2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2012, 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 "../../include/linux/lnet/lib-lnet.h"
27 #if 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_LARGE_MIN 256 /* min value for each CPT */
34 #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
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(lnet_ni_t *ni)
60 /* NI option overrides LNet default */
61 if (ni->ni_peerrtrcredits > 0)
62 return ni->ni_peerrtrcredits;
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 ni->ni_peertxcredits;
72 static int lnet_router_checker(void *);
76 lnet_peer_buffer_credits(lnet_ni_t *ni)
83 static int check_routers_before_use;
84 module_param(check_routers_before_use, int, 0444);
85 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
87 int avoid_asym_router_failure = 1;
88 module_param(avoid_asym_router_failure, int, 0644);
89 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
91 static int dead_router_check_interval = 60;
92 module_param(dead_router_check_interval, int, 0644);
93 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
95 static int live_router_check_interval = 60;
96 module_param(live_router_check_interval, int, 0644);
97 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
99 static int router_ping_timeout = 50;
100 module_param(router_ping_timeout, int, 0644);
101 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
104 lnet_peers_start_down(void)
106 return check_routers_before_use;
110 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive,
113 if (time_before(when, lp->lp_timestamp)) { /* out of date information */
114 CDEBUG(D_NET, "Out of date\n");
118 lp->lp_timestamp = when; /* update timestamp */
119 lp->lp_ping_deadline = 0; /* disable ping timeout */
121 if (lp->lp_alive_count != 0 && /* got old news */
122 (!lp->lp_alive) == (!alive)) { /* new date for old news */
123 CDEBUG(D_NET, "Old news\n");
127 /* Flag that notification is outstanding */
129 lp->lp_alive_count++;
130 lp->lp_alive = !(!alive); /* 1 bit! */
132 lp->lp_notifylnd |= notifylnd;
134 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
136 CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
140 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
145 /* Notify only in 1 thread at any time to ensure ordered notification.
146 * NB individual events can be missed; the only guarantee is that you
147 * always get the most recent news */
149 if (lp->lp_notifying || ni == NULL)
152 lp->lp_notifying = 1;
154 while (lp->lp_notify) {
155 alive = lp->lp_alive;
156 notifylnd = lp->lp_notifylnd;
158 lp->lp_notifylnd = 0;
161 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
162 lnet_net_unlock(lp->lp_cpt);
164 /* A new notification could happen now; I'll handle it
165 * when control returns to me */
167 (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
169 lnet_net_lock(lp->lp_cpt);
173 lp->lp_notifying = 0;
178 lnet_rtr_addref_locked(lnet_peer_t *lp)
180 LASSERT(lp->lp_refcount > 0);
181 LASSERT(lp->lp_rtr_refcount >= 0);
183 /* lnet_net_lock must be exclusively locked */
184 lp->lp_rtr_refcount++;
185 if (lp->lp_rtr_refcount == 1) {
186 struct list_head *pos;
188 /* a simple insertion sort */
189 list_for_each_prev(pos, &the_lnet.ln_routers) {
190 lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
193 if (rtr->lp_nid < lp->lp_nid)
197 list_add(&lp->lp_rtr_list, pos);
198 /* addref for the_lnet.ln_routers */
199 lnet_peer_addref_locked(lp);
200 the_lnet.ln_routers_version++;
205 lnet_rtr_decref_locked(lnet_peer_t *lp)
207 LASSERT(lp->lp_refcount > 0);
208 LASSERT(lp->lp_rtr_refcount > 0);
210 /* lnet_net_lock must be exclusively locked */
211 lp->lp_rtr_refcount--;
212 if (lp->lp_rtr_refcount == 0) {
213 LASSERT(list_empty(&lp->lp_routes));
215 if (lp->lp_rcd != NULL) {
216 list_add(&lp->lp_rcd->rcd_list,
