 pool_cfg->pl_pools[i].pl_mincredits =
578 rbp[i].rbp_mincredits;
582 lnet_net_unlock(LNET_LOCK_EX);
585 lnet_net_lock(LNET_LOCK_EX);
586 pool_cfg->pl_routing = the_lnet.ln_routing;
587 lnet_net_unlock(LNET_LOCK_EX);
593 lnet_get_route(int idx, __u32 *net, __u32 *hops,
594 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
596 struct list_head *e1;
597 struct list_head *e2;
598 lnet_remotenet_t *rnet;
602 struct list_head *rn_list;
604 cpt = lnet_net_lock_current();
606 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
607 rn_list = &the_lnet.ln_remote_nets_hash[i];
608 list_for_each(e1, rn_list) {
609 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
611 list_for_each(e2, &rnet->lrn_routes) {
612 route = list_entry(e2, lnet_route_t,
616 *net = rnet->lrn_net;
617 *hops = route->lr_hops;
618 *priority = route->lr_priority;
619 *gateway = route->lr_gateway->lp_nid;
621 route->lr_gateway->lp_alive &&
623 lnet_net_unlock(cpt);
630 lnet_net_unlock(cpt);
635 lnet_swap_pinginfo(lnet_ping_info_t *info)
638 lnet_ni_status_t *stat;
640 __swab32s(&info->pi_magic);
641 __swab32s(&info->pi_features);
642 __swab32s(&info->pi_pid);
643 __swab32s(&info->pi_nnis);
644 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
645 stat = &info->pi_ni[i];
646 __swab64s(&stat->ns_nid);
647 __swab32s(&stat->ns_status);
653 * parse router-checker pinginfo, record number of down NIs for remote
654 * networks on that router.
657 lnet_parse_rc_info(lnet_rc_data_t *rcd)
659 lnet_ping_info_t *info = rcd->rcd_pinginfo;
660 struct lnet_peer *gw = rcd->rcd_gateway;
666 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
667 lnet_swap_pinginfo(info);
669 /* NB always racing with network! */
670 if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
671 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
672 libcfs_nid2str(gw->lp_nid), info->pi_magic);
673 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
677 gw->lp_ping_feats = info->pi_features;
678 if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
679 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
680 libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
681 return; /* nothing I can understand */
684 if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
685 return; /* can't carry NI status info */
687 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
692 if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
697 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
698 lnet_ni_status_t *stat = &info->pi_ni[i];
699 lnet_nid_t nid = stat->ns_nid;
701 if (nid == LNET_NID_ANY) {
702 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
703 libcfs_nid2str(gw->lp_nid));
704 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
708 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
711 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
716 if (stat->ns_status == LNET_NI_STATUS_UP) {
717 if (LNET_NIDNET(nid) == rte->lr_net) {
724 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
725 libcfs_nid2str(gw->lp_nid), stat->ns_status);
726 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
730 if (up) { /* ignore downed NIs if NI for dest network is up */
734 rte->lr_downis = down;
739 lnet_router_checker_event(lnet_event_t *event)
741 lnet_rc_data_t *rcd = event->md.user_ptr;
742 struct lnet_peer *lp;
744 LASSERT(rcd != NULL);
746 if (event->unlinked) {
747 LNetInvalidateHandle(&rcd->rcd_mdh);
751 LASSERT(event->type == LNET_EVENT_SEND ||
752 event->type == LNET_EVENT_REPLY);
754 lp = rcd->rcd_gateway;
757 /* NB: it's called with holding lnet_res_lock, we have a few
758 * places need to hold both locks at the same time, please take
759 * care of lock ordering */
760 lnet_net_lock(lp->lp_cpt);
761 if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
762 /* ignore if no longer a router or rcd is replaced */
766 if (event->type == LNET_EVENT_SEND) {
767 lp->lp_ping_notsent = 0;
768 if (event->status == 0)
772 /* LNET_EVENT_REPLY */
773 /* A successful REPLY means the router is up. If _any_ comms
774 * to the router fail I assume it's down (this will happen if
775 * we ping alive routers to try to detect router death before
776 * apps get burned). */
778 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
779 /* The router checker will wake up very shortly and do the
780 * actual notification.
