2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2017, Intel Corporation.
6 * This file is part of Lustre, https://wiki.whamcloud.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
25 #include <linux/random.h>
26 #include <lnet/lib-lnet.h>
28 #define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
29 #define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4)
30 #define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */
31 #define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4)
32 #define LNET_NRB_SMALL_PAGES 1
33 #define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */
34 #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
35 #define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \
38 static char *forwarding = "";
39 module_param(forwarding, charp, 0444);
40 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
42 static int tiny_router_buffers;
43 module_param(tiny_router_buffers, int, 0444);
44 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
45 static int small_router_buffers;
46 module_param(small_router_buffers, int, 0444);
47 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
48 static int large_router_buffers;
49 module_param(large_router_buffers, int, 0444);
50 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
51 static int peer_buffer_credits;
52 module_param(peer_buffer_credits, int, 0444);
53 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
55 static int auto_down = 1;
56 module_param(auto_down, int, 0444);
57 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
60 lnet_peer_buffer_credits(struct lnet_net *net)
62 /* NI option overrides LNet default */
63 if (net->net_tunables.lct_peer_rtr_credits > 0)
64 return net->net_tunables.lct_peer_rtr_credits;
65 if (peer_buffer_credits > 0)
66 return peer_buffer_credits;
68 /* As an approximation, allow this peer the same number of router
69 * buffers as it is allowed outstanding sends */
70 return net->net_tunables.lct_peer_tx_credits;
73 static int check_routers_before_use;
74 module_param(check_routers_before_use, int, 0444);
75 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
77 int avoid_asym_router_failure = 1;
78 module_param(avoid_asym_router_failure, int, 0644);
79 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
81 int alive_router_check_interval = 60;
82 module_param(alive_router_check_interval, int, 0644);
83 MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
85 static int router_ping_timeout = 50;
86 module_param(router_ping_timeout, int, 0644);
87 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
90 * A value between 0 and 100. 0 meaning that even if router's interfaces
91 * have the worse health still consider the gateway usable.
92 * 100 means that at least one interface on the route's remote net is 100%
93 * healthy to consider the route alive.
94 * The default is set to 100 to ensure we maintain the original behavior.
96 unsigned int router_sensitivity_percentage = 100;
97 static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
98 static struct kernel_param_ops param_ops_rtr_sensitivity = {
99 .set = rtr_sensitivity_set,
100 .get = param_get_int,
102 #define param_check_rtr_sensitivity(name, p) \
103 __param_check(name, p, int)
104 #ifdef HAVE_KERNEL_PARAM_OPS
105 module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
107 module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
108 &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
110 MODULE_PARM_DESC(router_sensitivity_percentage,
111 "How healthy a gateway should be to be used in percent");
114 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
117 unsigned *sen = (unsigned *)kp->arg;
120 rc = kstrtoul(val, 0, &value);
122 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
126 if (value < 0 || value > 100) {
127 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
132 * The purpose of locking the api_mutex here is to ensure that
133 * the correct value ends up stored properly.
135 mutex_lock(&the_lnet.ln_api_mutex);
139 mutex_unlock(&the_lnet.ln_api_mutex);
145 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
147 struct lnet_route *route;
149 lnet_net_lock(LNET_LOCK_EX);
150 target->lp_rtr_refcount += src->lp_rtr_refcount;
151 /* move the list of queued messages to the new peer */
152 list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
153 /* move all the routes that reference the peer */
154 list_splice_init(&src->lp_routes, &target->lp_routes);
155 /* update all the routes to point to the new peer */
156 list_for_each_entry(route, &target->lp_routes, lr_gwlist)
157 route->lr_gateway = target;
158 /* remove the old peer from the ln_routers list */
159 list_del_init(&src->lp_rtr_list);
160 /* add the new peer to the ln_routers list */
161 if (list_empty(&target->lp_rtr_list)) {
162 lnet_peer_addref_locked(target);
163 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
165 /* reset the ref count on the old peer and decrement its ref count */
166 src->lp_rtr_refcount = 0;
167 lnet_peer_decref_locked(src);
168 /* update the router version */
169 the_lnet.ln_routers_version++;
170 lnet_net_unlock(LNET_LOCK_EX);
174 lnet_peers_start_down(void)
176 return check_routers_before_use;
180 * A net is alive if at least one gateway NI on the network is alive.
