Whamcloud - gitweb
LU-6209 lnet: Delete all obsolete LND drivers
[fs/lustre-release.git] / lnet / lnet / router.c
1 /*
2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
3  *
4  * Copyright (c) 2011, 2014, Intel Corporation.
5  *
6  *   This file is part of Portals
7  *   http://sourceforge.net/projects/sandiaportals/
8  *
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.
12  *
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.
17  *
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.
21  *
22  */
23
24 #define DEBUG_SUBSYSTEM S_LNET
25 #include <lnet/lib-lnet.h>
26
27 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
28 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
29 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
30 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
31 #define LNET_NRB_SMALL_PAGES    1
32 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
33 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
34 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \
35                                   PAGE_CACHE_SHIFT)
36
37 static char *forwarding = "";
38 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
39                 "Explicitly enable/disable forwarding between networks");
40
41 static int tiny_router_buffers;
42 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
43                 "# of 0 payload messages to buffer in the router");
44 static int small_router_buffers;
45 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
46                 "# of small (1 page) messages to buffer in the router");
47 static int large_router_buffers;
48 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
49                 "# of large messages to buffer in the router");
50 static int peer_buffer_credits = 0;
51 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
52                 "# router buffer credits per peer");
53
54 static int auto_down = 1;
55 CFS_MODULE_PARM(auto_down, "i", int, 0444,
56                 "Automatically mark peers down on comms error");
57
58 int
59 lnet_peer_buffer_credits(lnet_ni_t *ni)
60 {
61         /* NI option overrides LNet default */
62         if (ni->ni_peerrtrcredits > 0)
63                 return ni->ni_peerrtrcredits;
64         if (peer_buffer_credits > 0)
65                 return peer_buffer_credits;
66
67         /* As an approximation, allow this peer the same number of router
68          * buffers as it is allowed outstanding sends */
69         return ni->ni_peertxcredits;
70 }
71
72 /* forward ref's */
73 static int lnet_router_checker(void *);
74
75 static int check_routers_before_use = 0;
76 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
77                 "Assume routers are down and ping them before use");
78
79 int avoid_asym_router_failure = 1;
80 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
81                 "Avoid asymmetrical router failures (0 to disable)");
82
83 static int dead_router_check_interval = 60;
84 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
85                 "Seconds between dead router health checks (<= 0 to disable)");
86
87 static int live_router_check_interval = 60;
88 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
89                 "Seconds between live router health checks (<= 0 to disable)");
90
91 static int router_ping_timeout = 50;
92 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
93                 "Seconds to wait for the reply to a router health query");
94
95 int
96 lnet_peers_start_down(void)
97 {
98         return check_routers_before_use;
99 }
100
101 void
102 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
103 {
104         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
105                 CDEBUG(D_NET, "Out of date\n");
106                 return;
107         }
108
109         lp->lp_timestamp = when;                /* update timestamp */
110         lp->lp_ping_deadline = 0;               /* disable ping timeout */
111
112         if (lp->lp_alive_count != 0 &&          /* got old news */
113             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
114                 CDEBUG(D_NET, "Old news\n");
115                 return;
116         }
117
118         /* Flag that notification is outstanding */
119
120         lp->lp_alive_count++;
121         lp->lp_alive = !(!alive);               /* 1 bit! */
122         lp->lp_notify = 1;
123         lp->lp_notifylnd |= notifylnd;
124         if (lp->lp_alive)
125                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
126
127         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
128 }
129
130 static void
131 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
132 {
133         int        alive;
134         int        notifylnd;
135
136         /* Notify only in 1 thread at any time to ensure ordered notification.
137          * NB individual events can be missed; the only guarantee is that you
138          * always get the most recent news */
139
140         if (lp->lp_notifying)
141                 return;
142
143         lp->lp_notifying = 1;
144
145         while (lp->lp_notify) {
146                 alive     = lp->lp_alive;
147                 notifylnd = lp->lp_notifylnd;
148
149                 lp->lp_notifylnd = 0;
150                 lp->lp_notify    = 0;
151
152                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
153                         lnet_net_unlock(lp->lp_cpt);
154
155                         /* A new notification could happen now; I'll handle it
156                          * when control returns to me */
157
158                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
159
160                         lnet_net_lock(lp->lp_cpt);
161                 }
162         }
163
164         lp->lp_notifying = 0;
165 }
166
167
168 static void
169 lnet_rtr_addref_locked(lnet_peer_t *lp)
170 {
171         LASSERT(lp->lp_refcount > 0);
172         LASSERT(lp->lp_rtr_refcount >= 0);
173
174         /* lnet_net_lock must be exclusively locked */
175         lp->lp_rtr_refcount++;
176         if (lp->lp_rtr_refcount == 1) {
177                 struct list_head *pos;
178
179                 /* a simple insertion sort */
180                 list_for_each_prev(pos, &the_lnet.ln_routers) {
181                         lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
182                                                       lp_rtr_list);
183
184                         if (rtr->lp_nid < lp->lp_nid)
185                                 break;
186                 }
187
188                 list_add(&lp->lp_rtr_list, pos);
189                 /* addref for the_lnet.ln_routers */
190                 lnet_peer_addref_locked(lp);
191                 the_lnet.ln_routers_version++;
192         }
193 }
194
195 static void
196 lnet_rtr_decref_locked(lnet_peer_t *lp)
197 {
198         LASSERT(lp->lp_refcount > 0);
199         LASSERT(lp->lp_rtr_refcount > 0);
200
201         /* lnet_net_lock must be exclusively locked */
202         lp->lp_rtr_refcount--;
203         if (lp->lp_rtr_refcount == 0) {
204                 LASSERT(list_empty(&lp->lp_routes));
205
206                 if (lp->lp_rcd != NULL) {
207                         list_add(&lp->lp_rcd->rcd_list,
208                                  &the_lnet.ln_rcd_deathrow);
209                         lp->lp_rcd = NULL;
210                 }
211
212                 list_del(&lp->lp_rtr_list);
213                 /* decref for the_lnet.ln_routers */
214                 lnet_peer_decref_locked(lp);
215                 the_lnet.ln_routers_version++;
216         }
217 }
218
219 lnet_remotenet_t *
220 lnet_find_net_locked (__u32 net)
221 {
222         lnet_remotenet_t *rnet;