781 * XXX If 'lp' stops being a router before then, it will still
782 * have the notification pending!!! */
784 if (avoid_asym_router_failure && event->status == 0)
785 lnet_parse_rc_info(rcd);
788 lnet_net_unlock(lp->lp_cpt);
792 lnet_wait_known_routerstate(void)
795 struct list_head *entry;
798 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
801 int cpt = lnet_net_lock_current();
804 list_for_each(entry, &the_lnet.ln_routers) {
805 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
807 if (rtr->lp_alive_count == 0) {
813 lnet_net_unlock(cpt);
819 lnet_router_checker();
821 cfs_pause(cfs_time_seconds(1));
826 lnet_update_ni_status_locked(void)
832 LASSERT(the_lnet.ln_routing);
834 timeout = router_ping_timeout +
835 MAX(live_router_check_interval, dead_router_check_interval);
837 now = cfs_time_current_sec();
838 list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
839 if (ni->ni_lnd->lnd_type == LOLND)
842 if (now < ni->ni_last_alive + timeout)
846 /* re-check with lock */
847 if (now < ni->ni_last_alive + timeout) {
852 LASSERT(ni->ni_status != NULL);
854 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
855 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
856 libcfs_nid2str(ni->ni_nid), timeout);
857 /* NB: so far, this is the only place to set
858 * NI status to "down" */
859 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
866 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
868 LASSERT(list_empty(&rcd->rcd_list));
869 /* detached from network */
870 LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
872 if (rcd->rcd_gateway != NULL) {
873 int cpt = rcd->rcd_gateway->lp_cpt;
876 lnet_peer_decref_locked(rcd->rcd_gateway);
877 lnet_net_unlock(cpt);
880 if (rcd->rcd_pinginfo != NULL)
881 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
883 LIBCFS_FREE(rcd, sizeof(*rcd));
887 lnet_create_rc_data_locked(lnet_peer_t *gateway)
889 lnet_rc_data_t *rcd = NULL;
890 lnet_ping_info_t *pi;
894 lnet_net_unlock(gateway->lp_cpt);
896 LIBCFS_ALLOC(rcd, sizeof(*rcd));
900 LNetInvalidateHandle(&rcd->rcd_mdh);
901 INIT_LIST_HEAD(&rcd->rcd_list);
903 LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
907 for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
908 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
909 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
911 rcd->rcd_pinginfo = pi;
913 LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
914 rc = LNetMDBind((lnet_md_t){.start = pi,
916 .length = LNET_PINGINFO_SIZE,
917 .threshold = LNET_MD_THRESH_INF,
918 .options = LNET_MD_TRUNCATE,
919 .eq_handle = the_lnet.ln_rc_eqh},
923 CERROR("Can't bind MD: %d\n", rc);
928 lnet_net_lock(gateway->lp_cpt);
929 /* router table changed or someone has created rcd for this gateway */
930 if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
931 lnet_net_unlock(gateway->lp_cpt);
935 lnet_peer_addref_locked(gateway);
936 rcd->rcd_gateway = gateway;
937 gateway->lp_rcd = rcd;
938 gateway->lp_ping_notsent = 0;
944 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
945 rc = LNetMDUnlink(rcd->rcd_mdh);
948 lnet_destroy_rc_data(rcd);
951 lnet_net_lock(gateway->lp_cpt);
952 return gateway->lp_rcd;
956 lnet_router_check_interval (lnet_peer_t *rtr)
960 secs = rtr->lp_alive ? live_router_check_interval :
961 dead_router_check_interval;
969 lnet_ping_router_locked (lnet_peer_t *rtr)
971 lnet_rc_data_t *rcd = NULL;
972 cfs_time_t now = cfs_time_current();
975 lnet_peer_addref_locked(rtr);
977 if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
978 cfs_time_after(now, rtr->lp_ping_deadline))
979 lnet_notify_locked(rtr, 1, 0, now);
981 /* Run any outstanding notifications */
982 lnet_ni_notify_locked(rtr->lp_ni, rtr);
984 if (!lnet_isrouter(rtr) ||
985 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
986 /* router table changed or router checker is shutting down */
987 lnet_peer_decref_locked(rtr);
991 rcd = rtr->lp_rcd != NULL ?
