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 supported\n",
462 libcfs_nid2str(nid1),
463 libcfs_nid2str(nid2));
469 lnet_net_unlock(cpt);
474 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
476 struct lnet_peer *gateway;
477 lnet_remotenet_t *rnet;
479 struct list_head *e1;
480 struct list_head *e2;
482 struct list_head *rn_list;
485 CDEBUG(D_NET, "Del route: net %s : gw %s\n",
486 libcfs_net2str(net), libcfs_nid2str(gw_nid));
488 /* NB Caller may specify either all routes via the given gateway
489 * or a specific route entry actual NIDs) */
491 lnet_net_lock(LNET_LOCK_EX);
492 if (net == LNET_NIDNET(LNET_NID_ANY))
493 rn_list = &the_lnet.ln_remote_nets_hash[0];
495 rn_list = lnet_net2rnethash(net);
498 list_for_each(e1, rn_list) {
499 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
501 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
502 net == rnet->lrn_net))
505 list_for_each(e2, &rnet->lrn_routes) {
506 route = list_entry(e2, lnet_route_t, lr_list);
508 gateway = route->lr_gateway;
509 if (!(gw_nid == LNET_NID_ANY ||
510 gw_nid == gateway->lp_nid))
513 list_del(&route->lr_list);
514 list_del(&route->lr_gwlist);
515 the_lnet.ln_remote_nets_version++;
517 if (list_empty(&rnet->lrn_routes))
518 list_del(&rnet->lrn_list);
522 lnet_rtr_decref_locked(gateway);
523 lnet_peer_decref_locked(gateway);
525 lnet_net_unlock(LNET_LOCK_EX);
527 LIBCFS_FREE(route, sizeof(*route));
530 LIBCFS_FREE(rnet, sizeof(*rnet));
533 lnet_net_lock(LNET_LOCK_EX);
538 if (net == LNET_NIDNET(LNET_NID_ANY) &&
539 ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
540 rn_list = &the_lnet.ln_remote_nets_hash[idx];
543 lnet_net_unlock(LNET_LOCK_EX);
549 lnet_destroy_routes (void)
551 lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
555 lnet_get_route(int idx, __u32 *net, __u32 *hops,
556 lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
558 struct list_head *e1;
559 struct list_head *e2;
560 lnet_remotenet_t *rnet;
564 struct list_head *rn_list;
566 cpt = lnet_net_lock_current();
568 for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
569 rn_list = &the_lnet.ln_remote_nets_hash[i];
570 list_for_each(e1, rn_list) {
571 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
573 list_for_each(e2, &rnet->lrn_routes) {
574 route = list_entry(e2, lnet_route_t,
578 *net = rnet->lrn_net;
579 *hops = route->lr_hops;
580 *priority = route->lr_priority;
581 *gateway = route->lr_gateway->lp_nid;
582 *alive = route->lr_gateway->lp_alive;
583 lnet_net_unlock(cpt);
590 lnet_net_unlock(cpt);
595 lnet_swap_pinginfo(lnet_ping_info_t *info)
598 lnet_ni_status_t *stat;
600 __swab32s(&info->pi_magic);
601 __swab32s(&info->pi_features);
602 __swab32s(&info->pi_pid);
603 __swab32s(&info->pi_nnis);
604 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
605 stat = &info->pi_ni[i];
606 __swab64s(&stat->ns_nid);
607 __swab32s(&stat->ns_status);
613 * parse router-checker pinginfo, record number of down NIs for remote
614 * networks on that router.