217 &the_lnet.ln_rcd_deathrow);
221 list_del(&lp->lp_rtr_list);
222 /* decref for the_lnet.ln_routers */
223 lnet_peer_decref_locked(lp);
224 the_lnet.ln_routers_version++;
229 lnet_find_net_locked(__u32 net)
231 lnet_remotenet_t *rnet;
232 struct list_head *tmp;
233 struct list_head *rn_list;
235 LASSERT(!the_lnet.ln_shutdown);
237 rn_list = lnet_net2rnethash(net);
238 list_for_each(tmp, rn_list) {
239 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
241 if (rnet->lrn_net == net)
247 static void lnet_shuffle_seed(void)
250 int lnd_type, seed[2];
253 struct list_head *tmp;
258 cfs_get_random_bytes(seed, sizeof(seed));
260 /* Nodes with small feet have little entropy
261 * the NID for this node gives the most entropy in the low bits */
262 list_for_each(tmp, &the_lnet.ln_nis) {
263 ni = list_entry(tmp, lnet_ni_t, ni_list);
264 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
266 if (lnd_type != LOLND)
267 seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
270 do_gettimeofday(&tv);
271 cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
275 /* NB expects LNET_LOCK held */
277 lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
279 unsigned int len = 0;
280 unsigned int offset = 0;
285 list_for_each(e, &rnet->lrn_routes) {
289 /* len+1 positions to add a new entry, also prevents division by 0 */
290 offset = cfs_rand() % (len + 1);
291 list_for_each(e, &rnet->lrn_routes) {
296 list_add(&route->lr_list, e);
297 list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
299 the_lnet.ln_remote_nets_version++;
300 lnet_rtr_addref_locked(route->lr_gateway);
304 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
305 unsigned int priority)
308 lnet_remotenet_t *rnet;
309 lnet_remotenet_t *rnet2;
315 CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
316 libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
318 if (gateway == LNET_NID_ANY ||
319 LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
320 net == LNET_NIDNET(LNET_NID_ANY) ||
321 LNET_NETTYP(net) == LOLND ||
322 LNET_NIDNET(gateway) == net ||
323 hops < 1 || hops > 255)
326 if (lnet_islocalnet(net)) /* it's a local network */
327 return 0; /* ignore the route entry */
329 /* Assume net, route, all new */
330 LIBCFS_ALLOC(route, sizeof(*route));
331 LIBCFS_ALLOC(rnet, sizeof(*rnet));
332 if (route == NULL || rnet == NULL) {
333 CERROR("Out of memory creating route %s %d %s\n",
334 libcfs_net2str(net), hops, libcfs_nid2str(gateway));
336 LIBCFS_FREE(route, sizeof(*route));
338 LIBCFS_FREE(rnet, sizeof(*rnet));
342 INIT_LIST_HEAD(&rnet->lrn_routes);
344 route->lr_hops = hops;
346 route->lr_priority = priority;
348 lnet_net_lock(LNET_LOCK_EX);
350 rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
352 lnet_net_unlock(LNET_LOCK_EX);
354 LIBCFS_FREE(route, sizeof(*route));
355 LIBCFS_FREE(rnet, sizeof(*rnet));
357 if (rc == -EHOSTUNREACH) /* gateway is not on a local net */
358 return 0; /* ignore the route entry */
359 CERROR("Error %d creating route %s %d %s\n", rc,
360 libcfs_net2str(net), hops,
361 libcfs_nid2str(gateway));
366 LASSERT(!the_lnet.ln_shutdown);
368 rnet2 = lnet_find_net_locked(net);
371 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
375 /* Search for a duplicate route (it's a NOOP if it is) */
377 list_for_each(e, &rnet2->lrn_routes) {
378 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
380 if (route2->lr_gateway == route->lr_gateway) {
385 /* our lookups must be true */
386 LASSERT(route2->lr_gateway->lp_nid != gateway);
390 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
391 lnet_add_route_to_rnet(rnet2, route);
393 ni = route->lr_gateway->lp_ni;
394 lnet_net_unlock(LNET_LOCK_EX);
396 /* XXX Assume alive */
397 if (ni->ni_lnd->lnd_notify != NULL)
398 (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
400 lnet_net_lock(LNET_LOCK_EX);
