Whamcloud - gitweb
LU-13510 lnet: Correct the default LND timeout
[fs/lustre-release.git] / lnet / lnet / api-ni.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 #define DEBUG_SUBSYSTEM S_LNET
34
35 #include <linux/ctype.h>
36 #include <linux/log2.h>
37 #include <linux/ktime.h>
38 #include <linux/moduleparam.h>
39 #include <linux/uaccess.h>
40 #ifdef HAVE_SCHED_HEADERS
41 #include <linux/sched/signal.h>
42 #endif
43
44 #include <lnet/lib-lnet.h>
45
46 #define D_LNI D_CONSOLE
47
48 /*
49  * initialize ln_api_mutex statically, since it needs to be used in
50  * discovery_set callback. That module parameter callback can be called
51  * before module init completes. The mutex needs to be ready for use then.
52  */
53 struct lnet the_lnet = {
54         .ln_api_mutex = __MUTEX_INITIALIZER(the_lnet.ln_api_mutex),
55 };              /* THE state of the network */
56 EXPORT_SYMBOL(the_lnet);
57
58 static char *ip2nets = "";
59 module_param(ip2nets, charp, 0444);
60 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
61
62 static char *networks = "";
63 module_param(networks, charp, 0444);
64 MODULE_PARM_DESC(networks, "local networks");
65
66 static char *routes = "";
67 module_param(routes, charp, 0444);
68 MODULE_PARM_DESC(routes, "routes to non-local networks");
69
70 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
71 module_param(rnet_htable_size, int, 0444);
72 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
73
74 static int use_tcp_bonding = false;
75 module_param(use_tcp_bonding, int, 0444);
76 MODULE_PARM_DESC(use_tcp_bonding,
77                  "Set to 1 to use socklnd bonding. 0 to use Multi-Rail");
78
79 unsigned int lnet_numa_range = 0;
80 module_param(lnet_numa_range, uint, 0444);
81 MODULE_PARM_DESC(lnet_numa_range,
82                 "NUMA range to consider during Multi-Rail selection");
83
84 /*
85  * lnet_health_sensitivity determines by how much we decrement the health
86  * value on sending error. The value defaults to 100, which means health
87  * interface health is decremented by 100 points every failure.
88  */
89 unsigned int lnet_health_sensitivity = 100;
90 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
91 #ifdef HAVE_KERNEL_PARAM_OPS
92 static struct kernel_param_ops param_ops_health_sensitivity = {
93         .set = sensitivity_set,
94         .get = param_get_int,
95 };
96 #define param_check_health_sensitivity(name, p) \
97                 __param_check(name, p, int)
98 module_param(lnet_health_sensitivity, health_sensitivity, S_IRUGO|S_IWUSR);
99 #else
100 module_param_call(lnet_health_sensitivity, sensitivity_set, param_get_int,
101                   &lnet_health_sensitivity, S_IRUGO|S_IWUSR);
102 #endif
103 MODULE_PARM_DESC(lnet_health_sensitivity,
104                 "Value to decrement the health value by on error");
105
106 /*
107  * lnet_recovery_interval determines how often we should perform recovery
108  * on unhealthy interfaces.
109  */
110 unsigned int lnet_recovery_interval = 1;
111 static int recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp);
112 #ifdef HAVE_KERNEL_PARAM_OPS
113 static struct kernel_param_ops param_ops_recovery_interval = {
114         .set = recovery_interval_set,
115         .get = param_get_int,
116 };
117 #define param_check_recovery_interval(name, p) \
118                 __param_check(name, p, int)
119 module_param(lnet_recovery_interval, recovery_interval, S_IRUGO|S_IWUSR);
120 #else
121 module_param_call(lnet_recovery_interval, recovery_interval_set, param_get_int,
122                   &lnet_recovery_interval, S_IRUGO|S_IWUSR);
123 #endif
124 MODULE_PARM_DESC(lnet_recovery_interval,
125                 "Interval to recover unhealthy interfaces in seconds");
126
127 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
128 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
129
130 static struct kernel_param_ops param_ops_interfaces_max = {
131         .set = intf_max_set,
132         .get = param_get_int,
133 };
134
135 #define param_check_interfaces_max(name, p) \
136                 __param_check(name, p, int)
137
138 #ifdef HAVE_KERNEL_PARAM_OPS
139 module_param(lnet_interfaces_max, interfaces_max, 0644);
140 #else
141 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
142                   &param_ops_interfaces_max, 0644);
143 #endif
144 MODULE_PARM_DESC(lnet_interfaces_max,
145                 "Maximum number of interfaces in a node.");
146
147 unsigned lnet_peer_discovery_disabled = 0;
148 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
149
150 static struct kernel_param_ops param_ops_discovery_disabled = {
151         .set = discovery_set,
152         .get = param_get_int,
153 };
154
155 #define param_check_discovery_disabled(name, p) \
156                 __param_check(name, p, int)
157 #ifdef HAVE_KERNEL_PARAM_OPS
158 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
159 #else
160 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
161                   &param_ops_discovery_disabled, 0644);
162 #endif
163 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
164                 "Set to 1 to disable peer discovery on this node.");
165
166 unsigned int lnet_drop_asym_route;
167 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
168
169 static struct kernel_param_ops param_ops_drop_asym_route = {
170         .set = drop_asym_route_set,
171         .get = param_get_int,
172 };
173
174 #define param_check_drop_asym_route(name, p)    \
175         __param_check(name, p, int)
176 #ifdef HAVE_KERNEL_PARAM_OPS
177 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
178 #else
179 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
180                   &param_ops_drop_asym_route, 0644);
181 #endif
182 MODULE_PARM_DESC(lnet_drop_asym_route,
183                  "Set to 1 to drop asymmetrical route messages.");
184
185 #define LNET_TRANSACTION_TIMEOUT_NO_HEALTH_DEFAULT 50
186 #define LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT 50
187
188 unsigned lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT;
189 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
190 #ifdef HAVE_KERNEL_PARAM_OPS
191 static struct kernel_param_ops param_ops_transaction_timeout = {
192         .set = transaction_to_set,
193         .get = param_get_int,
194 };
195
196 #define param_check_transaction_timeout(name, p) \
197                 __param_check(name, p, int)
198 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
199 #else
200 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
201                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
202 #endif
203 MODULE_PARM_DESC(lnet_transaction_timeout,
204                 "Maximum number of seconds to wait for a peer response.");
205
206 #define LNET_RETRY_COUNT_HEALTH_DEFAULT 2
207 unsigned lnet_retry_count = LNET_RETRY_COUNT_HEALTH_DEFAULT;
208 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
209 #ifdef HAVE_KERNEL_PARAM_OPS
210 static struct kernel_param_ops param_ops_retry_count = {
211         .set = retry_count_set,
212         .get = param_get_int,
213 };
214
215 #define param_check_retry_count(name, p) \
216                 __param_check(name, p, int)
217 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
218 #else
219 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
220                   &lnet_retry_count, S_IRUGO|S_IWUSR);
221 #endif
222 MODULE_PARM_DESC(lnet_retry_count,
223                  "Maximum number of times to retry transmitting a message");
224
225
226 #define LNET_LND_TIMEOUT_DEFAULT ((LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT - 1) / \
227                                   (LNET_RETRY_COUNT_HEALTH_DEFAULT + 1))
228 unsigned int lnet_lnd_timeout = LNET_LND_TIMEOUT_DEFAULT;
229 static void lnet_set_lnd_timeout(void)
230 {
231         lnet_lnd_timeout = (lnet_transaction_timeout - 1) /
232                            (lnet_retry_count + 1);
233 }
234
235 unsigned int lnet_current_net_count;
236
237 /*
238  * This sequence number keeps track of how many times DLC was used to
239  * update the local NIs. It is incremented when a NI is added or
240  * removed and checked when sending a message to determine if there is
241  * a need to re-run the selection algorithm. See lnet_select_pathway()
242  * for more details on its usage.
243  */
244 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
245
246 static int lnet_ping(struct lnet_process_id id, signed long timeout,
247                      struct lnet_process_id __user *ids, int n_ids);
248
249 static int lnet_discover(struct lnet_process_id id, __u32 force,
250                          struct lnet_process_id __user *ids, int n_ids);
251
252 static int
253 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
254 {
255         int rc;
256         unsigned *sensitivity = (unsigned *)kp->arg;
257         unsigned long value;
258
259         rc = kstrtoul(val, 0, &value);
260         if (rc) {
261                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
262                 return rc;
263         }
264
265         /*
266          * The purpose of locking the api_mutex here is to ensure that
267          * the correct value ends up stored properly.
268          */
269         mutex_lock(&the_lnet.ln_api_mutex);
270
271         if (value > LNET_MAX_HEALTH_VALUE) {
272                 mutex_unlock(&the_lnet.ln_api_mutex);
273                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
274                        LNET_MAX_HEALTH_VALUE, value);
275                 return -EINVAL;
276         }
277
278         /*
279          * if we're turning on health then use the health timeout
280          * defaults.
281          */
282         if (*sensitivity == 0 && value != 0) {
283                 lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT;
284                 lnet_retry_count = LNET_RETRY_COUNT_HEALTH_DEFAULT;
285                 lnet_set_lnd_timeout();
286         /*
287          * if we're turning off health then use the no health timeout
288          * default.
289          */
290         } else if (*sensitivity != 0 && value == 0) {
291                 lnet_transaction_timeout =
292                         LNET_TRANSACTION_TIMEOUT_NO_HEALTH_DEFAULT;
293                 lnet_retry_count = 0;
294                 lnet_set_lnd_timeout();
295         }
296
297         *sensitivity = value;
298
299         mutex_unlock(&the_lnet.ln_api_mutex);
300
301         return 0;
302 }
303
304 static int
305 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
306 {
307         int rc;
308         unsigned *interval = (unsigned *)kp->arg;
309         unsigned long value;
310
311         rc = kstrtoul(val, 0, &value);
312         if (rc) {
313                 CERROR("Invalid module parameter value for 'lnet_recovery_interval'\n");
314                 return rc;
315         }
316
317         if (value < 1) {
318                 CERROR("lnet_recovery_interval must be at least 1 second\n");
319                 return -EINVAL;
320         }
321
322         /*
323          * The purpose of locking the api_mutex here is to ensure that
324          * the correct value ends up stored properly.
325          */
326         mutex_lock(&the_lnet.ln_api_mutex);
327
328         *interval = value;
329
330         mutex_unlock(&the_lnet.ln_api_mutex);
331
332         return 0;
333 }
334
335 static int
336 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
337 {
338         int rc;
339         unsigned *discovery_off = (unsigned *)kp->arg;
340         unsigned long value;
341         struct lnet_ping_buffer *pbuf;
342
343         rc = kstrtoul(val, 0, &value);
344         if (rc) {
345                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
346                 return rc;
347         }
348
349         value = (value) ? 1 : 0;
350
351         /*
352          * The purpose of locking the api_mutex here is to ensure that
353          * the correct value ends up stored properly.
354          */
355         mutex_lock(&the_lnet.ln_api_mutex);
356
357         if (value == *discovery_off) {
358                 mutex_unlock(&the_lnet.ln_api_mutex);
359                 return 0;
360         }
361
362         /*
363          * We still want to set the discovery value even when LNet is not
364          * running. This is the case when LNet is being loaded and we want
365          * the module parameters to take effect. Otherwise if we're
366          * changing the value dynamically, we want to set it after
367          * updating the peers
368          */
369         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
370                 *discovery_off = value;
371                 mutex_unlock(&the_lnet.ln_api_mutex);
372                 return 0;
373         }
374
375         /* tell peers that discovery setting has changed */
376         lnet_net_lock(LNET_LOCK_EX);
377         pbuf = the_lnet.ln_ping_target;
378         if (value)
379                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
380         else
381                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
382         lnet_net_unlock(LNET_LOCK_EX);
383
384         /* only send a push when we're turning off discovery */
385         if (*discovery_off <= 0 && value > 0)
386                 lnet_push_update_to_peers(1);
387         *discovery_off = value;
388
389         mutex_unlock(&the_lnet.ln_api_mutex);
390
391         return 0;
392 }
393
394 static int
395 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
396 {
397         int rc;
398         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
399         unsigned long value;
400
401         rc = kstrtoul(val, 0, &value);
402         if (rc) {
403                 CERROR("Invalid module parameter value for "
404                        "'lnet_drop_asym_route'\n");
405                 return rc;
406         }
407
408         /*
409          * The purpose of locking the api_mutex here is to ensure that
410          * the correct value ends up stored properly.
