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 = 0;
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 = 0;
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 static 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, unsigned long when)
112 if (time_before(when, lp->lp_timestamp)) { /* out of date information */
113 CDEBUG(D_NET, "Out of date\n");
117 lp->lp_timestamp = when; /* update timestamp */
118 lp->lp_ping_deadline = 0; /* disable ping timeout */
120 if (lp->lp_alive_count != 0 && /* got old news */
121 (!lp->lp_alive) == (!alive)) { /* new date for old news */
122 CDEBUG(D_NET, "Old news\n");
126 /* Flag that notification is outstanding */
128 lp->lp_alive_count++;
129 lp->lp_alive = !(!alive); /* 1 bit! */
131 lp->lp_notifylnd |= notifylnd;
133 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
135 CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
139 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
144 /* Notify only in 1 thread at any time to ensure ordered notification.
145 * NB individual events can be missed; the only guarantee is that you
146 * always get the most recent news */
148 if (lp->lp_notifying || ni == NULL)
151 lp->lp_notifying = 1;
153 while (lp->lp_notify) {
154 alive = lp->lp_alive;
155 notifylnd = lp->lp_notifylnd;
157 lp->lp_notifylnd = 0;
160 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
161 lnet_net_unlock(lp->lp_cpt);
163 /* A new notification could happen now; I'll handle it
164 * when control returns to me */
166 (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
168 lnet_net_lock(lp->lp_cpt);
172 lp->lp_notifying = 0;
177 lnet_rtr_addref_locked(lnet_peer_t *lp)
179 LASSERT(lp->lp_refcount > 0);
180 LASSERT(lp->lp_rtr_refcount >= 0);
182 /* lnet_net_lock must be exclusively locked */
183 lp->lp_rtr_refcount++;
184 if (lp->lp_rtr_refcount == 1) {
185 struct list_head *pos;
187 /* a simple insertion sort */
188 list_for_each_prev(pos, &the_lnet.ln_routers) {
189 lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
192 if (rtr->lp_nid < lp->lp_nid)
196 list_add(&lp->lp_rtr_list, pos);
197 /* addref for the_lnet.ln_routers */
198 lnet_peer_addref_locked(lp);
199 the_lnet.ln_routers_version++;
204 lnet_rtr_decref_locked(lnet_peer_t *lp)
206 LASSERT(lp->lp_refcount > 0);
207 LASSERT(lp->lp_rtr_refcount > 0);
209 /* lnet_net_lock must be exclusively locked */
210 lp->lp_rtr_refcount--;
211 if (lp->lp_rtr_refcount == 0) {
212 LASSERT(list_empty(&lp->lp_routes));
214 if (lp->lp_rcd != NULL) {
215 list_add(&lp->lp_rcd->rcd_list,
216 &the_lnet.ln_rcd_deathrow);
220 list_del(&lp->lp_rtr_list);
221 /* decref for the_lnet.ln_routers */
222 lnet_peer_decref_locked(lp);
223 the_lnet.ln_routers_version++;
228 lnet_find_net_locked (__u32 net)
230 lnet_remotenet_t *rnet;
231 struct list_head *tmp;
232 struct list_head *rn_list;
234 LASSERT(!the_lnet.ln_shutdown);
236 rn_list = lnet_net2rnethash(net);
237 list_for_each(tmp, rn_list) {
238 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
240 if (rnet->lrn_net == net)
246 static void lnet_shuffle_seed(void)
248 static int seeded = 0;
249 int lnd_type, seed[2];
252 struct list_head *tmp;
257 cfs_get_random_bytes(seed, sizeof(seed));
259 /* Nodes with small feet have little entropy
260 * the NID for this node gives the most entropy in the low bits */
261 list_for_each(tmp, &the_lnet.ln_nis) {
262 ni = list_entry(tmp, lnet_ni_t, ni_list);
263 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
265 if (lnd_type != LOLND)
266 seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
269 do_gettimeofday(&tv);
270 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 */
360 CERROR("Error %d creating route %s %d %s\n", rc,
361 libcfs_net2str(net), hops,
362 libcfs_nid2str(gateway));
367 LASSERT (!the_lnet.ln_shutdown);
369 rnet2 = lnet_find_net_locked(net);
372 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