223         struct list_head *tmp;
224         struct list_head *rn_list;
225
226         LASSERT(!the_lnet.ln_shutdown);
227
228         rn_list = lnet_net2rnethash(net);
229         list_for_each(tmp, rn_list) {
230                 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
231
232                 if (rnet->lrn_net == net)
233                         return rnet;
234         }
235         return NULL;
236 }
237
238 static void lnet_shuffle_seed(void)
239 {
240         static int seeded;
241         __u32 lnd_type;
242         __u32 seed[2];
243         struct timeval tv;
244         lnet_ni_t *ni;
245         struct list_head *tmp;
246
247         if (seeded)
248                 return;
249
250         cfs_get_random_bytes(seed, sizeof(seed));
251
252         /* Nodes with small feet have little entropy
253          * the NID for this node gives the most entropy in the low bits */
254         list_for_each(tmp, &the_lnet.ln_nis) {
255                 ni = list_entry(tmp, lnet_ni_t, ni_list);
256                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
257
258                 if (lnd_type != LOLND)
259                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
260         }
261
262         do_gettimeofday(&tv);
263         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
264         seeded = 1;
265         return;
266 }
267
268 /* NB expects LNET_LOCK held */
269 static void
270 lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
271 {
272         unsigned int      len = 0;
273         unsigned int      offset = 0;
274         struct list_head *e;
275
276         lnet_shuffle_seed();
277
278         list_for_each(e, &rnet->lrn_routes) {
279                 len++;
280         }
281
282         /* len+1 positions to add a new entry, also prevents division by 0 */
283         offset = cfs_rand() % (len + 1);
284         list_for_each(e, &rnet->lrn_routes) {
285                 if (offset == 0)
286                         break;
287                 offset--;
288         }
289         list_add(&route->lr_list, e);
290         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
291
292         the_lnet.ln_remote_nets_version++;
293         lnet_rtr_addref_locked(route->lr_gateway);
294 }
295
296 int
297 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
298                unsigned int priority)
299 {
300         struct list_head        *e;
301         lnet_remotenet_t        *rnet;
302         lnet_remotenet_t        *rnet2;
303         lnet_route_t            *route;
304         lnet_ni_t               *ni;
305         int                     add_route;
306         int                     rc;
307
308         CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
309                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
310
311         if (gateway == LNET_NID_ANY ||
312             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
313             net == LNET_NIDNET(LNET_NID_ANY) ||
314             LNET_NETTYP(net) == LOLND ||
315             LNET_NIDNET(gateway) == net ||
316             hops < 1 || hops > 255)
317                 return -EINVAL;
318
319         if (lnet_islocalnet(net))       /* it's a local network */
320                 return -EEXIST;
321
322         /* Assume net, route, all new */
323         LIBCFS_ALLOC(route, sizeof(*route));
324         LIBCFS_ALLOC(rnet, sizeof(*rnet));
325         if (route == NULL || rnet == NULL) {
326                 CERROR("Out of memory creating route %s %d %s\n",
327                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
328                 if (route != NULL)
329                         LIBCFS_FREE(route, sizeof(*route));
330                 if (rnet != NULL)
331                         LIBCFS_FREE(rnet, sizeof(*rnet));
332                 return -ENOMEM;
333         }
334
335         INIT_LIST_HEAD(&rnet->lrn_routes);
336         rnet->lrn_net = net;
337         route->lr_hops = hops;
338         route->lr_net = net;
339         route->lr_priority = priority;
340
341         lnet_net_lock(LNET_LOCK_EX);
342
343         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
344         if (rc != 0) {
345                 lnet_net_unlock(LNET_LOCK_EX);
346
347                 LIBCFS_FREE(route, sizeof(*route));
348                 LIBCFS_FREE(rnet, sizeof(*rnet));
349
350                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
351                         return rc;       /* ignore the route entry */
352                 CERROR("Error %d creating route %s %d %s\n", rc,
353                         libcfs_net2str(net), hops,
354                         libcfs_nid2str(gateway));
355                 return rc;
356         }
357
358         LASSERT(!the_lnet.ln_shutdown);
359
360         rnet2 = lnet_find_net_locked(net);
361         if (rnet2 == NULL) {
362                 /* new network */
363                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
364                 rnet2 = rnet;
365         }
366
367         /* Search for a duplicate route (it's a NOOP if it is) */
368         add_route = 1;
369         list_for_each(e, &rnet2->lrn_routes) {
370                 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
371
372                 if (route2->lr_gateway == route->lr_gateway) {
373                         add_route = 0;
374                         break;
375                 }
376
377                 /* our lookups must be true */
378                 LASSERT(route2->lr_gateway->lp_nid != gateway);
379         }
380
381         if (add_route) {
382                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
383                 lnet_add_route_to_rnet(rnet2, route);
384
385                 ni = route->lr_gateway->lp_ni;
386                 lnet_net_unlock(LNET_LOCK_EX);
387
388                 /* XXX Assume alive */
389                 if (ni->ni_lnd->lnd_notify != NULL)
390                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
391
392                 lnet_net_lock(LNET_LOCK_EX);
393         }
394
395         /* -1 for notify or !add_route */
396         lnet_peer_decref_locked(route->lr_gateway);
397         lnet_net_unlock(LNET_LOCK_EX);
398
399         rc = 0;
400
401         if (!add_route) {
402                 rc = -EEXIST;
403                 LIBCFS_FREE(route, sizeof(*route));
404         }
405
406         if (rnet != rnet2)
407                 LIBCFS_FREE(rnet, sizeof(*rnet));
408
409         /* indicate to startup the router checker if configured */
410         wake_up(&the_lnet.ln_rc_waitq);
411
412         return rc;
413 }
414
415 int
416 lnet_check_routes(void)
417 {
418         lnet_remotenet_t *rnet;
419         lnet_route_t     *route;
420         lnet_route_t     *route2;
421         struct list_head *e1;
422         struct list_head *e2;
423         int               cpt;
424         struct list_head *rn_list;
425         int               i;
426
427         cpt = lnet_net_lock_current();
428
429         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
430                 rn_list = &the_lnet.ln_remote_nets_hash[i];
431                 list_for_each(e1, rn_list) {
432                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
433
434                         route2 = NULL;
435                         list_for_each(e2, &rnet->lrn_routes) {
436                                 lnet_nid_t      nid1;
437                                 lnet_nid_t      nid2;
438                                 int             net;
439
440                                 route = list_entry(e2, lnet_route_t,
441                                                    lr_list);
442
443                                 if (route2 == NULL) {