617 lnet_parse_rc_info(lnet_rc_data_t *rcd)
619 lnet_ping_info_t *info = rcd->rcd_pinginfo;
620 struct lnet_peer *gw = rcd->rcd_gateway;
626 if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
627 lnet_swap_pinginfo(info);
629 /* NB always racing with network! */
630 if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
631 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
632 libcfs_nid2str(gw->lp_nid), info->pi_magic);
633 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
637 gw->lp_ping_feats = info->pi_features;
638 if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
639 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
640 libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
641 return; /* nothing I can understand */
644 if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
645 return; /* can't carry NI status info */
647 list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
648 int ptl_status = LNET_NI_STATUS_INVALID;
653 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
654 lnet_ni_status_t *stat = &info->pi_ni[i];
655 lnet_nid_t nid = stat->ns_nid;
657 if (nid == LNET_NID_ANY) {
658 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
659 libcfs_nid2str(gw->lp_nid));
660 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
664 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
667 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
668 if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
670 else if (ptl_status != LNET_NI_STATUS_UP)
671 ptl_status = LNET_NI_STATUS_DOWN;
675 if (stat->ns_status == LNET_NI_STATUS_UP) {
676 if (LNET_NIDNET(nid) == rtr->lr_net) {
680 /* ptl NIs are considered down only when
681 * they're all down */
682 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
683 ptl_status = LNET_NI_STATUS_UP;
687 CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
688 libcfs_nid2str(gw->lp_nid), stat->ns_status);
689 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
693 if (up) { /* ignore downed NIs if NI for dest network is up */
697 rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
702 lnet_router_checker_event(lnet_event_t *event)
704 lnet_rc_data_t *rcd = event->md.user_ptr;
705 struct lnet_peer *lp;
707 LASSERT(rcd != NULL);
709 if (event->unlinked) {
710 LNetInvalidateHandle(&rcd->rcd_mdh);
714 LASSERT(event->type == LNET_EVENT_SEND ||
715 event->type == LNET_EVENT_REPLY);
717 lp = rcd->rcd_gateway;
720 /* NB: it's called with holding lnet_res_lock, we have a few
721 * places need to hold both locks at the same time, please take
722 * care of lock ordering */
723 lnet_net_lock(lp->lp_cpt);
724 if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
725 /* ignore if no longer a router or rcd is replaced */
729 if (event->type == LNET_EVENT_SEND) {
730 lp->lp_ping_notsent = 0;
731 if (event->status == 0)
735 /* LNET_EVENT_REPLY */
736 /* A successful REPLY means the router is up. If _any_ comms
737 * to the router fail I assume it's down (this will happen if
738 * we ping alive routers to try to detect router death before
739 * apps get burned). */
741 lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
742 /* The router checker will wake up very shortly and do the
743 * actual notification.
744 * XXX If 'lp' stops being a router before then, it will still
745 * have the notification pending!!! */
747 if (avoid_asym_router_failure && event->status == 0)
748 lnet_parse_rc_info(rcd);
751 lnet_net_unlock(lp->lp_cpt);
755 lnet_wait_known_routerstate(void)
758 struct list_head *entry;
761 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
764 int cpt = lnet_net_lock_current();
767 list_for_each (entry, &the_lnet.ln_routers) {
768 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
770 if (rtr->lp_alive_count == 0) {
776 lnet_net_unlock(cpt);
781 set_current_state(TASK_UNINTERRUPTIBLE);
782 schedule_timeout(cfs_time_seconds(1));
787 lnet_update_ni_status_locked(void)
793 LASSERT(the_lnet.ln_routing);
795 timeout = router_ping_timeout +
796 MAX(live_router_check_interval, dead_router_check_interval);
799 list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
800 if (ni->ni_lnd->lnd_type == LOLND)
803 if (now < ni->ni_last_alive + timeout)
807 /* re-check with lock */
808 if (now < ni->ni_last_alive + timeout) {
813 LASSERT(ni->ni_status != NULL);
815 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
816 CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
817 libcfs_nid2str(ni->ni_nid), timeout);
