Whamcloud - gitweb
b4c5842f950917c0ec62861d03778ddb38846f6b
[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, 2016, Intel Corporation.
5  *
6  *   This file is part of Lustre, https://wiki.hpdd.intel.com/
7  *
8  *   Portals is free software; you can redistribute it and/or
9  *   modify it under the terms of version 2 of the GNU General Public
10  *   License as published by the Free Software Foundation.
11  *
12  *   Portals is distributed in the hope that it will be useful,
13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  *   GNU General Public License for more details.
16  *
17  *   You should have received a copy of the GNU General Public License
18  *   along with Portals; if not, write to the Free Software
19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  *
21  */
22
23 #define DEBUG_SUBSYSTEM S_LNET
24 #include <lnet/lib-lnet.h>
25
26 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
27 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
28 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
29 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
30 #define LNET_NRB_SMALL_PAGES    1
31 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
32 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
33 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_SIZE - 1) >> \
34                                   PAGE_SHIFT)
35
36 static char *forwarding = "";
37 module_param(forwarding, charp, 0444);
38 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
39
40 static int tiny_router_buffers;
41 module_param(tiny_router_buffers, int, 0444);
42 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
43 static int small_router_buffers;
44 module_param(small_router_buffers, int, 0444);
45 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
46 static int large_router_buffers;
47 module_param(large_router_buffers, int, 0444);
48 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
49 static int peer_buffer_credits;
50 module_param(peer_buffer_credits, int, 0444);
51 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
52
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");
56
57 int
58 lnet_peer_buffer_credits(struct lnet_net *net)
59 {
60         /* NI option overrides LNet default */
61         if (net->net_tunables.lct_peer_rtr_credits > 0)
62                 return net->net_tunables.lct_peer_rtr_credits;
63         if (peer_buffer_credits > 0)
64                 return peer_buffer_credits;
65
66         /* As an approximation, allow this peer the same number of router
67          * buffers as it is allowed outstanding sends */
68         return net->net_tunables.lct_peer_tx_credits;
69 }
70
71 /* forward ref's */
72 static int lnet_router_checker(void *);
73
74 static int check_routers_before_use;
75 module_param(check_routers_before_use, int, 0444);
76 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
77
78 int avoid_asym_router_failure = 1;
79 module_param(avoid_asym_router_failure, int, 0644);
80 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
81
82 static int dead_router_check_interval = 60;
83 module_param(dead_router_check_interval, int, 0644);
84 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
85
86 static int live_router_check_interval = 60;
87 module_param(live_router_check_interval, int, 0644);
88 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
89
90 static int router_ping_timeout = 50;
91 module_param(router_ping_timeout, int, 0644);
92 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
93
94 int
95 lnet_peers_start_down(void)
96 {
97         return check_routers_before_use;
98 }
99
100 void
101 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
102 {
103         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
104                 CDEBUG(D_NET, "Out of date\n");
105                 return;
106         }
107
108         lp->lp_timestamp = when;                /* update timestamp */
109         lp->lp_ping_deadline = 0;               /* disable ping timeout */
110
111         if (lp->lp_alive_count != 0 &&          /* got old news */
112             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
113                 CDEBUG(D_NET, "Old news\n");
114                 return;
115         }
116
117         /* Flag that notification is outstanding */
118
119         lp->lp_alive_count++;
120         lp->lp_alive = !(!alive);               /* 1 bit! */
121         lp->lp_notify = 1;
122         lp->lp_notifylnd |= notifylnd;
123         if (lp->lp_alive)
124                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
125
126         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
127 }
128
129 static void
130 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
131 {
132         int        alive;
133         int        notifylnd;
134
135         /* Notify only in 1 thread at any time to ensure ordered notification.
136          * NB individual events can be missed; the only guarantee is that you
137          * always get the most recent news */
138
139         if (lp->lp_notifying || ni == NULL)
140                 return;
141
142         lp->lp_notifying = 1;
143
144         while (lp->lp_notify) {
145                 alive     = lp->lp_alive;
146                 notifylnd = lp->lp_notifylnd;
147
148                 lp->lp_notifylnd = 0;
149                 lp->lp_notify    = 0;
150
151                 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
152                         lnet_net_unlock(lp->lp_cpt);
153
154                         /* A new notification could happen now; I'll handle it
155                          * when control returns to me */
156
157                         (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lp_nid,
158                                                           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_rnet_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 timespec64 ts;
244         lnet_ni_t *ni = NULL;
245
246         if (seeded)
247                 return;
248
249         cfs_get_random_bytes(seed, sizeof(seed));
250
251         /* Nodes with small feet have little entropy
252          * the NID for this node gives the most entropy in the low bits */
253         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
254                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
255
256                 if (lnd_type != LOLND)
257                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
258         }
259
260         ktime_get_ts64(&ts);
261         cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]);
262         seeded = 1;
263         return;
264 }
265
266 /* NB expects LNET_LOCK held */
267 static void
268 lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
269 {
270         unsigned int      len = 0;
271         unsigned int      offset = 0;
272         struct list_head *e;
273
274         lnet_shuffle_seed();
275
276         list_for_each(e, &rnet->lrn_routes) {
277                 len++;
278         }
279
280         /* len+1 positions to add a new entry, also prevents division by 0 */
281         offset = cfs_rand() % (len + 1);
282         list_for_each(e, &rnet->lrn_routes) {
283                 if (offset == 0)
284                         break;
285                 offset--;
286         }
287         list_add(&route->lr_list, e);
288         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
289
290         the_lnet.ln_remote_nets_version++;
291         lnet_rtr_addref_locked(route->lr_gateway);
292 }
293
294 int
295 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
296                unsigned int priority)
297 {
298         struct list_head        *e;
299         lnet_remotenet_t        *rnet;
300         lnet_remotenet_t        *rnet2;
301         lnet_route_t            *route;
302         lnet_ni_t               *ni;
303         int                     add_route;
304         int                     rc;
305
306         CDEBUG(D_NET, "Add route: net %s hops %d priority %u gw %s\n",
307                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
308
309         if (gateway == LNET_NID_ANY ||
310             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
311             net == LNET_NIDNET(LNET_NID_ANY) ||
312             LNET_NETTYP(net) == LOLND ||
313             LNET_NIDNET(gateway) == net ||
314             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
315                 return -EINVAL;
316
317         if (lnet_islocalnet(net))       /* it's a local network */
318                 return -EEXIST;
319
320         /* Assume net, route, all new */
321         LIBCFS_ALLOC(route, sizeof(*route));