403 /* -1 for notify or !add_route */
404 lnet_peer_decref_locked(route->lr_gateway);
405 lnet_net_unlock(LNET_LOCK_EX);
408 LIBCFS_FREE(route, sizeof(*route));
411 LIBCFS_FREE(rnet, sizeof(*rnet));
417 lnet_check_routes(void)
419 lnet_remotenet_t *rnet;
421 lnet_route_t *route2;
422 struct list_head *e1;
423 struct list_head *e2;
425 struct list_head *rn_list;
428 cpt = lnet_net_lock_current();
430 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
431 rn_list = &the_lnet.ln_remote_nets_hash[i];
432 list_for_each(e1, rn_list) {
433 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
436 list_for_each(e2, &rnet->lrn_routes) {
441 route = list_entry(e2, lnet_route_t,
444 if (route2 == NULL) {
449 if (route->lr_gateway->lp_ni ==
450 route2->lr_gateway->lp_ni)
453 nid1 = route->lr_gateway->lp_nid;
454 nid2 = route2->lr_gateway->lp_nid;
457 lnet_net_unlock(cpt);
459 CERROR("Routes to %s via %s and %s not supported\n",
461 libcfs_nid2str(nid1),
462 libcfs_nid2str(nid2));
468 lnet_net_unlock(cpt);
473 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
475 struct lnet_peer *gateway;
476 lnet_remotenet_t *rnet;
478 struct list_head *e1;
479 struct list_head *e2;
481 struct list_head *rn_list;
484 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
485 libcfs_net2str(net), libcfs_nid2str(gw_nid));
487 /* NB Caller may specify either all routes via the given gateway
488 * or a specific route entry actual NIDs) */
490 lnet_net_lock(LNET_LOCK_EX);
491 if (net == LNET_NIDNET(LNET_NID_ANY))
492 rn_list = &the_lnet.ln_remote_nets_hash[0];
494 rn_list = lnet_net2rnethash(net);
497 list_for_each(e1, rn_list) {
498 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
500 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
501 net == rnet->lrn_net))
504 list_for_each(e2, &rnet->lrn_routes) {
505 route = list_entry(e2, lnet_route_t, lr_list);
507 gateway = route->lr_gateway;
508 if (!(gw_nid == LNET_NID_ANY ||
509 gw_nid == gateway->lp_nid))
512 list_del(&route->lr_list);
513 list_del(&route->lr_gwlist);
514 the_lnet.ln_remote_nets_version++;
516 if (list_empty(&rnet->lrn_routes))
517 list_del(&rnet->lrn_list);
521 lnet_rtr_decref_locked(gateway);
522 lnet_peer_decref_locked(gateway);
524 lnet_net_unlock(LNET_LOCK_EX);
526 LIBCFS_FREE(route, sizeof(*route));
529 LIBCFS_FREE(rnet, sizeof(*rnet));
532 lnet_net_lock(LNET_LOCK_EX);
537 if (net == LNET_NIDNET(LNET_NID_ANY) &&
538 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
539 rn_list = &the_lnet.ln_remote_nets_hash[idx];
542 lnet_net_unlock(LNET_LOCK_EX);
548 lnet_destroy_routes(void)
550 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
554 lnet_get_route(int idx, __u32 *net, __u32 *hops,
555 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
557 struct list_head *e1;
558 struct list_head *e2;
559 lnet_remotenet_t *rnet;
563 struct list_head *rn_list;
565 cpt = lnet_net_lock_current();
567 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
568 rn_list = &the_lnet.ln_remote_nets_hash[i];
569 list_for_each(e1, rn_list) {
570 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
572 list_for_each(e2, &rnet->lrn_routes) {
573 route = list_entry(e2, lnet_route_t,
577 *net = rnet->lrn_net;
578 *hops = route->lr_hops;
579 *priority = route->lr_priority;
580 *gateway = route->lr_gateway->lp_nid;
581 *alive = route->lr_gateway->lp_alive;
582 lnet_net_unlock(cpt);
589 lnet_net_unlock(cpt);
594 lnet_swap_pinginfo(lnet_ping_info_t *info)
597 lnet_ni_status_t *stat;
599 __swab32s(&info->pi_magic);
600 __swab32s(&info->pi_features);
601 __swab32s(&info->pi_pid);
602 __swab32s(&info->pi_nnis);
603 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
604 stat = &info->pi_ni[i];
605 __swab64s(&stat->ns_nid);
606 __swab32s(&stat->ns_status);
611 * parse router-checker pinginfo, record number of down NIs for remote
612 * networks on that router.