183 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
185 struct lnet_peer_ni *lpni;
187 list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
188 if (lnet_is_peer_ni_alive(lpni))
196 * a gateway is alive only if all its nets are alive
197 * called with cpt lock held
199 bool lnet_is_gateway_alive(struct lnet_peer *gw)
201 struct lnet_peer_net *lpn;
203 list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
204 if (!lnet_is_gateway_net_alive(lpn))
212 * lnet_is_route_alive() needs to be called with cpt lock held
213 * A route is alive if the gateway can route between the local network and
214 * the remote network of the route.
215 * This means at least one NI is alive on each of the local and remote
216 * networks of the gateway.
218 bool lnet_is_route_alive(struct lnet_route *route)
220 struct lnet_peer *gw = route->lr_gateway;
221 struct lnet_peer_net *llpn;
222 struct lnet_peer_net *rlpn;
226 * check the gateway's interfaces on the route rnet to make sure
227 * that the gateway is viable.
229 llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
233 route_alive = lnet_is_gateway_net_alive(llpn);
235 if (avoid_asym_router_failure) {
236 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
239 route_alive = route_alive &&
240 lnet_is_gateway_net_alive(rlpn);
246 spin_lock(&gw->lp_lock);
247 if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
248 if (gw->lp_rtr_refcount > 0)
249 CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
250 libcfs_nid2str(gw->lp_primary_nid));
253 spin_unlock(&gw->lp_lock);
259 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
260 struct lnet_peer *new_lp)
262 struct lnet_peer_ni *lpni;
263 struct lnet_route *route;
266 * Although a route is correlated with a peer, but when it's added
267 * a specific NID is used. That NID refers to a peer_ni within
268 * a peer. There could be other peer_nis on the same net, which
269 * can be used to send to that gateway. However when we are
270 * consolidating gateways because of discovery, the nid used to
271 * add the route might've moved between gateway peers. In this
272 * case we want to move the route to the new gateway as well. The
273 * intent here is not to confuse the user who added the route.
275 list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
276 lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
278 lnet_net_lock(LNET_LOCK_EX);
279 list_move(&route->lr_gwlist, &new_lp->lp_routes);
280 lnet_net_unlock(LNET_LOCK_EX);
287 lnet_router_discovery_complete(struct lnet_peer *lp)
289 struct lnet_peer_ni *lpni = NULL;
291 spin_lock(&lp->lp_lock);
292 lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
293 spin_unlock(&lp->lp_lock);
296 * Router discovery successful? All peer information would've been
297 * updated already. No need to do any more processing
299 if (!lp->lp_dc_error)
302 * discovery failed? then we need to set the status of each lpni
303 * to DOWN. It will be updated the next time we discover the
304 * router. For router peer NIs not on local networks, we never send
305 * messages directly to them, so their health will always remain
306 * at maximum. We can only tell if they are up or down from the
307 * status returned in the PING response. If we fail to get that
308 * status in our scheduled router discovery, then we'll assume
309 * it's down until we're told otherwise.