444                                         route2 = route;
445                                         continue;
446                                 }
447
448                                 if (route->lr_gateway->lp_ni ==
449                                     route2->lr_gateway->lp_ni)
450                                         continue;
451
452                                 nid1 = route->lr_gateway->lp_nid;
453                                 nid2 = route2->lr_gateway->lp_nid;
454                                 net = rnet->lrn_net;
455
456                                 lnet_net_unlock(cpt);
457
458                                 CERROR("Routes to %s via %s and %s not "
459                                        "supported\n",
460                                        libcfs_net2str(net),
461                                        libcfs_nid2str(nid1),
462                                        libcfs_nid2str(nid2));
463                                 return -EINVAL;
464                         }
465                 }
466         }
467
468         lnet_net_unlock(cpt);
469         return 0;
470 }
471
472 int
473 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
474 {
475         struct lnet_peer        *gateway;
476         lnet_remotenet_t        *rnet;
477         lnet_route_t            *route;
478         struct list_head        *e1;
479         struct list_head        *e2;
480         int                     rc = -ENOENT;
481         struct list_head        *rn_list;
482         int                     idx = 0;
483
484         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
485                libcfs_net2str(net), libcfs_nid2str(gw_nid));
486
487         /* NB Caller may specify either all routes via the given gateway
488          * or a specific route entry actual NIDs) */
489
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];
493         else
494                 rn_list = lnet_net2rnethash(net);
495
496 again:
497         list_for_each(e1, rn_list) {
498                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
499
500                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
501                         net == rnet->lrn_net))
502                         continue;
503
504                 list_for_each(e2, &rnet->lrn_routes) {
505                         route = list_entry(e2, lnet_route_t, lr_list);
506
507                         gateway = route->lr_gateway;
508                         if (!(gw_nid == LNET_NID_ANY ||
509                               gw_nid == gateway->lp_nid))
510                                 continue;
511
512                         list_del(&route->lr_list);
513                         list_del(&route->lr_gwlist);
514                         the_lnet.ln_remote_nets_version++;
515
516                         if (list_empty(&rnet->lrn_routes))
517                                 list_del(&rnet->lrn_list);
518                         else
519                                 rnet = NULL;
520
521                         lnet_rtr_decref_locked(gateway);
522                         lnet_peer_decref_locked(gateway);
523
524                         lnet_net_unlock(LNET_LOCK_EX);
525
526                         LIBCFS_FREE(route, sizeof(*route));
527
528                         if (rnet != NULL)
529                                 LIBCFS_FREE(rnet, sizeof(*rnet));
530
531                         rc = 0;
532                         lnet_net_lock(LNET_LOCK_EX);
533                         goto again;
534                 }
535         }
536
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];
540                 goto again;
541         }
542         lnet_net_unlock(LNET_LOCK_EX);
543
544         return rc;
545 }
546
547 void
548 lnet_destroy_routes (void)
549 {
550         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
551 }
552
553 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
554 {
555         int i, rc = -ENOENT, lidx, j;
556
557         if (the_lnet.ln_rtrpools == NULL)
558                 return rc;
559
560         for (i = 0; i < LNET_NRBPOOLS; i++) {
561                 lnet_rtrbufpool_t *rbp;
562
563                 lnet_net_lock(LNET_LOCK_EX);
564                 lidx = idx;
565                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
566                         if (lidx-- == 0) {
567                                 rc = 0;
568                                 pool_cfg->pl_pools[i].pl_npages =
569                                         rbp[i].rbp_npages;
570                                 pool_cfg->pl_pools[i].pl_nbuffers =
571                                         rbp[i].rbp_nbuffers;
572                                 pool_cfg->pl_pools[i].pl_credits =
573                                         rbp[i].rbp_credits;
574                                 pool_cfg->pl_pools[i].pl_mincredits =
575                                         rbp[i].rbp_mincredits;
576                                 break;
577                         }
578                 }
579                 lnet_net_unlock(LNET_LOCK_EX);
580         }
581
582         lnet_net_lock(LNET_LOCK_EX);
583         pool_cfg->pl_routing = the_lnet.ln_routing;
584         lnet_net_unlock(LNET_LOCK_EX);
585
586         return rc;
587 }
588
589 int
590 lnet_get_route(int idx, __u32 *net, __u32 *hops,
591                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
592 {
593         struct list_head *e1;
594         struct list_head *e2;
595         lnet_remotenet_t *rnet;
596         lnet_route_t     *route;
597         int               cpt;
598         int               i;
599         struct list_head *rn_list;
600
601         cpt = lnet_net_lock_current();
602
603         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
604                 rn_list = &the_lnet.ln_remote_nets_hash[i];
605                 list_for_each(e1, rn_list) {
606                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
607
608                         list_for_each(e2, &rnet->lrn_routes) {
609                                 route = list_entry(e2, lnet_route_t,
610                                                    lr_list);
611
612                                 if (idx-- == 0) {
613                                         *net      = rnet->lrn_net;
614                                         *hops     = route->lr_hops;
615                                         *priority = route->lr_priority;
616                                         *gateway  = route->lr_gateway->lp_nid;
617                                         *alive    =
618                                                 route->lr_gateway->lp_alive &&
619                                                         !route->lr_downis;
620                                         lnet_net_unlock(cpt);
621                                         return 0;
622                                 }
623                         }
624                 }
625         }
626
627         lnet_net_unlock(cpt);
628         return -ENOENT;
629 }
630
631 void
632 lnet_swap_pinginfo(lnet_ping_info_t *info)
633 {
634         int               i;
635         lnet_ni_status_t *stat;
636
637         __swab32s(&info->pi_magic);
638         __swab32s(&info->pi_features);
639         __swab32s(&info->pi_pid);
640         __swab32s(&info->pi_nnis);
641         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
642                 stat = &info->pi_ni[i];
643                 __swab64s(&stat->ns_nid);
644                 __swab32s(&stat->ns_status);
645         }
646         return;
647 }
648
649 /**
650  * parse router-checker pinginfo, record number of down NIs for remote
651  * networks on that router.