615 lnet_parse_rc_info(lnet_rc_data_t *rcd)
617 lnet_ping_info_t *info = rcd->rcd_pinginfo;
618 struct lnet_peer *gw = rcd->rcd_gateway;
624 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
625 lnet_swap_pinginfo(info);
627 /* NB always racing with network! */
628 if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
629 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
630 libcfs_nid2str(gw->lp_nid), info->pi_magic);
631 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
635 gw->lp_ping_feats = info->pi_features;
636 if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
637 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
638 libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
639 return; /* nothing I can understand */
642 if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
643 return; /* can't carry NI status info */
645 list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
646 int ptl_status = LNET_NI_STATUS_INVALID;
651 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
652 lnet_ni_status_t *stat = &info->pi_ni[i];
653 lnet_nid_t nid = stat->ns_nid;
655 if (nid == LNET_NID_ANY) {
656 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
657 libcfs_nid2str(gw->lp_nid));
658 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
662 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
665 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
666 if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
668 else if (ptl_status != LNET_NI_STATUS_UP)
669 ptl_status = LNET_NI_STATUS_DOWN;
673 if (stat->ns_status == LNET_NI_STATUS_UP) {
674 if (LNET_NIDNET(nid) == rtr->lr_net) {
678 /* ptl NIs are considered down only when
679 * they're all down */
680 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
681 ptl_status = LNET_NI_STATUS_UP;
685 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
686 libcfs_nid2str(gw->lp_nid), stat->ns_status);
687 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
691 if (up) { /* ignore downed NIs if NI for dest network is up */
695 rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
700 lnet_router_checker_event(lnet_event_t *event)
702 lnet_rc_data_t *rcd = event->md.user_ptr;
703 struct lnet_peer *lp;
705 LASSERT(rcd != NULL);
707 if (event->unlinked) {
708 LNetInvalidateHandle(&rcd->rcd_mdh);
712 LASSERT(event->type == LNET_EVENT_SEND ||
713 event->type == LNET_EVENT_REPLY);
715 lp = rcd->rcd_gateway;
718 /* NB: it's called with holding lnet_res_lock, we have a few
719 * places need to hold both locks at the same time, please take
720 * care of lock ordering */
721 lnet_net_lock(lp->lp_cpt);
722 if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
723 /* ignore if no longer a router or rcd is replaced */
727 if (event->type == LNET_EVENT_SEND) {
728 lp->lp_ping_notsent = 0;
729 if (event->status == 0)
733 /* LNET_EVENT_REPLY */
734 /* A successful REPLY means the router is up. If _any_ comms
735 * to the router fail I assume it's down (this will happen if
736 * we ping alive routers to try to detect router death before
737 * apps get burned). */
739 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
740 /* The router checker will wake up very shortly and do the
741 * actual notification.
742 * XXX If 'lp' stops being a router before then, it will still
743 * have the notification pending!!! */
745 if (avoid_asym_router_failure && event->status == 0)
746 lnet_parse_rc_info(rcd);
749 lnet_net_unlock(lp->lp_cpt);
753 lnet_wait_known_routerstate(void)
756 struct list_head *entry;
759 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
762 int cpt = lnet_net_lock_current();
765 list_for_each(entry, &the_lnet.ln_routers) {
766 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
768 if (rtr->lp_alive_count == 0) {
774 lnet_net_unlock(cpt);
779 set_current_state(TASK_UNINTERRUPTIBLE);
780 schedule_timeout(cfs_time_seconds(1));
785 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
789 if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
790 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
791 if (rte->lr_net == net) {
800 lnet_update_ni_status_locked(void)
806 LASSERT(the_lnet.ln_routing);
808 timeout = router_ping_timeout +
809 max(live_router_check_interval, dead_router_check_interval);
812 list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
813 if (ni->ni_lnd->lnd_type == LOLND)
816 if (now < ni->ni_last_alive + timeout)
820 /* re-check with lock */
821 if (now < ni->ni_last_alive + timeout) {
826 LASSERT(ni->ni_status != NULL);