311 CDEBUG(D_NET, "%s: Router discovery failed %d\n",
312 libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
313 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
314 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
318 lnet_rtr_addref_locked(struct lnet_peer *lp)
320 LASSERT(lp->lp_rtr_refcount >= 0);
322 /* lnet_net_lock must be exclusively locked */
323 lp->lp_rtr_refcount++;
324 if (lp->lp_rtr_refcount == 1) {
325 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
326 /* addref for the_lnet.ln_routers */
327 lnet_peer_addref_locked(lp);
328 the_lnet.ln_routers_version++;
333 lnet_rtr_decref_locked(struct lnet_peer *lp)
335 LASSERT(atomic_read(&lp->lp_refcount) > 0);
336 LASSERT(lp->lp_rtr_refcount > 0);
338 /* lnet_net_lock must be exclusively locked */
339 lp->lp_rtr_refcount--;
340 if (lp->lp_rtr_refcount == 0) {
341 LASSERT(list_empty(&lp->lp_routes));
343 list_del(&lp->lp_rtr_list);
344 /* decref for the_lnet.ln_routers */
345 lnet_peer_decref_locked(lp);
346 the_lnet.ln_routers_version++;
350 struct lnet_remotenet *
351 lnet_find_rnet_locked(__u32 net)
353 struct lnet_remotenet *rnet;
354 struct list_head *tmp;
355 struct list_head *rn_list;
357 LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
359 rn_list = lnet_net2rnethash(net);
360 list_for_each(tmp, rn_list) {
361 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
363 if (rnet->lrn_net == net)
369 static void lnet_shuffle_seed(void)
372 struct lnet_ni *ni = NULL;
377 /* Nodes with small feet have little entropy
378 * the NID for this node gives the most entropy in the low bits */
379 while ((ni = lnet_get_next_ni_locked(NULL, ni)))
380 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
386 /* NB expects LNET_LOCK held */
388 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
390 unsigned int len = 0;
391 unsigned int offset = 0;
396 list_for_each(e, &rnet->lrn_routes)
400 * Randomly adding routes to the list is done to ensure that when
401 * different nodes are using the same list of routers, they end up
402 * preferring different routers.
404 offset = cfs_rand() % (len + 1);
405 list_for_each(e, &rnet->lrn_routes) {
410 list_add(&route->lr_list, e);
412 * force a router check on the gateway to make sure the route is
415 route->lr_gateway->lp_rtrcheck_timestamp = 0;
417 the_lnet.ln_remote_nets_version++;
419 /* add the route on the gateway list */
420 list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
422 /* take a router reference count on the gateway */
423 lnet_rtr_addref_locked(route->lr_gateway);
427 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
428 __u32 priority, __u32 sensitivity)
430 struct list_head *route_entry;
431 struct lnet_remotenet *rnet;
432 struct lnet_remotenet *rnet2;
433 struct lnet_route *route;
434 struct lnet_peer_ni *lpni;
435 struct lnet_peer *gw;
439 CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
440 libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
442 if (gateway == LNET_NID_ANY ||
443 LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
444 net == LNET_NIDNET(LNET_NID_ANY) ||
445 LNET_NETTYP(net) == LOLND ||
446 LNET_NIDNET(gateway) == net ||
447 (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
450 /* it's a local network */
451 if (lnet_islocalnet(net))
454 /* Assume net, route, all new */
455 LIBCFS_ALLOC(route, sizeof(*route));
456 LIBCFS_ALLOC(rnet, sizeof(*rnet));
457 if (route == NULL || rnet == NULL) {
458 CERROR("Out of memory creating route %s %d %s\n",
459 libcfs_net2str(net), hops, libcfs_nid2str(gateway));
461 LIBCFS_FREE(route, sizeof(*route));
463 LIBCFS_FREE(rnet, sizeof(*rnet));
467 INIT_LIST_HEAD(&rnet->lrn_routes);
469 /* store the local and remote net that the route represents */
470 route->lr_lnet = LNET_NIDNET(gateway);
472 route->lr_nid = gateway;
473 route->lr_priority = priority;
474 route->lr_hops = hops;
476 lnet_net_lock(LNET_LOCK_EX);
479 * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
480 * lose that once we're done
482 lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
484 lnet_net_unlock(LNET_LOCK_EX);
486 LIBCFS_FREE(route, sizeof(*route));
487 LIBCFS_FREE(rnet, sizeof(*rnet));
490 CERROR("Error %d creating route %s %d %s\n", rc,
491 libcfs_net2str(net), hops,
492 libcfs_nid2str(gateway));
496 LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
497 gw = lpni->lpni_peer_net->lpn_peer;
499 route->lr_gateway = gw;
501 rnet2 = lnet_find_rnet_locked(net);
504 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
508 /* Search for a duplicate route (it's a NOOP if it is) */
510 list_for_each(route_entry, &rnet2->lrn_routes) {
511 struct lnet_route *route2;
513 route2 = list_entry(route_entry, struct lnet_route, lr_list);
514 if (route2->lr_gateway == route->lr_gateway) {
519 /* our lookups must be true */
520 LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
524 * It is possible to add multiple routes through the same peer,
525 * but it'll be using a different NID of that peer. When the
526 * gateway is discovered, discovery will consolidate the different
527 * peers into one peer. In this case the discovery code will have
528 * to move the routes from the peer that's being deleted to the
529 * consolidated peer lp_routes list
532 gw->lp_health_sensitivity = sensitivity;
533 lnet_add_route_to_rnet(rnet2, route);
537 * get rid of the reference on the lpni.