411          */
412         mutex_lock(&the_lnet.ln_api_mutex);
413
414         if (value == *drop_asym_route) {
415                 mutex_unlock(&the_lnet.ln_api_mutex);
416                 return 0;
417         }
418
419         *drop_asym_route = value;
420
421         mutex_unlock(&the_lnet.ln_api_mutex);
422
423         return 0;
424 }
425
426 static int
427 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
428 {
429         int rc;
430         unsigned *transaction_to = (unsigned *)kp->arg;
431         unsigned long value;
432
433         rc = kstrtoul(val, 0, &value);
434         if (rc) {
435                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
436                 return rc;
437         }
438
439         /*
440          * The purpose of locking the api_mutex here is to ensure that
441          * the correct value ends up stored properly.
442          */
443         mutex_lock(&the_lnet.ln_api_mutex);
444
445         if (value < lnet_retry_count || value == 0) {
446                 mutex_unlock(&the_lnet.ln_api_mutex);
447                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
448                        "Has to be greater than lnet_retry_count (%u)\n",
449                        value, lnet_retry_count);
450                 return -EINVAL;
451         }
452
453         if (value == *transaction_to) {
454                 mutex_unlock(&the_lnet.ln_api_mutex);
455                 return 0;
456         }
457
458         *transaction_to = value;
459         /* Update the lnet_lnd_timeout now that we've modified the
460          * transaction timeout
461          */
462         lnet_set_lnd_timeout();
463
464         mutex_unlock(&the_lnet.ln_api_mutex);
465
466         return 0;
467 }
468
469 static int
470 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
471 {
472         int rc;
473         unsigned *retry_count = (unsigned *)kp->arg;
474         unsigned long value;
475
476         rc = kstrtoul(val, 0, &value);
477         if (rc) {
478                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
479                 return rc;
480         }
481
482         /*
483          * The purpose of locking the api_mutex here is to ensure that
484          * the correct value ends up stored properly.
485          */
486         mutex_lock(&the_lnet.ln_api_mutex);
487
488         if (lnet_health_sensitivity == 0) {
489                 mutex_unlock(&the_lnet.ln_api_mutex);
490                 CERROR("Can not set retry_count when health feature is turned off\n");
491                 return -EINVAL;
492         }
493
494         if (value > lnet_transaction_timeout) {
495                 mutex_unlock(&the_lnet.ln_api_mutex);
496                 CERROR("Invalid value for lnet_retry_count (%lu). "
497                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
498                        value, lnet_transaction_timeout);
499                 return -EINVAL;
500         }
501
502         *retry_count = value;
503
504         /* Update the lnet_lnd_timeout now that we've modified the
505          * retry count
506          */
507         lnet_set_lnd_timeout();
508
509         mutex_unlock(&the_lnet.ln_api_mutex);
510
511         return 0;
512 }
513
514 static int
515 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
516 {
517         int value, rc;
518
519         rc = kstrtoint(val, 0, &value);
520         if (rc) {
521                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
522                 return rc;
523         }
524
525         if (value < LNET_INTERFACES_MIN) {
526                 CWARN("max interfaces provided are too small, setting to %d\n",
527                       LNET_INTERFACES_MAX_DEFAULT);
528                 value = LNET_INTERFACES_MAX_DEFAULT;
529         }
530
531         *(int *)kp->arg = value;
532
533         return 0;
534 }
535
536 static const char *
537 lnet_get_routes(void)
538 {
539         return routes;
540 }
541
542 static const char *
543 lnet_get_networks(void)
544 {
545         const char *nets;
546         int rc;
547
548         if (*networks != 0 && *ip2nets != 0) {
549                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
550                                    "'ip2nets' but not both at once\n");
551                 return NULL;
552         }
553
554         if (*ip2nets != 0) {
555                 rc = lnet_parse_ip2nets(&nets, ip2nets);
556                 return (rc == 0) ? nets : NULL;
557         }
558
559         if (*networks != 0)
560                 return networks;
561
562         return "tcp";
563 }
564
565 static void
566 lnet_init_locks(void)
567 {
568         spin_lock_init(&the_lnet.ln_eq_wait_lock);
569         spin_lock_init(&the_lnet.ln_msg_resend_lock);
570         init_completion(&the_lnet.ln_mt_wait_complete);
571         mutex_init(&the_lnet.ln_lnd_mutex);
572 }
573
574 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
575 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
576                                             *  MDs kmem_cache */
577 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
578 struct kmem_cache *lnet_msg_cachep;
579
580 static int
581 lnet_slab_setup(void)
582 {
583         /* create specific kmem_cache for MEs and small MDs (i.e., originally
584          * allocated in <size-xxx> kmem_cache).
585          */
586         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
587                                             0, 0, NULL);
588         if (!lnet_mes_cachep)
589                 return -ENOMEM;
590
591         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
592                                                   LNET_SMALL_MD_SIZE, 0, 0,
593                                                   NULL);
594         if (!lnet_small_mds_cachep)
595                 return -ENOMEM;
596
597         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
598                                             0, 0, NULL);
599         if (!lnet_rspt_cachep)
600                 return -ENOMEM;
601
602         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
603                                             0, 0, NULL);
604         if (!lnet_msg_cachep)
605                 return -ENOMEM;
606
607         return 0;
608 }
609
610 static void
611 lnet_slab_cleanup(void)
612 {
613         if (lnet_msg_cachep) {
614                 kmem_cache_destroy(lnet_msg_cachep);
615                 lnet_msg_cachep = NULL;
616         }
617
618
619         if (lnet_rspt_cachep) {
620                 kmem_cache_destroy(lnet_rspt_cachep);
621                 lnet_rspt_cachep = NULL;
622         }
623
624         if (lnet_small_mds_cachep) {
625                 kmem_cache_destroy(lnet_small_mds_cachep);
626                 lnet_small_mds_cachep = NULL;
627         }
628
629         if (lnet_mes_cachep) {
630                 kmem_cache_destroy(lnet_mes_cachep);
631                 lnet_mes_cachep = NULL;
632         }
633 }
634
635 static int
636 lnet_create_remote_nets_table(void)
637 {
638         int               i;
639         struct list_head *hash;
640
641         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
642         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
643         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
644         if (hash == NULL) {
645                 CERROR("Failed to create remote nets hash table\n");
646                 return -ENOMEM;
647         }
648
649         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
650                 INIT_LIST_HEAD(&hash[i]);
651         the_lnet.ln_remote_nets_hash = hash;
652         return 0;
653 }
654
655 static void
656 lnet_destroy_remote_nets_table(void)
657 {
658         int i;
659
660         if (the_lnet.ln_remote_nets_hash == NULL)
661                 return;
662
663         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
664                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
665
666         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
667                            LNET_REMOTE_NETS_HASH_SIZE);
668         the_lnet.ln_remote_nets_hash = NULL;
669 }
670
671 static void
672 lnet_destroy_locks(void)
673 {
674         if (the_lnet.ln_res_lock != NULL) {
675                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
676                 the_lnet.ln_res_lock = NULL;
677         }
678
679         if (the_lnet.ln_net_lock != NULL) {
680                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
681                 the_lnet.ln_net_lock = NULL;
682         }
683 }
684
685 static int
686 lnet_create_locks(void)
687 {
688         lnet_init_locks();
689
690         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
691         if (the_lnet.ln_res_lock == NULL)
692                 goto failed;
693
694         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
695         if (the_lnet.ln_net_lock == NULL)
696                 goto failed;
697
698         return 0;
699
700  failed:
701         lnet_destroy_locks();
702         return -ENOMEM;
703 }
704
705 static void lnet_assert_wire_constants(void)
706 {
707         /* Wire protocol assertions generated by 'wirecheck'
708          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
709          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
710          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
711
712         /* Constants... */
713         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
714         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
715         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
716         BUILD_BUG_ON(LNET_MSG_ACK != 0);
717         BUILD_BUG_ON(LNET_MSG_PUT != 1);
718         BUILD_BUG_ON(LNET_MSG_GET != 2);
719         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
720         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
721
722         /* Checks for struct lnet_handle_wire */
723         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
724         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
725                                    wh_interface_cookie) != 0);
726         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
727         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
728                                    wh_object_cookie) != 8);
729         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
730
731         /* Checks for struct struct lnet_magicversion */
732         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
733         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
734         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
735         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
736         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
737         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
738                                    version_minor) != 6);
739         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
740
741         /* Checks for struct struct lnet_hdr */
742         BUILD_BUG_ON((int)sizeof(struct lnet_hdr) != 72);
743         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_nid) != 0);
744         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_nid) != 8);
745         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_nid) != 8);
746         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_nid) != 8);
747         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_pid) != 16);
748         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_pid) != 4);
749         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_pid) != 20);
750         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_pid) != 4);
751         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, type) != 24);
752         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->type) != 4);
753         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, payload_length) != 28);
754         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->payload_length) != 4);
755         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg) != 32);
756         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg) != 40);
757
758         /* Ack */
759         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) != 32);
760         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) != 16);
761         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.match_bits) != 48);
762         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) != 8);
763         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.mlength) != 56);
764         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) != 4);
765
766         /* Put */
767         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) != 32);
768         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) != 16);
769         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.match_bits) != 48);
770         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) != 8);
771         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.hdr_data) != 56);
772         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) != 8);
773         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ptl_index) != 64);
774         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) != 4);
775         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.offset) != 68);
776         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) != 4);
777
778         /* Get */
779         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.return_wmd) != 32);
780         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) != 16);
781         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.match_bits) != 48);
782         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) != 8);
783         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.ptl_index) != 56);
784         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) != 4);
785         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.src_offset) != 60);
786         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) != 4);
787         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.sink_length) != 64);
788         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) != 4);
789
790         /* Reply */
791         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) != 32);
792         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) != 16);
793
794         /* Hello */
795         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.incarnation) != 32);
796         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) != 8);
797         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.type) != 40);
798         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) != 4);
799
800         /* Checks for struct lnet_ni_status and related constants */
801         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
802         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
803         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
804
805         /* Checks for struct lnet_ni_status */
806         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
807         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
808         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
809         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
810         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
811         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_unused) != 12);
812         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) != 4);
813
814         /* Checks for struct lnet_ping_info and related constants */
815         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
816         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
817         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
818         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
819         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
820         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
821         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
822         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 31);
823
824         /* Checks for struct lnet_ping_info */
825         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
826         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
827         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
828         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
829         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
830         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
831         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
832         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
833         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
834         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
835         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) != 0);
836 }
837
838 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
839 {
840         const struct lnet_lnd *lnd;
841
842         /* holding lnd mutex */
843         if (type >= NUM_LNDS)
844                 return NULL;
845         lnd = the_lnet.ln_lnds[type];
846         LASSERT(!lnd || lnd->lnd_type == type);
847
848         return lnd;
849 }
850
851 unsigned int
852 lnet_get_lnd_timeout(void)
853 {
854         return lnet_lnd_timeout;
855 }
856 EXPORT_SYMBOL(lnet_get_lnd_timeout);
857
858 void
859 lnet_register_lnd(const struct lnet_lnd *lnd)
860 {
861         mutex_lock(&the_lnet.ln_lnd_mutex);
862
863         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
864         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