828 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
829 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
830 libcfs_nid2str(ni->ni_nid), timeout);
831 /* NB: so far, this is the only place to set
832 * NI status to "down" */
833 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
840 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
842 LASSERT(list_empty(&rcd->rcd_list));
843 /* detached from network */
844 LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
846 if (rcd->rcd_gateway != NULL) {
847 int cpt = rcd->rcd_gateway->lp_cpt;
850 lnet_peer_decref_locked(rcd->rcd_gateway);
851 lnet_net_unlock(cpt);
854 if (rcd->rcd_pinginfo != NULL)
855 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
857 LIBCFS_FREE(rcd, sizeof(*rcd));
860 static lnet_rc_data_t *
861 lnet_create_rc_data_locked(lnet_peer_t *gateway)
863 lnet_rc_data_t *rcd = NULL;
864 lnet_ping_info_t *pi;
868 lnet_net_unlock(gateway->lp_cpt);
870 LIBCFS_ALLOC(rcd, sizeof(*rcd));
874 LNetInvalidateHandle(&rcd->rcd_mdh);
875 INIT_LIST_HEAD(&rcd->rcd_list);
877 LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
881 for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
882 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
883 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
885 rcd->rcd_pinginfo = pi;
887 LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
888 rc = LNetMDBind((lnet_md_t){.start = pi,
890 .length = LNET_PINGINFO_SIZE,
891 .threshold = LNET_MD_THRESH_INF,
892 .options = LNET_MD_TRUNCATE,
893 .eq_handle = the_lnet.ln_rc_eqh},
897 CERROR("Can't bind MD: %d\n", rc);
902 lnet_net_lock(gateway->lp_cpt);
903 /* router table changed or someone has created rcd for this gateway */
904 if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
905 lnet_net_unlock(gateway->lp_cpt);
909 lnet_peer_addref_locked(gateway);
910 rcd->rcd_gateway = gateway;
911 gateway->lp_rcd = rcd;
912 gateway->lp_ping_notsent = 0;
918 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
919 rc = LNetMDUnlink(rcd->rcd_mdh);
922 lnet_destroy_rc_data(rcd);
925 lnet_net_lock(gateway->lp_cpt);
926 return gateway->lp_rcd;
930 lnet_router_check_interval(lnet_peer_t *rtr)
934 secs = rtr->lp_alive ? live_router_check_interval :
935 dead_router_check_interval;
943 lnet_ping_router_locked(lnet_peer_t *rtr)
945 lnet_rc_data_t *rcd = NULL;
946 unsigned long now = cfs_time_current();
949 lnet_peer_addref_locked(rtr);
951 if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
952 cfs_time_after(now, rtr->lp_ping_deadline))
953 lnet_notify_locked(rtr, 1, 0, now);
955 /* Run any outstanding notifications */
956 lnet_ni_notify_locked(rtr->lp_ni, rtr);
958 if (!lnet_isrouter(rtr) ||
959 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
960 /* router table changed or router checker is shutting down */
961 lnet_peer_decref_locked(rtr);
965 rcd = rtr->lp_rcd != NULL ?
966 rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
971 secs = lnet_router_check_interval(rtr);
974 "rtr %s %d: deadline %lu ping_notsent %d alive %d alive_count %d lp_ping_timestamp %lu\n",
975 libcfs_nid2str(rtr->lp_nid), secs,
976 rtr->lp_ping_deadline, rtr->lp_ping_notsent,
977 rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
979 if (secs != 0 && !rtr->lp_ping_notsent &&
980 cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
981 cfs_time_seconds(secs)))) {
983 lnet_process_id_t id;
984 lnet_handle_md_t mdh;
986 id.nid = rtr->lp_nid;
987 id.pid = LUSTRE_SRV_LNET_PID;
988 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
990 rtr->lp_ping_notsent = 1;
991 rtr->lp_ping_timestamp = now;
995 if (rtr->lp_ping_deadline == 0) {
996 rtr->lp_ping_deadline =
997 cfs_time_shift(router_ping_timeout);
1000 lnet_net_unlock(rtr->lp_cpt);
1002 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1003 LNET_PROTO_PING_MATCHBITS, 0);
1005 lnet_net_lock(rtr->lp_cpt);
1007 rtr->lp_ping_notsent = 0; /* no event pending */
1010 lnet_peer_decref_locked(rtr);
1014 lnet_router_checker_start(void)
1019 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1021 if (check_routers_before_use &&
1022 dead_router_check_interval <= 0) {
1023 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be set if 'check_routers_before_use' is set\n");
1027 if (!the_lnet.ln_routing &&
1028 live_router_check_interval <= 0 &&
1029 dead_router_check_interval <= 0)
1032 sema_init(&the_lnet.ln_rc_signal, 0);
1033 /* EQ size doesn't matter; the callback is guaranteed to get every