539 lnet_peer_ni_decref_locked(lpni);
540 lnet_net_unlock(LNET_LOCK_EX);
546 LIBCFS_FREE(route, sizeof(*route));
550 LIBCFS_FREE(rnet, sizeof(*rnet));
552 /* kick start the monitor thread to handle the added route */
553 wake_up(&the_lnet.ln_mt_waitq);
559 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
560 struct list_head *zombies)
562 struct lnet_peer *gateway;
563 struct lnet_route *route;
564 struct lnet_route *tmp;
566 list_for_each_entry_safe(route, tmp, route_list, lr_list) {
567 gateway = route->lr_gateway;
568 if (gw_nid != LNET_NID_ANY &&
569 gw_nid != gateway->lp_primary_nid)
573 * move to zombie to delete outside the lock
574 * Note that this function is called with the
575 * ln_api_mutex held as well as the exclusive net
576 * lock. Adding to the remote net list happens
577 * under the same conditions. Same goes for the
578 * gateway router list
580 list_move(&route->lr_list, zombies);
581 the_lnet.ln_remote_nets_version++;
583 list_del(&route->lr_gwlist);
584 lnet_rtr_decref_locked(gateway);
589 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
591 struct list_head rnet_zombies;
592 struct lnet_remotenet *rnet;
593 struct lnet_remotenet *tmp;
594 struct list_head *rn_list;
595 struct lnet_peer_ni *lpni;
596 struct lnet_route *route;
597 struct list_head zombies;
598 struct lnet_peer *lp;
601 INIT_LIST_HEAD(&rnet_zombies);
602 INIT_LIST_HEAD(&zombies);
604 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
605 libcfs_net2str(net), libcfs_nid2str(gw_nid));
607 /* NB Caller may specify either all routes via the given gateway
608 * or a specific route entry actual NIDs) */
610 lnet_net_lock(LNET_LOCK_EX);
612 lpni = lnet_find_peer_ni_locked(gw_nid);
614 lp = lpni->lpni_peer_net->lpn_peer;
616 gw_nid = lp->lp_primary_nid;
617 lnet_peer_ni_decref_locked(lpni);
620 if (net != LNET_NIDNET(LNET_NID_ANY)) {
621 rnet = lnet_find_rnet_locked(net);
623 lnet_net_unlock(LNET_LOCK_EX);
626 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
628 if (list_empty(&rnet->lrn_routes))
629 list_move(&rnet->lrn_list, &rnet_zombies);
633 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
634 rn_list = &the_lnet.ln_remote_nets_hash[i];
636 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
637 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
639 if (list_empty(&rnet->lrn_routes))
640 list_move(&rnet->lrn_list, &rnet_zombies);
645 lnet_net_unlock(LNET_LOCK_EX);
647 while (!list_empty(&zombies)) {
648 route = list_first_entry(&zombies, struct lnet_route, lr_list);
649 list_del(&route->lr_list);
650 LIBCFS_FREE(route, sizeof(*route));
653 while (!list_empty(&rnet_zombies)) {
654 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
656 list_del(&rnet->lrn_list);
657 LIBCFS_FREE(rnet, sizeof(*rnet));
664 lnet_destroy_routes (void)
666 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
669 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
671 struct lnet_rtrbufpool *rbp;
672 int i, rc = -ENOENT, j;
674 if (the_lnet.ln_rtrpools == NULL)
678 cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
683 for (j = 0; j < LNET_NRBPOOLS; j++) {
684 pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
685 pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
686 pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
687 pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
694 lnet_net_lock(LNET_LOCK_EX);
695 pool_cfg->pl_routing = the_lnet.ln_routing;
696 lnet_net_unlock(LNET_LOCK_EX);