865
866         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
867
868         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
869
870         mutex_unlock(&the_lnet.ln_lnd_mutex);
871 }
872 EXPORT_SYMBOL(lnet_register_lnd);
873
874 void
875 lnet_unregister_lnd(const struct lnet_lnd *lnd)
876 {
877         mutex_lock(&the_lnet.ln_lnd_mutex);
878
879         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
880
881         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
882         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
883
884         mutex_unlock(&the_lnet.ln_lnd_mutex);
885 }
886 EXPORT_SYMBOL(lnet_unregister_lnd);
887
888 void
889 lnet_counters_get_common(struct lnet_counters_common *common)
890 {
891         struct lnet_counters *ctr;
892         int i;
893
894         memset(common, 0, sizeof(*common));
895
896         lnet_net_lock(LNET_LOCK_EX);
897
898         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
899                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
900                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
901                 common->lcc_errors       += ctr->lct_common.lcc_errors;
902                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
903                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
904                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
905                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
906                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
907                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
908                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
909                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
910         }
911         lnet_net_unlock(LNET_LOCK_EX);
912 }
913 EXPORT_SYMBOL(lnet_counters_get_common);
914
915 void
916 lnet_counters_get(struct lnet_counters *counters)
917 {
918         struct lnet_counters *ctr;
919         struct lnet_counters_health *health = &counters->lct_health;
920         int             i;
921
922         memset(counters, 0, sizeof(*counters));
923
924         lnet_counters_get_common(&counters->lct_common);
925
926         lnet_net_lock(LNET_LOCK_EX);
927
928         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
929                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
930                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
931                 health->lch_response_timeout_count +=
932                                 ctr->lct_health.lch_response_timeout_count;
933                 health->lch_local_interrupt_count +=
934                                 ctr->lct_health.lch_local_interrupt_count;
935                 health->lch_local_dropped_count +=
936                                 ctr->lct_health.lch_local_dropped_count;
937                 health->lch_local_aborted_count +=
938                                 ctr->lct_health.lch_local_aborted_count;
939                 health->lch_local_no_route_count +=
940                                 ctr->lct_health.lch_local_no_route_count;
941                 health->lch_local_timeout_count +=
942                                 ctr->lct_health.lch_local_timeout_count;
943                 health->lch_local_error_count +=
944                                 ctr->lct_health.lch_local_error_count;
945                 health->lch_remote_dropped_count +=
946                                 ctr->lct_health.lch_remote_dropped_count;
947                 health->lch_remote_error_count +=
948                                 ctr->lct_health.lch_remote_error_count;
949                 health->lch_remote_timeout_count +=
950                                 ctr->lct_health.lch_remote_timeout_count;
951                 health->lch_network_timeout_count +=
952                                 ctr->lct_health.lch_network_timeout_count;
953         }
954         lnet_net_unlock(LNET_LOCK_EX);
955 }
956 EXPORT_SYMBOL(lnet_counters_get);
957
958 void
959 lnet_counters_reset(void)
960 {
961         struct lnet_counters *counters;
962         int             i;
963
964         lnet_net_lock(LNET_LOCK_EX);
965
966         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
967                 memset(counters, 0, sizeof(struct lnet_counters));
968
969         lnet_net_unlock(LNET_LOCK_EX);
970 }
971
972 static char *
973 lnet_res_type2str(int type)
974 {
975         switch (type) {
976         default:
977                 LBUG();
978         case LNET_COOKIE_TYPE_MD:
979                 return "MD";
980         case LNET_COOKIE_TYPE_ME:
981                 return "ME";
982         case LNET_COOKIE_TYPE_EQ:
983                 return "EQ";
984         }
985 }
986
987 static void
988 lnet_res_container_cleanup(struct lnet_res_container *rec)
989 {
990         int     count = 0;
991
992         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
993                 return;
994
995         while (!list_empty(&rec->rec_active)) {
996                 struct list_head *e = rec->rec_active.next;
997
998                 list_del_init(e);
999                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
1000                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
1001
1002                 } else { /* NB: Active MEs should be attached on portals */
1003                         LBUG();
1004                 }
1005                 count++;
1006         }
1007
1008         if (count > 0) {
1009                 /* Found alive MD/ME/EQ, user really should unlink/free
1010                  * all of them before finalize LNet, but if someone didn't,
1011                  * we have to recycle garbage for him */
1012                 CERROR("%d active elements on exit of %s container\n",
1013                        count, lnet_res_type2str(rec->rec_type));
1014         }
1015
1016         if (rec->rec_lh_hash != NULL) {
1017                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
1018                 rec->rec_lh_hash = NULL;
1019         }
1020
1021         rec->rec_type = 0; /* mark it as finalized */
1022 }
1023
1024 static int
1025 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1026 {
1027         int     rc = 0;
1028         int     i;
1029
1030         LASSERT(rec->rec_type == 0);
1031
1032         rec->rec_type = type;
1033         INIT_LIST_HEAD(&rec->rec_active);
1034
1035         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1036
1037         /* Arbitrary choice of hash table size */
1038         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1039                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1040         if (rec->rec_lh_hash == NULL) {
1041                 rc = -ENOMEM;
1042                 goto out;
1043         }
1044
1045         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1046                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1047
1048         return 0;
1049
1050 out:
1051         CERROR("Failed to setup %s resource container\n",
1052                lnet_res_type2str(type));
1053         lnet_res_container_cleanup(rec);
1054         return rc;
1055 }
1056
1057 static void
1058 lnet_res_containers_destroy(struct lnet_res_container **recs)
1059 {
1060         struct lnet_res_container       *rec;
1061         int                             i;
1062
1063         cfs_percpt_for_each(rec, i, recs)
1064                 lnet_res_container_cleanup(rec);
1065
1066         cfs_percpt_free(recs);
1067 }
1068
1069 static struct lnet_res_container **
1070 lnet_res_containers_create(int type)
1071 {
1072         struct lnet_res_container       **recs;
1073         struct lnet_res_container       *rec;
1074         int                             rc;
1075         int                             i;
1076
1077         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1078         if (recs == NULL) {
1079                 CERROR("Failed to allocate %s resource containers\n",
1080                        lnet_res_type2str(type));
1081                 return NULL;
1082         }
1083
1084         cfs_percpt_for_each(rec, i, recs) {
1085                 rc = lnet_res_container_setup(rec, i, type);
1086                 if (rc != 0) {
1087                         lnet_res_containers_destroy(recs);
1088                         return NULL;
1089                 }
1090         }
1091
1092         return recs;
1093 }
1094
1095 struct lnet_libhandle *
1096 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1097 {
1098         /* ALWAYS called with lnet_res_lock held */
1099         struct list_head        *head;
1100         struct lnet_libhandle   *lh;
1101         unsigned int            hash;
1102
1103         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1104                 return NULL;
1105
1106         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1107         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1108
1109         list_for_each_entry(lh, head, lh_hash_chain) {
1110                 if (lh->lh_cookie == cookie)
1111                         return lh;
1112         }
1113
1114         return NULL;
1115 }
1116
1117 void
1118 lnet_res_lh_initialize(struct lnet_res_container *rec,
1119                        struct lnet_libhandle *lh)
1120 {
1121         /* ALWAYS called with lnet_res_lock held */
1122         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1123         unsigned int    hash;
1124
1125         lh->lh_cookie = rec->rec_lh_cookie;
1126         rec->rec_lh_cookie += 1 << ibits;
1127
1128         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1129
1130         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1131 }
1132
1133 struct list_head **
1134 lnet_create_array_of_queues(void)
1135 {
1136         struct list_head **qs;
1137         struct list_head *q;
1138         int i;
1139
1140         qs = cfs_percpt_alloc(lnet_cpt_table(),
1141                               sizeof(struct list_head));
1142         if (!qs) {
1143                 CERROR("Failed to allocate queues\n");
1144                 return NULL;
1145         }
1146
1147         cfs_percpt_for_each(q, i, qs)
1148                 INIT_LIST_HEAD(q);
1149
1150         return qs;
1151 }
1152
1153 static int lnet_unprepare(void);
1154
1155 static int
1156 lnet_prepare(lnet_pid_t requested_pid)
1157 {
1158         /* Prepare to bring up the network */
1159         struct lnet_res_container **recs;
1160         int                       rc = 0;
1161
1162         if (requested_pid == LNET_PID_ANY) {
1163                 /* Don't instantiate LNET just for me */
1164                 return -ENETDOWN;
1165         }
1166
1167         LASSERT(the_lnet.ln_refcount == 0);
1168
1169         the_lnet.ln_routing = 0;
1170
1171         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1172         the_lnet.ln_pid = requested_pid;
1173
1174         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1175         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1176         INIT_LIST_HEAD(&the_lnet.ln_nets);
1177         INIT_LIST_HEAD(&the_lnet.ln_routers);
1178         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1179         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1180         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1181         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1182         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1183         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1184         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1185         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1186         the_lnet.ln_mt_handler = NULL;
1187         init_completion(&the_lnet.ln_started);
1188
1189         rc = lnet_slab_setup();
1190         if (rc != 0)
1191                 goto failed;
1192
1193         rc = lnet_create_remote_nets_table();
1194         if (rc != 0)
1195                 goto failed;
1196
1197         /*
1198          * NB the interface cookie in wire handles guards against delayed
1199          * replies and ACKs appearing valid after reboot.
1200          */
1201         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1202
1203         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1204                                                 sizeof(struct lnet_counters));
1205         if (the_lnet.ln_counters == NULL) {
1206                 CERROR("Failed to allocate counters for LNet\n");
1207                 rc = -ENOMEM;
1208                 goto failed;
1209         }
1210
1211         rc = lnet_peer_tables_create();
1212         if (rc != 0)
1213                 goto failed;
1214
1215         rc = lnet_msg_containers_create();
1216         if (rc != 0)
1217                 goto failed;
1218
1219         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1220                                       LNET_COOKIE_TYPE_EQ);
1221         if (rc != 0)
1222                 goto failed;
1223
1224         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1225         if (recs == NULL) {
1226                 rc = -ENOMEM;
1227                 goto failed;
1228         }
1229
1230         the_lnet.ln_md_containers = recs;
1231
1232         rc = lnet_portals_create();
1233         if (rc != 0) {
1234                 CERROR("Failed to create portals for LNet: %d\n", rc);
1235                 goto failed;
1236         }
1237
1238         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1239         if (!the_lnet.ln_mt_zombie_rstqs) {
1240                 rc = -ENOMEM;
1241                 goto failed;
1242         }
1243
1244         return 0;
1245
1246  failed:
1247         lnet_unprepare();
1248         return rc;
1249 }
1250
1251 static int
1252 lnet_unprepare (void)
1253 {
1254         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1255          * have shut down already, so it is safe to unlink and free all
1256          * descriptors, even those that appear committed to a network op (eg MD
1257          * with non-zero pending count) */
1258
1259         lnet_fail_nid(LNET_NID_ANY, 0);
1260
1261         LASSERT(the_lnet.ln_refcount == 0);
1262         LASSERT(list_empty(&the_lnet.ln_test_peers));
1263         LASSERT(list_empty(&the_lnet.ln_nets));
1264
1265         if (the_lnet.ln_mt_zombie_rstqs) {
1266                 lnet_clean_zombie_rstqs();
1267                 the_lnet.ln_mt_zombie_rstqs = NULL;
1268         }
1269
1270         lnet_assert_handler_unused(the_lnet.ln_mt_handler);
1271         the_lnet.ln_mt_handler = NULL;
1272
1273         lnet_portals_destroy();
1274
1275         if (the_lnet.ln_md_containers != NULL) {
1276                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1277                 the_lnet.ln_md_containers = NULL;
1278         }
1279
1280         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1281
1282         lnet_msg_containers_destroy();
1283         lnet_peer_uninit();
1284         lnet_rtrpools_free(0);
1285
1286         if (the_lnet.ln_counters != NULL) {
1287                 cfs_percpt_free(the_lnet.ln_counters);
1288                 the_lnet.ln_counters = NULL;
1289         }
1290         lnet_destroy_remote_nets_table();
1291         lnet_slab_cleanup();
1292
1293         return 0;
1294 }
1295
1296 struct lnet_ni  *
1297 lnet_net2ni_locked(__u32 net_id, int cpt)
1298 {
1299         struct lnet_ni   *ni;
1300         struct lnet_net  *net;
1301
1302         LASSERT(cpt != LNET_LOCK_EX);
1303
1304         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1305                 if (net->net_id == net_id) {
1306                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1307                                         ni_netlist);
1308                         return ni;
1309                 }
1310         }
1311
1312         return NULL;
1313 }
1314
1315 struct lnet_ni *
1316 lnet_net2ni_addref(__u32 net)
1317 {
1318         struct lnet_ni *ni;
1319
1320         lnet_net_lock(0);
1321         ni = lnet_net2ni_locked(net, 0);
1322         if (ni)
1323                 lnet_ni_addref_locked(ni, 0);
1324         lnet_net_unlock(0);
1325
1326         return ni;
1327 }
1328 EXPORT_SYMBOL(lnet_net2ni_addref);
1329
1330 struct lnet_net *
1331 lnet_get_net_locked(__u32 net_id)
1332 {
1333         struct lnet_net  *net;
1334
1335         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1336                 if (net->net_id == net_id)
1337                         return net;
1338         }
1339
1340         return NULL;
1341 }
1342
1343 unsigned int
1344 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
1345 {
1346         __u64           key = nid;
1347         unsigned int    val;
1348
1349         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1350
1351         if (number == 1)
1352                 return 0;
1353
1354         val = hash_long(key, LNET_CPT_BITS);
1355         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1356         if (val < number)
1357                 return val;
1358
1359         return (unsigned int)(key + val + (val >> 1)) % number;
1360 }
1361
1362 int
1363 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
1364 {
1365         struct lnet_net *net;
1366
1367         /* must called with hold of lnet_net_lock */
1368         if (LNET_CPT_NUMBER == 1)
1369                 return 0; /* the only one */
1370
1371         /*
1372          * If NI is provided then use the CPT identified in the NI cpt
1373          * list if one exists. If one doesn't exist, then that NI is
1374          * associated with all CPTs and it follows that the net it belongs
1375          * to is implicitly associated with all CPTs, so just hash the nid
1376          * and return that.