1036 rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1037 &the_lnet.ln_rc_eqh);
1039 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1043 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1044 rc = PTR_ERR(kthread_run(lnet_router_checker,
1045 NULL, "router_checker"));
1046 if (IS_ERR_VALUE(rc)) {
1047 CERROR("Can't start router checker thread: %d\n", rc);
1048 /* block until event callback signals exit */
1049 down(&the_lnet.ln_rc_signal);
1050 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1052 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1056 if (check_routers_before_use) {
1057 /* Note that a helpful side-effect of pinging all known routers
1058 * at startup is that it makes them drop stale connections they
1059 * may have to a previous instance of me. */
1060 lnet_wait_known_routerstate();
1067 lnet_router_checker_stop(void)
1071 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1074 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1075 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1077 /* block until event callback signals exit */
1078 down(&the_lnet.ln_rc_signal);
1079 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1081 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1086 lnet_prune_rc_data(int wait_unlink)
1088 lnet_rc_data_t *rcd;
1089 lnet_rc_data_t *tmp;
1091 struct list_head head;
1094 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1095 list_empty(&the_lnet.ln_rcd_deathrow) &&
1096 list_empty(&the_lnet.ln_rcd_zombie)))
1099 INIT_LIST_HEAD(&head);
1101 lnet_net_lock(LNET_LOCK_EX);
1103 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1104 /* router checker is stopping, prune all */
1105 list_for_each_entry(lp, &the_lnet.ln_routers,
1107 if (lp->lp_rcd == NULL)
1110 LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1111 list_add(&lp->lp_rcd->rcd_list,
1112 &the_lnet.ln_rcd_deathrow);
1117 /* unlink all RCDs on deathrow list */
1118 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1120 if (!list_empty(&head)) {
1121 lnet_net_unlock(LNET_LOCK_EX);
1123 list_for_each_entry(rcd, &head, rcd_list)
1124 LNetMDUnlink(rcd->rcd_mdh);
1126 lnet_net_lock(LNET_LOCK_EX);
1129 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1131 /* release all zombie RCDs */
1132 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1133 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1135 if (LNetHandleIsInvalid(rcd->rcd_mdh))
1136 list_move(&rcd->rcd_list, &head);
1139 wait_unlink = wait_unlink &&
1140 !list_empty(&the_lnet.ln_rcd_zombie);
1142 lnet_net_unlock(LNET_LOCK_EX);
1144 while (!list_empty(&head)) {
1145 rcd = list_entry(head.next,
1146 lnet_rc_data_t, rcd_list);
1147 list_del_init(&rcd->rcd_list);
1148 lnet_destroy_rc_data(rcd);
1155 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1156 "Waiting for rc buffers to unlink\n");
1157 set_current_state(TASK_UNINTERRUPTIBLE);
1158 schedule_timeout(cfs_time_seconds(1) / 4);
1160 lnet_net_lock(LNET_LOCK_EX);
1163 lnet_net_unlock(LNET_LOCK_EX);
1167 #if defined(LNET_ROUTER)
1170 lnet_router_checker(void *arg)
1173 struct list_head *entry;
1175 cfs_block_allsigs();
1177 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1179 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1184 cpt = lnet_net_lock_current();
1186 version = the_lnet.ln_routers_version;
1188 list_for_each(entry, &the_lnet.ln_routers) {
1189 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1191 cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1193 lnet_net_unlock(cpt);
1196 /* the routers list has changed */
1197 if (version != the_lnet.ln_routers_version)
1201 lnet_ping_router_locked(rtr);
1203 /* NB dropped lock */
1204 if (version != the_lnet.ln_routers_version) {
1205 /* the routers list has changed */
1210 if (the_lnet.ln_routing)
1211 lnet_update_ni_status_locked();
1213 lnet_net_unlock(cpt);
1215 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1217 /* Call schedule_timeout() here always adds 1 to load average
1218 * because kernel counts # active tasks as nr_running
1219 * + nr_uninterruptible. */
1220 set_current_state(TASK_INTERRUPTIBLE);
1221 schedule_timeout(cfs_time_seconds(1));
1224 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1226 lnet_prune_rc_data(1); /* wait for UNLINK */
1228 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1229 up(&the_lnet.ln_rc_signal);
1230 /* The unlink event callback will signal final completion */
1235 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1237 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1239 while (--npages >= 0)