376 /* Search for a duplicate route (it's a NOOP if it is) */
378 list_for_each (e, &rnet2->lrn_routes) {
379 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
381 if (route2->lr_gateway == route->lr_gateway) {
386 /* our lookups must be true */
387 LASSERT (route2->lr_gateway->lp_nid != gateway);
391 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
392 lnet_add_route_to_rnet(rnet2, route);
394 ni = route->lr_gateway->lp_ni;
395 lnet_net_unlock(LNET_LOCK_EX);
397 /* XXX Assume alive */
398 if (ni->ni_lnd->lnd_notify != NULL)
399 (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
401 lnet_net_lock(LNET_LOCK_EX);
404 /* -1 for notify or !add_route */
405 lnet_peer_decref_locked(route->lr_gateway);
406 lnet_net_unlock(LNET_LOCK_EX);
409 LIBCFS_FREE(route, sizeof(*route));
412 LIBCFS_FREE(rnet, sizeof(*rnet));
418 lnet_check_routes(void)
420 lnet_remotenet_t *rnet;
422 lnet_route_t *route2;
423 struct list_head *e1;
424 struct list_head *e2;
426 struct list_head *rn_list;
429 cpt = lnet_net_lock_current();
431 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
432 rn_list = &the_lnet.ln_remote_nets_hash[i];
433 list_for_each(e1, rn_list) {
434 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
437 list_for_each(e2, &rnet->lrn_routes) {
442 route = list_entry(e2, lnet_route_t,
445 if (route2 == NULL) {
450 if (route->lr_gateway->lp_ni ==
451 route2->lr_gateway->lp_ni)
454 nid1 = route->lr_gateway->lp_nid;
455 nid2 = route2->lr_gateway->lp_nid;
458 lnet_net_unlock(cpt);
460 CERROR("Routes to %s via %s and %s not "
463 libcfs_nid2str(nid1),
464 libcfs_nid2str(nid2));
470 lnet_net_unlock(cpt);
475 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
477 struct lnet_peer *gateway;
478 lnet_remotenet_t *rnet;
480 struct list_head *e1;
481 struct list_head *e2;
483 struct list_head *rn_list;
486 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
487 libcfs_net2str(net), libcfs_nid2str(gw_nid));
489 /* NB Caller may specify either all routes via the given gateway
490 * or a specific route entry actual NIDs) */
492 lnet_net_lock(LNET_LOCK_EX);
493 if (net == LNET_NIDNET(LNET_NID_ANY))
494 rn_list = &the_lnet.ln_remote_nets_hash[0];
496 rn_list = lnet_net2rnethash(net);
499 list_for_each(e1, rn_list) {
500 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
502 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
503 net == rnet->lrn_net))
506 list_for_each(e2, &rnet->lrn_routes) {
507 route = list_entry(e2, lnet_route_t, lr_list);
509 gateway = route->lr_gateway;
510 if (!(gw_nid == LNET_NID_ANY ||
511 gw_nid == gateway->lp_nid))
514 list_del(&route->lr_list);
515 list_del(&route->lr_gwlist);
516 the_lnet.ln_remote_nets_version++;
518 if (list_empty(&rnet->lrn_routes))
519 list_del(&rnet->lrn_list);
523 lnet_rtr_decref_locked(gateway);
524 lnet_peer_decref_locked(gateway);
526 lnet_net_unlock(LNET_LOCK_EX);
528 LIBCFS_FREE(route, sizeof(*route));
531 LIBCFS_FREE(rnet, sizeof(*rnet));
534 lnet_net_lock(LNET_LOCK_EX);
539 if (net == LNET_NIDNET(LNET_NID_ANY) &&
540 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
541 rn_list = &the_lnet.ln_remote_nets_hash[idx];
544 lnet_net_unlock(LNET_LOCK_EX);
550 lnet_destroy_routes (void)
552 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
556 lnet_get_route(int idx, __u32 *net, __u32 *hops,
557 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
559 struct list_head *e1;
560 struct list_head *e2;
561 lnet_remotenet_t *rnet;
565 struct list_head *rn_list;
567 cpt = lnet_net_lock_current();
569 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
570 rn_list = &the_lnet.ln_remote_nets_hash[i];
571 list_for_each(e1, rn_list) {
572 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
574 list_for_each(e2, &rnet->lrn_routes) {
575 route = list_entry(e2, lnet_route_t,
579 *net = rnet->lrn_net;
580 *hops = route->lr_hops;