652  */
653 static void
654 lnet_parse_rc_info(lnet_rc_data_t *rcd)
655 {
656         lnet_ping_info_t        *info = rcd->rcd_pinginfo;
657         struct lnet_peer        *gw   = rcd->rcd_gateway;
658         lnet_route_t            *rte;
659
660         if (!gw->lp_alive)
661                 return;
662
663         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
664                 lnet_swap_pinginfo(info);
665
666         /* NB always racing with network! */
667         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
668                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
669                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
670                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
671                 return;
672         }
673
674         gw->lp_ping_feats = info->pi_features;
675         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
676                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
677                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
678                 return; /* nothing I can understand */
679         }
680
681         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
682                 return; /* can't carry NI status info */
683
684         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
685                 int     down = 0;
686                 int     up = 0;
687                 int     i;
688
689                 if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
690                         rte->lr_downis = 1;
691                         continue;
692                 }
693
694                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
695                         lnet_ni_status_t *stat = &info->pi_ni[i];
696                         lnet_nid_t       nid = stat->ns_nid;
697
698                         if (nid == LNET_NID_ANY) {
699                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
700                                        libcfs_nid2str(gw->lp_nid));
701                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
702                                 return;
703                         }
704
705                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
706                                 continue;
707
708                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
709                                 down++;
710                                 continue;
711                         }
712
713                         if (stat->ns_status == LNET_NI_STATUS_UP) {
714                                 if (LNET_NIDNET(nid) == rte->lr_net) {
715                                         up = 1;
716                                         break;
717                                 }
718                                 continue;
719                         }
720
721                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
722                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
723                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
724                         return;
725                 }
726
727                 if (up) { /* ignore downed NIs if NI for dest network is up */
728                         rte->lr_downis = 0;
729                         continue;
730                 }
731                 /* if @down is zero and this route is single-hop, it means
732                  * we can't find NI for target network */
733                 if (down == 0 && rte->lr_hops == 1)
734                         down = 1;
735
736                 rte->lr_downis = down;
737         }
738 }
739
740 static void
741 lnet_router_checker_event(lnet_event_t *event)
742 {
743         lnet_rc_data_t          *rcd = event->md.user_ptr;
744         struct lnet_peer        *lp;
745
746         LASSERT(rcd != NULL);
747
748         if (event->unlinked) {
749                 LNetInvalidateHandle(&rcd->rcd_mdh);
750                 return;
751         }
752
753         LASSERT(event->type == LNET_EVENT_SEND ||
754                 event->type == LNET_EVENT_REPLY);
755
756         lp = rcd->rcd_gateway;
757         LASSERT(lp != NULL);
758
759          /* NB: it's called with holding lnet_res_lock, we have a few
760           * places need to hold both locks at the same time, please take
761           * care of lock ordering */
762         lnet_net_lock(lp->lp_cpt);
763         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
764                 /* ignore if no longer a router or rcd is replaced */
765                 goto out;
766         }
767
768         if (event->type == LNET_EVENT_SEND) {
769                 lp->lp_ping_notsent = 0;
770                 if (event->status == 0)
771                         goto out;
772         }
773
774         /* LNET_EVENT_REPLY */
775         /* A successful REPLY means the router is up.  If _any_ comms
776          * to the router fail I assume it's down (this will happen if
777          * we ping alive routers to try to detect router death before
778          * apps get burned). */
779
780         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
781         /* The router checker will wake up very shortly and do the
782          * actual notification.
783          * XXX If 'lp' stops being a router before then, it will still
784          * have the notification pending!!! */
785
786         if (avoid_asym_router_failure && event->status == 0)
787                 lnet_parse_rc_info(rcd);
788
789  out:
790         lnet_net_unlock(lp->lp_cpt);
791 }
792
793 static void
794 lnet_wait_known_routerstate(void)
795 {
796         lnet_peer_t      *rtr;
797         struct list_head *entry;
798         int               all_known;
799
800         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
801
802         for (;;) {
803                 int cpt = lnet_net_lock_current();
804
805                 all_known = 1;
806                 list_for_each(entry, &the_lnet.ln_routers) {
807                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
808
809                         if (rtr->lp_alive_count == 0) {
810                                 all_known = 0;
811                                 break;
812                         }
813                 }
814
815                 lnet_net_unlock(cpt);
816
817                 if (all_known)
818                         return;
819
820                 cfs_pause(cfs_time_seconds(1));
821         }
822 }
823
824 void
825 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
826 {
827         lnet_route_t *rte;
828
829         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
830                 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
831                         if (rte->lr_net == net) {
832                                 rte->lr_downis = 0;
833                                 break;
834                         }
835                 }
836         }
837 }
838
839 static void
840 lnet_update_ni_status_locked(void)
841 {
842         lnet_ni_t       *ni;
843         long            now;
844         int             timeout;
845
846         LASSERT(the_lnet.ln_routing);
847
848         timeout = router_ping_timeout +
849                   MAX(live_router_check_interval, dead_router_check_interval);
850
851         now = cfs_time_current_sec();
852         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
853                 if (ni->ni_lnd->lnd_type == LOLND)
854                         continue;
855
856                 if (now < ni->ni_last_alive + timeout)
857                         continue;
858
859                 lnet_ni_lock(ni);
860                 /* re-check with lock */
861                 if (now < ni->ni_last_alive + timeout) {
862                         lnet_ni_unlock(ni);
863                         continue;
864                 }
865
866                 LASSERT(ni->ni_status != NULL);
867
868                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
869                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
870                                libcfs_nid2str(ni->ni_nid), timeout);
871                         /* NB: so far, this is the only place to set
872                          * NI status to "down" */
873                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
874                 }
875                 lnet_ni_unlock(ni);
876         }
877 }
878
879 static void
880 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
881 {
882         LASSERT(list_empty(&rcd->rcd_list));
883         /* detached from network */
884         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
885
886         if (rcd->rcd_gateway != NULL) {
887                 int cpt = rcd->rcd_gateway->lp_cpt;
888
889                 lnet_net_lock(cpt);
890                 lnet_peer_decref_locked(rcd->rcd_gateway);
891                 lnet_net_unlock(cpt);
892         }
893
894         if (rcd->rcd_pinginfo != NULL)
895                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
896
897         LIBCFS_FREE(rcd, sizeof(*rcd));
898 }
899
900 static lnet_rc_data_t *
901 lnet_create_rc_data_locked(lnet_peer_t *gateway)
902 {
903         lnet_rc_data_t          *rcd = NULL;
904         lnet_ping_info_t        *pi;
905         int                     rc;
906         int                     i;
907
908         lnet_net_unlock(gateway->lp_cpt);
909
910         LIBCFS_ALLOC(rcd, sizeof(*rcd));
911         if (rcd == NULL)
912                 goto out;
913
914         LNetInvalidateHandle(&rcd->rcd_mdh);
915         INIT_LIST_HEAD(&rcd->rcd_list);
916
917         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
918         if (pi == NULL)
919                 goto out;
920
921         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
922                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
923                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
924         }
925         rcd->rcd_pinginfo = pi;
926
927         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
928         rc = LNetMDBind((lnet_md_t){.start     = pi,
929                                     .user_ptr  = rcd,
930                                     .length    = LNET_PINGINFO_SIZE,