322         LIBCFS_ALLOC(rnet, sizeof(*rnet));
323         if (route == NULL || rnet == NULL) {
324                 CERROR("Out of memory creating route %s %d %s\n",
325                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
326                 if (route != NULL)
327                         LIBCFS_FREE(route, sizeof(*route));
328                 if (rnet != NULL)
329                         LIBCFS_FREE(rnet, sizeof(*rnet));
330                 return -ENOMEM;
331         }
332
333         INIT_LIST_HEAD(&rnet->lrn_routes);
334         rnet->lrn_net = net;
335         route->lr_hops = hops;
336         route->lr_net = net;
337         route->lr_priority = priority;
338
339         lnet_net_lock(LNET_LOCK_EX);
340
341         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
342         if (rc != 0) {
343                 lnet_net_unlock(LNET_LOCK_EX);
344
345                 LIBCFS_FREE(route, sizeof(*route));
346                 LIBCFS_FREE(rnet, sizeof(*rnet));
347
348                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
349                         return rc;       /* ignore the route entry */
350                 CERROR("Error %d creating route %s %d %s\n", rc,
351                         libcfs_net2str(net), hops,
352                         libcfs_nid2str(gateway));
353                 return rc;
354         }
355
356         LASSERT(!the_lnet.ln_shutdown);
357
358         rnet2 = lnet_find_rnet_locked(net);
359         if (rnet2 == NULL) {
360                 /* new network */
361                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
362                 rnet2 = rnet;
363         }
364
365         /* Search for a duplicate route (it's a NOOP if it is) */
366         add_route = 1;
367         list_for_each(e, &rnet2->lrn_routes) {
368                 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
369
370                 if (route2->lr_gateway == route->lr_gateway) {
371                         add_route = 0;
372                         break;
373                 }
374
375                 /* our lookups must be true */
376                 LASSERT(route2->lr_gateway->lp_nid != gateway);
377         }
378
379         if (add_route) {
380                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
381                 lnet_add_route_to_rnet(rnet2, route);
382
383                 ni = lnet_get_next_ni_locked(route->lr_gateway->lp_net, NULL);
384                 lnet_net_unlock(LNET_LOCK_EX);
385
386                 /* XXX Assume alive */
387                 if (ni->ni_net->net_lnd->lnd_notify != NULL)
388                         (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
389
390                 lnet_net_lock(LNET_LOCK_EX);
391         }
392
393         /* -1 for notify or !add_route */
394         lnet_peer_decref_locked(route->lr_gateway);
395         lnet_net_unlock(LNET_LOCK_EX);
396
397         rc = 0;
398
399         if (!add_route) {
400                 rc = -EEXIST;
401                 LIBCFS_FREE(route, sizeof(*route));
402         }
403
404         if (rnet != rnet2)
405                 LIBCFS_FREE(rnet, sizeof(*rnet));
406
407         /* indicate to startup the router checker if configured */
408         wake_up(&the_lnet.ln_rc_waitq);
409
410         return rc;
411 }
412
413 int
414 lnet_check_routes(void)
415 {
416         lnet_remotenet_t *rnet;
417         lnet_route_t     *route;
418         lnet_route_t     *route2;
419         struct list_head *e1;
420         struct list_head *e2;
421         int               cpt;
422         struct list_head *rn_list;
423         int               i;
424
425         cpt = lnet_net_lock_current();
426
427         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
428                 rn_list = &the_lnet.ln_remote_nets_hash[i];
429                 list_for_each(e1, rn_list) {
430                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
431
432                         route2 = NULL;
433                         list_for_each(e2, &rnet->lrn_routes) {
434                                 lnet_nid_t      nid1;
435                                 lnet_nid_t      nid2;
436                                 int             net;
437
438                                 route = list_entry(e2, lnet_route_t,
439                                                    lr_list);
440
441                                 if (route2 == NULL) {
442                                         route2 = route;
443                                         continue;
444                                 }
445
446                                 if (route->lr_gateway->lp_net ==
447                                     route2->lr_gateway->lp_net)
448                                         continue;
449
450                                 nid1 = route->lr_gateway->lp_nid;
451                                 nid2 = route2->lr_gateway->lp_nid;
452                                 net = rnet->lrn_net;
453
454                                 lnet_net_unlock(cpt);
455
456                                 CERROR("Routes to %s via %s and %s not "
457                                        "supported\n",
458                                        libcfs_net2str(net),
459                                        libcfs_nid2str(nid1),
460                                        libcfs_nid2str(nid2));
461                                 return -EINVAL;
462                         }
463                 }
464         }
465
466         lnet_net_unlock(cpt);
467         return 0;
468 }
469
470 int
471 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
472 {
473         struct lnet_peer        *gateway;
474         lnet_remotenet_t        *rnet;
475         lnet_route_t            *route;
476         struct list_head        *e1;
477         struct list_head        *e2;
478         int                     rc = -ENOENT;
479         struct list_head        *rn_list;
480         int                     idx = 0;
481
482         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
483                libcfs_net2str(net), libcfs_nid2str(gw_nid));
484
485         /* NB Caller may specify either all routes via the given gateway
486          * or a specific route entry actual NIDs) */
487
488         lnet_net_lock(LNET_LOCK_EX);
489         if (net == LNET_NIDNET(LNET_NID_ANY))
490                 rn_list = &the_lnet.ln_remote_nets_hash[0];
491         else
492                 rn_list = lnet_net2rnethash(net);
493
494 again:
495         list_for_each(e1, rn_list) {
496                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
497
498                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
499                         net == rnet->lrn_net))
500                         continue;
501
502                 list_for_each(e2, &rnet->lrn_routes) {
503                         route = list_entry(e2, lnet_route_t, lr_list);
504
505                         gateway = route->lr_gateway;
506                         if (!(gw_nid == LNET_NID_ANY ||
507                               gw_nid == gateway->lp_nid))
508                                 continue;
509
510                         list_del(&route->lr_list);
511                         list_del(&route->lr_gwlist);
512                         the_lnet.ln_remote_nets_version++;
513
514                         if (list_empty(&rnet->lrn_routes))
515                                 list_del(&rnet->lrn_list);
516                         else
517                                 rnet = NULL;
518
519                         lnet_rtr_decref_locked(gateway);
520                         lnet_peer_decref_locked(gateway);
521
522                         lnet_net_unlock(LNET_LOCK_EX);
523
524                         LIBCFS_FREE(route, sizeof(*route));
525
526                         if (rnet != NULL)
527                                 LIBCFS_FREE(rnet, sizeof(*rnet));
528
529                         rc = 0;
530                         lnet_net_lock(LNET_LOCK_EX);
531                         goto again;
532                 }
533         }
534
535         if (net == LNET_NIDNET(LNET_NID_ANY) &&
536             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
537                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
538                 goto again;
539         }
540         lnet_net_unlock(LNET_LOCK_EX);
541
542         return rc;
543 }
544
545 void
546 lnet_destroy_routes (void)
547 {
548         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
549 }
550
551 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
552 {
553         int i, rc = -ENOENT, j;
554