581 *priority = route->lr_priority;
582 *gateway = route->lr_gateway->lp_nid;
583 *alive = route->lr_gateway->lp_alive;
584 lnet_net_unlock(cpt);
591 lnet_net_unlock(cpt);
596 lnet_swap_pinginfo(lnet_ping_info_t *info)
599 lnet_ni_status_t *stat;
601 __swab32s(&info->pi_magic);
602 __swab32s(&info->pi_features);
603 __swab32s(&info->pi_pid);
604 __swab32s(&info->pi_nnis);
605 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
606 stat = &info->pi_ni[i];
607 __swab64s(&stat->ns_nid);
608 __swab32s(&stat->ns_status);
614 * parse router-checker pinginfo, record number of down NIs for remote
615 * networks on that router.
618 lnet_parse_rc_info(lnet_rc_data_t *rcd)
620 lnet_ping_info_t *info = rcd->rcd_pinginfo;
621 struct lnet_peer *gw = rcd->rcd_gateway;
627 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
628 lnet_swap_pinginfo(info);
630 /* NB always racing with network! */
631 if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
632 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
633 libcfs_nid2str(gw->lp_nid), info->pi_magic);
634 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
638 gw->lp_ping_feats = info->pi_features;
639 if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
640 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
641 libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
642 return; /* nothing I can understand */
645 if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
646 return; /* can't carry NI status info */
648 list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
649 int ptl_status = LNET_NI_STATUS_INVALID;
654 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
655 lnet_ni_status_t *stat = &info->pi_ni[i];
656 lnet_nid_t nid = stat->ns_nid;
658 if (nid == LNET_NID_ANY) {
659 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
660 libcfs_nid2str(gw->lp_nid));
661 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
665 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
668 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
669 if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
671 else if (ptl_status != LNET_NI_STATUS_UP)
672 ptl_status = LNET_NI_STATUS_DOWN;
676 if (stat->ns_status == LNET_NI_STATUS_UP) {
677 if (LNET_NIDNET(nid) == rtr->lr_net) {
681 /* ptl NIs are considered down only when
682 * they're all down */
683 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
684 ptl_status = LNET_NI_STATUS_UP;
688 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
689 libcfs_nid2str(gw->lp_nid), stat->ns_status);
690 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
694 if (up) { /* ignore downed NIs if NI for dest network is up */
698 rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
703 lnet_router_checker_event(lnet_event_t *event)
705 lnet_rc_data_t *rcd = event->md.user_ptr;
706 struct lnet_peer *lp;
708 LASSERT(rcd != NULL);
710 if (event->unlinked) {
711 LNetInvalidateHandle(&rcd->rcd_mdh);
715 LASSERT(event->type == LNET_EVENT_SEND ||
716 event->type == LNET_EVENT_REPLY);
718 lp = rcd->rcd_gateway;
721 /* NB: it's called with holding lnet_res_lock, we have a few
722 * places need to hold both locks at the same time, please take
723 * care of lock ordering */
724 lnet_net_lock(lp->lp_cpt);
725 if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
726 /* ignore if no longer a router or rcd is replaced */
730 if (event->type == LNET_EVENT_SEND) {
731 lp->lp_ping_notsent = 0;
732 if (event->status == 0)
736 /* LNET_EVENT_REPLY */
737 /* A successful REPLY means the router is up. If _any_ comms
738 * to the router fail I assume it's down (this will happen if
739 * we ping alive routers to try to detect router death before
740 * apps get burned). */
742 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
743 /* The router checker will wake up very shortly and do the
744 * actual notification.