931                                     .threshold = LNET_MD_THRESH_INF,
932                                     .options   = LNET_MD_TRUNCATE,
933                                     .eq_handle = the_lnet.ln_rc_eqh},
934                         LNET_UNLINK,
935                         &rcd->rcd_mdh);
936         if (rc < 0) {
937                 CERROR("Can't bind MD: %d\n", rc);
938                 goto out;
939         }
940         LASSERT(rc == 0);
941
942         lnet_net_lock(gateway->lp_cpt);
943         /* router table changed or someone has created rcd for this gateway */
944         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
945                 lnet_net_unlock(gateway->lp_cpt);
946                 goto out;
947         }
948
949         lnet_peer_addref_locked(gateway);
950         rcd->rcd_gateway = gateway;
951         gateway->lp_rcd = rcd;
952         gateway->lp_ping_notsent = 0;
953
954         return rcd;
955
956  out:
957         if (rcd != NULL) {
958                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
959                         rc = LNetMDUnlink(rcd->rcd_mdh);
960                         LASSERT(rc == 0);
961                 }
962                 lnet_destroy_rc_data(rcd);
963         }
964
965         lnet_net_lock(gateway->lp_cpt);
966         return gateway->lp_rcd;
967 }
968
969 static int
970 lnet_router_check_interval (lnet_peer_t *rtr)
971 {
972         int secs;
973
974         secs = rtr->lp_alive ? live_router_check_interval :
975                                dead_router_check_interval;
976         if (secs < 0)
977                 secs = 0;
978
979         return secs;
980 }
981
982 static void
983 lnet_ping_router_locked (lnet_peer_t *rtr)
984 {
985         lnet_rc_data_t *rcd = NULL;
986         cfs_time_t      now = cfs_time_current();
987         int             secs;
988
989         lnet_peer_addref_locked(rtr);
990
991         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
992             cfs_time_after(now, rtr->lp_ping_deadline))
993                 lnet_notify_locked(rtr, 1, 0, now);
994
995         /* Run any outstanding notifications */
996         lnet_ni_notify_locked(rtr->lp_ni, rtr);
997
998         if (!lnet_isrouter(rtr) ||
999             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1000                 /* router table changed or router checker is shutting down */
1001                 lnet_peer_decref_locked(rtr);
1002                 return;
1003         }
1004
1005         rcd = rtr->lp_rcd != NULL ?
1006               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
1007
1008         if (rcd == NULL)
1009                 return;
1010
1011         secs = lnet_router_check_interval(rtr);
1012
1013         CDEBUG(D_NET,
1014                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1015                "alive_count %d lp_ping_timestamp %lu\n",
1016                libcfs_nid2str(rtr->lp_nid), secs,
1017                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1018                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1019
1020         if (secs != 0 && !rtr->lp_ping_notsent &&
1021             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1022                                              cfs_time_seconds(secs)))) {
1023                 int               rc;
1024                 lnet_process_id_t id;
1025                 lnet_handle_md_t  mdh;
1026
1027                 id.nid = rtr->lp_nid;
1028                 id.pid = LNET_PID_LUSTRE;
1029                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1030
1031                 rtr->lp_ping_notsent   = 1;
1032                 rtr->lp_ping_timestamp = now;
1033
1034                 mdh = rcd->rcd_mdh;
1035
1036                 if (rtr->lp_ping_deadline == 0) {
1037                         rtr->lp_ping_deadline =
1038                                 cfs_time_shift(router_ping_timeout);
1039                 }
1040
1041                 lnet_net_unlock(rtr->lp_cpt);
1042
1043                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1044                              LNET_PROTO_PING_MATCHBITS, 0);
1045
1046                 lnet_net_lock(rtr->lp_cpt);
1047                 if (rc != 0)
1048                         rtr->lp_ping_notsent = 0; /* no event pending */
1049         }
1050
1051         lnet_peer_decref_locked(rtr);
1052         return;
1053 }
1054
1055 int
1056 lnet_router_checker_start(void)
1057 {
1058         int                     rc;
1059         int                     eqsz = 0;
1060         struct task_struct     *task;
1061
1062         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1063
1064         if (check_routers_before_use &&
1065             dead_router_check_interval <= 0) {
1066                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1067                                    " set if 'check_routers_before_use' is set"
1068                                    "\n");
1069                 return -EINVAL;
1070         }
1071
1072         sema_init(&the_lnet.ln_rc_signal, 0);
1073
1074         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1075         if (rc != 0) {
1076                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1077                 return -ENOMEM;
1078         }
1079
1080         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1081         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1082         if (IS_ERR(task)) {
1083                 rc = PTR_ERR(task);
1084                 CERROR("Can't start router checker thread: %d\n", rc);
1085                 /* block until event callback signals exit */
1086                 down(&the_lnet.ln_rc_signal);
1087                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1088                 LASSERT(rc == 0);
1089                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1090                 return -ENOMEM;
1091         }
1092
1093         if (check_routers_before_use) {
1094                 /* Note that a helpful side-effect of pinging all known routers
1095                  * at startup is that it makes them drop stale connections they
1096                  * may have to a previous instance of me. */
1097                 lnet_wait_known_routerstate();
1098         }
1099
1100         return 0;
1101 }
1102
1103 void
1104 lnet_router_checker_stop (void)
1105 {
1106         int rc;
1107
1108         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1109                 return;
1110
1111         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1112         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1113         /* wakeup the RC thread if it's sleeping */
1114         wake_up(&the_lnet.ln_rc_waitq);
1115
1116         /* block until event callback signals exit */
1117         down(&the_lnet.ln_rc_signal);
1118         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1119
1120         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1121         LASSERT (rc == 0);
1122         return;
1123 }
1124
1125 static void
1126 lnet_prune_rc_data(int wait_unlink)
1127 {
1128         lnet_rc_data_t          *rcd;
1129         lnet_rc_data_t          *tmp;
1130         lnet_peer_t             *lp;
1131         struct list_head         head;
1132         int                      i = 2;
1133
1134         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1135                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1136                    list_empty(&the_lnet.ln_rcd_zombie)))
1137                 return;
1138
1139         INIT_LIST_HEAD(&head);
1140
1141         lnet_net_lock(LNET_LOCK_EX);
1142
1143         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1144                 /* router checker is stopping, prune all */
1145                 list_for_each_entry(lp, &the_lnet.ln_routers,
1146                                     lp_rtr_list) {
1147                         if (lp->lp_rcd == NULL)
1148                                 continue;
1149
1150                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1151                         list_add(&lp->lp_rcd->rcd_list,
1152                                  &the_lnet.ln_rcd_deathrow);
1153                         lp->lp_rcd = NULL;
1154                 }
1155         }
1156
1157         /* unlink all RCDs on deathrow list */
1158         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1159
1160         if (!list_empty(&head)) {
1161                 lnet_net_unlock(LNET_LOCK_EX);
1162
1163                 list_for_each_entry(rcd, &head, rcd_list)
1164                         LNetMDUnlink(rcd->rcd_mdh);
1165
1166                 lnet_net_lock(LNET_LOCK_EX);
1167         }
1168
1169         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1170
1171         /* release all zombie RCDs */
1172         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1173                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1174                                          rcd_list) {
1175                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1176                                 list_move(&rcd->rcd_list, &head);
1177                 }
1178
1179                 wait_unlink = wait_unlink &&
1180                               !list_empty(&the_lnet.ln_rcd_zombie);
1181
1182                 lnet_net_unlock(LNET_LOCK_EX);
1183
1184                 while (!list_empty(&head)) {
1185                         rcd = list_entry(head.next,
1186                                          lnet_rc_data_t, rcd_list);
1187                         list_del_init(&rcd->rcd_list);
1188                         lnet_destroy_rc_data(rcd);
1189                 }
1190
1191                 if (!wait_unlink)
1192                         return;
1193
1194                 i++;
1195                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1196                        "Waiting for rc buffers to unlink\n");
1197                 cfs_pause(cfs_time_seconds(1) / 4);
1198
1199                 lnet_net_lock(LNET_LOCK_EX);
1200         }
1201
1202         lnet_net_unlock(LNET_LOCK_EX);
1203 }
1204
1205 /*
1206  * This function is called to check if the RC should block indefinitely.