555         if (the_lnet.ln_rtrpools == NULL)
556                 return rc;
557
558         for (i = 0; i < LNET_NRBPOOLS; i++) {
559                 lnet_rtrbufpool_t *rbp;
560
561                 lnet_net_lock(LNET_LOCK_EX);
562                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
563                         if (i++ != idx)
564                                 continue;
565
566                         pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
567                         pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
568                         pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
569                         pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
570                         rc = 0;
571                         break;
572                 }
573                 lnet_net_unlock(LNET_LOCK_EX);
574         }
575
576         lnet_net_lock(LNET_LOCK_EX);
577         pool_cfg->pl_routing = the_lnet.ln_routing;
578         lnet_net_unlock(LNET_LOCK_EX);
579
580         return rc;
581 }
582
583 int
584 lnet_get_route(int idx, __u32 *net, __u32 *hops,
585                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
586 {
587         struct list_head *e1;
588         struct list_head *e2;
589         lnet_remotenet_t *rnet;
590         lnet_route_t     *route;
591         int               cpt;
592         int               i;
593         struct list_head *rn_list;
594
595         cpt = lnet_net_lock_current();
596
597         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
598                 rn_list = &the_lnet.ln_remote_nets_hash[i];
599                 list_for_each(e1, rn_list) {
600                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
601
602                         list_for_each(e2, &rnet->lrn_routes) {
603                                 route = list_entry(e2, lnet_route_t,
604                                                    lr_list);
605
606                                 if (idx-- == 0) {
607                                         *net      = rnet->lrn_net;
608                                         *hops     = route->lr_hops;
609                                         *priority = route->lr_priority;
610                                         *gateway  = route->lr_gateway->lp_nid;
611                                         *alive    = lnet_is_route_alive(route);
612                                         lnet_net_unlock(cpt);
613                                         return 0;
614                                 }
615                         }
616                 }
617         }
618
619         lnet_net_unlock(cpt);
620         return -ENOENT;
621 }
622
623 void
624 lnet_swap_pinginfo(struct lnet_ping_info *info)
625 {
626         int               i;
627         struct lnet_ni_status *stat;
628
629         __swab32s(&info->pi_magic);
630         __swab32s(&info->pi_features);
631         __swab32s(&info->pi_pid);
632         __swab32s(&info->pi_nnis);
633         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
634                 stat = &info->pi_ni[i];
635                 __swab64s(&stat->ns_nid);
636                 __swab32s(&stat->ns_status);
637         }
638         return;
639 }
640
641 /**
642  * parse router-checker pinginfo, record number of down NIs for remote
643  * networks on that router.
644  */
645 static void
646 lnet_parse_rc_info(lnet_rc_data_t *rcd)
647 {
648         struct lnet_ping_info   *info = rcd->rcd_pinginfo;
649         struct lnet_peer        *gw   = rcd->rcd_gateway;
650         lnet_route_t            *rte;
651
652         if (!gw->lp_alive)
653                 return;
654
655         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
656                 lnet_swap_pinginfo(info);
657
658         /* NB always racing with network! */
659         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
660                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
661                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
662                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
663                 return;
664         }
665
666         gw->lp_ping_feats = info->pi_features;
667         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
668                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
669                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
670                 return; /* nothing I can understand */
671         }
672
673         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
674                 return; /* can't carry NI status info */
675
676         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
677                 int     down = 0;
678                 int     up = 0;
679                 int     i;
680
681                 if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
682                         rte->lr_downis = 1;
683                         continue;
684                 }
685
686                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
687                         struct lnet_ni_status *stat = &info->pi_ni[i];
688                         lnet_nid_t       nid = stat->ns_nid;
689
690                         if (nid == LNET_NID_ANY) {
691                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
692                                        libcfs_nid2str(gw->lp_nid));
693                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
694                                 return;
695                         }
696
697                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
698                                 continue;
699
700                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
701                                 down++;
702                                 continue;
703                         }
704
705                         if (stat->ns_status == LNET_NI_STATUS_UP) {
706                                 if (LNET_NIDNET(nid) == rte->lr_net) {
707                                         up = 1;
708                                         break;
709                                 }
710                                 continue;
711                         }
712
713                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
714                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
715                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
716                         return;
717                 }
718
719                 if (up) { /* ignore downed NIs if NI for dest network is up */
720                         rte->lr_downis = 0;
721                         continue;
722                 }
723                 /* if @down is zero and this route is single-hop, it means
724                  * we can't find NI for target network */
725                 if (down == 0 && rte->lr_hops == 1)
726                         down = 1;
727
728                 rte->lr_downis = down;
729         }
730 }
731
732 static void
733 lnet_router_checker_event(lnet_event_t *event)
734 {
735         lnet_rc_data_t          *rcd = event->md.user_ptr;
736         struct lnet_peer        *lp;
737
738         LASSERT(rcd != NULL);
739
740         if (event->unlinked) {
741                 LNetInvalidateHandle(&rcd->rcd_mdh);
742                 return;
743         }
744
745         LASSERT(event->type == LNET_EVENT_SEND ||
746                 event->type == LNET_EVENT_REPLY);
747
748         lp = rcd->rcd_gateway;
749         LASSERT(lp != NULL);
750
751          /* NB: it's called with holding lnet_res_lock, we have a few
752           * places need to hold both locks at the same time, please take
753           * care of lock ordering */
754         lnet_net_lock(lp->lp_cpt);
755         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
756                 /* ignore if no longer a router or rcd is replaced */
757                 goto out;
758         }
759
760         if (event->type == LNET_EVENT_SEND) {
761                 lp->lp_ping_notsent = 0;
762                 if (event->status == 0)
763                         goto out;
764         }
765
766         /* LNET_EVENT_REPLY */
767         /* A successful REPLY means the router is up.  If _any_ comms
768          * to the router fail I assume it's down (this will happen if
769          * we ping alive routers to try to detect router death before
770          * apps get burned). */
771
772         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
773         /* The router checker will wake up very shortly and do the
774          * actual notification.