702 lnet_get_route(int idx, __u32 *net, __u32 *hops,
703 lnet_nid_t *gateway, __u32 *alive, __u32 *priority, __u32 *sensitivity)
705 struct lnet_remotenet *rnet;
706 struct list_head *rn_list;
707 struct lnet_route *route;
708 struct list_head *e1;
709 struct list_head *e2;
713 cpt = lnet_net_lock_current();
715 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
716 rn_list = &the_lnet.ln_remote_nets_hash[i];
717 list_for_each(e1, rn_list) {
718 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
720 list_for_each(e2, &rnet->lrn_routes) {
721 route = list_entry(e2, struct lnet_route,
725 *net = rnet->lrn_net;
726 *gateway = route->lr_nid;
727 *hops = route->lr_hops;
728 *priority = route->lr_priority;
729 *sensitivity = route->lr_gateway->
730 lp_health_sensitivity;
731 *alive = lnet_is_route_alive(route);
732 lnet_net_unlock(cpt);
739 lnet_net_unlock(cpt);
744 lnet_wait_known_routerstate(void)
746 struct lnet_peer *rtr;
747 struct list_head *entry;
750 LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
753 int cpt = lnet_net_lock_current();
756 list_for_each(entry, &the_lnet.ln_routers) {
757 rtr = list_entry(entry, struct lnet_peer,
760 spin_lock(&rtr->lp_lock);
762 if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
764 spin_unlock(&rtr->lp_lock);
767 spin_unlock(&rtr->lp_lock);
770 lnet_net_unlock(cpt);
775 set_current_state(TASK_UNINTERRUPTIBLE);
776 schedule_timeout(cfs_time_seconds(1));
781 lnet_update_ni_status_locked(void)
783 struct lnet_ni *ni = NULL;
787 LASSERT(the_lnet.ln_routing);
789 timeout = router_ping_timeout + alive_router_check_interval;
791 now = ktime_get_real_seconds();
792 while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
793 if (ni->ni_net->net_lnd->lnd_type == LOLND)
796 if (now < ni->ni_last_alive + timeout)
800 /* re-check with lock */
801 if (now < ni->ni_last_alive + timeout) {
806 LASSERT(ni->ni_status != NULL);
808 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
809 CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
810 libcfs_nid2str(ni->ni_nid), timeout);
811 /* NB: so far, this is the only place to set
812 * NI status to "down" */
813 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
819 void lnet_wait_router_start(void)
821 if (check_routers_before_use) {
822 /* Note that a helpful side-effect of pinging all known routers
823 * at startup is that it makes them drop stale connections they
824 * may have to a previous instance of me. */
825 lnet_wait_known_routerstate();
830 * This function is called from the monitor thread to check if there are
831 * any active routers that need to be checked.
834 lnet_router_checker_active(void)
836 /* Router Checker thread needs to run when routing is enabled in
837 * order to call lnet_update_ni_status_locked() */
838 if (the_lnet.ln_routing)
841 return !list_empty(&the_lnet.ln_routers) &&
842 alive_router_check_interval > 0;
846 lnet_check_routers(void)
848 struct lnet_peer_ni *lpni;
849 struct list_head *entry;
850 struct lnet_peer *rtr;
856 cpt = lnet_net_lock_current();
858 version = the_lnet.ln_routers_version;
860 list_for_each(entry, &the_lnet.ln_routers) {
861 rtr = list_entry(entry, struct lnet_peer,
864 now = ktime_get_real_seconds();
867 * only discover the router if we've passed
868 * alive_router_check_interval seconds. Some of the router
869 * interfaces could be down and in that case they would be
870 * undergoing recovery separately from this discovery.