745 * XXX If 'lp' stops being a router before then, it will still
746 * have the notification pending!!! */
748 if (avoid_asym_router_failure && event->status == 0)
749 lnet_parse_rc_info(rcd);
752 lnet_net_unlock(lp->lp_cpt);
756 lnet_wait_known_routerstate(void)
759 struct list_head *entry;
762 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
765 int cpt = lnet_net_lock_current();
768 list_for_each (entry, &the_lnet.ln_routers) {
769 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
771 if (rtr->lp_alive_count == 0) {
777 lnet_net_unlock(cpt);
782 set_current_state(TASK_UNINTERRUPTIBLE);
783 schedule_timeout(cfs_time_seconds(1));
788 lnet_update_ni_status_locked(void)
794 LASSERT(the_lnet.ln_routing);
796 timeout = router_ping_timeout +
797 MAX(live_router_check_interval, dead_router_check_interval);
800 list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
801 if (ni->ni_lnd->lnd_type == LOLND)
804 if (now < ni->ni_last_alive + timeout)
808 /* re-check with lock */
809 if (now < ni->ni_last_alive + timeout) {
814 LASSERT(ni->ni_status != NULL);
816 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
817 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
818 libcfs_nid2str(ni->ni_nid), timeout);
819 /* NB: so far, this is the only place to set
820 * NI status to "down" */
821 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
828 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
830 LASSERT(list_empty(&rcd->rcd_list));
831 /* detached from network */
832 LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
834 if (rcd->rcd_gateway != NULL) {
835 int cpt = rcd->rcd_gateway->lp_cpt;
838 lnet_peer_decref_locked(rcd->rcd_gateway);
839 lnet_net_unlock(cpt);
842 if (rcd->rcd_pinginfo != NULL)
843 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
845 LIBCFS_FREE(rcd, sizeof(*rcd));
849 lnet_create_rc_data_locked(lnet_peer_t *gateway)
851 lnet_rc_data_t *rcd = NULL;
852 lnet_ping_info_t *pi;
856 lnet_net_unlock(gateway->lp_cpt);
858 LIBCFS_ALLOC(rcd, sizeof(*rcd));
862 LNetInvalidateHandle(&rcd->rcd_mdh);
863 INIT_LIST_HEAD(&rcd->rcd_list);
865 LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
869 for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
870 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
871 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
873 rcd->rcd_pinginfo = pi;
875 LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
876 rc = LNetMDBind((lnet_md_t){.start = pi,
878 .length = LNET_PINGINFO_SIZE,
879 .threshold = LNET_MD_THRESH_INF,
880 .options = LNET_MD_TRUNCATE,
881 .eq_handle = the_lnet.ln_rc_eqh},
885 CERROR("Can't bind MD: %d\n", rc);
890 lnet_net_lock(gateway->lp_cpt);
891 /* router table changed or someone has created rcd for this gateway */
892 if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
893 lnet_net_unlock(gateway->lp_cpt);
897 lnet_peer_addref_locked(gateway);
898 rcd->rcd_gateway = gateway;
899 gateway->lp_rcd = rcd;
900 gateway->lp_ping_notsent = 0;
906 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
907 rc = LNetMDUnlink(rcd->rcd_mdh);
910 lnet_destroy_rc_data(rcd);
913 lnet_net_lock(gateway->lp_cpt);
914 return gateway->lp_rcd;
918 lnet_router_check_interval (lnet_peer_t *rtr)
922 secs = rtr->lp_alive ? live_router_check_interval :
923 dead_router_check_interval;
931 lnet_ping_router_locked (lnet_peer_t *rtr)
933 lnet_rc_data_t *rcd = NULL;
934 unsigned long now = cfs_time_current();
937 lnet_peer_addref_locked(rtr);
939 if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
940 cfs_time_after(now, rtr->lp_ping_deadline))
941 lnet_notify_locked(rtr, 1, 0, now);
943 /* Run any outstanding notifications */
944 lnet_ni_notify_locked(rtr->lp_ni, rtr);
946 if (!lnet_isrouter(rtr) ||
947 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
948 /* router table changed or router checker is shutting down */
949 lnet_peer_decref_locked(rtr);
953 rcd = rtr->lp_rcd != NULL ?