818 /* NB: so far, this is the only place to set
819 * NI status to "down" */
820 ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
827 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
829 LASSERT(list_empty(&rcd->rcd_list));
830 /* detached from network */
831 LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
833 if (rcd->rcd_gateway != NULL) {
834 int cpt = rcd->rcd_gateway->lp_cpt;
837 lnet_peer_decref_locked(rcd->rcd_gateway);
838 lnet_net_unlock(cpt);
841 if (rcd->rcd_pinginfo != NULL)
842 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
844 LIBCFS_FREE(rcd, sizeof(*rcd));
848 lnet_create_rc_data_locked(lnet_peer_t *gateway)
850 lnet_rc_data_t *rcd = NULL;
851 lnet_ping_info_t *pi;
855 lnet_net_unlock(gateway->lp_cpt);
857 LIBCFS_ALLOC(rcd, sizeof(*rcd));
861 LNetInvalidateHandle(&rcd->rcd_mdh);
862 INIT_LIST_HEAD(&rcd->rcd_list);
864 LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
868 for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
869 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
870 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
872 rcd->rcd_pinginfo = pi;
874 LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
875 rc = LNetMDBind((lnet_md_t){.start = pi,
877 .length = LNET_PINGINFO_SIZE,
878 .threshold = LNET_MD_THRESH_INF,
879 .options = LNET_MD_TRUNCATE,
880 .eq_handle = the_lnet.ln_rc_eqh},
884 CERROR("Can't bind MD: %d\n", rc);
889 lnet_net_lock(gateway->lp_cpt);
890 /* router table changed or someone has created rcd for this gateway */
891 if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
892 lnet_net_unlock(gateway->lp_cpt);
896 lnet_peer_addref_locked(gateway);
897 rcd->rcd_gateway = gateway;
898 gateway->lp_rcd = rcd;
899 gateway->lp_ping_notsent = 0;
905 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
906 rc = LNetMDUnlink(rcd->rcd_mdh);
909 lnet_destroy_rc_data(rcd);
912 lnet_net_lock(gateway->lp_cpt);
913 return gateway->lp_rcd;
917 lnet_router_check_interval (lnet_peer_t *rtr)
921 secs = rtr->lp_alive ? live_router_check_interval :
922 dead_router_check_interval;
930 lnet_ping_router_locked (lnet_peer_t *rtr)
932 lnet_rc_data_t *rcd = NULL;
933 unsigned long now = cfs_time_current();
936 lnet_peer_addref_locked(rtr);
938 if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
939 cfs_time_after(now, rtr->lp_ping_deadline))
940 lnet_notify_locked(rtr, 1, 0, now);
942 /* Run any outstanding notifications */
943 lnet_ni_notify_locked(rtr->lp_ni, rtr);
945 if (!lnet_isrouter(rtr) ||
946 the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
947 /* router table changed or router checker is shutting down */
948 lnet_peer_decref_locked(rtr);
952 rcd = rtr->lp_rcd != NULL ?
953 rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
958 secs = lnet_router_check_interval(rtr);
961 "rtr %s %d: deadline %lu ping_notsent %d alive %d alive_count %d lp_ping_timestamp %lu\n",
962 libcfs_nid2str(rtr->lp_nid), secs,
963 rtr->lp_ping_deadline, rtr->lp_ping_notsent,
964 rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
966 if (secs != 0 && !rtr->lp_ping_notsent &&
967 cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
968 cfs_time_seconds(secs)))) {
970 lnet_process_id_t id;
971 lnet_handle_md_t mdh;
973 id.nid = rtr->lp_nid;
974 id.pid = LUSTRE_SRV_LNET_PID;
975 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
977 rtr->lp_ping_notsent = 1;
978 rtr->lp_ping_timestamp = now;
982 if (rtr->lp_ping_deadline == 0) {
983 rtr->lp_ping_deadline =
984 cfs_time_shift(router_ping_timeout);
987 lnet_net_unlock(rtr->lp_cpt);
989 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
990 LNET_PROTO_PING_MATCHBITS, 0);
992 lnet_net_lock(rtr->lp_cpt);
994 rtr->lp_ping_notsent = 0; /* no event pending */
997 lnet_peer_decref_locked(rtr);
1002 lnet_router_checker_start(void)
1007 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1009 if (check_routers_before_use &&
1010 dead_router_check_interval <= 0) {
1011 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be set if 'check_routers_before_use' is set\n");
1015 if (!the_lnet.ln_routing &&
1016 live_router_check_interval <= 0 &&
1017 dead_router_check_interval <= 0)
1020 sema_init(&the_lnet.ln_rc_signal, 0);
1021 /* EQ size doesn't matter; the callback is guaranteed to get every
1024 rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1025 &the_lnet.ln_rc_eqh);
1027 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1031 the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1032 rc = PTR_ERR(kthread_run(lnet_router_checker,