872 if (now - rtr->lp_rtrcheck_timestamp <
873 alive_router_check_interval)
877 * If we're currently discovering the peer then don't
878 * issue another discovery
880 spin_lock(&rtr->lp_lock);
881 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
882 spin_unlock(&rtr->lp_lock);
885 /* make sure we actively discover the router */
886 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
887 rtr->lp_state |= LNET_PEER_RTR_DISCOVERY;
888 spin_unlock(&rtr->lp_lock);
890 /* find the peer_ni associated with the primary NID */
891 lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
893 CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
894 libcfs_nid2str(rtr->lp_primary_nid));
897 lnet_peer_ni_addref_locked(lpni);
899 /* discover the router */
900 CDEBUG(D_NET, "discover %s, cpt = %d\n",
901 libcfs_nid2str(lpni->lpni_nid), cpt);
902 rc = lnet_discover_peer_locked(lpni, cpt, false);
904 /* decrement ref count acquired by find_peer_ni_locked() */
905 lnet_peer_ni_decref_locked(lpni);
908 rtr->lp_rtrcheck_timestamp = now;
910 CERROR("Failed to discover router %s\n",
911 libcfs_nid2str(rtr->lp_primary_nid));
913 /* NB dropped lock */
914 if (version != the_lnet.ln_routers_version) {
915 /* the routers list has changed */
920 if (the_lnet.ln_routing)
921 lnet_update_ni_status_locked();
923 lnet_net_unlock(cpt);
927 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
929 int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
931 while (--npages >= 0)
932 __free_page(rb->rb_kiov[npages].kiov_page);
937 static struct lnet_rtrbuf *
938 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
940 int npages = rbp->rbp_npages;
941 int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
943 struct lnet_rtrbuf *rb;
946 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
952 for (i = 0; i < npages; i++) {
953 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
954 GFP_KERNEL | __GFP_ZERO);
957 __free_page(rb->rb_kiov[i].kiov_page);
963 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
964 rb->rb_kiov[i].kiov_offset = 0;
965 rb->rb_kiov[i].kiov_page = page;
972 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
974 int npages = rbp->rbp_npages;
975 struct lnet_rtrbuf *rb;
976 struct list_head tmp;
978 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
981 INIT_LIST_HEAD(&tmp);
984 list_splice_init(&rbp->rbp_msgs, &tmp);
985 lnet_drop_routed_msgs_locked(&tmp, cpt);
986 list_splice_init(&rbp->rbp_bufs, &tmp);
987 rbp->rbp_req_nbuffers = 0;
988 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
989 rbp->rbp_mincredits = 0;
990 lnet_net_unlock(cpt);
992 /* Free buffers on the free list. */
993 while (!list_empty(&tmp)) {
994 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
995 list_del(&rb->rb_list);
996 lnet_destroy_rtrbuf(rb, npages);
1001 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1003 struct list_head rb_list;
1004 struct lnet_rtrbuf *rb;
1006 int num_buffers = 0;
1008 int npages = rbp->rbp_npages;
1011 /* If we are called for less buffers than already in the pool, we
1012 * just lower the req_nbuffers number and excess buffers will be
1013 * thrown away as they are returned to the free list. Credits
1014 * then get adjusted as well.
1015 * If we already have enough buffers allocated to serve the
1016 * increase requested, then we can treat that the same way as we
1017 * do the decrease. */
1018 num_rb = nbufs - rbp->rbp_nbuffers;
1019 if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1020 rbp->rbp_req_nbuffers = nbufs;
1021 lnet_net_unlock(cpt);
1024 /* store the older value of rbp_req_nbuffers and then set it to
1025 * the new request to prevent lnet_return_rx_credits_locked() from
1026 * freeing buffers that we need to keep around */
1027 old_req_nbufs = rbp->rbp_req_nbuffers;
1028 rbp->rbp_req_nbuffers = nbufs;
1029 lnet_net_unlock(cpt);
1031 INIT_LIST_HEAD(&rb_list);
1033 /* allocate the buffers on a local list first. If all buffers are
1034 * allocated successfully then join this list to the rbp buffer