1377          */
1378         if (ni != NULL) {
1379                 if (ni->ni_cpts != NULL)
1380                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1381                                                              ni->ni_ncpts)];
1382                 else
1383                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1384         }
1385
1386         /* no NI provided so look at the net */
1387         net = lnet_get_net_locked(LNET_NIDNET(nid));
1388
1389         if (net != NULL && net->net_cpts != NULL) {
1390                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1391         }
1392
1393         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1394 }
1395
1396 int
1397 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
1398 {
1399         int     cpt;
1400         int     cpt2;
1401
1402         if (LNET_CPT_NUMBER == 1)
1403                 return 0; /* the only one */
1404
1405         cpt = lnet_net_lock_current();
1406
1407         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1408
1409         lnet_net_unlock(cpt);
1410
1411         return cpt2;
1412 }
1413 EXPORT_SYMBOL(lnet_cpt_of_nid);
1414
1415 int
1416 lnet_islocalnet_locked(__u32 net_id)
1417 {
1418         struct lnet_net *net;
1419         bool local;
1420
1421         net = lnet_get_net_locked(net_id);
1422
1423         local = net != NULL;
1424
1425         return local;
1426 }
1427
1428 int
1429 lnet_islocalnet(__u32 net_id)
1430 {
1431         int cpt;
1432         bool local;
1433
1434         cpt = lnet_net_lock_current();
1435
1436         local = lnet_islocalnet_locked(net_id);
1437
1438         lnet_net_unlock(cpt);
1439
1440         return local;
1441 }
1442
1443 struct lnet_ni  *
1444 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
1445 {
1446         struct lnet_net  *net;
1447         struct lnet_ni   *ni;
1448
1449         LASSERT(cpt != LNET_LOCK_EX);
1450
1451         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1452                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1453                         if (ni->ni_nid == nid)
1454                                 return ni;
1455                 }
1456         }
1457
1458         return NULL;
1459 }
1460
1461 struct lnet_ni *
1462 lnet_nid2ni_addref(lnet_nid_t nid)
1463 {
1464         struct lnet_ni *ni;
1465
1466         lnet_net_lock(0);
1467         ni = lnet_nid2ni_locked(nid, 0);
1468         if (ni)
1469                 lnet_ni_addref_locked(ni, 0);
1470         lnet_net_unlock(0);
1471
1472         return ni;
1473 }
1474 EXPORT_SYMBOL(lnet_nid2ni_addref);
1475
1476 int
1477 lnet_islocalnid(lnet_nid_t nid)
1478 {
1479         struct lnet_ni  *ni;
1480         int             cpt;
1481
1482         cpt = lnet_net_lock_current();
1483         ni = lnet_nid2ni_locked(nid, cpt);
1484         lnet_net_unlock(cpt);
1485
1486         return ni != NULL;
1487 }
1488
1489 int
1490 lnet_count_acceptor_nets(void)
1491 {
1492         /* Return the # of NIs that need the acceptor. */
1493         int              count = 0;
1494         struct lnet_net  *net;
1495         int              cpt;
1496
1497         cpt = lnet_net_lock_current();
1498         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1499                 /* all socklnd type networks should have the acceptor
1500                  * thread started */
1501                 if (net->net_lnd->lnd_accept != NULL)
1502                         count++;
1503         }
1504
1505         lnet_net_unlock(cpt);
1506
1507         return count;
1508 }
1509
1510 struct lnet_ping_buffer *
1511 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1512 {
1513         struct lnet_ping_buffer *pbuf;
1514
1515         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1516         if (pbuf) {
1517                 pbuf->pb_nnis = nnis;
1518                 pbuf->pb_needs_post = false;
1519                 atomic_set(&pbuf->pb_refcnt, 1);
1520         }
1521
1522         return pbuf;
1523 }
1524
1525 void
1526 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1527 {
1528         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1529         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1530 }
1531
1532 static struct lnet_ping_buffer *
1533 lnet_ping_target_create(int nnis)
1534 {
1535         struct lnet_ping_buffer *pbuf;
1536
1537         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1538         if (pbuf == NULL) {
1539                 CERROR("Can't allocate ping source [%d]\n", nnis);
1540                 return NULL;
1541         }
1542
1543         pbuf->pb_info.pi_nnis = nnis;
1544         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1545         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1546         pbuf->pb_info.pi_features =
1547                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1548
1549         return pbuf;
1550 }
1551
1552 static inline int
1553 lnet_get_net_ni_count_locked(struct lnet_net *net)
1554 {
1555         struct lnet_ni  *ni;
1556         int             count = 0;
1557
1558         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1559                 count++;
1560
1561         return count;
1562 }
1563
1564 static inline int
1565 lnet_get_net_ni_count_pre(struct lnet_net *net)
1566 {
1567         struct lnet_ni  *ni;
1568         int             count = 0;
1569
1570         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1571                 count++;
1572
1573         return count;
1574 }
1575
1576 static inline int
1577 lnet_get_ni_count(void)
1578 {
1579         struct lnet_ni  *ni;
1580         struct lnet_net *net;
1581         int             count = 0;
1582
1583         lnet_net_lock(0);
1584
1585         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1586                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1587                         count++;
1588         }
1589
1590         lnet_net_unlock(0);
1591
1592         return count;
1593 }
1594
1595 int
1596 lnet_get_net_count(void)
1597 {
1598         struct lnet_net *net;
1599         int count = 0;
1600
1601         lnet_net_lock(0);
1602
1603         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1604                 count++;
1605         }
1606
1607         lnet_net_unlock(0);
1608
1609         return count;
1610 }
1611
1612 void
1613 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1614 {
1615         struct lnet_ni_status *stat;
1616         int nnis;
1617         int i;
1618
1619         __swab32s(&pbuf->pb_info.pi_magic);
1620         __swab32s(&pbuf->pb_info.pi_features);
1621         __swab32s(&pbuf->pb_info.pi_pid);
1622         __swab32s(&pbuf->pb_info.pi_nnis);
1623         nnis = pbuf->pb_info.pi_nnis;
1624         if (nnis > pbuf->pb_nnis)
1625                 nnis = pbuf->pb_nnis;
1626         for (i = 0; i < nnis; i++) {
1627                 stat = &pbuf->pb_info.pi_ni[i];
1628                 __swab64s(&stat->ns_nid);
1629                 __swab32s(&stat->ns_status);
1630         }
1631 }
1632
1633 int
1634 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1635 {
1636         if (!pinfo)
1637                 return -EINVAL;
1638         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1639                 return -EPROTO;
1640         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1641                 return -EPROTO;
1642         /* Loopback is guaranteed to be present */
1643         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1644                 return -ERANGE;
1645         if (LNET_PING_INFO_LONI(pinfo) != LNET_NID_LO_0)
1646                 return -EPROTO;
1647         return 0;
1648 }
1649
1650 static void
1651 lnet_ping_target_destroy(void)
1652 {
1653         struct lnet_net *net;
1654         struct lnet_ni  *ni;
1655
1656         lnet_net_lock(LNET_LOCK_EX);
1657
1658         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1659                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1660                         lnet_ni_lock(ni);
1661                         ni->ni_status = NULL;
1662                         lnet_ni_unlock(ni);
1663                 }
1664         }
1665
1666         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1667         the_lnet.ln_ping_target = NULL;
1668
1669         lnet_net_unlock(LNET_LOCK_EX);
1670 }
1671
1672 static void
1673 lnet_ping_target_event_handler(struct lnet_event *event)
1674 {
1675         struct lnet_ping_buffer *pbuf = event->md_user_ptr;
1676
1677         if (event->unlinked)
1678                 lnet_ping_buffer_decref(pbuf);
1679 }
1680
1681 static int
1682 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1683                        struct lnet_handle_md *ping_mdh,
1684                        int ni_count, bool set_eq)
1685 {
1686         struct lnet_process_id id = {
1687                 .nid = LNET_NID_ANY,
1688                 .pid = LNET_PID_ANY
1689         };
1690         struct lnet_me *me;
1691         struct lnet_md md = { NULL };
1692         int rc;
1693
1694         if (set_eq)
1695                 the_lnet.ln_ping_target_handler =
1696                         lnet_ping_target_event_handler;
1697
1698         *ppbuf = lnet_ping_target_create(ni_count);
1699         if (*ppbuf == NULL) {
1700                 rc = -ENOMEM;
1701                 goto fail_free_eq;
1702         }
1703
1704         /* Ping target ME/MD */
1705         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1706                           LNET_PROTO_PING_MATCHBITS, 0,
1707                           LNET_UNLINK, LNET_INS_AFTER);
1708         if (IS_ERR(me)) {
1709                 rc = PTR_ERR(me);
1710                 CERROR("Can't create ping target ME: %d\n", rc);
1711                 goto fail_decref_ping_buffer;
1712         }
1713
1714         /* initialize md content */
1715         md.start     = &(*ppbuf)->pb_info;
1716         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1717         md.threshold = LNET_MD_THRESH_INF;
1718         md.max_size  = 0;
1719         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1720                        LNET_MD_MANAGE_REMOTE;
1721         md.handler   = the_lnet.ln_ping_target_handler;
1722         md.user_ptr  = *ppbuf;
1723
1724         rc = LNetMDAttach(me, &md, LNET_RETAIN, ping_mdh);
1725         if (rc != 0) {
1726                 CERROR("Can't attach ping target MD: %d\n", rc);
1727                 goto fail_unlink_ping_me;
1728         }
1729         lnet_ping_buffer_addref(*ppbuf);
1730
1731         return 0;
1732
1733 fail_unlink_ping_me:
1734         LNetMEUnlink(me);
1735 fail_decref_ping_buffer:
1736         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
1737         lnet_ping_buffer_decref(*ppbuf);
1738         *ppbuf = NULL;
1739 fail_free_eq:
1740         return rc;
1741 }
1742
1743 static void
1744 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1745                     struct lnet_handle_md *ping_mdh)
1746 {
1747         LNetMDUnlink(*ping_mdh);
1748         LNetInvalidateMDHandle(ping_mdh);
1749
1750         /* NB the MD could be busy; this just starts the unlink */
1751         wait_var_event_warning(&pbuf->pb_refcnt,
1752                                atomic_read(&pbuf->pb_refcnt) <= 1,
1753                                "Still waiting for ping data MD to unlink\n");
1754 }
1755
1756 static void
1757 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1758 {
1759         struct lnet_ni          *ni;
1760         struct lnet_net         *net;
1761         struct lnet_ni_status *ns;
1762         int                     i;
1763         int                     rc;
1764
1765         i = 0;
1766         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1767                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1768                         LASSERT(i < pbuf->pb_nnis);
1769
1770                         ns = &pbuf->pb_info.pi_ni[i];
1771
1772                         ns->ns_nid = ni->ni_nid;
1773
1774                         lnet_ni_lock(ni);
1775                         ns->ns_status = (ni->ni_status != NULL) ?
1776                                          ni->ni_status->ns_status :
1777                                                 LNET_NI_STATUS_UP;
1778                         ni->ni_status = ns;
1779                         lnet_ni_unlock(ni);
1780
1781                         i++;
1782                 }
1783         }
1784         /*
1785          * We (ab)use the ns_status of the loopback interface to
1786          * transmit the sequence number. The first interface listed
1787          * must be the loopback interface.
