Whamcloud - gitweb
LU-13265 lnet: Modify push MD to have single threshold
[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 unsigned lnet_lnd_timeout = LNET_LND_DEFAULT_TIMEOUT;
227 unsigned int lnet_current_net_count;
228
229 /*
230  * This sequence number keeps track of how many times DLC was used to
231  * update the local NIs. It is incremented when a NI is added or
232  * removed and checked when sending a message to determine if there is
233  * a need to re-run the selection algorithm. See lnet_select_pathway()
234  * for more details on its usage.
235  */
236 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
237
238 static int lnet_ping(struct lnet_process_id id, signed long timeout,
239                      struct lnet_process_id __user *ids, int n_ids);
240
241 static int lnet_discover(struct lnet_process_id id, __u32 force,
242                          struct lnet_process_id __user *ids, int n_ids);
243
244 static int
245 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
246 {
247         int rc;
248         unsigned *sensitivity = (unsigned *)kp->arg;
249         unsigned long value;
250
251         rc = kstrtoul(val, 0, &value);
252         if (rc) {
253                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
254                 return rc;
255         }
256
257         /*
258          * The purpose of locking the api_mutex here is to ensure that
259          * the correct value ends up stored properly.
260          */
261         mutex_lock(&the_lnet.ln_api_mutex);
262
263         if (value > LNET_MAX_HEALTH_VALUE) {
264                 mutex_unlock(&the_lnet.ln_api_mutex);
265                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
266                        LNET_MAX_HEALTH_VALUE, value);
267                 return -EINVAL;
268         }
269
270         /*
271          * if we're turning on health then use the health timeout
272          * defaults.
273          */
274         if (*sensitivity == 0 && value != 0) {
275                 lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT;
276                 lnet_retry_count = LNET_RETRY_COUNT_HEALTH_DEFAULT;
277         /*
278          * if we're turning off health then use the no health timeout
279          * default.
280          */
281         } else if (*sensitivity != 0 && value == 0) {
282                 lnet_transaction_timeout =
283                         LNET_TRANSACTION_TIMEOUT_NO_HEALTH_DEFAULT;
284                 lnet_retry_count = 0;
285         }
286
287         *sensitivity = value;
288
289         mutex_unlock(&the_lnet.ln_api_mutex);
290
291         return 0;
292 }
293
294 static int
295 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
296 {
297         int rc;
298         unsigned *interval = (unsigned *)kp->arg;
299         unsigned long value;
300
301         rc = kstrtoul(val, 0, &value);
302         if (rc) {
303                 CERROR("Invalid module parameter value for 'lnet_recovery_interval'\n");
304                 return rc;
305         }
306
307         if (value < 1) {
308                 CERROR("lnet_recovery_interval must be at least 1 second\n");
309                 return -EINVAL;
310         }
311
312         /*
313          * The purpose of locking the api_mutex here is to ensure that
314          * the correct value ends up stored properly.
315          */
316         mutex_lock(&the_lnet.ln_api_mutex);
317
318         *interval = value;
319
320         mutex_unlock(&the_lnet.ln_api_mutex);
321
322         return 0;
323 }
324
325 static int
326 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
327 {
328         int rc;
329         unsigned *discovery = (unsigned *)kp->arg;
330         unsigned long value;
331         struct lnet_ping_buffer *pbuf;
332
333         rc = kstrtoul(val, 0, &value);
334         if (rc) {
335                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
336                 return rc;
337         }
338
339         value = (value) ? 1 : 0;
340
341         /*
342          * The purpose of locking the api_mutex here is to ensure that
343          * the correct value ends up stored properly.
344          */
345         mutex_lock(&the_lnet.ln_api_mutex);
346
347         if (value == *discovery) {
348                 mutex_unlock(&the_lnet.ln_api_mutex);
349                 return 0;
350         }
351
352         *discovery = value;
353
354         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
355                 mutex_unlock(&the_lnet.ln_api_mutex);
356                 return 0;
357         }
358
359         /* tell peers that discovery setting has changed */
360         lnet_net_lock(LNET_LOCK_EX);
361         pbuf = the_lnet.ln_ping_target;
362         if (value)
363                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
364         else
365                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
366         lnet_net_unlock(LNET_LOCK_EX);
367
368         lnet_push_update_to_peers(1);
369
370         mutex_unlock(&the_lnet.ln_api_mutex);
371
372         return 0;
373 }
374
375 static int
376 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
377 {
378         int rc;
379         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
380         unsigned long value;
381
382         rc = kstrtoul(val, 0, &value);
383         if (rc) {
384                 CERROR("Invalid module parameter value for "
385                        "'lnet_drop_asym_route'\n");
386                 return rc;
387         }
388
389         /*
390          * The purpose of locking the api_mutex here is to ensure that
391          * the correct value ends up stored properly.