992 rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
997 secs = lnet_router_check_interval(rtr);
1000 "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1001 "alive_count %d lp_ping_timestamp %lu\n",
1002 libcfs_nid2str(rtr->lp_nid), secs,
1003 rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1004 rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1006 if (secs != 0 && !rtr->lp_ping_notsent &&
1007 cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1008 cfs_time_seconds(secs)))) {
1010 lnet_process_id_t id;
1011 lnet_handle_md_t mdh;
1013 id.nid = rtr->lp_nid;
1014 id.pid = LNET_PID_LUSTRE;
1015 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1017 rtr->lp_ping_notsent = 1;
1018 rtr->lp_ping_timestamp = now;
1022 if (rtr->lp_ping_deadline == 0) {
1023 rtr->lp_ping_deadline =
1024 cfs_time_shift(router_ping_timeout);
1027 lnet_net_unlock(rtr->lp_cpt);
1029 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1030 LNET_PROTO_PING_MATCHBITS, 0);
1032 lnet_net_lock(rtr->lp_cpt);
1034 rtr->lp_ping_notsent = 0; /* no event pending */
1037 lnet_peer_decref_locked(rtr);
1042 lnet_router_checker_start(void)
1047 struct task_struct *task;
1048 #else /* __KERNEL__ */
1052 int router_checker_max_eqsize = 10240;
1054 LASSERT (check_routers_before_use);
1055 LASSERT (dead_router_check_interval > 0);
1059 /* As an approximation, allow each router the same number of
1060 * outstanding events as it is allowed outstanding sends */
1062 version = the_lnet.ln_routers_version;
1063 list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1064 lnet_ni_t *ni = rtr->lp_ni;
1065 lnet_process_id_t id;
1068 eqsz += ni->ni_peertxcredits;
1070 /* one async ping reply per router */
1071 id.nid = rtr->lp_nid;
1072 id.pid = LNET_PID_LUSTRE;
1076 rc = LNetSetAsync(id, 1);
1078 CWARN("LNetSetAsync %s failed: %d\n",
1079 libcfs_id2str(id), rc);
1084 /* NB router list doesn't change in userspace */
1085 LASSERT(version == the_lnet.ln_routers_version);
1092 "No router found, not starting router checker\n");
1096 /* at least allow a SENT and a REPLY per router */
1097 if (router_checker_max_eqsize < 2 * nrtr)
1098 router_checker_max_eqsize = 2 * nrtr;
1101 if (eqsz > router_checker_max_eqsize)
1102 eqsz = router_checker_max_eqsize;
1105 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1107 if (check_routers_before_use &&
1108 dead_router_check_interval <= 0) {
1109 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1110 " set if 'check_routers_before_use' is set"
1115 if (!the_lnet.ln_routing &&
1116 live_router_check_interval <= 0 &&
1117 dead_router_check_interval <= 0)
1121 sema_init(&the_lnet.ln_rc_signal, 0);
1122 /* EQ size doesn't matter; the callback is guaranteed to get every
1125 rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1126 &the_lnet.ln_rc_eqh);
1128 rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1129 &the_lnet.ln_rc_eqh);
1132 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1136 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1138 task = kthread_run(lnet_router_checker, NULL, "router_checker");
1141 CERROR("Can't start router checker thread: %d\n", rc);
1142 /* block until event callback signals exit */
1143 down(&the_lnet.ln_rc_signal);
1144 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1146 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1151 if (check_routers_before_use) {
1152 /* Note that a helpful side-effect of pinging all known routers
1153 * at startup is that it makes them drop stale connections they
1154 * may have to a previous instance of me. */
1155 lnet_wait_known_routerstate();
1162 lnet_router_checker_stop (void)
1166 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1169 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1170 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1173 /* block until event callback signals exit */