775          * XXX If 'lp' stops being a router before then, it will still
776          * have the notification pending!!! */
777
778         if (avoid_asym_router_failure && event->status == 0)
779                 lnet_parse_rc_info(rcd);
780
781  out:
782         lnet_net_unlock(lp->lp_cpt);
783 }
784
785 static void
786 lnet_wait_known_routerstate(void)
787 {
788         lnet_peer_t      *rtr;
789         struct list_head *entry;
790         int               all_known;
791
792         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
793
794         for (;;) {
795                 int cpt = lnet_net_lock_current();
796
797                 all_known = 1;
798                 list_for_each(entry, &the_lnet.ln_routers) {
799                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
800
801                         if (rtr->lp_alive_count == 0) {
802                                 all_known = 0;
803                                 break;
804                         }
805                 }
806
807                 lnet_net_unlock(cpt);
808
809                 if (all_known)
810                         return;
811
812                 set_current_state(TASK_UNINTERRUPTIBLE);
813                 schedule_timeout(cfs_time_seconds(1));
814         }
815 }
816
817 void
818 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
819 {
820         lnet_route_t *rte;
821
822         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
823                 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
824                         if (rte->lr_net == net) {
825                                 rte->lr_downis = 0;
826                                 break;
827                         }
828                 }
829         }
830 }
831
832 static void
833 lnet_update_ni_status_locked(void)
834 {
835         lnet_ni_t       *ni = NULL;
836         time64_t        now;
837         int             timeout;
838
839         LASSERT(the_lnet.ln_routing);
840
841         timeout = router_ping_timeout +
842                   MAX(live_router_check_interval, dead_router_check_interval);
843
844         now = ktime_get_real_seconds();
845         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
846                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
847                         continue;
848
849                 if (now < ni->ni_last_alive + timeout)
850                         continue;
851
852                 lnet_ni_lock(ni);
853                 /* re-check with lock */
854                 if (now < ni->ni_last_alive + timeout) {
855                         lnet_ni_unlock(ni);
856                         continue;
857                 }
858
859                 LASSERT(ni->ni_status != NULL);
860
861                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
862                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
863                                libcfs_nid2str(ni->ni_nid), timeout);
864                         /* NB: so far, this is the only place to set
865                          * NI status to "down" */
866                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
867                 }
868                 lnet_ni_unlock(ni);
869         }
870 }
871
872 static void
873 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
874 {
875         LASSERT(list_empty(&rcd->rcd_list));
876         /* detached from network */
877         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
878
879         if (rcd->rcd_gateway != NULL) {
880                 int cpt = rcd->rcd_gateway->lp_cpt;
881
882                 lnet_net_lock(cpt);
883                 lnet_peer_decref_locked(rcd->rcd_gateway);
884                 lnet_net_unlock(cpt);
885         }
886
887         if (rcd->rcd_pinginfo != NULL)
888                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
889
890         LIBCFS_FREE(rcd, sizeof(*rcd));
891 }
892
893 static lnet_rc_data_t *
894 lnet_create_rc_data_locked(lnet_peer_t *gateway)
895 {
896         lnet_rc_data_t          *rcd = NULL;
897         struct lnet_ping_info   *pi;
898         int                     rc;
899         int                     i;
900
901         lnet_net_unlock(gateway->lp_cpt);
902
903         LIBCFS_ALLOC(rcd, sizeof(*rcd));
904         if (rcd == NULL)
905                 goto out;
906
907         LNetInvalidateHandle(&rcd->rcd_mdh);
908         INIT_LIST_HEAD(&rcd->rcd_list);
909
910         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
911         if (pi == NULL)
912                 goto out;
913
914         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
915                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
916                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
917         }
918         rcd->rcd_pinginfo = pi;
919
920         LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
921         rc = LNetMDBind((lnet_md_t){.start     = pi,
922                                     .user_ptr  = rcd,
923                                     .length    = LNET_PINGINFO_SIZE,
924                                     .threshold = LNET_MD_THRESH_INF,
925                                     .options   = LNET_MD_TRUNCATE,
926                                     .eq_handle = the_lnet.ln_rc_eqh},
927                         LNET_UNLINK,
928                         &rcd->rcd_mdh);
929         if (rc < 0) {
930                 CERROR("Can't bind MD: %d\n", rc);
931                 goto out;
932         }
933         LASSERT(rc == 0);
934
935         lnet_net_lock(gateway->lp_cpt);
936         /* router table changed or someone has created rcd for this gateway */
937         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
938                 lnet_net_unlock(gateway->lp_cpt);
939                 goto out;
940         }
941
942         lnet_peer_addref_locked(gateway);
943         rcd->rcd_gateway = gateway;
944         gateway->lp_rcd = rcd;
945         gateway->lp_ping_notsent = 0;
946
947         return rcd;
948
949  out:
950         if (rcd != NULL) {
951                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
952                         rc = LNetMDUnlink(rcd->rcd_mdh);
953                         LASSERT(rc == 0);
954                 }
955                 lnet_destroy_rc_data(rcd);
956         }
957
958         lnet_net_lock(gateway->lp_cpt);
959         return gateway->lp_rcd;
960 }
961
962 static int
963 lnet_router_check_interval (lnet_peer_t *rtr)
964 {
965         int secs;
966
967         secs = rtr->lp_alive ? live_router_check_interval :
968                                dead_router_check_interval;
969         if (secs < 0)
970                 secs = 0;
971
972         return secs;
973 }
974
975 static void
976 lnet_ping_router_locked (lnet_peer_t *rtr)
977 {
978         lnet_rc_data_t *rcd = NULL;
979         cfs_time_t      now = cfs_time_current();
980         int             secs;
981         struct lnet_ni  *ni;
982
983         lnet_peer_addref_locked(rtr);
984
985         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
986             cfs_time_after(now, rtr->lp_ping_deadline))
987                 lnet_notify_locked(rtr, 1, 0, now);
988
989         /* Run any outstanding notifications */
990         ni = lnet_get_next_ni_locked(rtr->lp_net, NULL);
991         lnet_ni_notify_locked(ni, rtr);
992
993         if (!lnet_isrouter(rtr) ||
994             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
995                 /* router table changed or router checker is shutting down */
996                 lnet_peer_decref_locked(rtr);
997                 return;
998         }
999
1000         rcd = rtr->lp_rcd != NULL ?