1240 __free_page(rb->rb_kiov[npages].kiov_page);
1242 LIBCFS_FREE(rb, sz);
1245 static lnet_rtrbuf_t *
1246 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1248 int npages = rbp->rbp_npages;
1249 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1254 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1260 for (i = 0; i < npages; i++) {
1261 page = alloc_pages_node(
1262 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1263 __GFP_ZERO | GFP_IOFS, 0);
1266 __free_page(rb->rb_kiov[i].kiov_page);
1268 LIBCFS_FREE(rb, sz);
1272 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1273 rb->rb_kiov[i].kiov_offset = 0;
1274 rb->rb_kiov[i].kiov_page = page;
1281 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1283 int npages = rbp->rbp_npages;
1287 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1290 LASSERT(list_empty(&rbp->rbp_msgs));
1291 LASSERT(rbp->rbp_credits == rbp->rbp_nbuffers);
1293 while (!list_empty(&rbp->rbp_bufs)) {
1294 LASSERT(rbp->rbp_credits > 0);
1296 rb = list_entry(rbp->rbp_bufs.next,
1297 lnet_rtrbuf_t, rb_list);
1298 list_del(&rb->rb_list);
1299 lnet_destroy_rtrbuf(rb, npages);
1303 LASSERT(rbp->rbp_nbuffers == nbuffers);
1304 LASSERT(rbp->rbp_credits == nbuffers);
1306 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1310 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1315 if (rbp->rbp_nbuffers != 0) {
1316 LASSERT(rbp->rbp_nbuffers == nbufs);
1320 for (i = 0; i < nbufs; i++) {
1321 rb = lnet_new_rtrbuf(rbp, cpt);
1324 CERROR("Failed to allocate %d router bufs of %d pages\n",
1325 nbufs, rbp->rbp_npages);
1329 rbp->rbp_nbuffers++;
1331 rbp->rbp_mincredits++;
1332 list_add(&rb->rb_list, &rbp->rbp_bufs);
1334 /* No allocation "under fire" */
1335 /* Otherwise we'd need code to schedule blocked msgs etc */
1336 LASSERT(!the_lnet.ln_routing);
1339 LASSERT(rbp->rbp_credits == nbufs);
1344 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1346 INIT_LIST_HEAD(&rbp->rbp_msgs);
1347 INIT_LIST_HEAD(&rbp->rbp_bufs);
1349 rbp->rbp_npages = npages;
1350 rbp->rbp_credits = 0;
1351 rbp->rbp_mincredits = 0;
1355 lnet_rtrpools_free(void)
1357 lnet_rtrbufpool_t *rtrp;
1360 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1363 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1364 lnet_rtrpool_free_bufs(&rtrp[0]);
1365 lnet_rtrpool_free_bufs(&rtrp[1]);
1366 lnet_rtrpool_free_bufs(&rtrp[2]);
1369 cfs_percpt_free(the_lnet.ln_rtrpools);
1370 the_lnet.ln_rtrpools = NULL;
1374 lnet_nrb_tiny_calculate(int npages)
1376 int nrbs = LNET_NRB_TINY;
1378 if (tiny_router_buffers < 0) {
1379 LCONSOLE_ERROR_MSG(0x10c,
1380 "tiny_router_buffers=%d invalid when routing enabled\n",
1381 tiny_router_buffers);
1385 if (tiny_router_buffers > 0)
1386 nrbs = tiny_router_buffers;
1388 nrbs /= LNET_CPT_NUMBER;
1389 return max(nrbs, LNET_NRB_TINY_MIN);
1393 lnet_nrb_small_calculate(int npages)
1395 int nrbs = LNET_NRB_SMALL;
1397 if (small_router_buffers < 0) {
1398 LCONSOLE_ERROR_MSG(0x10c,
1399 "small_router_buffers=%d invalid when routing enabled\n",
1400 small_router_buffers);
1404 if (small_router_buffers > 0)
1405 nrbs = small_router_buffers;
1407 nrbs /= LNET_CPT_NUMBER;
1408 return max(nrbs, LNET_NRB_SMALL_MIN);
1412 lnet_nrb_large_calculate(int npages)
1414 int nrbs = LNET_NRB_LARGE;
1416 if (large_router_buffers < 0) {
1417 LCONSOLE_ERROR_MSG(0x10c,
1418 "large_router_buffers=%d invalid when routing enabled\n",
1419 large_router_buffers);
1423 if (large_router_buffers > 0)
1424 nrbs = large_router_buffers;
1426 nrbs /= LNET_CPT_NUMBER;
1427 return max(nrbs, LNET_NRB_LARGE_MIN);
1431 lnet_rtrpools_alloc(int im_a_router)
1433 lnet_rtrbufpool_t *rtrp;
1435 int small_pages = 1;
1442 large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1444 if (!strcmp(forwarding, "")) {
1445 /* not set either way */
1448 } else if (!strcmp(forwarding, "disabled")) {
1449 /* explicitly disabled */
1451 } else if (!strcmp(forwarding, "enabled")) {
1452 /* explicitly enabled */
1454 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either 'enabled' or 'disabled'\n");
1458 nrb_tiny = lnet_nrb_tiny_calculate(0);
1462 nrb_small = lnet_nrb_small_calculate(small_pages);
1466 nrb_large = lnet_nrb_large_calculate(large_pages);
1470 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1472 sizeof(lnet_rtrbufpool_t));
1473 if (the_lnet.ln_rtrpools == NULL) {
1474 LCONSOLE_ERROR_MSG(0x10c,
1475 "Failed to initialize router buffe pool\n");