392          */
393         mutex_lock(&the_lnet.ln_api_mutex);
394
395         if (value == *drop_asym_route) {
396                 mutex_unlock(&the_lnet.ln_api_mutex);
397                 return 0;
398         }
399
400         *drop_asym_route = value;
401
402         mutex_unlock(&the_lnet.ln_api_mutex);
403
404         return 0;
405 }
406
407 static int
408 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
409 {
410         int rc;
411         unsigned *transaction_to = (unsigned *)kp->arg;
412         unsigned long value;
413
414         rc = kstrtoul(val, 0, &value);
415         if (rc) {
416                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
417                 return rc;
418         }
419
420         /*
421          * The purpose of locking the api_mutex here is to ensure that
422          * the correct value ends up stored properly.
423          */
424         mutex_lock(&the_lnet.ln_api_mutex);
425
426         if (value < lnet_retry_count || value == 0) {
427                 mutex_unlock(&the_lnet.ln_api_mutex);
428                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
429                        "Has to be greater than lnet_retry_count (%u)\n",
430                        value, lnet_retry_count);
431                 return -EINVAL;
432         }
433
434         if (value == *transaction_to) {
435                 mutex_unlock(&the_lnet.ln_api_mutex);
436                 return 0;
437         }
438
439         *transaction_to = value;
440         if (lnet_retry_count == 0)
441                 lnet_lnd_timeout = value;
442         else
443                 lnet_lnd_timeout = value / lnet_retry_count;
444
445         mutex_unlock(&the_lnet.ln_api_mutex);
446
447         return 0;
448 }
449
450 static int
451 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
452 {
453         int rc;
454         unsigned *retry_count = (unsigned *)kp->arg;
455         unsigned long value;
456
457         rc = kstrtoul(val, 0, &value);
458         if (rc) {
459                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
460                 return rc;
461         }
462
463         /*
464          * The purpose of locking the api_mutex here is to ensure that
465          * the correct value ends up stored properly.
466          */
467         mutex_lock(&the_lnet.ln_api_mutex);
468
469         if (lnet_health_sensitivity == 0) {
470                 mutex_unlock(&the_lnet.ln_api_mutex);
471                 CERROR("Can not set retry_count when health feature is turned off\n");
472                 return -EINVAL;
473         }
474
475         if (value > lnet_transaction_timeout) {
476                 mutex_unlock(&the_lnet.ln_api_mutex);
477                 CERROR("Invalid value for lnet_retry_count (%lu). "
478                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
479                        value, lnet_transaction_timeout);
480                 return -EINVAL;
481         }
482
483         *retry_count = value;
484
485         if (value == 0)
486                 lnet_lnd_timeout = lnet_transaction_timeout;
487         else
488                 lnet_lnd_timeout = lnet_transaction_timeout / value;
489
490         mutex_unlock(&the_lnet.ln_api_mutex);
491
492         return 0;
493 }
494
495 static int
496 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
497 {
498         int value, rc;
499
500         rc = kstrtoint(val, 0, &value);
501         if (rc) {
502                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
503                 return rc;
504         }
505
506         if (value < LNET_INTERFACES_MIN) {
507                 CWARN("max interfaces provided are too small, setting to %d\n",
508                       LNET_INTERFACES_MAX_DEFAULT);
509                 value = LNET_INTERFACES_MAX_DEFAULT;
510         }
511
512         *(int *)kp->arg = value;
513
514         return 0;
515 }
516
517 static char *
518 lnet_get_routes(void)
519 {
520         return routes;
521 }
522
523 static char *
524 lnet_get_networks(void)
525 {
526         char   *nets;
527         int     rc;
528
529         if (*networks != 0 && *ip2nets != 0) {
530                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
531                                    "'ip2nets' but not both at once\n");
532                 return NULL;
533         }
534
535         if (*ip2nets != 0) {
536                 rc = lnet_parse_ip2nets(&nets, ip2nets);
537                 return (rc == 0) ? nets : NULL;
538         }
539
540         if (*networks != 0)
541                 return networks;
542
543         return "tcp";
544 }
545
546 static void
547 lnet_init_locks(void)
548 {
549         spin_lock_init(&the_lnet.ln_eq_wait_lock);
550         spin_lock_init(&the_lnet.ln_msg_resend_lock);
551         init_completion(&the_lnet.ln_mt_wait_complete);
552         mutex_init(&the_lnet.ln_lnd_mutex);
553 }
554
555 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
556 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
557                                             *  MDs kmem_cache */
558 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
559 struct kmem_cache *lnet_msg_cachep;
560
561 static int
562 lnet_slab_setup(void)
563 {
564         /* create specific kmem_cache for MEs and small MDs (i.e., originally
565          * allocated in <size-xxx> kmem_cache).