1001               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
1002
1003         if (rcd == NULL)
1004                 return;
1005
1006         secs = lnet_router_check_interval(rtr);
1007
1008         CDEBUG(D_NET,
1009                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1010                "alive_count %d lp_ping_timestamp %lu\n",
1011                libcfs_nid2str(rtr->lp_nid), secs,
1012                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1013                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1014
1015         if (secs != 0 && !rtr->lp_ping_notsent &&
1016             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1017                                              cfs_time_seconds(secs)))) {
1018                 int               rc;
1019                 lnet_process_id_t id;
1020                 lnet_handle_md_t  mdh;
1021
1022                 id.nid = rtr->lp_nid;
1023                 id.pid = LNET_PID_LUSTRE;
1024                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1025
1026                 rtr->lp_ping_notsent   = 1;
1027                 rtr->lp_ping_timestamp = now;
1028
1029                 mdh = rcd->rcd_mdh;
1030
1031                 if (rtr->lp_ping_deadline == 0) {
1032                         rtr->lp_ping_deadline =
1033                                 cfs_time_shift(router_ping_timeout);
1034                 }
1035
1036                 lnet_net_unlock(rtr->lp_cpt);
1037
1038                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1039                              LNET_PROTO_PING_MATCHBITS, 0);
1040
1041                 lnet_net_lock(rtr->lp_cpt);
1042                 if (rc != 0)
1043                         rtr->lp_ping_notsent = 0; /* no event pending */
1044         }
1045
1046         lnet_peer_decref_locked(rtr);
1047         return;
1048 }
1049
1050 int
1051 lnet_router_checker_start(void)
1052 {
1053         int                     rc;
1054         int                     eqsz = 0;
1055         struct task_struct     *task;
1056
1057         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1058
1059         if (check_routers_before_use &&
1060             dead_router_check_interval <= 0) {
1061                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1062                                    " set if 'check_routers_before_use' is set"
1063                                    "\n");
1064                 return -EINVAL;
1065         }
1066
1067         sema_init(&the_lnet.ln_rc_signal, 0);
1068
1069         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1070         if (rc != 0) {
1071                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1072                 return -ENOMEM;
1073         }
1074
1075         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1076         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1077         if (IS_ERR(task)) {
1078                 rc = PTR_ERR(task);
1079                 CERROR("Can't start router checker thread: %d\n", rc);
1080                 /* block until event callback signals exit */
1081                 down(&the_lnet.ln_rc_signal);
1082                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1083                 LASSERT(rc == 0);
1084                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1085                 return -ENOMEM;
1086         }
1087
1088         if (check_routers_before_use) {
1089                 /* Note that a helpful side-effect of pinging all known routers
1090                  * at startup is that it makes them drop stale connections they
1091                  * may have to a previous instance of me. */
1092                 lnet_wait_known_routerstate();
1093         }
1094
1095         return 0;
1096 }
1097
1098 void
1099 lnet_router_checker_stop (void)
1100 {
1101         int rc;
1102
1103         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1104                 return;
1105
1106         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1107         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1108         /* wakeup the RC thread if it's sleeping */
1109         wake_up(&the_lnet.ln_rc_waitq);
1110
1111         /* block until event callback signals exit */
1112         down(&the_lnet.ln_rc_signal);
1113         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1114
1115         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1116         LASSERT(rc == 0);
1117         return;
1118 }
1119
1120 static void
1121 lnet_prune_rc_data(int wait_unlink)
1122 {
1123         lnet_rc_data_t          *rcd;
1124         lnet_rc_data_t          *tmp;
1125         lnet_peer_t             *lp;
1126         struct list_head         head;
1127         int                      i = 2;
1128
1129         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1130                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1131                    list_empty(&the_lnet.ln_rcd_zombie)))
1132                 return;
1133
1134         INIT_LIST_HEAD(&head);
1135
1136         lnet_net_lock(LNET_LOCK_EX);
1137
1138         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1139                 /* router checker is stopping, prune all */
1140                 list_for_each_entry(lp, &the_lnet.ln_routers,
1141                                     lp_rtr_list) {
1142                         if (lp->lp_rcd == NULL)
1143                                 continue;
1144
1145                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1146                         list_add(&lp->lp_rcd->rcd_list,
1147                                  &the_lnet.ln_rcd_deathrow);
1148                         lp->lp_rcd = NULL;
1149                 }
1150         }
1151
1152         /* unlink all RCDs on deathrow list */
1153         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1154
1155         if (!list_empty(&head)) {
1156                 lnet_net_unlock(LNET_LOCK_EX);
1157
1158                 list_for_each_entry(rcd, &head, rcd_list)
1159                         LNetMDUnlink(rcd->rcd_mdh);
1160
1161                 lnet_net_lock(LNET_LOCK_EX);
1162         }
1163
1164         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1165
1166         /* release all zombie RCDs */
1167         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1168                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1169                                          rcd_list) {
1170                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1171                                 list_move(&rcd->rcd_list, &head);
1172                 }
1173
1174                 wait_unlink = wait_unlink &&
1175                               !list_empty(&the_lnet.ln_rcd_zombie);
1176
1177                 lnet_net_unlock(LNET_LOCK_EX);
1178
1179                 while (!list_empty(&head)) {
1180                         rcd = list_entry(head.next,
1181                                          lnet_rc_data_t, rcd_list);
1182                         list_del_init(&rcd->rcd_list);
1183                         lnet_destroy_rc_data(rcd);
1184                 }
1185
1186                 if (!wait_unlink)
1187                         return;
1188
1189                 i++;
1190                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1191                        "Waiting for rc buffers to unlink\n");
1192                 set_current_state(TASK_UNINTERRUPTIBLE);
1193                 schedule_timeout(cfs_time_seconds(1) / 4);
1194
1195                 lnet_net_lock(LNET_LOCK_EX);
1196         }
1197
1198         lnet_net_unlock(LNET_LOCK_EX);
1199 }
1200
1201 /*
1202  * This function is called to check if the RC should block indefinitely.