954 rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
959 secs = lnet_router_check_interval(rtr);
962 "rtr %s %d: deadline %lu ping_notsent %d alive %d "
963 "alive_count %d lp_ping_timestamp %lu\n",
964 libcfs_nid2str(rtr->lp_nid), secs,
965 rtr->lp_ping_deadline, rtr->lp_ping_notsent,
966 rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
968 if (secs != 0 && !rtr->lp_ping_notsent &&
969 cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
970 cfs_time_seconds(secs)))) {
972 lnet_process_id_t id;
973 lnet_handle_md_t mdh;
975 id.nid = rtr->lp_nid;
976 id.pid = LUSTRE_SRV_LNET_PID;
977 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
979 rtr->lp_ping_notsent = 1;
980 rtr->lp_ping_timestamp = now;
984 if (rtr->lp_ping_deadline == 0) {
985 rtr->lp_ping_deadline =
986 cfs_time_shift(router_ping_timeout);
989 lnet_net_unlock(rtr->lp_cpt);
991 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
992 LNET_PROTO_PING_MATCHBITS, 0);
994 lnet_net_lock(rtr->lp_cpt);
996 rtr->lp_ping_notsent = 0; /* no event pending */
999 lnet_peer_decref_locked(rtr);
1004 lnet_router_checker_start(void)
1009 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1011 if (check_routers_before_use &&
1012 dead_router_check_interval <= 0) {
1013 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1014 " set if 'check_routers_before_use' is set"
1019 if (!the_lnet.ln_routing &&
1020 live_router_check_interval <= 0 &&
1021 dead_router_check_interval <= 0)
1024 sema_init(&the_lnet.ln_rc_signal, 0);
1025 /* EQ size doesn't matter; the callback is guaranteed to get every
1028 rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1029 &the_lnet.ln_rc_eqh);
1031 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1035 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1036 rc = PTR_ERR(kthread_run(lnet_router_checker,
1037 NULL, "router_checker"));
1038 if (IS_ERR_VALUE(rc)) {
1039 CERROR("Can't start router checker thread: %d\n", rc);
1040 /* block until event callback signals exit */
1041 down(&the_lnet.ln_rc_signal);
1042 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1044 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1048 if (check_routers_before_use) {
1049 /* Note that a helpful side-effect of pinging all known routers
1050 * at startup is that it makes them drop stale connections they
1051 * may have to a previous instance of me. */
1052 lnet_wait_known_routerstate();
1059 lnet_router_checker_stop (void)
1063 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1066 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1067 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1069 /* block until event callback signals exit */
1070 down(&the_lnet.ln_rc_signal);
1071 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1073 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1079 lnet_prune_rc_data(int wait_unlink)
1081 lnet_rc_data_t *rcd;
1082 lnet_rc_data_t *tmp;
1084 struct list_head head;
1087 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1088 list_empty(&the_lnet.ln_rcd_deathrow) &&
1089 list_empty(&the_lnet.ln_rcd_zombie)))
1092 INIT_LIST_HEAD(&head);
1094 lnet_net_lock(LNET_LOCK_EX);
1096 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1097 /* router checker is stopping, prune all */
1098 list_for_each_entry(lp, &the_lnet.ln_routers,
1100 if (lp->lp_rcd == NULL)
1103 LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1104 list_add(&lp->lp_rcd->rcd_list,
1105 &the_lnet.ln_rcd_deathrow);
1110 /* unlink all RCDs on deathrow list */
1111 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1113 if (!list_empty(&head)) {
1114 lnet_net_unlock(LNET_LOCK_EX);
1116 list_for_each_entry(rcd, &head, rcd_list)
1117 LNetMDUnlink(rcd->rcd_mdh);
1119 lnet_net_lock(LNET_LOCK_EX);
1122 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1124 /* release all zombie RCDs */