1207  * It's called from lnet_router_checker() as well as being passed to
1208  * wait_event_interruptible() to avoid the lost wake_up problem.
1209  *
1210  * When it's called from wait_event_interruptible() it is necessary to
1211  * also not sleep if the rc state is not running to avoid a deadlock
1212  * when the system is shutting down
1213  */
1214 static inline bool
1215 lnet_router_checker_active(void)
1216 {
1217         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1218                 return true;
1219
1220         /* Router Checker thread needs to run when routing is enabled in
1221          * order to call lnet_update_ni_status_locked() */
1222         if (the_lnet.ln_routing)
1223                 return true;
1224
1225         return !list_empty(&the_lnet.ln_routers) &&
1226                 (live_router_check_interval > 0 ||
1227                  dead_router_check_interval > 0);
1228 }
1229
1230 static int
1231 lnet_router_checker(void *arg)
1232 {
1233         lnet_peer_t       *rtr;
1234         struct list_head  *entry;
1235
1236         cfs_block_allsigs();
1237
1238         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1239
1240         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1241                 __u64   version;
1242                 int     cpt;
1243                 int     cpt2;
1244
1245                 cpt = lnet_net_lock_current();
1246 rescan:
1247                 version = the_lnet.ln_routers_version;
1248
1249                 list_for_each(entry, &the_lnet.ln_routers) {
1250                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1251
1252                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1253                         if (cpt != cpt2) {
1254                                 lnet_net_unlock(cpt);
1255                                 cpt = cpt2;
1256                                 lnet_net_lock(cpt);
1257                                 /* the routers list has changed */
1258                                 if (version != the_lnet.ln_routers_version)
1259                                         goto rescan;
1260                         }
1261
1262                         lnet_ping_router_locked(rtr);
1263
1264                         /* NB dropped lock */
1265                         if (version != the_lnet.ln_routers_version) {
1266                                 /* the routers list has changed */
1267                                 goto rescan;
1268                         }
1269                 }
1270
1271                 if (the_lnet.ln_routing)
1272                         lnet_update_ni_status_locked();
1273
1274                 lnet_net_unlock(cpt);
1275
1276                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1277
1278                 /* Call cfs_pause() here always adds 1 to load average
1279                  * because kernel counts # active tasks as nr_running
1280                  * + nr_uninterruptible. */
1281                 /* if there are any routes then wakeup every second.  If
1282                  * there are no routes then sleep indefinitely until woken
1283                  * up by a user adding a route */
1284                 if (!lnet_router_checker_active())
1285                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1286                                                  lnet_router_checker_active());
1287                 else
1288                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1289                                                          false,
1290                                                          cfs_time_seconds(1));
1291         }
1292
1293         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1294
1295         lnet_prune_rc_data(1); /* wait for UNLINK */
1296
1297         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1298         up(&the_lnet.ln_rc_signal);
1299         /* The unlink event callback will signal final completion */
1300         return 0;
1301 }
1302
1303 void
1304 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1305 {
1306         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1307
1308         while (--npages >= 0)
1309                 __free_page(rb->rb_kiov[npages].kiov_page);
1310
1311         LIBCFS_FREE(rb, sz);
1312 }
1313
1314 static lnet_rtrbuf_t *
1315 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1316 {
1317         int            npages = rbp->rbp_npages;
1318         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1319         struct page   *page;
1320         lnet_rtrbuf_t *rb;
1321         int            i;
1322
1323         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1324         if (rb == NULL)
1325                 return NULL;
1326
1327         rb->rb_pool = rbp;
1328
1329         for (i = 0; i < npages; i++) {
1330                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1331                                           __GFP_ZERO | GFP_IOFS);
1332                 if (page == NULL) {
1333                         while (--i >= 0)
1334                                 __free_page(rb->rb_kiov[i].kiov_page);
1335
1336                         LIBCFS_FREE(rb, sz);
1337                         return NULL;
1338                 }
1339
1340                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1341                 rb->rb_kiov[i].kiov_offset = 0;
1342                 rb->rb_kiov[i].kiov_page = page;
1343         }
1344
1345         return rb;
1346 }
1347
1348 static void
1349 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1350 {
1351         int              npages = rbp->rbp_npages;
1352         lnet_rtrbuf_t    *rb;
1353         struct list_head tmp;
1354
1355         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1356                 return;
1357
1358         INIT_LIST_HEAD(&tmp);
1359
1360         lnet_net_lock(cpt);
1361         lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1362         list_splice_init(&rbp->rbp_bufs, &tmp);
1363         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1364         rbp->rbp_mincredits = 0;
1365         lnet_net_unlock(cpt);
1366
1367         /* Free buffers on the free list. */
1368         while (!list_empty(&tmp)) {
1369                 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1370                 list_del(&rb->rb_list);
1371                 lnet_destroy_rtrbuf(rb, npages);
1372         }
1373 }
1374
1375 static int
1376 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1377 {
1378         struct list_head rb_list;
1379         lnet_rtrbuf_t   *rb;
1380         int             num_rb;
1381         int             num_buffers = 0;
1382         int             npages = rbp->rbp_npages;
1383
1384         /* If we are called for less buffers than already in the pool, we
1385          * just lower the nbuffers number and excess buffers will be
1386          * thrown away as they are returned to the free list.  Credits
1387          * then get adjusted as well. */
1388         if (nbufs <= rbp->rbp_nbuffers) {