1174 down(&the_lnet.ln_rc_signal);
1176 lnet_router_checker();
1178 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1180 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1186 lnet_prune_rc_data(int wait_unlink)
1188 lnet_rc_data_t *rcd;
1189 lnet_rc_data_t *tmp;
1191 struct list_head head;
1194 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1195 list_empty(&the_lnet.ln_rcd_deathrow) &&
1196 list_empty(&the_lnet.ln_rcd_zombie)))
1199 INIT_LIST_HEAD(&head);
1201 lnet_net_lock(LNET_LOCK_EX);
1203 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1204 /* router checker is stopping, prune all */
1205 list_for_each_entry(lp, &the_lnet.ln_routers,
1207 if (lp->lp_rcd == NULL)
1210 LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1211 list_add(&lp->lp_rcd->rcd_list,
1212 &the_lnet.ln_rcd_deathrow);
1217 /* unlink all RCDs on deathrow list */
1218 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1220 if (!list_empty(&head)) {
1221 lnet_net_unlock(LNET_LOCK_EX);
1223 list_for_each_entry(rcd, &head, rcd_list)
1224 LNetMDUnlink(rcd->rcd_mdh);
1226 lnet_net_lock(LNET_LOCK_EX);
1229 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1231 /* release all zombie RCDs */
1232 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1233 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1235 if (LNetHandleIsInvalid(rcd->rcd_mdh))
1236 list_move(&rcd->rcd_list, &head);
1239 wait_unlink = wait_unlink &&
1240 !list_empty(&the_lnet.ln_rcd_zombie);
1242 lnet_net_unlock(LNET_LOCK_EX);
1244 while (!list_empty(&head)) {
1245 rcd = list_entry(head.next,
1246 lnet_rc_data_t, rcd_list);
1247 list_del_init(&rcd->rcd_list);
1248 lnet_destroy_rc_data(rcd);
1255 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1256 "Waiting for rc buffers to unlink\n");
1257 cfs_pause(cfs_time_seconds(1) / 4);
1259 lnet_net_lock(LNET_LOCK_EX);
1262 lnet_net_unlock(LNET_LOCK_EX);
1266 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1269 lnet_router_checker(void *arg)
1272 struct list_head *entry;
1274 cfs_block_allsigs();
1276 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1278 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1283 cpt = lnet_net_lock_current();
1285 version = the_lnet.ln_routers_version;
1287 list_for_each(entry, &the_lnet.ln_routers) {
1288 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1290 cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1292 lnet_net_unlock(cpt);
1295 /* the routers list has changed */
1296 if (version != the_lnet.ln_routers_version)
1300 lnet_ping_router_locked(rtr);
1302 /* NB dropped lock */
1303 if (version != the_lnet.ln_routers_version) {
1304 /* the routers list has changed */
1309 if (the_lnet.ln_routing)
1310 lnet_update_ni_status_locked();
1312 lnet_net_unlock(cpt);
1314 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1316 /* Call cfs_pause() here always adds 1 to load average
1317 * because kernel counts # active tasks as nr_running
1318 * + nr_uninterruptible. */
1319 schedule_timeout_and_set_state(TASK_INTERRUPTIBLE,
1320 cfs_time_seconds(1));
1323 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1325 lnet_prune_rc_data(1); /* wait for UNLINK */
1327 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1328 up(&the_lnet.ln_rc_signal);
1329 /* The unlink event callback will signal final completion */
1334 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1336 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1338 while (--npages >= 0)
1339 __free_page(rb->rb_kiov[npages].kiov_page);
1341 LIBCFS_FREE(rb, sz);
1345 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1347 int npages = rbp->rbp_npages;
1348 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1353 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1359 for (i = 0; i < npages; i++) {
1360 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1361 __GFP_ZERO | GFP_IOFS);