1125 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1126 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1128 if (LNetHandleIsInvalid(rcd->rcd_mdh))
1129 list_move(&rcd->rcd_list, &head);
1132 wait_unlink = wait_unlink &&
1133 !list_empty(&the_lnet.ln_rcd_zombie);
1135 lnet_net_unlock(LNET_LOCK_EX);
1137 while (!list_empty(&head)) {
1138 rcd = list_entry(head.next,
1139 lnet_rc_data_t, rcd_list);
1140 list_del_init(&rcd->rcd_list);
1141 lnet_destroy_rc_data(rcd);
1148 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1149 "Waiting for rc buffers to unlink\n");
1150 set_current_state(TASK_UNINTERRUPTIBLE);
1151 schedule_timeout(cfs_time_seconds(1) / 4);
1153 lnet_net_lock(LNET_LOCK_EX);
1156 lnet_net_unlock(LNET_LOCK_EX);
1160 #if defined(LNET_ROUTER)
1163 lnet_router_checker(void *arg)
1166 struct list_head *entry;
1168 cfs_block_allsigs();
1170 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1172 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1177 cpt = lnet_net_lock_current();
1179 version = the_lnet.ln_routers_version;
1181 list_for_each(entry, &the_lnet.ln_routers) {
1182 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1184 cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1186 lnet_net_unlock(cpt);
1189 /* the routers list has changed */
1190 if (version != the_lnet.ln_routers_version)
1194 lnet_ping_router_locked(rtr);
1196 /* NB dropped lock */
1197 if (version != the_lnet.ln_routers_version) {
1198 /* the routers list has changed */
1203 if (the_lnet.ln_routing)
1204 lnet_update_ni_status_locked();
1206 lnet_net_unlock(cpt);
1208 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1210 /* Call schedule_timeout() here always adds 1 to load average
1211 * because kernel counts # active tasks as nr_running
1212 * + nr_uninterruptible. */
1213 set_current_state(TASK_INTERRUPTIBLE);
1214 schedule_timeout(cfs_time_seconds(1));
1217 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1219 lnet_prune_rc_data(1); /* wait for UNLINK */
1221 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1222 up(&the_lnet.ln_rc_signal);
1223 /* The unlink event callback will signal final completion */
1228 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1230 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1232 while (--npages >= 0)
1233 __free_page(rb->rb_kiov[npages].kiov_page);
1235 LIBCFS_FREE(rb, sz);
1239 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1241 int npages = rbp->rbp_npages;
1242 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1247 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1253 for (i = 0; i < npages; i++) {
1254 page = alloc_pages_node(
1255 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1256 __GFP_ZERO | GFP_IOFS, 0);
1259 __free_page(rb->rb_kiov[i].kiov_page);
1261 LIBCFS_FREE(rb, sz);
1265 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1266 rb->rb_kiov[i].kiov_offset = 0;
1267 rb->rb_kiov[i].kiov_page = page;
1274 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1276 int npages = rbp->rbp_npages;
1280 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1283 LASSERT (list_empty(&rbp->rbp_msgs));
1284 LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1286 while (!list_empty(&rbp->rbp_bufs)) {
1287 LASSERT (rbp->rbp_credits > 0);
1289 rb = list_entry(rbp->rbp_bufs.next,
1290 lnet_rtrbuf_t, rb_list);
1291 list_del(&rb->rb_list);
1292 lnet_destroy_rtrbuf(rb, npages);
1296 LASSERT (rbp->rbp_nbuffers == nbuffers);
1297 LASSERT (rbp->rbp_credits == nbuffers);
1299 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1303 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1308 if (rbp->rbp_nbuffers != 0) {
1309 LASSERT (rbp->rbp_nbuffers == nbufs);
1313 for (i = 0; i < nbufs; i++) {
1314 rb = lnet_new_rtrbuf(rbp, cpt);