1203  * It's called from lnet_router_checker() as well as being passed to
1204  * wait_event_interruptible() to avoid the lost wake_up problem.
1205  *
1206  * When it's called from wait_event_interruptible() it is necessary to
1207  * also not sleep if the rc state is not running to avoid a deadlock
1208  * when the system is shutting down
1209  */
1210 static inline bool
1211 lnet_router_checker_active(void)
1212 {
1213         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1214                 return true;
1215
1216         /* Router Checker thread needs to run when routing is enabled in
1217          * order to call lnet_update_ni_status_locked() */
1218         if (the_lnet.ln_routing)
1219                 return true;
1220
1221         return !list_empty(&the_lnet.ln_routers) &&
1222                 (live_router_check_interval > 0 ||
1223                  dead_router_check_interval > 0);
1224 }
1225
1226 static int
1227 lnet_router_checker(void *arg)
1228 {
1229         lnet_peer_t       *rtr;
1230         struct list_head  *entry;
1231
1232         cfs_block_allsigs();
1233
1234         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1235                 __u64   version;
1236                 int     cpt;
1237                 int     cpt2;
1238
1239                 cpt = lnet_net_lock_current();
1240 rescan:
1241                 version = the_lnet.ln_routers_version;
1242
1243                 list_for_each(entry, &the_lnet.ln_routers) {
1244                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1245
1246                         cpt2 = rtr->lp_cpt;
1247                         if (cpt != cpt2) {
1248                                 lnet_net_unlock(cpt);
1249                                 cpt = cpt2;
1250                                 lnet_net_lock(cpt);
1251                                 /* the routers list has changed */
1252                                 if (version != the_lnet.ln_routers_version)
1253                                         goto rescan;
1254                         }
1255
1256                         lnet_ping_router_locked(rtr);
1257
1258                         /* NB dropped lock */
1259                         if (version != the_lnet.ln_routers_version) {
1260                                 /* the routers list has changed */
1261                                 goto rescan;
1262                         }
1263                 }
1264
1265                 if (the_lnet.ln_routing)
1266                         lnet_update_ni_status_locked();
1267
1268                 lnet_net_unlock(cpt);
1269
1270                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1271
1272                 /* Call schedule_timeout() here always adds 1 to load average
1273                  * because kernel counts # active tasks as nr_running
1274                  * + nr_uninterruptible. */
1275                 /* if there are any routes then wakeup every second.  If
1276                  * there are no routes then sleep indefinitely until woken
1277                  * up by a user adding a route */
1278                 if (!lnet_router_checker_active())
1279                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1280                                                  lnet_router_checker_active());
1281                 else
1282                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1283                                                          false,
1284                                                          cfs_time_seconds(1));
1285         }
1286
1287         lnet_prune_rc_data(1); /* wait for UNLINK */
1288
1289         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1290         up(&the_lnet.ln_rc_signal);
1291         /* The unlink event callback will signal final completion */
1292         return 0;
1293 }
1294
1295 void
1296 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1297 {
1298         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1299
1300         while (--npages >= 0)
1301                 __free_page(rb->rb_kiov[npages].kiov_page);
1302
1303         LIBCFS_FREE(rb, sz);
1304 }
1305
1306 static lnet_rtrbuf_t *
1307 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1308 {
1309         int            npages = rbp->rbp_npages;
1310         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1311         struct page   *page;
1312         lnet_rtrbuf_t *rb;
1313         int            i;
1314
1315         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1316         if (rb == NULL)
1317                 return NULL;
1318
1319         rb->rb_pool = rbp;
1320
1321         for (i = 0; i < npages; i++) {
1322                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1323                                           GFP_KERNEL | __GFP_ZERO);
1324                 if (page == NULL) {
1325                         while (--i >= 0)
1326                                 __free_page(rb->rb_kiov[i].kiov_page);
1327
1328                         LIBCFS_FREE(rb, sz);
1329                         return NULL;
1330                 }
1331
1332                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1333                 rb->rb_kiov[i].kiov_offset = 0;
1334                 rb->rb_kiov[i].kiov_page = page;
1335         }
1336
1337         return rb;
1338 }
1339
1340 static void
1341 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1342 {
1343         int              npages = rbp->rbp_npages;
1344         lnet_rtrbuf_t    *rb;
1345         struct list_head tmp;
1346
1347         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1348                 return;
1349
1350         INIT_LIST_HEAD(&tmp);
1351
1352         lnet_net_lock(cpt);
1353         lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1354         list_splice_init(&rbp->rbp_bufs, &tmp);
1355         rbp->rbp_req_nbuffers = 0;
1356         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1357         rbp->rbp_mincredits = 0;
1358         lnet_net_unlock(cpt);
1359
1360         /* Free buffers on the free list. */
1361         while (!list_empty(&tmp)) {
1362                 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1363                 list_del(&rb->rb_list);
1364                 lnet_destroy_rtrbuf(rb, npages);
1365         }
1366 }
1367
1368 static int
1369 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1370 {
1371         struct list_head rb_list;
1372         lnet_rtrbuf_t   *rb;
1373         int             num_rb;
1374         int             num_buffers = 0;
1375         int             old_req_nbufs;
1376         int             npages = rbp->rbp_npages;
1377
1378         lnet_net_lock(cpt);
1379         /* If we are called for less buffers than already in the pool, we
1380          * just lower the req_nbuffers number and excess buffers will be
1381          * thrown away as they are returned to the free list.  Credits
1382          * then get adjusted as well.