1033 NULL, "router_checker"));
1034 if (IS_ERR_VALUE(rc)) {
1035 CERROR("Can't start router checker thread: %d\n", rc);
1036 /* block until event callback signals exit */
1037 down(&the_lnet.ln_rc_signal);
1038 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1040 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1044 if (check_routers_before_use) {
1045 /* Note that a helpful side-effect of pinging all known routers
1046 * at startup is that it makes them drop stale connections they
1047 * may have to a previous instance of me. */
1048 lnet_wait_known_routerstate();
1055 lnet_router_checker_stop (void)
1059 if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1062 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1063 the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1065 /* block until event callback signals exit */
1066 down(&the_lnet.ln_rc_signal);
1067 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1069 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1075 lnet_prune_rc_data(int wait_unlink)
1077 lnet_rc_data_t *rcd;
1078 lnet_rc_data_t *tmp;
1080 struct list_head head;
1083 if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1084 list_empty(&the_lnet.ln_rcd_deathrow) &&
1085 list_empty(&the_lnet.ln_rcd_zombie)))
1088 INIT_LIST_HEAD(&head);
1090 lnet_net_lock(LNET_LOCK_EX);
1092 if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1093 /* router checker is stopping, prune all */
1094 list_for_each_entry(lp, &the_lnet.ln_routers,
1096 if (lp->lp_rcd == NULL)
1099 LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1100 list_add(&lp->lp_rcd->rcd_list,
1101 &the_lnet.ln_rcd_deathrow);
1106 /* unlink all RCDs on deathrow list */
1107 list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1109 if (!list_empty(&head)) {
1110 lnet_net_unlock(LNET_LOCK_EX);
1112 list_for_each_entry(rcd, &head, rcd_list)
1113 LNetMDUnlink(rcd->rcd_mdh);
1115 lnet_net_lock(LNET_LOCK_EX);
1118 list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1120 /* release all zombie RCDs */
1121 while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1122 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1124 if (LNetHandleIsInvalid(rcd->rcd_mdh))
1125 list_move(&rcd->rcd_list, &head);
1128 wait_unlink = wait_unlink &&
1129 !list_empty(&the_lnet.ln_rcd_zombie);
1131 lnet_net_unlock(LNET_LOCK_EX);
1133 while (!list_empty(&head)) {
1134 rcd = list_entry(head.next,
1135 lnet_rc_data_t, rcd_list);
1136 list_del_init(&rcd->rcd_list);
1137 lnet_destroy_rc_data(rcd);
1144 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1145 "Waiting for rc buffers to unlink\n");
1146 set_current_state(TASK_UNINTERRUPTIBLE);
1147 schedule_timeout(cfs_time_seconds(1) / 4);
1149 lnet_net_lock(LNET_LOCK_EX);
1152 lnet_net_unlock(LNET_LOCK_EX);
1156 #if defined(LNET_ROUTER)
1159 lnet_router_checker(void *arg)
1162 struct list_head *entry;
1164 cfs_block_allsigs();
1166 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1168 while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1173 cpt = lnet_net_lock_current();
1175 version = the_lnet.ln_routers_version;
1177 list_for_each(entry, &the_lnet.ln_routers) {
1178 rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1180 cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1182 lnet_net_unlock(cpt);
1185 /* the routers list has changed */
1186 if (version != the_lnet.ln_routers_version)
1190 lnet_ping_router_locked(rtr);
1192 /* NB dropped lock */
1193 if (version != the_lnet.ln_routers_version) {
1194 /* the routers list has changed */
1199 if (the_lnet.ln_routing)
1200 lnet_update_ni_status_locked();
1202 lnet_net_unlock(cpt);
1204 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1206 /* Call schedule_timeout() here always adds 1 to load average
1207 * because kernel counts # active tasks as nr_running
1208 * + nr_uninterruptible. */
1209 set_current_state(TASK_INTERRUPTIBLE);
1210 schedule_timeout(cfs_time_seconds(1));
1213 LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1215 lnet_prune_rc_data(1); /* wait for UNLINK */
1217 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1218 up(&the_lnet.ln_rc_signal);
1219 /* The unlink event callback will signal final completion */
1224 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1226 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1228 while (--npages >= 0)
1229 __free_page(rb->rb_kiov[npages].kiov_page);
1231 LIBCFS_FREE(rb, sz);