1364 __free_page(rb->rb_kiov[i].kiov_page);
1366 LIBCFS_FREE(rb, sz);
1370 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1371 rb->rb_kiov[i].kiov_offset = 0;
1372 rb->rb_kiov[i].kiov_page = page;
1379 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1381 int npages = rbp->rbp_npages;
1383 struct list_head tmp;
1385 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1388 INIT_LIST_HEAD(&tmp);
1391 lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1392 list_splice_init(&rbp->rbp_bufs, &tmp);
1393 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1394 rbp->rbp_mincredits = 0;
1395 lnet_net_unlock(cpt);
1397 /* Free buffers on the free list. */
1398 while (!list_empty(&tmp)) {
1399 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1400 list_del(&rb->rb_list);
1401 lnet_destroy_rtrbuf(rb, npages);
1406 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1408 struct list_head rb_list;
1411 int num_buffers = 0;
1412 int npages = rbp->rbp_npages;
1414 /* If we are called for less buffers than already in the pool, we
1415 * just lower the nbuffers number and excess buffers will be
1416 * thrown away as they are returned to the free list. Credits
1417 * then get adjusted as well. */
1418 if (nbufs <= rbp->rbp_nbuffers) {
1420 rbp->rbp_nbuffers = nbufs;
1421 lnet_net_unlock(cpt);
1425 INIT_LIST_HEAD(&rb_list);
1427 /* allocate the buffers on a local list first. If all buffers are
1428 * allocated successfully then join this list to the rbp buffer
1429 * list. If not then free all allocated buffers. */
1430 num_rb = rbp->rbp_nbuffers;
1432 while (num_rb < nbufs) {
1433 rb = lnet_new_rtrbuf(rbp, cpt);
1435 CERROR("Failed to allocate %d route bufs of %d pages\n",
1440 list_add(&rb->rb_list, &rb_list);
1447 list_splice_tail(&rb_list, &rbp->rbp_bufs);
1448 rbp->rbp_nbuffers += num_buffers;
1449 rbp->rbp_credits += num_buffers;
1450 rbp->rbp_mincredits = rbp->rbp_credits;
1451 /* We need to schedule blocked msg using the newly
1453 while (!list_empty(&rbp->rbp_bufs) &&
1454 !list_empty(&rbp->rbp_msgs))
1455 lnet_schedule_blocked_locked(rbp);
1457 lnet_net_unlock(cpt);
1462 while (!list_empty(&rb_list)) {
1463 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1464 list_del(&rb->rb_list);
1465 lnet_destroy_rtrbuf(rb, npages);
1472 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1474 INIT_LIST_HEAD(&rbp->rbp_msgs);
1475 INIT_LIST_HEAD(&rbp->rbp_bufs);
1477 rbp->rbp_npages = npages;
1478 rbp->rbp_credits = 0;
1479 rbp->rbp_mincredits = 0;
1483 lnet_rtrpools_free(int keep_pools)
1485 lnet_rtrbufpool_t *rtrp;
1488 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1491 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1492 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1493 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1494 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1498 cfs_percpt_free(the_lnet.ln_rtrpools);
1499 the_lnet.ln_rtrpools = NULL;
1504 lnet_nrb_tiny_calculate(void)
1506 int nrbs = LNET_NRB_TINY;
1508 if (tiny_router_buffers < 0) {
1509 LCONSOLE_ERROR_MSG(0x10c,
1510 "tiny_router_buffers=%d invalid when "
1511 "routing enabled\n", tiny_router_buffers);
1515 if (tiny_router_buffers > 0)
1516 nrbs = tiny_router_buffers;
1518 nrbs /= LNET_CPT_NUMBER;
1519 return max(nrbs, LNET_NRB_TINY_MIN);
1523 lnet_nrb_small_calculate(void)
1525 int nrbs = LNET_NRB_SMALL;
1527 if (small_router_buffers < 0) {
1528 LCONSOLE_ERROR_MSG(0x10c,
1529 "small_router_buffers=%d invalid when "
1530 "routing enabled\n", small_router_buffers);
1534 if (small_router_buffers > 0)
1535 nrbs = small_router_buffers;
1537 nrbs /= LNET_CPT_NUMBER;
1538 return max(nrbs, LNET_NRB_SMALL_MIN);
1542 lnet_nrb_large_calculate(void)
1544 int nrbs = LNET_NRB_LARGE;
1546 if (large_router_buffers < 0) {
1547 LCONSOLE_ERROR_MSG(0x10c,