1383          * If we already have enough buffers allocated to serve the
1384          * increase requested, then we can treat that the same way as we
1385          * do the decrease. */
1386         num_rb = nbufs - rbp->rbp_nbuffers;
1387         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1388                 rbp->rbp_req_nbuffers = nbufs;
1389                 lnet_net_unlock(cpt);
1390                 return 0;
1391         }
1392         /* store the older value of rbp_req_nbuffers and then set it to
1393          * the new request to prevent lnet_return_rx_credits_locked() from
1394          * freeing buffers that we need to keep around */
1395         old_req_nbufs = rbp->rbp_req_nbuffers;
1396         rbp->rbp_req_nbuffers = nbufs;
1397         lnet_net_unlock(cpt);
1398
1399         INIT_LIST_HEAD(&rb_list);
1400
1401         /* allocate the buffers on a local list first.  If all buffers are
1402          * allocated successfully then join this list to the rbp buffer
1403          * list.  If not then free all allocated buffers. */
1404         while (num_rb-- > 0) {
1405                 rb = lnet_new_rtrbuf(rbp, cpt);
1406                 if (rb == NULL) {
1407                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1408                                nbufs, npages);
1409
1410                         lnet_net_lock(cpt);
1411                         rbp->rbp_req_nbuffers = old_req_nbufs;
1412                         lnet_net_unlock(cpt);
1413
1414                         goto failed;
1415                 }
1416
1417                 list_add(&rb->rb_list, &rb_list);
1418                 num_buffers++;
1419         }
1420
1421         lnet_net_lock(cpt);
1422
1423         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1424         rbp->rbp_nbuffers += num_buffers;
1425         rbp->rbp_credits += num_buffers;
1426         rbp->rbp_mincredits = rbp->rbp_credits;
1427         /* We need to schedule blocked msg using the newly
1428          * added buffers. */
1429         while (!list_empty(&rbp->rbp_bufs) &&
1430                !list_empty(&rbp->rbp_msgs))
1431                 lnet_schedule_blocked_locked(rbp);
1432
1433         lnet_net_unlock(cpt);
1434
1435         return 0;
1436
1437 failed:
1438         while (!list_empty(&rb_list)) {
1439                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1440                 list_del(&rb->rb_list);
1441                 lnet_destroy_rtrbuf(rb, npages);
1442         }
1443
1444         return -ENOMEM;
1445 }
1446
1447 static void
1448 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1449 {
1450         INIT_LIST_HEAD(&rbp->rbp_msgs);
1451         INIT_LIST_HEAD(&rbp->rbp_bufs);
1452
1453         rbp->rbp_npages = npages;
1454         rbp->rbp_credits = 0;
1455         rbp->rbp_mincredits = 0;
1456 }
1457
1458 void
1459 lnet_rtrpools_free(int keep_pools)
1460 {
1461         lnet_rtrbufpool_t *rtrp;
1462         int               i;
1463
1464         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1465                 return;
1466
1467         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1468                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1469                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1470                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1471         }
1472
1473         if (!keep_pools) {
1474                 cfs_percpt_free(the_lnet.ln_rtrpools);
1475                 the_lnet.ln_rtrpools = NULL;
1476         }
1477 }
1478
1479 static int
1480 lnet_nrb_tiny_calculate(void)
1481 {
1482         int     nrbs = LNET_NRB_TINY;
1483
1484         if (tiny_router_buffers < 0) {
1485                 LCONSOLE_ERROR_MSG(0x10c,
1486                                    "tiny_router_buffers=%d invalid when "
1487                                    "routing enabled\n", tiny_router_buffers);
1488                 return -EINVAL;
1489         }
1490
1491         if (tiny_router_buffers > 0)
1492                 nrbs = tiny_router_buffers;
1493
1494         nrbs /= LNET_CPT_NUMBER;
1495         return max(nrbs, LNET_NRB_TINY_MIN);
1496 }
1497
1498 static int
1499 lnet_nrb_small_calculate(void)
1500 {
1501         int     nrbs = LNET_NRB_SMALL;
1502
1503         if (small_router_buffers < 0) {
1504                 LCONSOLE_ERROR_MSG(0x10c,
1505                                    "small_router_buffers=%d invalid when "
1506                                    "routing enabled\n", small_router_buffers);
1507                 return -EINVAL;
1508         }
1509
1510         if (small_router_buffers > 0)
1511                 nrbs = small_router_buffers;
1512
1513         nrbs /= LNET_CPT_NUMBER;
1514         return max(nrbs, LNET_NRB_SMALL_MIN);
1515 }
1516
1517 static int
1518 lnet_nrb_large_calculate(void)
1519 {
1520         int     nrbs = LNET_NRB_LARGE;
1521
1522         if (large_router_buffers < 0) {
1523                 LCONSOLE_ERROR_MSG(0x10c,
1524                                    "large_router_buffers=%d invalid when "
1525                                    "routing enabled\n", large_router_buffers);
1526                 return -EINVAL;
1527         }
1528
1529         if (large_router_buffers > 0)
1530                 nrbs = large_router_buffers;
1531
1532         nrbs /= LNET_CPT_NUMBER;
1533         return max(nrbs, LNET_NRB_LARGE_MIN);
1534 }
1535
1536 int
1537 lnet_rtrpools_alloc(int im_a_router)
1538 {
1539         lnet_rtrbufpool_t *rtrp;
1540         int     nrb_tiny;
1541         int     nrb_small;
1542         int     nrb_large;
1543         int     rc;
1544         int     i;
1545
1546         if (!strcmp(forwarding, "")) {
1547                 /* not set either way */
1548                 if (!im_a_router)
1549                         return 0;
1550         } else if (!strcmp(forwarding, "disabled")) {
1551                 /* explicitly disabled */
1552                 return 0;
1553         } else if (!strcmp(forwarding, "enabled")) {
1554                 /* explicitly enabled */
1555         } else {
1556                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1557                                    "'enabled' or 'disabled'\n");
1558                 return -EINVAL;
1559         }
1560
1561         nrb_tiny = lnet_nrb_tiny_calculate();
1562         if (nrb_tiny < 0)
1563                 return -EINVAL;
1564
1565         nrb_small = lnet_nrb_small_calculate();
1566         if (nrb_small < 0)
1567                 return -EINVAL;
1568
1569         nrb_large = lnet_nrb_large_calculate();
1570         if (nrb_large < 0)
1571                 return -EINVAL;
1572
1573         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1574                                                 LNET_NRBPOOLS *
1575                                                 sizeof(lnet_rtrbufpool_t));
1576         if (the_lnet.ln_rtrpools == NULL) {
1577                 LCONSOLE_ERROR_MSG(0x10c,
1578                                    "Failed to initialize router buffe pool\n");
1579                 return -ENOMEM;
1580         }
1581
1582         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1583                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1584                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1585                                               nrb_tiny, i);
1586                 if (rc != 0)
1587                         goto failed;