1035 * list. If not then free all allocated buffers. */
1036 while (num_rb-- > 0) {
1037 rb = lnet_new_rtrbuf(rbp, cpt);
1039 CERROR("Failed to allocate %d route bufs of %d pages\n",
1043 rbp->rbp_req_nbuffers = old_req_nbufs;
1044 lnet_net_unlock(cpt);
1049 list_add(&rb->rb_list, &rb_list);
1055 list_splice_tail(&rb_list, &rbp->rbp_bufs);
1056 rbp->rbp_nbuffers += num_buffers;
1057 rbp->rbp_credits += num_buffers;
1058 rbp->rbp_mincredits = rbp->rbp_credits;
1059 /* We need to schedule blocked msg using the newly
1061 while (!list_empty(&rbp->rbp_bufs) &&
1062 !list_empty(&rbp->rbp_msgs))
1063 lnet_schedule_blocked_locked(rbp);
1065 lnet_net_unlock(cpt);
1070 while (!list_empty(&rb_list)) {
1071 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1072 list_del(&rb->rb_list);
1073 lnet_destroy_rtrbuf(rb, npages);
1080 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1082 INIT_LIST_HEAD(&rbp->rbp_msgs);
1083 INIT_LIST_HEAD(&rbp->rbp_bufs);
1085 rbp->rbp_npages = npages;
1086 rbp->rbp_credits = 0;
1087 rbp->rbp_mincredits = 0;
1091 lnet_rtrpools_free(int keep_pools)
1093 struct lnet_rtrbufpool *rtrp;
1096 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1099 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1100 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1101 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1102 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1106 cfs_percpt_free(the_lnet.ln_rtrpools);
1107 the_lnet.ln_rtrpools = NULL;
1112 lnet_nrb_tiny_calculate(void)
1114 int nrbs = LNET_NRB_TINY;
1116 if (tiny_router_buffers < 0) {
1117 LCONSOLE_ERROR_MSG(0x10c,
1118 "tiny_router_buffers=%d invalid when "
1119 "routing enabled\n", tiny_router_buffers);
1123 if (tiny_router_buffers > 0)
1124 nrbs = tiny_router_buffers;
1126 nrbs /= LNET_CPT_NUMBER;
1127 return max(nrbs, LNET_NRB_TINY_MIN);
1131 lnet_nrb_small_calculate(void)
1133 int nrbs = LNET_NRB_SMALL;
1135 if (small_router_buffers < 0) {
1136 LCONSOLE_ERROR_MSG(0x10c,
1137 "small_router_buffers=%d invalid when "
1138 "routing enabled\n", small_router_buffers);
1142 if (small_router_buffers > 0)
1143 nrbs = small_router_buffers;
1145 nrbs /= LNET_CPT_NUMBER;
1146 return max(nrbs, LNET_NRB_SMALL_MIN);
1150 lnet_nrb_large_calculate(void)
1152 int nrbs = LNET_NRB_LARGE;
1154 if (large_router_buffers < 0) {
1155 LCONSOLE_ERROR_MSG(0x10c,
1156 "large_router_buffers=%d invalid when "
1157 "routing enabled\n", large_router_buffers);
1161 if (large_router_buffers > 0)
1162 nrbs = large_router_buffers;
1164 nrbs /= LNET_CPT_NUMBER;
1165 return max(nrbs, LNET_NRB_LARGE_MIN);
1169 lnet_rtrpools_alloc(int im_a_router)
1171 struct lnet_rtrbufpool *rtrp;
1178 if (!strcmp(forwarding, "")) {
1179 /* not set either way */
1182 } else if (!strcmp(forwarding, "disabled")) {
1183 /* explicitly disabled */
1185 } else if (!strcmp(forwarding, "enabled")) {
1186 /* explicitly enabled */
1188 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1189 "'enabled' or 'disabled'\n");
1193 nrb_tiny = lnet_nrb_tiny_calculate();
1197 nrb_small = lnet_nrb_small_calculate();
1201 nrb_large = lnet_nrb_large_calculate();
1205 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1207 sizeof(struct lnet_rtrbufpool));
1208 if (the_lnet.ln_rtrpools == NULL) {
1209 LCONSOLE_ERROR_MSG(0x10c,
1210 "Failed to initialize router buffe pool\n");
1214 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1215 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1216 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1221 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1222 LNET_NRB_SMALL_PAGES);
1223 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1228 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1229 LNET_NRB_LARGE_PAGES);
1230 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1236 lnet_net_lock(LNET_LOCK_EX);
1237 the_lnet.ln_routing = 1;
1238 lnet_net_unlock(LNET_LOCK_EX);
1239 wake_up(&the_lnet.ln_mt_waitq);
1243 lnet_rtrpools_free(0);