1788          */
1789         rc = lnet_ping_info_validate(&pbuf->pb_info);
1790         if (rc) {
1791                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1792                 LBUG();
1793         }
1794         LNET_PING_BUFFER_SEQNO(pbuf) =
1795                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1796 }
1797
1798 static void
1799 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1800                         struct lnet_handle_md ping_mdh)
1801 {
1802         struct lnet_ping_buffer *old_pbuf = NULL;
1803         struct lnet_handle_md old_ping_md;
1804
1805         /* switch the NIs to point to the new ping info created */
1806         lnet_net_lock(LNET_LOCK_EX);
1807
1808         if (!the_lnet.ln_routing)
1809                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1810         if (!lnet_peer_discovery_disabled)
1811                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1812
1813         /* Ensure only known feature bits have been set. */
1814         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1815         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1816
1817         lnet_ping_target_install_locked(pbuf);
1818
1819         if (the_lnet.ln_ping_target) {
1820                 old_pbuf = the_lnet.ln_ping_target;
1821                 old_ping_md = the_lnet.ln_ping_target_md;
1822         }
1823         the_lnet.ln_ping_target_md = ping_mdh;
1824         the_lnet.ln_ping_target = pbuf;
1825
1826         lnet_net_unlock(LNET_LOCK_EX);
1827
1828         if (old_pbuf) {
1829                 /* unlink and free the old ping info */
1830                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1831                 lnet_ping_buffer_decref(old_pbuf);
1832         }
1833
1834         lnet_push_update_to_peers(0);
1835 }
1836
1837 static void
1838 lnet_ping_target_fini(void)
1839 {
1840         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1841                             &the_lnet.ln_ping_target_md);
1842
1843         lnet_assert_handler_unused(the_lnet.ln_ping_target_handler);
1844         lnet_ping_target_destroy();
1845 }
1846
1847 /* Resize the push target. */
1848 int lnet_push_target_resize(void)
1849 {
1850         struct lnet_handle_md mdh;
1851         struct lnet_handle_md old_mdh;
1852         struct lnet_ping_buffer *pbuf;
1853         struct lnet_ping_buffer *old_pbuf;
1854         int nnis;
1855         int rc;
1856
1857 again:
1858         nnis = the_lnet.ln_push_target_nnis;
1859         if (nnis <= 0) {
1860                 CDEBUG(D_NET, "Invalid nnis %d\n", nnis);
1861                 return -EINVAL;
1862         }
1863
1864         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
1865          * dropped when we need to resize again (see "old_pbuf" below) or when
1866          * LNet is shutdown (see lnet_push_target_fini())
1867          */
1868         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1869         if (!pbuf) {
1870                 CDEBUG(D_NET, "Can't allocate pbuf for nnis %d\n", nnis);
1871                 return -ENOMEM;
1872         }
1873
1874         rc = lnet_push_target_post(pbuf, &mdh);
1875         if (rc) {
1876                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
1877                 lnet_ping_buffer_decref(pbuf);
1878                 return rc;
1879         }
1880
1881         lnet_net_lock(LNET_LOCK_EX);
1882         old_pbuf = the_lnet.ln_push_target;
1883         old_mdh = the_lnet.ln_push_target_md;
1884         the_lnet.ln_push_target = pbuf;
1885         the_lnet.ln_push_target_md = mdh;
1886         lnet_net_unlock(LNET_LOCK_EX);
1887
1888         if (old_pbuf) {
1889                 LNetMDUnlink(old_mdh);
1890                 /* Drop ref set by lnet_ping_buffer_alloc() */
1891                 lnet_ping_buffer_decref(old_pbuf);
1892         }
1893
1894         /* Received another push or reply that requires a larger buffer */
1895         if (nnis < the_lnet.ln_push_target_nnis)
1896                 goto again;
1897
1898         CDEBUG(D_NET, "nnis %d success\n", nnis);
1899         return 0;
1900 }
1901
1902 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
1903                           struct lnet_handle_md *mdhp)
1904 {
1905         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
1906         struct lnet_md md = { NULL };
1907         struct lnet_me *me;
1908         int rc;
1909
1910         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1911                           LNET_PROTO_PING_MATCHBITS, 0,
1912                           LNET_UNLINK, LNET_INS_AFTER);
1913         if (IS_ERR(me)) {
1914                 rc = PTR_ERR(me);
1915                 CERROR("Can't create push target ME: %d\n", rc);
1916                 return rc;
1917         }
1918
1919         pbuf->pb_needs_post = false;
1920
1921         /* This reference is dropped by lnet_push_target_event_handler() */
1922         lnet_ping_buffer_addref(pbuf);
1923
1924         /* initialize md content */
1925         md.start     = &pbuf->pb_info;
1926         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
1927         md.threshold = 1;
1928         md.max_size  = 0;
1929         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
1930         md.user_ptr  = pbuf;
1931         md.handler   = the_lnet.ln_push_target_handler;
1932
1933         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdhp);
1934         if (rc) {
1935                 CERROR("Can't attach push MD: %d\n", rc);
1936                 LNetMEUnlink(me);
1937                 lnet_ping_buffer_decref(pbuf);
1938                 pbuf->pb_needs_post = true;
1939                 return rc;
1940         }
1941
1942         CDEBUG(D_NET, "posted push target %p\n", pbuf);
1943
1944         return 0;
1945 }
1946
1947 static void lnet_push_target_event_handler(struct lnet_event *ev)
1948 {
1949         struct lnet_ping_buffer *pbuf = ev->md_user_ptr;
1950
1951         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
1952                ev->unlinked);
1953
1954         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
1955                 lnet_swap_pinginfo(pbuf);
1956
1957         if (ev->type == LNET_EVENT_UNLINK) {
1958                 /* Drop ref added by lnet_push_target_post() */
1959                 lnet_ping_buffer_decref(pbuf);
1960                 return;
1961         }
1962
1963         lnet_peer_push_event(ev);
1964         if (ev->unlinked)
1965                 /* Drop ref added by lnet_push_target_post */
1966                 lnet_ping_buffer_decref(pbuf);
1967 }
1968
1969 /* Initialize the push target. */
1970 static int lnet_push_target_init(void)
1971 {
1972         int rc;
1973
1974         if (the_lnet.ln_push_target)
1975                 return -EALREADY;
1976
1977         the_lnet.ln_push_target_handler =
1978                 lnet_push_target_event_handler;
1979
1980         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
1981         LASSERT(rc == 0);
1982
1983         /* Start at the required minimum, we'll enlarge if required. */
1984         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
1985
1986         rc = lnet_push_target_resize();
1987
1988         if (rc) {
1989                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
1990                 the_lnet.ln_push_target_handler = NULL;
1991         }
1992
1993         return rc;
1994 }
1995
1996 /* Clean up the push target. */
1997 static void lnet_push_target_fini(void)
1998 {
1999         if (!the_lnet.ln_push_target)
2000                 return;
2001
2002         /* Unlink and invalidate to prevent new references. */
2003         LNetMDUnlink(the_lnet.ln_push_target_md);
2004         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2005
2006         /* Wait for the unlink to complete. */
2007         wait_var_event_warning(&the_lnet.ln_push_target->pb_refcnt,
2008                                atomic_read(&the_lnet.ln_push_target->pb_refcnt) <= 1,
2009                                "Still waiting for ping data MD to unlink\n");
2010
2011         /* Drop ref set by lnet_ping_buffer_alloc() */
2012         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2013         the_lnet.ln_push_target = NULL;
2014         the_lnet.ln_push_target_nnis = 0;
2015
2016         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2017         lnet_assert_handler_unused(the_lnet.ln_push_target_handler);
2018         the_lnet.ln_push_target_handler = NULL;
2019 }
2020
2021 static int
2022 lnet_ni_tq_credits(struct lnet_ni *ni)
2023 {
2024         int     credits;
2025
2026         LASSERT(ni->ni_ncpts >= 1);
2027
2028         if (ni->ni_ncpts == 1)
2029                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2030
2031         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2032         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2033         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2034
2035         return credits;
2036 }
2037
2038 static void
2039 lnet_ni_unlink_locked(struct lnet_ni *ni)
2040 {
2041         /* move it to zombie list and nobody can find it anymore */
2042         LASSERT(!list_empty(&ni->ni_netlist));
2043         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2044         lnet_ni_decref_locked(ni, 0);
2045 }
2046
2047 static void
2048 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2049 {
2050         int             i;
2051         int             islo;
2052         struct lnet_ni  *ni;
2053         struct list_head *zombie_list = &net->net_ni_zombie;
2054
2055         /*
2056          * Now wait for the NIs I just nuked to show up on the zombie
2057          * list and shut them down in guaranteed thread context
2058          */
2059         i = 2;
2060         while (!list_empty(zombie_list)) {
2061                 int     *ref;
2062                 int     j;
2063
2064                 ni = list_entry(zombie_list->next,
2065                                 struct lnet_ni, ni_netlist);
2066                 list_del_init(&ni->ni_netlist);
2067                 /* the ni should be in deleting state. If it's not it's
2068                  * a bug */
2069                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2070                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2071                         if (*ref == 0)
2072                                 continue;
2073                         /* still busy, add it back to zombie list */
2074                         list_add(&ni->ni_netlist, zombie_list);
2075                         break;
2076                 }
2077
2078                 if (!list_empty(&ni->ni_netlist)) {
2079                         lnet_net_unlock(LNET_LOCK_EX);
2080                         ++i;
2081                         if ((i & (-i)) == i) {
2082                                 CDEBUG(D_WARNING,
2083                                        "Waiting for zombie LNI %s\n",
2084                                        libcfs_nid2str(ni->ni_nid));
2085                         }
2086                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2087                         lnet_net_lock(LNET_LOCK_EX);
2088                         continue;
2089                 }
2090
2091                 lnet_net_unlock(LNET_LOCK_EX);
2092
2093                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2094
2095                 LASSERT(!in_interrupt());
2096                 /* Holding the mutex makes it safe for lnd_shutdown
2097                  * to call module_put(). Module unload cannot finish
2098                  * until lnet_unregister_lnd() completes, and that
2099                  * requires the mutex.
2100                  */
2101                 mutex_lock(&the_lnet.ln_lnd_mutex);
2102                 (net->net_lnd->lnd_shutdown)(ni);
2103                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2104
2105                 if (!islo)
2106                         CDEBUG(D_LNI, "Removed LNI %s\n",
2107                               libcfs_nid2str(ni->ni_nid));
2108
2109                 lnet_ni_free(ni);
2110                 i = 2;
2111                 lnet_net_lock(LNET_LOCK_EX);
2112         }
2113 }
2114
2115 /* shutdown down the NI and release refcount */
2116 static void
2117 lnet_shutdown_lndni(struct lnet_ni *ni)
2118 {
2119         int i;
2120         struct lnet_net *net = ni->ni_net;
2121
2122         lnet_net_lock(LNET_LOCK_EX);
2123         lnet_ni_lock(ni);
2124         ni->ni_state = LNET_NI_STATE_DELETING;
2125         lnet_ni_unlock(ni);
2126         lnet_ni_unlink_locked(ni);
2127         lnet_incr_dlc_seq();
2128         lnet_net_unlock(LNET_LOCK_EX);
2129
2130         /* clear messages for this NI on the lazy portal */
2131         for (i = 0; i < the_lnet.ln_nportals; i++)
2132                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2133
2134         lnet_net_lock(LNET_LOCK_EX);
2135         lnet_clear_zombies_nis_locked(net);
2136         lnet_net_unlock(LNET_LOCK_EX);
2137 }
2138
2139 static void
2140 lnet_shutdown_lndnet(struct lnet_net *net)
2141 {
2142         struct lnet_ni *ni;
2143
2144         lnet_net_lock(LNET_LOCK_EX);
2145
2146         list_del_init(&net->net_list);
2147
2148         while (!list_empty(&net->net_ni_list)) {
2149                 ni = list_entry(net->net_ni_list.next,
2150                                 struct lnet_ni, ni_netlist);
2151                 lnet_net_unlock(LNET_LOCK_EX);
2152                 lnet_shutdown_lndni(ni);
2153                 lnet_net_lock(LNET_LOCK_EX);
2154         }
2155
2156         lnet_net_unlock(LNET_LOCK_EX);
2157
2158         /* Do peer table cleanup for this net */
2159         lnet_peer_tables_cleanup(net);
2160
2161         lnet_net_free(net);
2162 }
2163
2164 static void
2165 lnet_shutdown_lndnets(void)
2166 {
2167         struct lnet_net *net;
2168         LIST_HEAD(resend);
2169         struct lnet_msg *msg, *tmp;
2170
2171         /* NB called holding the global mutex */
2172
2173         /* All quiet on the API front */
2174         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2175         LASSERT(the_lnet.ln_refcount == 0);
2176
2177         lnet_net_lock(LNET_LOCK_EX);
2178         the_lnet.ln_state = LNET_STATE_STOPPING;
2179
2180         /*
2181          * move the nets to the zombie list to avoid them being
2182          * picked up for new work. LONET is also included in the
2183          * Nets that will be moved to the zombie list
2184          */
2185         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2186
2187         /* Drop the cached loopback Net. */
2188         if (the_lnet.ln_loni != NULL) {
2189                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2190                 the_lnet.ln_loni = NULL;
2191         }
2192         lnet_net_unlock(LNET_LOCK_EX);
2193
2194         /* iterate through the net zombie list and delete each net */
2195         while (!list_empty(&the_lnet.ln_net_zombie)) {
2196                 net = list_entry(the_lnet.ln_net_zombie.next,
2197                                  struct lnet_net, net_list);
2198                 lnet_shutdown_lndnet(net);
2199         }
2200
2201         spin_lock(&the_lnet.ln_msg_resend_lock);
2202         list_splice(&the_lnet.ln_msg_resend, &resend);
2203         spin_unlock(&the_lnet.ln_msg_resend_lock);
2204
2205         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2206                 list_del_init(&msg->msg_list);
2207                 msg->msg_no_resend = true;
2208                 lnet_finalize(msg, -ECANCELED);
2209         }
2210
2211         lnet_net_lock(LNET_LOCK_EX);
2212         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2213         lnet_net_unlock(LNET_LOCK_EX);
2214 }
2215
2216 static int
2217 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2218 {
2219         int                     rc = -EINVAL;
2220         struct lnet_tx_queue    *tq;
2221         int                     i;
2222         struct lnet_net         *net = ni->ni_net;
2223
2224         mutex_lock(&the_lnet.ln_lnd_mutex);
2225
2226         if (tun) {
2227                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2228                 ni->ni_lnd_tunables_set = true;
2229         }
2230
2231         rc = (net->net_lnd->lnd_startup)(ni);
2232
2233         mutex_unlock(&the_lnet.ln_lnd_mutex);
2234
2235         if (rc != 0) {
2236                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2237                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2238                 goto failed0;
2239         }
2240
2241         lnet_ni_lock(ni);
2242         ni->ni_state = LNET_NI_STATE_ACTIVE;
2243         lnet_ni_unlock(ni);
2244
2245         /* We keep a reference on the loopback net through the loopback NI */
2246         if (net->net_lnd->lnd_type == LOLND) {
2247                 lnet_ni_addref(ni);
2248                 LASSERT(the_lnet.ln_loni == NULL);
2249                 the_lnet.ln_loni = ni;
2250                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2251                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2252                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2253                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2254                 return 0;
2255         }
2256
2257         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2258             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2259                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2260                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2261                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2262                                         "" : "per-peer ");
2263                 /* shutdown the NI since if we get here then it must've already
2264                  * been started
2265                  */
2266                 lnet_shutdown_lndni(ni);
2267                 return -EINVAL;
2268         }
2269
2270         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2271                 tq->tq_credits_min =
2272                 tq->tq_credits_max =
2273                 tq->tq_credits = lnet_ni_tq_credits(ni);
2274         }
2275
2276         atomic_set(&ni->ni_tx_credits,
2277                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2278         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2279
2280         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2281                 libcfs_nid2str(ni->ni_nid),
2282                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2283                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2284                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2285                 ni->ni_net->net_tunables.lct_peer_timeout);
2286
2287         return 0;
2288 failed0:
2289         lnet_ni_free(ni);
2290         return rc;
2291 }
2292
2293 static int
2294 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2295 {
2296         struct lnet_ni *ni;
2297         struct lnet_net *net_l = NULL;
2298         LIST_HEAD(local_ni_list);
2299         int rc;
2300         int ni_count = 0;
2301         __u32 lnd_type;
2302         const struct lnet_lnd  *lnd;
2303         int peer_timeout =
2304                 net->net_tunables.lct_peer_timeout;
2305         int maxtxcredits =
2306                 net->net_tunables.lct_max_tx_credits;
2307         int peerrtrcredits =
2308                 net->net_tunables.lct_peer_rtr_credits;
2309
2310         /*
2311          * make sure that this net is unique. If it isn't then
2312          * we are adding interfaces to an already existing network, and
2313          * 'net' is just a convenient way to pass in the list.