1588
1589                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1590                                   LNET_NRB_SMALL_PAGES);
1591                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1592                                               nrb_small, i);
1593                 if (rc != 0)
1594                         goto failed;
1595
1596                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1597                                   LNET_NRB_LARGE_PAGES);
1598                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1599                                               nrb_large, i);
1600                 if (rc != 0)
1601                         goto failed;
1602         }
1603
1604         lnet_net_lock(LNET_LOCK_EX);
1605         the_lnet.ln_routing = 1;
1606         lnet_net_unlock(LNET_LOCK_EX);
1607         return 0;
1608
1609  failed:
1610         lnet_rtrpools_free(0);
1611         return rc;
1612 }
1613
1614 static int
1615 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1616 {
1617         int nrb = 0;
1618         int rc = 0;
1619         int i;
1620         lnet_rtrbufpool_t *rtrp;
1621
1622         /* If the provided values for each buffer pool are different than the
1623          * configured values, we need to take action. */
1624         if (tiny >= 0) {
1625                 tiny_router_buffers = tiny;
1626                 nrb = lnet_nrb_tiny_calculate();
1627                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1628                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1629                                                       nrb, i);
1630                         if (rc != 0)
1631                                 return rc;
1632                 }
1633         }
1634         if (small >= 0) {
1635                 small_router_buffers = small;
1636                 nrb = lnet_nrb_small_calculate();
1637                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1638                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1639                                                       nrb, i);
1640                         if (rc != 0)
1641                                 return rc;
1642                 }
1643         }
1644         if (large >= 0) {
1645                 large_router_buffers = large;
1646                 nrb = lnet_nrb_large_calculate();
1647                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1648                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1649                                                       nrb, i);
1650                         if (rc != 0)
1651                                 return rc;
1652                 }
1653         }
1654
1655         return 0;
1656 }
1657
1658 int
1659 lnet_rtrpools_adjust(int tiny, int small, int large)
1660 {
1661         /* this function doesn't revert the changes if adding new buffers
1662          * failed.  It's up to the user space caller to revert the
1663          * changes. */
1664
1665         if (!the_lnet.ln_routing)
1666                 return 0;
1667
1668         return lnet_rtrpools_adjust_helper(tiny, small, large);
1669 }
1670
1671 int
1672 lnet_rtrpools_enable(void)
1673 {
1674         int rc = 0;
1675
1676         if (the_lnet.ln_routing)
1677                 return 0;
1678
1679         if (the_lnet.ln_rtrpools == NULL)
1680                 /* If routing is turned off, and we have never
1681                  * initialized the pools before, just call the
1682                  * standard buffer pool allocation routine as
1683                  * if we are just configuring this for the first
1684                  * time. */
1685                 rc = lnet_rtrpools_alloc(1);
1686         else
1687                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1688         if (rc != 0)
1689                 return rc;
1690
1691         lnet_net_lock(LNET_LOCK_EX);
1692         the_lnet.ln_routing = 1;
1693
1694         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1695         lnet_net_unlock(LNET_LOCK_EX);
1696
1697         return rc;
1698 }
1699
1700 void
1701 lnet_rtrpools_disable(void)
1702 {
1703         if (!the_lnet.ln_routing)
1704                 return;
1705
1706         lnet_net_lock(LNET_LOCK_EX);
1707         the_lnet.ln_routing = 0;
1708         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1709
1710         tiny_router_buffers = 0;
1711         small_router_buffers = 0;
1712         large_router_buffers = 0;
1713         lnet_net_unlock(LNET_LOCK_EX);
1714         lnet_rtrpools_free(1);
1715 }
1716
1717 int
1718 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1719 {
1720         struct lnet_peer        *lp = NULL;
1721         cfs_time_t              now = cfs_time_current();
1722         int                     cpt = lnet_cpt_of_nid(nid, ni);
1723
1724         LASSERT (!in_interrupt ());
1725
1726         CDEBUG (D_NET, "%s notifying %s: %s\n",
1727                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1728                 libcfs_nid2str(nid),
1729                 alive ? "up" : "down");
1730
1731         if (ni != NULL &&
1732             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1733                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1734                       libcfs_nid2str(nid), alive ? "birth" : "death",
1735                       libcfs_nid2str(ni->ni_nid));
1736                 return -EINVAL;
1737         }
1738
1739         /* can't do predictions... */
1740         if (cfs_time_after(when, now)) {
1741                 CWARN("Ignoring prediction from %s of %s %s "
1742                       "%ld seconds in the future\n",
1743                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1744                       libcfs_nid2str(nid), alive ? "up" : "down",
1745                       cfs_duration_sec(cfs_time_sub(when, now)));
1746                 return -EINVAL;
1747         }
1748
1749         if (ni != NULL && !alive &&             /* LND telling me she's down */
1750             !auto_down) {                       /* auto-down disabled */
1751                 CDEBUG(D_NET, "Auto-down disabled\n");
1752                 return 0;
1753         }
1754
1755         lnet_net_lock(cpt);
1756
1757         if (the_lnet.ln_shutdown) {
1758                 lnet_net_unlock(cpt);
1759                 return -ESHUTDOWN;
1760         }
1761
1762         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1763         if (lp == NULL) {
1764                 /* nid not found */
1765                 lnet_net_unlock(cpt);
1766                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1767                 return 0;
1768         }
1769
1770         /* We can't fully trust LND on reporting exact peer last_alive
1771          * if he notifies us about dead peer. For example ksocklnd can
1772          * call us with when == _time_when_the_node_was_booted_ if
1773          * no connections were successfully established */
1774         if (ni != NULL && !alive && when < lp->lp_last_alive)
1775                 when = lp->lp_last_alive;
1776
1777         lnet_notify_locked(lp, ni == NULL, alive, when);
1778
1779         if (ni != NULL)
1780                 lnet_ni_notify_locked(ni, lp);
1781
1782         lnet_peer_decref_locked(lp);
1783
1784         lnet_net_unlock(cpt);
1785         return 0;
1786 }
1787 EXPORT_SYMBOL(lnet_notify);