1479 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1480 lnet_rtrpool_init(&rtrp[0], 0);
1481 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1485 lnet_rtrpool_init(&rtrp[1], small_pages);
1486 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1490 lnet_rtrpool_init(&rtrp[2], large_pages);
1491 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1496 lnet_net_lock(LNET_LOCK_EX);
1497 the_lnet.ln_routing = 1;
1498 lnet_net_unlock(LNET_LOCK_EX);
1503 lnet_rtrpools_free();
1508 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1510 struct lnet_peer *lp = NULL;
1511 unsigned long now = cfs_time_current();
1512 int cpt = lnet_cpt_of_nid(nid);
1514 LASSERT(!in_interrupt ());
1516 CDEBUG(D_NET, "%s notifying %s: %s\n",
1517 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1518 libcfs_nid2str(nid),
1519 alive ? "up" : "down");
1522 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1523 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1524 libcfs_nid2str(nid), alive ? "birth" : "death",
1525 libcfs_nid2str(ni->ni_nid));
1529 /* can't do predictions... */
1530 if (cfs_time_after(when, now)) {
1531 CWARN("Ignoring prediction from %s of %s %s %ld seconds in the future\n",
1532 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1533 libcfs_nid2str(nid), alive ? "up" : "down",
1534 cfs_duration_sec(cfs_time_sub(when, now)));
1538 if (ni != NULL && !alive && /* LND telling me she's down */
1539 !auto_down) { /* auto-down disabled */
1540 CDEBUG(D_NET, "Auto-down disabled\n");
1546 if (the_lnet.ln_shutdown) {
1547 lnet_net_unlock(cpt);
1551 lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1554 lnet_net_unlock(cpt);
1555 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1559 /* We can't fully trust LND on reporting exact peer last_alive
1560 * if he notifies us about dead peer. For example ksocklnd can
1561 * call us with when == _time_when_the_node_was_booted_ if
1562 * no connections were successfully established */
1563 if (ni != NULL && !alive && when < lp->lp_last_alive)
1564 when = lp->lp_last_alive;
1566 lnet_notify_locked(lp, ni == NULL, alive, when);
1568 lnet_ni_notify_locked(ni, lp);
1570 lnet_peer_decref_locked(lp);
1572 lnet_net_unlock(cpt);
1575 EXPORT_SYMBOL(lnet_notify);
1578 lnet_get_tunables(void)
1585 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1591 lnet_router_checker(void)
1596 time_t now = get_seconds();
1597 int interval = now - last;
1602 /* It's no use to call me again within a sec - all intervals and
1603 * timeouts are measured in seconds */
1604 if (last != 0 && interval < 2)
1608 interval > max(live_router_check_interval,
1609 dead_router_check_interval))
1610 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1611 live_router_check_interval, dead_router_check_interval,
1614 LASSERT(LNET_CPT_NUMBER == 1);
1617 LASSERT(!running); /* recursion check */
1623 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1624 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1626 /* consume all pending events */
1631 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1632 * recursion breaker in LNetEQPoll would fail */
1633 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1634 if (rc == 0) /* no event pending */
1637 /* NB a lost SENT prevents me from pinging a router again */
1638 if (rc == -EOVERFLOW) {
1639 CERROR("Dropped an event!!!\n");
1645 lnet_router_checker_event(&ev);
1648 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1649 lnet_prune_rc_data(1); /* release rcd */
1650 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1655 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1659 version = the_lnet.ln_routers_version;
1660 list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1661 lnet_ping_router_locked(rtr);
1662 LASSERT(version == the_lnet.ln_routers_version);
1667 running = 0; /* lock only needed for the recursion check */
1670 /* NB lnet_peers_start_down depends on me,
1671 * so must be called before any peer creation */
1673 lnet_get_tunables(void)
1677 s = getenv("LNET_ROUTER_PING_TIMEOUT");
1679 router_ping_timeout = atoi(s);
1681 s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1683 live_router_check_interval = atoi(s);
1685 s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1687 dead_router_check_interval = atoi(s);
1689 /* This replaces old lnd_notify mechanism */
1690 check_routers_before_use = 1;
1691 if (dead_router_check_interval <= 0)
1692 dead_router_check_interval = 30;
1696 lnet_rtrpools_free(void)
1701 lnet_rtrpools_alloc(int im_a_arouter)