2314          * if it is unique we need to find the LND and load it if
2315          * necessary.
2316          */
2317         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2318                 lnd_type = LNET_NETTYP(net->net_id);
2319
2320                 mutex_lock(&the_lnet.ln_lnd_mutex);
2321                 lnd = lnet_find_lnd_by_type(lnd_type);
2322
2323                 if (lnd == NULL) {
2324                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2325                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2326                         mutex_lock(&the_lnet.ln_lnd_mutex);
2327
2328                         lnd = lnet_find_lnd_by_type(lnd_type);
2329                         if (lnd == NULL) {
2330                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2331                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2332                                 libcfs_lnd2str(lnd_type),
2333                                 libcfs_lnd2modname(lnd_type), rc);
2334 #ifndef HAVE_MODULE_LOADING_SUPPORT
2335                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2336                                                 "compiled with kernel module "
2337                                                 "loading support.");
2338 #endif
2339                                 rc = -EINVAL;
2340                                 goto failed0;
2341                         }
2342                 }
2343
2344                 net->net_lnd = lnd;
2345
2346                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2347
2348                 net_l = net;
2349         }
2350
2351         /*
2352          * net_l: if the network being added is unique then net_l
2353          *        will point to that network
2354          *        if the network being added is not unique then
2355          *        net_l points to the existing network.
2356          *
2357          * When we enter the loop below, we'll pick NIs off he
2358          * network beign added and start them up, then add them to
2359          * a local ni list. Once we've successfully started all
2360          * the NIs then we join the local NI list (of started up
2361          * networks) with the net_l->net_ni_list, which should
2362          * point to the correct network to add the new ni list to
2363          *
2364          * If any of the new NIs fail to start up, then we want to
2365          * iterate through the local ni list, which should include
2366          * any NIs which were successfully started up, and shut
2367          * them down.
2368          *
2369          * After than we want to delete the network being added,
2370          * to avoid a memory leak.
2371          */
2372
2373         /*
2374          * When a network uses TCP bonding then all its interfaces
2375          * must be specified when the network is first defined: the
2376          * TCP bonding code doesn't allow for interfaces to be added
2377          * or removed.
2378          */
2379         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2380             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2381                 rc = -EINVAL;
2382                 goto failed0;
2383         }
2384
2385         while (!list_empty(&net->net_ni_added)) {
2386                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2387                                 ni_netlist);
2388                 list_del_init(&ni->ni_netlist);
2389
2390                 /* make sure that the the NI we're about to start
2391                  * up is actually unique. if it's not fail. */
2392                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2393                                         ni->ni_interfaces[0])) {
2394                         rc = -EEXIST;
2395                         goto failed1;
2396                 }
2397
2398                 /* adjust the pointer the parent network, just in case it
2399                  * the net is a duplicate */
2400                 ni->ni_net = net_l;
2401
2402                 rc = lnet_startup_lndni(ni, tun);
2403
2404                 if (rc < 0)
2405                         goto failed1;
2406
2407                 lnet_ni_addref(ni);
2408                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2409
2410                 ni_count++;
2411         }
2412
2413         lnet_net_lock(LNET_LOCK_EX);
2414         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2415         lnet_incr_dlc_seq();
2416         lnet_net_unlock(LNET_LOCK_EX);
2417
2418         /* if the network is not unique then we don't want to keep
2419          * it around after we're done. Free it. Otherwise add that
2420          * net to the global the_lnet.ln_nets */
2421         if (net_l != net && net_l != NULL) {
2422                 /*
2423                  * TODO - note. currently the tunables can not be updated
2424                  * once added
2425                  */
2426                 lnet_net_free(net);
2427         } else {
2428                 /*
2429                  * restore tunables after it has been overwitten by the
2430                  * lnd
2431                  */
2432                 if (peer_timeout != -1)
2433                         net->net_tunables.lct_peer_timeout = peer_timeout;
2434                 if (maxtxcredits != -1)
2435                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2436                 if (peerrtrcredits != -1)
2437                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2438
2439                 lnet_net_lock(LNET_LOCK_EX);
2440                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2441                 lnet_net_unlock(LNET_LOCK_EX);
2442         }
2443
2444         /* update net count */
2445         lnet_current_net_count = lnet_get_net_count();
2446
2447         return ni_count;
2448
2449 failed1:
2450         /*
2451          * shutdown the new NIs that are being started up
2452          * free the NET being started
2453          */
2454         while (!list_empty(&local_ni_list)) {
2455                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2456                                 ni_netlist);
2457
2458                 lnet_shutdown_lndni(ni);
2459         }
2460
2461 failed0:
2462         lnet_net_free(net);
2463
2464         return rc;
2465 }
2466
2467 static int
2468 lnet_startup_lndnets(struct list_head *netlist)
2469 {
2470         struct lnet_net         *net;
2471         int                     rc;
2472         int                     ni_count = 0;
2473
2474         /*
2475          * Change to running state before bringing up the LNDs. This
2476          * allows lnet_shutdown_lndnets() to assert that we've passed
2477          * through here.
2478          */
2479         lnet_net_lock(LNET_LOCK_EX);
2480         the_lnet.ln_state = LNET_STATE_RUNNING;
2481         lnet_net_unlock(LNET_LOCK_EX);
2482
2483         while (!list_empty(netlist)) {
2484                 net = list_entry(netlist->next, struct lnet_net, net_list);
2485                 list_del_init(&net->net_list);
2486
2487                 rc = lnet_startup_lndnet(net, NULL);
2488
2489                 if (rc < 0)
2490                         goto failed;
2491
2492                 ni_count += rc;
2493         }
2494
2495         return ni_count;
2496 failed:
2497         lnet_shutdown_lndnets();
2498
2499         return rc;
2500 }
2501
2502 /**
2503  * Initialize LNet library.
2504  *
2505  * Automatically called at module loading time. Caller has to call
2506  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2507  * latter returned 0. It must be called exactly once.
2508  *
2509  * \retval 0 on success
2510  * \retval -ve on failures.
2511  */
2512 int lnet_lib_init(void)
2513 {
2514         int rc;
2515
2516         lnet_assert_wire_constants();
2517
2518         /* refer to global cfs_cpt_table for now */
2519         the_lnet.ln_cpt_table = cfs_cpt_tab;
2520         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2521
2522         LASSERT(the_lnet.ln_cpt_number > 0);
2523         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2524                 /* we are under risk of consuming all lh_cookie */
2525                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2526                        "please change setting of CPT-table and retry\n",
2527                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2528                 return -E2BIG;
2529         }
2530
2531         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2532                 the_lnet.ln_cpt_bits++;
2533
2534         rc = lnet_create_locks();
2535         if (rc != 0) {
2536                 CERROR("Can't create LNet global locks: %d\n", rc);
2537                 return rc;
2538         }
2539
2540         the_lnet.ln_refcount = 0;
2541         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2542         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2543
2544         /* The hash table size is the number of bits it takes to express the set
2545          * ln_num_routes, minus 1 (better to under estimate than over so we
2546          * don't waste memory). */
2547         if (rnet_htable_size <= 0)
2548                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2549         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2550                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2551         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2552                                            order_base_2(rnet_htable_size) - 1);
2553
2554         /* All LNDs apart from the LOLND are in separate modules.  They
2555          * register themselves when their module loads, and unregister
2556          * themselves when their module is unloaded. */
2557         lnet_register_lnd(&the_lolnd);
2558         return 0;
2559 }
2560
2561 /**
2562  * Finalize LNet library.
2563  *
2564  * \pre lnet_lib_init() called with success.
2565  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2566  *
2567  * As this happens at module-unload, all lnds must already be unloaded,
2568  * so they must already be unregistered.
2569  */
2570 void lnet_lib_exit(void)
2571 {
2572         int i;
2573
2574         LASSERT(the_lnet.ln_refcount == 0);
2575         lnet_unregister_lnd(&the_lolnd);
2576         for (i = 0; i < NUM_LNDS; i++)
2577                 LASSERT(!the_lnet.ln_lnds[i]);
2578         lnet_destroy_locks();
2579 }
2580
2581 /**
2582  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2583  *
2584  * Users must call this function at least once before any other functions.
2585  * For each successful call there must be a corresponding call to
2586  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2587  * ignored.
2588  *
2589  * The PID used by LNet may be different from the one requested.
2590  * See LNetGetId().
2591  *
2592  * \param requested_pid PID requested by the caller.
2593  *
2594  * \return >= 0 on success, and < 0 error code on failures.
2595  */
2596 int
2597 LNetNIInit(lnet_pid_t requested_pid)
2598 {
2599         int                     im_a_router = 0;
2600         int                     rc;
2601         int                     ni_count;
2602         struct lnet_ping_buffer *pbuf;
2603         struct lnet_handle_md   ping_mdh;
2604         LIST_HEAD(net_head);
2605         struct lnet_net         *net;
2606
2607         mutex_lock(&the_lnet.ln_api_mutex);
2608
2609         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2610
2611         if (the_lnet.ln_refcount > 0) {
2612                 rc = the_lnet.ln_refcount++;
2613                 mutex_unlock(&the_lnet.ln_api_mutex);
2614                 return rc;
2615         }
2616
2617         rc = lnet_prepare(requested_pid);
2618         if (rc != 0) {
2619                 mutex_unlock(&the_lnet.ln_api_mutex);
2620                 return rc;
2621         }
2622
2623         /* create a network for Loopback network */
2624         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2625         if (net == NULL) {
2626                 rc = -ENOMEM;
2627                 goto err_empty_list;
2628         }
2629
2630         /* Add in the loopback NI */
2631         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2632                 rc = -ENOMEM;
2633                 goto err_empty_list;
2634         }
2635
2636         /* If LNet is being initialized via DLC it is possible
2637          * that the user requests not to load module parameters (ones which
2638          * are supported by DLC) on initialization.  Therefore, make sure not
2639          * to load networks, routes and forwarding from module parameters
2640          * in this case.  On cleanup in case of failure only clean up
2641          * routes if it has been loaded */
2642         if (!the_lnet.ln_nis_from_mod_params) {
2643                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2644                                          use_tcp_bonding);
2645                 if (rc < 0)
2646                         goto err_empty_list;
2647         }
2648
2649         ni_count = lnet_startup_lndnets(&net_head);
2650         if (ni_count < 0) {
2651                 rc = ni_count;
2652                 goto err_empty_list;
2653         }
2654
2655         if (!the_lnet.ln_nis_from_mod_params) {
2656                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2657                 if (rc != 0)
2658                         goto err_shutdown_lndnis;
2659
2660                 rc = lnet_rtrpools_alloc(im_a_router);
2661                 if (rc != 0)
2662                         goto err_destroy_routes;
2663         }
2664
2665         rc = lnet_acceptor_start();
2666         if (rc != 0)
2667                 goto err_destroy_routes;
2668
2669         the_lnet.ln_refcount = 1;
2670         /* Now I may use my own API functions... */
2671
2672         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2673         if (rc != 0)
2674                 goto err_acceptor_stop;
2675
2676         lnet_ping_target_update(pbuf, ping_mdh);
2677
2678         the_lnet.ln_mt_handler = lnet_mt_event_handler;
2679
2680         rc = lnet_push_target_init();
2681         if (rc != 0)
2682                 goto err_stop_ping;
2683
2684         rc = lnet_peer_discovery_start();
2685         if (rc != 0)
2686                 goto err_destroy_push_target;
2687
2688         rc = lnet_monitor_thr_start();
2689         if (rc != 0)
2690                 goto err_stop_discovery_thr;
2691
2692         lnet_fault_init();
2693         lnet_router_debugfs_init();
2694
2695         mutex_unlock(&the_lnet.ln_api_mutex);
2696
2697         complete_all(&the_lnet.ln_started);
2698
2699         /* wait for all routers to start */
2700         lnet_wait_router_start();
2701
2702         return 0;
2703
2704 err_stop_discovery_thr:
2705         lnet_peer_discovery_stop();
2706 err_destroy_push_target:
2707         lnet_push_target_fini();
2708 err_stop_ping:
2709         lnet_ping_target_fini();
2710 err_acceptor_stop:
2711         the_lnet.ln_refcount = 0;
2712         lnet_acceptor_stop();
2713 err_destroy_routes:
2714         if (!the_lnet.ln_nis_from_mod_params)
2715                 lnet_destroy_routes();
2716 err_shutdown_lndnis:
2717         lnet_shutdown_lndnets();
2718 err_empty_list:
2719         lnet_unprepare();
2720         LASSERT(rc < 0);
2721         mutex_unlock(&the_lnet.ln_api_mutex);
2722         while (!list_empty(&net_head)) {
2723                 struct lnet_net *net;
2724
2725                 net = list_entry(net_head.next, struct lnet_net, net_list);
2726                 list_del_init(&net->net_list);
2727                 lnet_net_free(net);
2728         }
2729         return rc;
2730 }
2731 EXPORT_SYMBOL(LNetNIInit);
2732
2733 /**
2734  * Stop LNet interfaces, routing, and forwarding.