1235 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1237 int npages = rbp->rbp_npages;
1238 int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1243 LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1249 for (i = 0; i < npages; i++) {
1250 page = alloc_pages_node(
1251 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1252 __GFP_ZERO | GFP_IOFS, 0);
1255 __free_page(rb->rb_kiov[i].kiov_page);
1257 LIBCFS_FREE(rb, sz);
1261 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1262 rb->rb_kiov[i].kiov_offset = 0;
1263 rb->rb_kiov[i].kiov_page = page;
1270 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1272 int npages = rbp->rbp_npages;
1276 if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1279 LASSERT (list_empty(&rbp->rbp_msgs));
1280 LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1282 while (!list_empty(&rbp->rbp_bufs)) {
1283 LASSERT (rbp->rbp_credits > 0);
1285 rb = list_entry(rbp->rbp_bufs.next,
1286 lnet_rtrbuf_t, rb_list);
1287 list_del(&rb->rb_list);
1288 lnet_destroy_rtrbuf(rb, npages);
1292 LASSERT (rbp->rbp_nbuffers == nbuffers);
1293 LASSERT (rbp->rbp_credits == nbuffers);
1295 rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1299 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1304 if (rbp->rbp_nbuffers != 0) {
1305 LASSERT (rbp->rbp_nbuffers == nbufs);
1309 for (i = 0; i < nbufs; i++) {
1310 rb = lnet_new_rtrbuf(rbp, cpt);
1313 CERROR("Failed to allocate %d router bufs of %d pages\n",
1314 nbufs, rbp->rbp_npages);
1318 rbp->rbp_nbuffers++;
1320 rbp->rbp_mincredits++;
1321 list_add(&rb->rb_list, &rbp->rbp_bufs);
1323 /* No allocation "under fire" */
1324 /* Otherwise we'd need code to schedule blocked msgs etc */
1325 LASSERT (!the_lnet.ln_routing);
1328 LASSERT (rbp->rbp_credits == nbufs);
1333 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1335 INIT_LIST_HEAD(&rbp->rbp_msgs);
1336 INIT_LIST_HEAD(&rbp->rbp_bufs);
1338 rbp->rbp_npages = npages;
1339 rbp->rbp_credits = 0;
1340 rbp->rbp_mincredits = 0;
1344 lnet_rtrpools_free(void)
1346 lnet_rtrbufpool_t *rtrp;
1349 if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1352 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1353 lnet_rtrpool_free_bufs(&rtrp[0]);
1354 lnet_rtrpool_free_bufs(&rtrp[1]);
1355 lnet_rtrpool_free_bufs(&rtrp[2]);
1358 cfs_percpt_free(the_lnet.ln_rtrpools);
1359 the_lnet.ln_rtrpools = NULL;
1363 lnet_nrb_tiny_calculate(int npages)
1365 int nrbs = LNET_NRB_TINY;
1367 if (tiny_router_buffers < 0) {
1368 LCONSOLE_ERROR_MSG(0x10c,
1369 "tiny_router_buffers=%d invalid when routing enabled\n",
1370 tiny_router_buffers);
1374 if (tiny_router_buffers > 0)
1375 nrbs = tiny_router_buffers;
1377 nrbs /= LNET_CPT_NUMBER;
1378 return max(nrbs, LNET_NRB_TINY_MIN);
1382 lnet_nrb_small_calculate(int npages)
1384 int nrbs = LNET_NRB_SMALL;
1386 if (small_router_buffers < 0) {
1387 LCONSOLE_ERROR_MSG(0x10c,
1388 "small_router_buffers=%d invalid when routing enabled\n",
1389 small_router_buffers);
1393 if (small_router_buffers > 0)
1394 nrbs = small_router_buffers;
1396 nrbs /= LNET_CPT_NUMBER;
1397 return max(nrbs, LNET_NRB_SMALL_MIN);
1401 lnet_nrb_large_calculate(int npages)
1403 int nrbs = LNET_NRB_LARGE;
1405 if (large_router_buffers < 0) {
1406 LCONSOLE_ERROR_MSG(0x10c,
1407 "large_router_buffers=%d invalid when routing enabled\n",
1408 large_router_buffers);
1412 if (large_router_buffers > 0)
1413 nrbs = large_router_buffers;
1415 nrbs /= LNET_CPT_NUMBER;
1416 return max(nrbs, LNET_NRB_LARGE_MIN);
1420 lnet_rtrpools_alloc(int im_a_router)
1422 lnet_rtrbufpool_t *rtrp;
1423 int large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1424 int small_pages = 1;
1431 if (!strcmp(forwarding, "")) {
1432 /* not set either way */
1435 } else if (!strcmp(forwarding, "disabled")) {
1436 /* explicitly disabled */
1438 } else if (!strcmp(forwarding, "enabled")) {
1439 /* explicitly enabled */
1441 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either 'enabled' or 'disabled'\n");
1445 nrb_tiny = lnet_nrb_tiny_calculate(0);
1449 nrb_small = lnet_nrb_small_calculate(small_pages);
1453 nrb_large = lnet_nrb_large_calculate(large_pages);
1457 the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1459 sizeof(lnet_rtrbufpool_t));
1460 if (the_lnet.ln_rtrpools == NULL) {
1461 LCONSOLE_ERROR_MSG(0x10c,
1462 "Failed to initialize router buffe pool\n");
1466 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1467 lnet_rtrpool_init(&rtrp[0], 0);
1468 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1472 lnet_rtrpool_init(&rtrp[1], small_pages);