1317 CERROR("Failed to allocate %d router bufs of %d pages\n",
1318 nbufs, rbp->rbp_npages);
1322 rbp->rbp_nbuffers++;
1324 rbp->rbp_mincredits++;
1325 list_add(&rb->rb_list, &rbp->rbp_bufs);
1327 /* No allocation "under fire" */
1328 /* Otherwise we'd need code to schedule blocked msgs etc */
1329 LASSERT (!the_lnet.ln_routing);
1332 LASSERT (rbp->rbp_credits == nbufs);
1337 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1339 INIT_LIST_HEAD(&rbp->rbp_msgs);
1340 INIT_LIST_HEAD(&rbp->rbp_bufs);
1342 rbp->rbp_npages = npages;
1343 rbp->rbp_credits = 0;
1344 rbp->rbp_mincredits = 0;
1348 lnet_rtrpools_free(void)
1350 lnet_rtrbufpool_t *rtrp;
1353 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1356 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1357 lnet_rtrpool_free_bufs(&rtrp[0]);
1358 lnet_rtrpool_free_bufs(&rtrp[1]);
1359 lnet_rtrpool_free_bufs(&rtrp[2]);
1362 cfs_percpt_free(the_lnet.ln_rtrpools);
1363 the_lnet.ln_rtrpools = NULL;
1367 lnet_nrb_tiny_calculate(int npages)
1369 int nrbs = LNET_NRB_TINY;
1371 if (tiny_router_buffers < 0) {
1372 LCONSOLE_ERROR_MSG(0x10c,
1373 "tiny_router_buffers=%d invalid when "
1374 "routing enabled\n", tiny_router_buffers);
1378 if (tiny_router_buffers > 0)
1379 nrbs = tiny_router_buffers;
1381 nrbs /= LNET_CPT_NUMBER;
1382 return max(nrbs, LNET_NRB_TINY_MIN);
1386 lnet_nrb_small_calculate(int npages)
1388 int nrbs = LNET_NRB_SMALL;
1390 if (small_router_buffers < 0) {
1391 LCONSOLE_ERROR_MSG(0x10c,
1392 "small_router_buffers=%d invalid when "
1393 "routing enabled\n", small_router_buffers);
1397 if (small_router_buffers > 0)
1398 nrbs = small_router_buffers;
1400 nrbs /= LNET_CPT_NUMBER;
1401 return max(nrbs, LNET_NRB_SMALL_MIN);
1405 lnet_nrb_large_calculate(int npages)
1407 int nrbs = LNET_NRB_LARGE;
1409 if (large_router_buffers < 0) {
1410 LCONSOLE_ERROR_MSG(0x10c,
1411 "large_router_buffers=%d invalid when "
1412 "routing enabled\n", large_router_buffers);
1416 if (large_router_buffers > 0)
1417 nrbs = large_router_buffers;
1419 nrbs /= LNET_CPT_NUMBER;
1420 return max(nrbs, LNET_NRB_LARGE_MIN);
1424 lnet_rtrpools_alloc(int im_a_router)
1426 lnet_rtrbufpool_t *rtrp;
1427 int large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1428 int small_pages = 1;
1435 if (!strcmp(forwarding, "")) {
1436 /* not set either way */
1439 } else if (!strcmp(forwarding, "disabled")) {
1440 /* explicitly disabled */
1442 } else if (!strcmp(forwarding, "enabled")) {
1443 /* explicitly enabled */
1445 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1446 "'enabled' or 'disabled'\n");
1450 nrb_tiny = lnet_nrb_tiny_calculate(0);
1454 nrb_small = lnet_nrb_small_calculate(small_pages);
1458 nrb_large = lnet_nrb_large_calculate(large_pages);
1462 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1464 sizeof(lnet_rtrbufpool_t));
1465 if (the_lnet.ln_rtrpools == NULL) {
1466 LCONSOLE_ERROR_MSG(0x10c,
1467 "Failed to initialize router buffe pool\n");
1471 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1472 lnet_rtrpool_init(&rtrp[0], 0);
1473 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1477 lnet_rtrpool_init(&rtrp[1], small_pages);
1478 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1482 lnet_rtrpool_init(&rtrp[2], large_pages);
1483 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1488 lnet_net_lock(LNET_LOCK_EX);
1489 the_lnet.ln_routing = 1;
1490 lnet_net_unlock(LNET_LOCK_EX);
1495 lnet_rtrpools_free();
1500 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1502 struct lnet_peer *lp = NULL;
1503 unsigned long now = cfs_time_current();
1504 int cpt = lnet_cpt_of_nid(nid);
1506 LASSERT (!in_interrupt ());
1508 CDEBUG (D_NET, "%s notifying %s: %s\n",
1509 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1510 libcfs_nid2str(nid),
1511 alive ? "up" : "down");