2735  *
2736  * Users must call this function once for each successful call to LNetNIInit().
2737  * Once the LNetNIFini() operation has been started, the results of pending
2738  * API operations are undefined.
2739  *
2740  * \return always 0 for current implementation.
2741  */
2742 int
2743 LNetNIFini(void)
2744 {
2745         mutex_lock(&the_lnet.ln_api_mutex);
2746
2747         LASSERT(the_lnet.ln_refcount > 0);
2748
2749         if (the_lnet.ln_refcount != 1) {
2750                 the_lnet.ln_refcount--;
2751         } else {
2752                 LASSERT(!the_lnet.ln_niinit_self);
2753
2754                 lnet_fault_fini();
2755
2756                 lnet_router_debugfs_fini();
2757                 lnet_monitor_thr_stop();
2758                 lnet_peer_discovery_stop();
2759                 lnet_push_target_fini();
2760                 lnet_ping_target_fini();
2761
2762                 /* Teardown fns that use my own API functions BEFORE here */
2763                 the_lnet.ln_refcount = 0;
2764
2765                 lnet_acceptor_stop();
2766                 lnet_destroy_routes();
2767                 lnet_shutdown_lndnets();
2768                 lnet_unprepare();
2769         }
2770
2771         mutex_unlock(&the_lnet.ln_api_mutex);
2772         return 0;
2773 }
2774 EXPORT_SYMBOL(LNetNIFini);
2775
2776 /**
2777  * Grabs the ni data from the ni structure and fills the out
2778  * parameters
2779  *
2780  * \param[in] ni network        interface structure
2781  * \param[out] cfg_ni           NI config information
2782  * \param[out] tun              network and LND tunables
2783  */
2784 static void
2785 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2786                    struct lnet_ioctl_config_lnd_tunables *tun,
2787                    struct lnet_ioctl_element_stats *stats,
2788                    __u32 tun_size)
2789 {
2790         size_t min_size = 0;
2791         int i;
2792
2793         if (!ni || !cfg_ni || !tun)
2794                 return;
2795
2796         if (ni->ni_interfaces[0] != NULL) {
2797                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2798                         if (ni->ni_interfaces[i] != NULL) {
2799                                 strncpy(cfg_ni->lic_ni_intf[i],
2800                                         ni->ni_interfaces[i],
2801                                         sizeof(cfg_ni->lic_ni_intf[i]));
2802                         }
2803                 }
2804         }
2805
2806         cfg_ni->lic_nid = ni->ni_nid;
2807         if (ni->ni_nid == LNET_NID_LO_0)
2808                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2809         else
2810                 cfg_ni->lic_status = ni->ni_status->ns_status;
2811         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2812         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2813
2814         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2815
2816         if (stats) {
2817                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
2818                                                        LNET_STATS_TYPE_SEND);
2819                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
2820                                                        LNET_STATS_TYPE_RECV);
2821                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
2822                                                        LNET_STATS_TYPE_DROP);
2823         }
2824
2825         /*
2826          * tun->lt_tun will always be present, but in order to be
2827          * backwards compatible, we need to deal with the cases when
2828          * tun->lt_tun is smaller than what the kernel has, because it
2829          * comes from an older version of a userspace program, then we'll
2830          * need to copy as much information as we have available space.
2831          */
2832         min_size = tun_size - sizeof(tun->lt_cmn);
2833         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2834
2835         /* copy over the cpts */
2836         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2837             ni->ni_cpts == NULL)  {
2838                 for (i = 0; i < ni->ni_ncpts; i++)
2839                         cfg_ni->lic_cpts[i] = i;
2840         } else {
2841                 for (i = 0;
2842                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2843                      i < LNET_MAX_SHOW_NUM_CPT;
2844                      i++)
2845                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2846         }
2847         cfg_ni->lic_ncpts = ni->ni_ncpts;
2848 }
2849
2850 /**
2851  * NOTE: This is a legacy function left in the code to be backwards
2852  * compatible with older userspace programs. It should eventually be
2853  * removed.
2854  *
2855  * Grabs the ni data from the ni structure and fills the out
2856  * parameters
2857  *
2858  * \param[in] ni network        interface structure
2859  * \param[out] config           config information
2860  */
2861 static void
2862 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2863                          struct lnet_ioctl_config_data *config)
2864 {
2865         struct lnet_ioctl_net_config *net_config;
2866         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2867         size_t min_size, tunable_size = 0;
2868         int i;
2869
2870         if (!ni || !config)
2871                 return;
2872
2873         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2874         if (!net_config)
2875                 return;
2876
2877         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2878                      ARRAY_SIZE(net_config->ni_interfaces));
2879
2880         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2881                 if (!ni->ni_interfaces[i])
2882                         break;
2883
2884                 strncpy(net_config->ni_interfaces[i],
2885                         ni->ni_interfaces[i],
2886                         sizeof(net_config->ni_interfaces[i]));
2887         }
2888
2889         config->cfg_nid = ni->ni_nid;
2890         config->cfg_config_u.cfg_net.net_peer_timeout =
2891                 ni->ni_net->net_tunables.lct_peer_timeout;
2892         config->cfg_config_u.cfg_net.net_max_tx_credits =
2893                 ni->ni_net->net_tunables.lct_max_tx_credits;
2894         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2895                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2896         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2897                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2898
2899         if (ni->ni_nid == LNET_NID_LO_0)
2900                 net_config->ni_status = LNET_NI_STATUS_UP;
2901         else
2902                 net_config->ni_status = ni->ni_status->ns_status;
2903
2904         if (ni->ni_cpts) {
2905                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2906
2907                 for (i = 0; i < num_cpts; i++)
2908                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2909
2910                 config->cfg_ncpts = num_cpts;
2911         }
2912
2913         /*
2914          * See if user land tools sent in a newer and larger version
2915          * of struct lnet_tunables than what the kernel uses.
2916          */
2917         min_size = sizeof(*config) + sizeof(*net_config);
2918
2919         if (config->cfg_hdr.ioc_len > min_size)
2920                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2921
2922         /* Don't copy too much data to user space */
2923         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2924         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2925
2926         if (lnd_cfg && min_size) {
2927                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2928                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2929
2930                 /* Tell user land that kernel side has less data */
2931                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2932                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2933                         config->cfg_hdr.ioc_len -= min_size;
2934                 }
2935         }
2936 }
2937
2938 struct lnet_ni *
2939 lnet_get_ni_idx_locked(int idx)
2940 {
2941         struct lnet_ni          *ni;
2942         struct lnet_net         *net;
2943
2944         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2945                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2946                         if (idx-- == 0)
2947                                 return ni;
2948                 }
2949         }
2950
2951         return NULL;
2952 }
2953
2954 struct lnet_ni *
2955 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2956 {
2957         struct lnet_ni          *ni;
2958         struct lnet_net         *net = mynet;
2959
2960         /*
2961          * It is possible that the net has been cleaned out while there is
2962          * a message being sent. This function accessed the net without
2963          * checking if the list is empty
2964          */
2965         if (prev == NULL) {
2966                 if (net == NULL)
2967                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2968                                         net_list);
2969                 if (list_empty(&net->net_ni_list))
2970                         return NULL;
2971                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2972                                 ni_netlist);
2973
2974                 return ni;
2975         }
2976
2977         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2978                 /* if you reached the end of the ni list and the net is
2979                  * specified, then there are no more nis in that net */
2980                 if (net != NULL)
2981                         return NULL;
2982
2983                 /* we reached the end of this net ni list. move to the
2984                  * next net */
2985                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2986                         /* no more nets and no more NIs. */
2987                         return NULL;
2988
2989                 /* get the next net */
2990                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2991                                  net_list);
2992                 if (list_empty(&net->net_ni_list))
2993                         return NULL;
2994                 /* get the ni on it */
2995                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2996                                 ni_netlist);
2997
2998                 return ni;
2999         }
3000
3001         if (list_empty(&prev->ni_netlist))
3002                 return NULL;
3003
3004         /* there are more nis left */
3005         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
3006
3007         return ni;
3008 }
3009
3010 int
3011 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3012 {
3013         struct lnet_ni *ni;
3014         int cpt;
3015         int rc = -ENOENT;
3016         int idx = config->cfg_count;
3017
3018         cpt = lnet_net_lock_current();
3019
3020         ni = lnet_get_ni_idx_locked(idx);
3021
3022         if (ni != NULL) {
3023                 rc = 0;
3024                 lnet_ni_lock(ni);
3025                 lnet_fill_ni_info_legacy(ni, config);
3026                 lnet_ni_unlock(ni);
3027         }
3028
3029         lnet_net_unlock(cpt);
3030         return rc;
3031 }
3032
3033 int
3034 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3035                    struct lnet_ioctl_config_lnd_tunables *tun,
3036                    struct lnet_ioctl_element_stats *stats,
3037                    __u32 tun_size)
3038 {
3039         struct lnet_ni          *ni;
3040         int                     cpt;
3041         int                     rc = -ENOENT;
3042
3043         if (!cfg_ni || !tun || !stats)
3044                 return -EINVAL;
3045
3046         cpt = lnet_net_lock_current();
3047
3048         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3049
3050         if (ni) {
3051                 rc = 0;
3052                 lnet_ni_lock(ni);
3053                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3054                 lnet_ni_unlock(ni);
3055         }
3056
3057         lnet_net_unlock(cpt);
3058         return rc;
3059 }
3060
3061 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3062 {
3063         struct lnet_ni *ni;
3064         int cpt;
3065         int rc = -ENOENT;
3066
3067         if (!msg_stats)
3068                 return -EINVAL;
3069
3070         cpt = lnet_net_lock_current();
3071
3072         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3073
3074         if (ni) {
3075                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3076                 rc = 0;
3077         }
3078
3079         lnet_net_unlock(cpt);
3080
3081         return rc;
3082 }
3083
3084 static int lnet_add_net_common(struct lnet_net *net,
3085                                struct lnet_ioctl_config_lnd_tunables *tun)
3086 {
3087         __u32                   net_id;
3088         struct lnet_ping_buffer *pbuf;
3089         struct lnet_handle_md   ping_mdh;
3090         int                     rc;
3091         struct lnet_remotenet *rnet;
3092         int                     net_ni_count;
3093
3094         lnet_net_lock(LNET_LOCK_EX);
3095         rnet = lnet_find_rnet_locked(net->net_id);
3096         lnet_net_unlock(LNET_LOCK_EX);
3097         /*
3098          * make sure that the net added doesn't invalidate the current
3099          * configuration LNet is keeping
3100          */
3101         if (rnet) {
3102                 CERROR("Adding net %s will invalidate routing configuration\n",
3103                        libcfs_net2str(net->net_id));
3104                 lnet_net_free(net);
3105                 return -EUSERS;
3106         }
3107
3108         /*
3109          * make sure you calculate the correct number of slots in the ping
3110          * buffer. Since the ping info is a flattened list of all the NIs,
3111          * we should allocate enough slots to accomodate the number of NIs
3112          * which will be added.