1389                 lnet_net_lock(cpt);
1390                 rbp->rbp_nbuffers = nbufs;
1391                 lnet_net_unlock(cpt);
1392                 return 0;
1393         }
1394
1395         INIT_LIST_HEAD(&rb_list);
1396
1397         /* allocate the buffers on a local list first.  If all buffers are
1398          * allocated successfully then join this list to the rbp buffer
1399          * list.  If not then free all allocated buffers. */
1400         num_rb = rbp->rbp_nbuffers;
1401
1402         while (num_rb < nbufs) {
1403                 rb = lnet_new_rtrbuf(rbp, cpt);
1404                 if (rb == NULL) {
1405                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1406                                nbufs, npages);
1407                         goto failed;
1408                 }
1409
1410                 list_add(&rb->rb_list, &rb_list);
1411                 num_buffers++;
1412                 num_rb++;
1413         }
1414
1415         lnet_net_lock(cpt);
1416
1417         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1418         rbp->rbp_nbuffers += num_buffers;
1419         rbp->rbp_credits += num_buffers;
1420         rbp->rbp_mincredits = rbp->rbp_credits;
1421         /* We need to schedule blocked msg using the newly
1422          * added buffers. */
1423         while (!list_empty(&rbp->rbp_bufs) &&
1424                !list_empty(&rbp->rbp_msgs))
1425                 lnet_schedule_blocked_locked(rbp);
1426
1427         lnet_net_unlock(cpt);
1428
1429         return 0;
1430
1431 failed:
1432         while (!list_empty(&rb_list)) {
1433                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1434                 list_del(&rb->rb_list);
1435                 lnet_destroy_rtrbuf(rb, npages);
1436         }
1437
1438         return -ENOMEM;
1439 }
1440
1441 static void
1442 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1443 {
1444         INIT_LIST_HEAD(&rbp->rbp_msgs);
1445         INIT_LIST_HEAD(&rbp->rbp_bufs);
1446
1447         rbp->rbp_npages = npages;
1448         rbp->rbp_credits = 0;
1449         rbp->rbp_mincredits = 0;
1450 }
1451
1452 void
1453 lnet_rtrpools_free(int keep_pools)
1454 {
1455         lnet_rtrbufpool_t *rtrp;
1456         int               i;
1457
1458         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1459                 return;
1460
1461         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1462                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1463                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1464                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1465         }
1466
1467         if (!keep_pools) {
1468                 cfs_percpt_free(the_lnet.ln_rtrpools);
1469                 the_lnet.ln_rtrpools = NULL;
1470         }
1471 }
1472
1473 static int
1474 lnet_nrb_tiny_calculate(void)
1475 {
1476         int     nrbs = LNET_NRB_TINY;
1477
1478         if (tiny_router_buffers < 0) {
1479                 LCONSOLE_ERROR_MSG(0x10c,
1480                                    "tiny_router_buffers=%d invalid when "
1481                                    "routing enabled\n", tiny_router_buffers);
1482                 return -1;
1483         }
1484
1485         if (tiny_router_buffers > 0)
1486                 nrbs = tiny_router_buffers;
1487
1488         nrbs /= LNET_CPT_NUMBER;
1489         return max(nrbs, LNET_NRB_TINY_MIN);
1490 }
1491
1492 static int
1493 lnet_nrb_small_calculate(void)
1494 {
1495         int     nrbs = LNET_NRB_SMALL;
1496
1497         if (small_router_buffers < 0) {
1498                 LCONSOLE_ERROR_MSG(0x10c,
1499                                    "small_router_buffers=%d invalid when "
1500                                    "routing enabled\n", small_router_buffers);
1501                 return -1;
1502         }
1503
1504         if (small_router_buffers > 0)
1505                 nrbs = small_router_buffers;
1506
1507         nrbs /= LNET_CPT_NUMBER;
1508         return max(nrbs, LNET_NRB_SMALL_MIN);
1509 }
1510
1511 static int
1512 lnet_nrb_large_calculate(void)
1513 {
1514         int     nrbs = LNET_NRB_LARGE;
1515
1516         if (large_router_buffers < 0) {
1517                 LCONSOLE_ERROR_MSG(0x10c,
1518                                    "large_router_buffers=%d invalid when "
1519                                    "routing enabled\n", large_router_buffers);
1520                 return -1;
1521         }
1522
1523         if (large_router_buffers > 0)
1524                 nrbs = large_router_buffers;
1525
1526         nrbs /= LNET_CPT_NUMBER;
1527         return max(nrbs, LNET_NRB_LARGE_MIN);
1528 }
1529
1530 int
1531 lnet_rtrpools_alloc(int im_a_router)
1532 {
1533         lnet_rtrbufpool_t *rtrp;
1534         int     nrb_tiny;
1535         int     nrb_small;
1536         int     nrb_large;
1537         int     rc;
1538         int     i;
1539
1540         if (!strcmp(forwarding, "")) {
1541                 /* not set either way */
1542                 if (!im_a_router)
1543                         return 0;
1544         } else if (!strcmp(forwarding, "disabled")) {
1545                 /* explicitly disabled */
1546                 return 0;
1547         } else if (!strcmp(forwarding, "enabled")) {
1548                 /* explicitly enabled */
1549         } else {
1550                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1551                                    "'enabled' or 'disabled'\n");
1552                 return -EINVAL;
1553         }
1554
1555         nrb_tiny = lnet_nrb_tiny_calculate();
1556         if (nrb_tiny < 0)
1557                 return -EINVAL;
1558
1559         nrb_small = lnet_nrb_small_calculate();
1560         if (nrb_small < 0)
1561                 return -EINVAL;
1562
1563         nrb_large = lnet_nrb_large_calculate();
1564         if (nrb_large < 0)
1565                 return -EINVAL;
1566
1567         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1568                                                 LNET_NRBPOOLS *
1569                                                 sizeof(lnet_rtrbufpool_t));
1570         if (the_lnet.ln_rtrpools == NULL) {
1571                 LCONSOLE_ERROR_MSG(0x10c,
1572                                    "Failed to initialize router buffe pool\n");
1573                 return -ENOMEM;
1574         }
1575
1576         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1577                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1578                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1579                                               nrb_tiny, i);
1580                 if (rc != 0)
1581                         goto failed;
1582
1583                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1584                                   LNET_NRB_SMALL_PAGES);
1585                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1586                                               nrb_small, i);
1587                 if (rc != 0)
1588                         goto failed;