1548 "large_router_buffers=%d invalid when "
1549 "routing enabled\n", large_router_buffers);
1553 if (large_router_buffers > 0)
1554 nrbs = large_router_buffers;
1556 nrbs /= LNET_CPT_NUMBER;
1557 return max(nrbs, LNET_NRB_LARGE_MIN);
1561 lnet_rtrpools_alloc(int im_a_router)
1563 lnet_rtrbufpool_t *rtrp;
1570 if (!strcmp(forwarding, "")) {
1571 /* not set either way */
1574 } else if (!strcmp(forwarding, "disabled")) {
1575 /* explicitly disabled */
1577 } else if (!strcmp(forwarding, "enabled")) {
1578 /* explicitly enabled */
1580 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1581 "'enabled' or 'disabled'\n");
1585 nrb_tiny = lnet_nrb_tiny_calculate();
1589 nrb_small = lnet_nrb_small_calculate();
1593 nrb_large = lnet_nrb_large_calculate();
1597 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1599 sizeof(lnet_rtrbufpool_t));
1600 if (the_lnet.ln_rtrpools == NULL) {
1601 LCONSOLE_ERROR_MSG(0x10c,
1602 "Failed to initialize router buffe pool\n");
1606 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1607 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1608 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1613 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1614 LNET_NRB_SMALL_PAGES);
1615 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1620 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1621 LNET_NRB_LARGE_PAGES);
1622 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1628 lnet_net_lock(LNET_LOCK_EX);
1629 the_lnet.ln_routing = 1;
1630 lnet_net_unlock(LNET_LOCK_EX);
1634 lnet_rtrpools_free(0);
1639 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1644 lnet_rtrbufpool_t *rtrp;
1646 /* If the provided values for each buffer pool are different than the
1647 * configured values, we need to take action. */
1649 tiny_router_buffers = tiny;
1650 nrb = lnet_nrb_tiny_calculate();
1651 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1652 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1659 small_router_buffers = small;
1660 nrb = lnet_nrb_small_calculate();
1661 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1662 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1669 large_router_buffers = large;
1670 nrb = lnet_nrb_large_calculate();
1671 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1672 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1683 lnet_rtrpools_adjust(int tiny, int small, int large)
1685 /* this function doesn't revert the changes if adding new buffers
1686 * failed. It's up to the user space caller to revert the
1689 if (!the_lnet.ln_routing)
1692 return lnet_rtrpools_adjust_helper(tiny, small, large);
1696 lnet_rtrpools_enable(void)
1700 if (the_lnet.ln_routing)
1703 if (the_lnet.ln_rtrpools == NULL)
1704 /* If routing is turned off, and we have never
1705 * initialized the pools before, just call the
1706 * standard buffer pool allocation routine as
1707 * if we are just configuring this for the first
1709 return lnet_rtrpools_alloc(1);
1711 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1715 lnet_net_lock(LNET_LOCK_EX);
1716 the_lnet.ln_routing = 1;
1718 the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1719 lnet_net_unlock(LNET_LOCK_EX);
1725 lnet_rtrpools_disable(void)
1727 if (!the_lnet.ln_routing)
1730 lnet_net_lock(LNET_LOCK_EX);
1731 the_lnet.ln_routing = 0;
1732 the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1734 tiny_router_buffers = 0;
1735 small_router_buffers = 0;
1736 large_router_buffers = 0;
1737 lnet_net_unlock(LNET_LOCK_EX);
1738 lnet_rtrpools_free(1);
1742 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1744 struct lnet_peer *lp = NULL;
1745 cfs_time_t now = cfs_time_current();
1746 int cpt = lnet_cpt_of_nid(nid);
1748 LASSERT (!in_interrupt ());
1750 CDEBUG (D_NET, "%s notifying %s: %s\n",
1751 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1752 libcfs_nid2str(nid),