1248 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1253 struct lnet_rtrbufpool *rtrp;
1255 /* If the provided values for each buffer pool are different than the
1256 * configured values, we need to take action. */
1258 tiny_router_buffers = tiny;
1259 nrb = lnet_nrb_tiny_calculate();
1260 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1261 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1268 small_router_buffers = small;
1269 nrb = lnet_nrb_small_calculate();
1270 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1271 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1278 large_router_buffers = large;
1279 nrb = lnet_nrb_large_calculate();
1280 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1281 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1292 lnet_rtrpools_adjust(int tiny, int small, int large)
1294 /* this function doesn't revert the changes if adding new buffers
1295 * failed. It's up to the user space caller to revert the
1298 if (!the_lnet.ln_routing)
1301 return lnet_rtrpools_adjust_helper(tiny, small, large);
1305 lnet_rtrpools_enable(void)
1309 if (the_lnet.ln_routing)
1312 if (the_lnet.ln_rtrpools == NULL)
1313 /* If routing is turned off, and we have never
1314 * initialized the pools before, just call the
1315 * standard buffer pool allocation routine as
1316 * if we are just configuring this for the first
1318 rc = lnet_rtrpools_alloc(1);
1320 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1324 lnet_net_lock(LNET_LOCK_EX);
1325 the_lnet.ln_routing = 1;
1327 the_lnet.ln_ping_target->pb_info.pi_features &=
1328 ~LNET_PING_FEAT_RTE_DISABLED;
1329 lnet_net_unlock(LNET_LOCK_EX);
1335 lnet_rtrpools_disable(void)
1337 if (!the_lnet.ln_routing)
1340 lnet_net_lock(LNET_LOCK_EX);
1341 the_lnet.ln_routing = 0;
1342 the_lnet.ln_ping_target->pb_info.pi_features |=
1343 LNET_PING_FEAT_RTE_DISABLED;
1345 tiny_router_buffers = 0;
1346 small_router_buffers = 0;
1347 large_router_buffers = 0;
1348 lnet_net_unlock(LNET_LOCK_EX);
1349 lnet_rtrpools_free(1);
1353 lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
1355 if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1356 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1360 * ni: local NI used to communicate with the peer
1362 * alive: true if peer is alive, false otherwise
1363 * reset: reset health value. This is requested by the LND.
1364 * when: notificaiton time.
1367 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1370 struct lnet_peer_ni *lpni = NULL;
1371 time64_t now = ktime_get_seconds();
1374 LASSERT (!in_interrupt ());
1376 CDEBUG (D_NET, "%s notifying %s: %s\n",
1377 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1378 libcfs_nid2str(nid),
1379 alive ? "up" : "down");
1382 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1383 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1384 libcfs_nid2str(nid), alive ? "birth" : "death",
1385 libcfs_nid2str(ni->ni_nid));
1389 /* can't do predictions... */
1391 CWARN("Ignoring prediction from %s of %s %s "
1392 "%lld seconds in the future\n",
1393 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1394 libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1398 if (ni != NULL && !alive && /* LND telling me she's down */
1399 !auto_down) { /* auto-down disabled */
1400 CDEBUG(D_NET, "Auto-down disabled\n");
1404 /* must lock 0 since this is used for synchronization */
1407 if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1412 lpni = lnet_find_peer_ni_locked(nid);
1416 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1422 lnet_set_healthv(&lpni->lpni_healthv,
1423 LNET_MAX_HEALTH_VALUE);
1425 lnet_inc_healthv(&lpni->lpni_healthv);
1427 lnet_handle_remote_failure_locked(lpni);
1430 /* recalculate aliveness */
1431 alive = lnet_is_peer_ni_alive(lpni);
1434 if (ni != NULL && !alive)
1435 lnet_notify_peer_down(ni, lpni->lpni_nid);
1437 cpt = lpni->lpni_cpt;
1439 lnet_peer_ni_decref_locked(lpni);
1440 lnet_net_unlock(cpt);
1444 EXPORT_SYMBOL(lnet_notify);