1589
1590                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1591                                   LNET_NRB_LARGE_PAGES);
1592                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1593                                               nrb_large, i);
1594                 if (rc != 0)
1595                         goto failed;
1596         }
1597
1598         lnet_net_lock(LNET_LOCK_EX);
1599         the_lnet.ln_routing = 1;
1600         lnet_net_unlock(LNET_LOCK_EX);
1601         return 0;
1602
1603  failed:
1604         lnet_rtrpools_free(0);
1605         return rc;
1606 }
1607
1608 static int
1609 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1610 {
1611         int nrb = 0;
1612         int rc = 0;
1613         int i;
1614         lnet_rtrbufpool_t *rtrp;
1615
1616         /* If the provided values for each buffer pool are different than the
1617          * configured values, we need to take action. */
1618         if (tiny >= 0) {
1619                 tiny_router_buffers = tiny;
1620                 nrb = lnet_nrb_tiny_calculate();
1621                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1622                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1623                                                       nrb, i);
1624                         if (rc != 0)
1625                                 return rc;
1626                 }
1627         }
1628         if (small >= 0) {
1629                 small_router_buffers = small;
1630                 nrb = lnet_nrb_small_calculate();
1631                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1632                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1633                                                       nrb, i);
1634                         if (rc != 0)
1635                                 return rc;
1636                 }
1637         }
1638         if (large >= 0) {
1639                 large_router_buffers = large;
1640                 nrb = lnet_nrb_large_calculate();
1641                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1642                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1643                                                       nrb, i);
1644                         if (rc != 0)
1645                                 return rc;
1646                 }
1647         }
1648
1649         return 0;
1650 }
1651
1652 int
1653 lnet_rtrpools_adjust(int tiny, int small, int large)
1654 {
1655         /* this function doesn't revert the changes if adding new buffers
1656          * failed.  It's up to the user space caller to revert the
1657          * changes. */
1658
1659         if (!the_lnet.ln_routing)
1660                 return 0;
1661
1662         return lnet_rtrpools_adjust_helper(tiny, small, large);
1663 }
1664
1665 int
1666 lnet_rtrpools_enable(void)
1667 {
1668         int rc;
1669
1670         if (the_lnet.ln_routing)
1671                 return 0;
1672
1673         if (the_lnet.ln_rtrpools == NULL)
1674                 /* If routing is turned off, and we have never
1675                  * initialized the pools before, just call the
1676                  * standard buffer pool allocation routine as
1677                  * if we are just configuring this for the first
1678                  * time. */
1679                 return lnet_rtrpools_alloc(1);
1680
1681         rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1682         if (rc != 0)
1683                 return rc;
1684
1685         lnet_net_lock(LNET_LOCK_EX);
1686         the_lnet.ln_routing = 1;
1687
1688         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1689         lnet_net_unlock(LNET_LOCK_EX);
1690
1691         return 0;
1692 }
1693
1694 void
1695 lnet_rtrpools_disable(void)
1696 {
1697         if (!the_lnet.ln_routing)
1698                 return;
1699
1700         lnet_net_lock(LNET_LOCK_EX);
1701         the_lnet.ln_routing = 0;
1702         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1703
1704         tiny_router_buffers = 0;
1705         small_router_buffers = 0;
1706         large_router_buffers = 0;
1707         lnet_net_unlock(LNET_LOCK_EX);
1708         lnet_rtrpools_free(1);
1709 }
1710
1711 int
1712 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1713 {
1714         struct lnet_peer        *lp = NULL;
1715         cfs_time_t              now = cfs_time_current();
1716         int                     cpt = lnet_cpt_of_nid(nid);
1717
1718         LASSERT (!in_interrupt ());
1719
1720         CDEBUG (D_NET, "%s notifying %s: %s\n",
1721                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1722                 libcfs_nid2str(nid),
1723                 alive ? "up" : "down");
1724
1725         if (ni != NULL &&
1726             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1727                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1728                         libcfs_nid2str(nid), alive ? "birth" : "death",
1729                         libcfs_nid2str(ni->ni_nid));
1730                 return -EINVAL;
1731         }
1732
1733         /* can't do predictions... */
1734         if (cfs_time_after(when, now)) {
1735                 CWARN ("Ignoring prediction from %s of %s %s "
1736                        "%ld seconds in the future\n",
1737                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1738                        libcfs_nid2str(nid), alive ? "up" : "down",
1739                        cfs_duration_sec(cfs_time_sub(when, now)));
1740                 return -EINVAL;
1741         }
1742
1743         if (ni != NULL && !alive &&             /* LND telling me she's down */
1744             !auto_down) {                       /* auto-down disabled */
1745                 CDEBUG(D_NET, "Auto-down disabled\n");
1746                 return 0;
1747         }
1748
1749         lnet_net_lock(cpt);
1750
1751         if (the_lnet.ln_shutdown) {
1752                 lnet_net_unlock(cpt);
1753                 return -ESHUTDOWN;
1754         }
1755
1756         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1757         if (lp == NULL) {
1758                 /* nid not found */
1759                 lnet_net_unlock(cpt);
1760                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1761                 return 0;
1762         }
1763
1764         /* We can't fully trust LND on reporting exact peer last_alive
1765          * if he notifies us about dead peer. For example ksocklnd can
1766          * call us with when == _time_when_the_node_was_booted_ if
1767          * no connections were successfully established */
1768         if (ni != NULL && !alive && when < lp->lp_last_alive)
1769                 when = lp->lp_last_alive;
1770
1771         lnet_notify_locked(lp, ni == NULL, alive, when);
1772
1773         if (ni != NULL)
1774                 lnet_ni_notify_locked(ni, lp);
1775
1776         lnet_peer_decref_locked(lp);
1777
1778         lnet_net_unlock(cpt);
1779         return 0;
1780 }
1781 EXPORT_SYMBOL(lnet_notify);