1753 alive ? "up" : "down");
1756 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1757 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1758 libcfs_nid2str(nid), alive ? "birth" : "death",
1759 libcfs_nid2str(ni->ni_nid));
1763 /* can't do predictions... */
1764 if (cfs_time_after(when, now)) {
1765 CWARN ("Ignoring prediction from %s of %s %s "
1766 "%ld seconds in the future\n",
1767 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1768 libcfs_nid2str(nid), alive ? "up" : "down",
1769 cfs_duration_sec(cfs_time_sub(when, now)));
1773 if (ni != NULL && !alive && /* LND telling me she's down */
1774 !auto_down) { /* auto-down disabled */
1775 CDEBUG(D_NET, "Auto-down disabled\n");
1781 if (the_lnet.ln_shutdown) {
1782 lnet_net_unlock(cpt);
1786 lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1789 lnet_net_unlock(cpt);
1790 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1794 /* We can't fully trust LND on reporting exact peer last_alive
1795 * if he notifies us about dead peer. For example ksocklnd can
1796 * call us with when == _time_when_the_node_was_booted_ if
1797 * no connections were successfully established */
1798 if (ni != NULL && !alive && when < lp->lp_last_alive)
1799 when = lp->lp_last_alive;
1801 lnet_notify_locked(lp, ni == NULL, alive, when);
1804 lnet_ni_notify_locked(ni, lp);
1806 lnet_peer_decref_locked(lp);
1808 lnet_net_unlock(cpt);
1811 EXPORT_SYMBOL(lnet_notify);
1814 lnet_get_tunables (void)
1822 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1828 lnet_router_checker (void)
1830 static time_t last = 0;
1831 static int running = 0;
1833 time_t now = cfs_time_current_sec();
1834 int interval = now - last;
1839 /* It's no use to call me again within a sec - all intervals and
1840 * timeouts are measured in seconds */
1841 if (last != 0 && interval < 2)
1845 interval > MAX(live_router_check_interval,
1846 dead_router_check_interval))
1847 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1848 live_router_check_interval, dead_router_check_interval,
1851 LASSERT(LNET_CPT_NUMBER == 1);
1854 LASSERT(!running); /* recursion check */
1860 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1861 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1863 /* consume all pending events */
1868 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1869 * recursion breaker in LNetEQPoll would fail */
1870 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1871 if (rc == 0) /* no event pending */
1874 /* NB a lost SENT prevents me from pinging a router again */
1875 if (rc == -EOVERFLOW) {
1876 CERROR("Dropped an event!!!\n");
1882 lnet_router_checker_event(&ev);
1885 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1886 lnet_prune_rc_data(1); /* release rcd */
1887 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1892 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1896 version = the_lnet.ln_routers_version;
1897 list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1898 lnet_ping_router_locked(rtr);
1899 LASSERT(version == the_lnet.ln_routers_version);
1904 running = 0; /* lock only needed for the recursion check */
1908 /* NB lnet_peers_start_down depends on me,
1909 * so must be called before any peer creation */
1911 lnet_get_tunables (void)
1915 s = getenv("LNET_ROUTER_PING_TIMEOUT");
1916 if (s != NULL) router_ping_timeout = atoi(s);
1918 s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1919 if (s != NULL) live_router_check_interval = atoi(s);
1921 s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1922 if (s != NULL) dead_router_check_interval = atoi(s);
1924 /* This replaces old lnd_notify mechanism */
1925 check_routers_before_use = 1;
1926 if (dead_router_check_interval <= 0)
1927 dead_router_check_interval = 30;
1931 lnet_rtrpools_free(int keep_pools)
1936 lnet_rtrpools_alloc(int im_a_arouter)