1473 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1477 lnet_rtrpool_init(&rtrp[2], large_pages);
1478 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1483 lnet_net_lock(LNET_LOCK_EX);
1484 the_lnet.ln_routing = 1;
1485 lnet_net_unlock(LNET_LOCK_EX);
1490 lnet_rtrpools_free();
1495 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1497 struct lnet_peer *lp = NULL;
1498 unsigned long now = cfs_time_current();
1499 int cpt = lnet_cpt_of_nid(nid);
1501 LASSERT (!in_interrupt ());
1503 CDEBUG (D_NET, "%s notifying %s: %s\n",
1504 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1505 libcfs_nid2str(nid),
1506 alive ? "up" : "down");
1509 LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1510 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1511 libcfs_nid2str(nid), alive ? "birth" : "death",
1512 libcfs_nid2str(ni->ni_nid));
1516 /* can't do predictions... */
1517 if (cfs_time_after(when, now)) {
1518 CWARN("Ignoring prediction from %s of %s %s %ld seconds in the future\n",
1519 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1520 libcfs_nid2str(nid), alive ? "up" : "down",
1521 cfs_duration_sec(cfs_time_sub(when, now)));
1525 if (ni != NULL && !alive && /* LND telling me she's down */
1526 !auto_down) { /* auto-down disabled */
1527 CDEBUG(D_NET, "Auto-down disabled\n");
1533 if (the_lnet.ln_shutdown) {
1534 lnet_net_unlock(cpt);
1538 lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1541 lnet_net_unlock(cpt);
1542 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1546 /* We can't fully trust LND on reporting exact peer last_alive
1547 * if he notifies us about dead peer. For example ksocklnd can
1548 * call us with when == _time_when_the_node_was_booted_ if
1549 * no connections were successfully established */
1550 if (ni != NULL && !alive && when < lp->lp_last_alive)
1551 when = lp->lp_last_alive;
1553 lnet_notify_locked(lp, ni == NULL, alive, when);
1555 lnet_ni_notify_locked(ni, lp);
1557 lnet_peer_decref_locked(lp);
1559 lnet_net_unlock(cpt);
1562 EXPORT_SYMBOL(lnet_notify);
1565 lnet_get_tunables (void)
1573 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1579 lnet_router_checker (void)
1581 static time_t last = 0;
1582 static int running = 0;
1584 time_t now = get_seconds();
1585 int interval = now - last;
1590 /* It's no use to call me again within a sec - all intervals and
1591 * timeouts are measured in seconds */
1592 if (last != 0 && interval < 2)
1596 interval > MAX(live_router_check_interval,
1597 dead_router_check_interval))
1598 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1599 live_router_check_interval, dead_router_check_interval,
1602 LASSERT(LNET_CPT_NUMBER == 1);
1605 LASSERT(!running); /* recursion check */
1611 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1612 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1614 /* consume all pending events */
1619 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1620 * recursion breaker in LNetEQPoll would fail */
1621 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1622 if (rc == 0) /* no event pending */
1625 /* NB a lost SENT prevents me from pinging a router again */
1626 if (rc == -EOVERFLOW) {
1627 CERROR("Dropped an event!!!\n");
1633 lnet_router_checker_event(&ev);
1636 if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1637 lnet_prune_rc_data(1); /* release rcd */
1638 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1643 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1647 version = the_lnet.ln_routers_version;
1648 list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1649 lnet_ping_router_locked(rtr);
1650 LASSERT (version == the_lnet.ln_routers_version);
1655 running = 0; /* lock only needed for the recursion check */
1659 /* NB lnet_peers_start_down depends on me,
1660 * so must be called before any peer creation */
1662 lnet_get_tunables (void)
1666 s = getenv("LNET_ROUTER_PING_TIMEOUT");
1667 if (s != NULL) router_ping_timeout = atoi(s);
1669 s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1670 if (s != NULL) live_router_check_interval = atoi(s);
1672 s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1673 if (s != NULL) dead_router_check_interval = atoi(s);
1675 /* This replaces old lnd_notify mechanism */
1676 check_routers_before_use = 1;
1677 if (dead_router_check_interval <= 0)
1678 dead_router_check_interval = 30;
1682 lnet_rtrpools_free(void)
1687 lnet_rtrpools_alloc(int im_a_arouter)