3113          *
3114          * since ni hasn't been configured yet, use
3115          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3116          */
3117         net_ni_count = lnet_get_net_ni_count_pre(net);
3118
3119         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3120                                     net_ni_count + lnet_get_ni_count(),
3121                                     false);
3122         if (rc < 0) {
3123                 lnet_net_free(net);
3124                 return rc;
3125         }
3126
3127         if (tun)
3128                 memcpy(&net->net_tunables,
3129                        &tun->lt_cmn, sizeof(net->net_tunables));
3130         else
3131                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3132
3133         net_id = net->net_id;
3134
3135         rc = lnet_startup_lndnet(net,
3136                                  (tun) ? &tun->lt_tun : NULL);
3137         if (rc < 0)
3138                 goto failed;
3139
3140         lnet_net_lock(LNET_LOCK_EX);
3141         net = lnet_get_net_locked(net_id);
3142         lnet_net_unlock(LNET_LOCK_EX);
3143
3144         LASSERT(net);
3145
3146         /*
3147          * Start the acceptor thread if this is the first network
3148          * being added that requires the thread.
3149          */
3150         if (net->net_lnd->lnd_accept) {
3151                 rc = lnet_acceptor_start();
3152                 if (rc < 0) {
3153                         /* shutdown the net that we just started */
3154                         CERROR("Failed to start up acceptor thread\n");
3155                         lnet_shutdown_lndnet(net);
3156                         goto failed;
3157                 }
3158         }
3159
3160         lnet_net_lock(LNET_LOCK_EX);
3161         lnet_peer_net_added(net);
3162         lnet_net_unlock(LNET_LOCK_EX);
3163
3164         lnet_ping_target_update(pbuf, ping_mdh);
3165
3166         return 0;
3167
3168 failed:
3169         lnet_ping_md_unlink(pbuf, &ping_mdh);
3170         lnet_ping_buffer_decref(pbuf);
3171         return rc;
3172 }
3173
3174 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3175                                       struct lnet_ioctl_config_lnd_tunables *tun)
3176 {
3177         struct lnet_net *net;
3178         const char *nets;
3179         int rc;
3180         LIST_HEAD(net_head);
3181
3182         rc = lnet_parse_ip2nets(&nets, ip2nets);
3183         if (rc < 0)
3184                 return rc;
3185
3186         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3187         if (rc < 0)
3188                 return rc;
3189
3190         mutex_lock(&the_lnet.ln_api_mutex);
3191         while (!list_empty(&net_head)) {
3192                 net = list_entry(net_head.next, struct lnet_net, net_list);
3193                 list_del_init(&net->net_list);
3194                 rc = lnet_add_net_common(net, tun);
3195                 if (rc < 0)
3196                         goto out;
3197         }
3198
3199 out:
3200         mutex_unlock(&the_lnet.ln_api_mutex);
3201
3202         while (!list_empty(&net_head)) {
3203                 net = list_entry(net_head.next, struct lnet_net, net_list);
3204                 list_del_init(&net->net_list);
3205                 lnet_net_free(net);
3206         }
3207         return rc;
3208 }
3209
3210 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3211 {
3212         struct lnet_net *net;
3213         struct lnet_ni *ni;
3214         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3215         int rc, i;
3216         __u32 net_id, lnd_type;
3217
3218         /* get the tunables if they are available */
3219         if (conf->lic_cfg_hdr.ioc_len >=
3220             sizeof(*conf) + sizeof(*tun))
3221                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3222                         conf->lic_bulk;
3223
3224         /* handle legacy ip2nets from DLC */
3225         if (conf->lic_legacy_ip2nets[0] != '\0')
3226                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3227                                                   tun);
3228
3229         net_id = LNET_NIDNET(conf->lic_nid);
3230         lnd_type = LNET_NETTYP(net_id);
3231
3232         if (!libcfs_isknown_lnd(lnd_type)) {
3233                 CERROR("No valid net and lnd information provided\n");
3234                 return -EINVAL;
3235         }
3236
3237         net = lnet_net_alloc(net_id, NULL);
3238         if (!net)
3239                 return -ENOMEM;
3240
3241         for (i = 0; i < conf->lic_ncpts; i++) {
3242                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3243                         return -EINVAL;
3244         }
3245
3246         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3247                                        conf->lic_ni_intf[0]);
3248         if (!ni)
3249                 return -ENOMEM;
3250
3251         mutex_lock(&the_lnet.ln_api_mutex);
3252
3253         rc = lnet_add_net_common(net, tun);
3254
3255         mutex_unlock(&the_lnet.ln_api_mutex);
3256
3257         return rc;
3258 }
3259
3260 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3261 {
3262         struct lnet_net  *net;
3263         struct lnet_ni *ni;
3264         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3265         struct lnet_ping_buffer *pbuf;
3266         struct lnet_handle_md  ping_mdh;
3267         int               rc;
3268         int               net_count;
3269         __u32             addr;
3270
3271         /* don't allow userspace to shutdown the LOLND */
3272         if (LNET_NETTYP(net_id) == LOLND)
3273                 return -EINVAL;
3274
3275         mutex_lock(&the_lnet.ln_api_mutex);
3276
3277         lnet_net_lock(0);
3278
3279         net = lnet_get_net_locked(net_id);
3280         if (!net) {
3281                 CERROR("net %s not found\n",
3282                        libcfs_net2str(net_id));
3283                 rc = -ENOENT;
3284                 goto unlock_net;
3285         }
3286
3287         addr = LNET_NIDADDR(conf->lic_nid);
3288         if (addr == 0) {
3289                 /* remove the entire net */
3290                 net_count = lnet_get_net_ni_count_locked(net);
3291
3292                 lnet_net_unlock(0);
3293
3294                 /* create and link a new ping info, before removing the old one */
3295                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3296                                         lnet_get_ni_count() - net_count,
3297                                         false);
3298                 if (rc != 0)
3299                         goto unlock_api_mutex;
3300
3301                 lnet_shutdown_lndnet(net);
3302
3303                 lnet_acceptor_stop();
3304
3305                 lnet_ping_target_update(pbuf, ping_mdh);
3306
3307                 goto unlock_api_mutex;
3308         }
3309
3310         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3311         if (!ni) {
3312                 CERROR("nid %s not found\n",
3313                        libcfs_nid2str(conf->lic_nid));
3314                 rc = -ENOENT;
3315                 goto unlock_net;
3316         }
3317
3318         net_count = lnet_get_net_ni_count_locked(net);
3319
3320         lnet_net_unlock(0);
3321
3322         /* create and link a new ping info, before removing the old one */
3323         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3324                                   lnet_get_ni_count() - 1, false);
3325         if (rc != 0)
3326                 goto unlock_api_mutex;
3327
3328         lnet_shutdown_lndni(ni);
3329
3330         lnet_acceptor_stop();
3331
3332         lnet_ping_target_update(pbuf, ping_mdh);
3333
3334         /* check if the net is empty and remove it if it is */
3335         if (net_count == 1)
3336                 lnet_shutdown_lndnet(net);
3337
3338         goto unlock_api_mutex;
3339
3340 unlock_net:
3341         lnet_net_unlock(0);
3342 unlock_api_mutex:
3343         mutex_unlock(&the_lnet.ln_api_mutex);
3344
3345         return rc;
3346 }
3347
3348 /*
3349  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3350  * They are only expected to be called for unique networks.
3351  * That can be as a result of older DLC library
3352  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3353  */
3354 int
3355 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3356 {
3357         struct lnet_net *net;
3358         LIST_HEAD(net_head);
3359         int rc;
3360         struct lnet_ioctl_config_lnd_tunables tun;
3361         const char *nets = conf->cfg_config_u.cfg_net.net_intf;
3362
3363         /* Create a net/ni structures for the network string */
3364         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3365         if (rc <= 0)
3366                 return rc == 0 ? -EINVAL : rc;
3367
3368         mutex_lock(&the_lnet.ln_api_mutex);
3369
3370         if (rc > 1) {
3371                 rc = -EINVAL; /* only add one network per call */
3372                 goto out_unlock_clean;
3373         }
3374
3375         net = list_entry(net_head.next, struct lnet_net, net_list);
3376         list_del_init(&net->net_list);
3377
3378         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3379
3380         memset(&tun, 0, sizeof(tun));
3381
3382         tun.lt_cmn.lct_peer_timeout =
3383           conf->cfg_config_u.cfg_net.net_peer_timeout;
3384         tun.lt_cmn.lct_peer_tx_credits =
3385           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3386         tun.lt_cmn.lct_peer_rtr_credits =
3387           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3388         tun.lt_cmn.lct_max_tx_credits =
3389           conf->cfg_config_u.cfg_net.net_max_tx_credits;
3390
3391         rc = lnet_add_net_common(net, &tun);
3392
3393 out_unlock_clean:
3394         mutex_unlock(&the_lnet.ln_api_mutex);
3395         while (!list_empty(&net_head)) {
3396                 /* net_head list is empty in success case */
3397                 net = list_entry(net_head.next, struct lnet_net, net_list);
3398                 list_del_init(&net->net_list);
3399                 lnet_net_free(net);
3400         }
3401         return rc;
3402 }
3403
3404 int
3405 lnet_dyn_del_net(__u32 net_id)
3406 {
3407         struct lnet_net  *net;
3408         struct lnet_ping_buffer *pbuf;
3409         struct lnet_handle_md ping_mdh;
3410         int               rc;
3411         int               net_ni_count;
3412
3413         /* don't allow userspace to shutdown the LOLND */
3414         if (LNET_NETTYP(net_id) == LOLND)
3415                 return -EINVAL;
3416
3417         mutex_lock(&the_lnet.ln_api_mutex);
3418
3419         lnet_net_lock(0);
3420
3421         net = lnet_get_net_locked(net_id);
3422         if (net == NULL) {
3423                 lnet_net_unlock(0);
3424                 rc = -EINVAL;
3425                 goto out;
3426         }
3427
3428         net_ni_count = lnet_get_net_ni_count_locked(net);
3429
3430         lnet_net_unlock(0);
3431
3432         /* create and link a new ping info, before removing the old one */
3433         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3434                                     lnet_get_ni_count() - net_ni_count, false);
3435         if (rc != 0)
3436                 goto out;
3437
3438         lnet_shutdown_lndnet(net);
3439
3440         lnet_acceptor_stop();
3441
3442         lnet_ping_target_update(pbuf, ping_mdh);
3443
3444 out:
3445         mutex_unlock(&the_lnet.ln_api_mutex);
3446
3447         return rc;
3448 }
3449
3450 void lnet_incr_dlc_seq(void)
3451 {
3452         atomic_inc(&lnet_dlc_seq_no);
3453 }
3454
3455 __u32 lnet_get_dlc_seq_locked(void)
3456 {
3457         return atomic_read(&lnet_dlc_seq_no);
3458 }
3459
3460 static void
3461 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3462 {
3463         struct lnet_net *net;
3464         struct lnet_ni *ni;
3465
3466         lnet_net_lock(LNET_LOCK_EX);
3467         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3468                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3469                         if (ni->ni_nid == nid || all) {
3470                                 atomic_set(&ni->ni_healthv, value);
3471                                 if (list_empty(&ni->ni_recovery) &&
3472                                     value < LNET_MAX_HEALTH_VALUE) {
3473                                         CERROR("manually adding local NI %s to recovery\n",
3474                                                libcfs_nid2str(ni->ni_nid));
3475                                         list_add_tail(&ni->ni_recovery,
3476                                                       &the_lnet.ln_mt_localNIRecovq);
3477                                         lnet_ni_addref_locked(ni, 0);
3478                                 }
3479                                 if (!all) {
3480                                         lnet_net_unlock(LNET_LOCK_EX);
3481                                         return;
3482                                 }
3483                         }
3484                 }
3485         }
3486         lnet_net_unlock(LNET_LOCK_EX);
3487 }
3488
3489 static int
3490 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3491 {
3492         int cpt, rc = 0;
3493         struct lnet_ni *ni;
3494         lnet_nid_t nid = stats->hlni_nid;
3495
3496         cpt = lnet_net_lock_current();
3497         ni = lnet_nid2ni_locked(nid, cpt);
3498
3499         if (!ni) {
3500                 rc = -ENOENT;
3501                 goto unlock;
3502         }
3503
3504         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3505         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3506         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3507         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3508         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3509         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3510         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3511
3512 unlock:
3513         lnet_net_unlock(cpt);
3514
3515         return rc;
3516 }
3517
3518 static int
3519 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)