1514 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1515 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1516 libcfs_nid2str(nid), alive ? "birth" : "death",
1517 libcfs_nid2str(ni->ni_nid));
1521 /* can't do predictions... */
1522 if (cfs_time_after(when, now)) {
1523 CWARN ("Ignoring prediction from %s of %s %s "
1524 "%ld seconds in the future\n",
1525 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1526 libcfs_nid2str(nid), alive ? "up" : "down",
1527 cfs_duration_sec(cfs_time_sub(when, now)));
1531 if (ni != NULL && !alive && /* LND telling me she's down */
1532 !auto_down) { /* auto-down disabled */
1533 CDEBUG(D_NET, "Auto-down disabled\n");
1539 if (the_lnet.ln_shutdown) {
1540 lnet_net_unlock(cpt);
1544 lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1547 lnet_net_unlock(cpt);
1548 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1552 /* We can't fully trust LND on reporting exact peer last_alive
1553 * if he notifies us about dead peer. For example ksocklnd can
1554 * call us with when == _time_when_the_node_was_booted_ if
1555 * no connections were successfully established */
1556 if (ni != NULL && !alive && when < lp->lp_last_alive)
1557 when = lp->lp_last_alive;
1559 lnet_notify_locked(lp, ni == NULL, alive, when);
1561 lnet_ni_notify_locked(ni, lp);
1563 lnet_peer_decref_locked(lp);
1565 lnet_net_unlock(cpt);
1568 EXPORT_SYMBOL(lnet_notify);
1571 lnet_get_tunables (void)
1579 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1585 lnet_router_checker (void)
1587 static time_t last = 0;
1588 static int running = 0;
1590 time_t now = get_seconds();
1591 int interval = now - last;
1596 /* It's no use to call me again within a sec - all intervals and
1597 * timeouts are measured in seconds */
1598 if (last != 0 && interval < 2)
1602 interval > MAX(live_router_check_interval,
1603 dead_router_check_interval))
1604 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1605 live_router_check_interval, dead_router_check_interval,
1608 LASSERT(LNET_CPT_NUMBER == 1);
1611 LASSERT(!running); /* recursion check */
1617 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1618 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1620 /* consume all pending events */
1625 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1626 * recursion breaker in LNetEQPoll would fail */
1627 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1628 if (rc == 0) /* no event pending */
1631 /* NB a lost SENT prevents me from pinging a router again */
1632 if (rc == -EOVERFLOW) {
1633 CERROR("Dropped an event!!!\n");
1639 lnet_router_checker_event(&ev);
1642 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1643 lnet_prune_rc_data(1); /* release rcd */
1644 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1649 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1653 version = the_lnet.ln_routers_version;
1654 list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1655 lnet_ping_router_locked(rtr);
1656 LASSERT (version == the_lnet.ln_routers_version);
1661 running = 0; /* lock only needed for the recursion check */
1665 /* NB lnet_peers_start_down depends on me,
1666 * so must be called before any peer creation */
1668 lnet_get_tunables (void)
1672 s = getenv("LNET_ROUTER_PING_TIMEOUT");
1673 if (s != NULL) router_ping_timeout = atoi(s);
1675 s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1676 if (s != NULL) live_router_check_interval = atoi(s);
1678 s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1679 if (s != NULL) dead_router_check_interval = atoi(s);
1681 /* This replaces old lnd_notify mechanism */
1682 check_routers_before_use = 1;
1683 if (dead_router_check_interval <= 0)
1684 dead_router_check_interval = 30;
1688 lnet_rtrpools_free(void)
1693 lnet_rtrpools_alloc(int im_a_arouter)