566          */
567         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
568                                             0, 0, NULL);
569         if (!lnet_mes_cachep)
570                 return -ENOMEM;
571
572         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
573                                                   LNET_SMALL_MD_SIZE, 0, 0,
574                                                   NULL);
575         if (!lnet_small_mds_cachep)
576                 return -ENOMEM;
577
578         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
579                                             0, 0, NULL);
580         if (!lnet_rspt_cachep)
581                 return -ENOMEM;
582
583         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
584                                             0, 0, NULL);
585         if (!lnet_msg_cachep)
586                 return -ENOMEM;
587
588         return 0;
589 }
590
591 static void
592 lnet_slab_cleanup(void)
593 {
594         if (lnet_msg_cachep) {
595                 kmem_cache_destroy(lnet_msg_cachep);
596                 lnet_msg_cachep = NULL;
597         }
598
599
600         if (lnet_rspt_cachep) {
601                 kmem_cache_destroy(lnet_rspt_cachep);
602                 lnet_rspt_cachep = NULL;
603         }
604
605         if (lnet_small_mds_cachep) {
606                 kmem_cache_destroy(lnet_small_mds_cachep);
607                 lnet_small_mds_cachep = NULL;
608         }
609
610         if (lnet_mes_cachep) {
611                 kmem_cache_destroy(lnet_mes_cachep);
612                 lnet_mes_cachep = NULL;
613         }
614 }
615
616 static int
617 lnet_create_remote_nets_table(void)
618 {
619         int               i;
620         struct list_head *hash;
621
622         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
623         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
624         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
625         if (hash == NULL) {
626                 CERROR("Failed to create remote nets hash table\n");
627                 return -ENOMEM;
628         }
629
630         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
631                 INIT_LIST_HEAD(&hash[i]);
632         the_lnet.ln_remote_nets_hash = hash;
633         return 0;
634 }
635
636 static void
637 lnet_destroy_remote_nets_table(void)
638 {
639         int i;
640
641         if (the_lnet.ln_remote_nets_hash == NULL)
642                 return;
643
644         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
645                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
646
647         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
648                            LNET_REMOTE_NETS_HASH_SIZE);
649         the_lnet.ln_remote_nets_hash = NULL;
650 }
651
652 static void
653 lnet_destroy_locks(void)
654 {
655         if (the_lnet.ln_res_lock != NULL) {
656                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
657                 the_lnet.ln_res_lock = NULL;
658         }
659
660         if (the_lnet.ln_net_lock != NULL) {
661                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
662                 the_lnet.ln_net_lock = NULL;
663         }
664 }
665
666 static int
667 lnet_create_locks(void)
668 {
669         lnet_init_locks();
670
671         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
672         if (the_lnet.ln_res_lock == NULL)
673                 goto failed;
674
675         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
676         if (the_lnet.ln_net_lock == NULL)
677                 goto failed;
678
679         return 0;
680
681  failed:
682         lnet_destroy_locks();
683         return -ENOMEM;
684 }
685
686 static void lnet_assert_wire_constants(void)
687 {
688         /* Wire protocol assertions generated by 'wirecheck'
689          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
690          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
691          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
692
693         /* Constants... */
694         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
695         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
696         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
697         BUILD_BUG_ON(LNET_MSG_ACK != 0);
698         BUILD_BUG_ON(LNET_MSG_PUT != 1);
699         BUILD_BUG_ON(LNET_MSG_GET != 2);
700         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
701         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
702
703         /* Checks for struct lnet_handle_wire */
704         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
705         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
706                                    wh_interface_cookie) != 0);
707         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
708         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
709                                    wh_object_cookie) != 8);
710         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
711
712         /* Checks for struct struct lnet_magicversion */
713         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
714         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
715         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
716         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
717         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
718         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
719                                    version_minor) != 6);
720         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
721
722         /* Checks for struct struct lnet_hdr */
723         BUILD_BUG_ON((int)sizeof(struct lnet_hdr) != 72);
724         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_nid) != 0);
725         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_nid) != 8);
726         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_nid) != 8);
727         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_nid) != 8);
728         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_pid) != 16);
729         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_pid) != 4);
730         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_pid) != 20);
731         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_pid) != 4);
732         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, type) != 24);
733         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->type) != 4);
734         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, payload_length) != 28);
735         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->payload_length) != 4);
736         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg) != 32);
737         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg) != 40);
738
739         /* Ack */
740         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) != 32);
741         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) != 16);
742         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.match_bits) != 48);
743         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) != 8);
744         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.mlength) != 56);
745         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) != 4);
746
747         /* Put */
748         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) != 32);
749         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) != 16);
750         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.match_bits) != 48);
751         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) != 8);
752         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.hdr_data) != 56);
753         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) != 8);
754         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ptl_index) != 64);
755         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) != 4);
756         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.offset) != 68);
757         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) != 4);
758
759         /* Get */
760         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.return_wmd) != 32);
761         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) != 16);
762         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.match_bits) != 48);
763         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) != 8);
764         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.ptl_index) != 56);
765         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) != 4);
766         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.src_offset) != 60);
767         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) != 4);
768         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.sink_length) != 64);
769         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) != 4);
770
771         /* Reply */
772         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) != 32);
773         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) != 16);
774
775         /* Hello */
776         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.incarnation) != 32);
777         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) != 8);
778         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.type) != 40);
779         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) != 4);
780
781         /* Checks for struct lnet_ni_status and related constants */
782         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
783         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
784         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
785
786         /* Checks for struct lnet_ni_status */
787         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
788         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
789         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
790         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
791         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
792         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_unused) != 12);
793         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) != 4);
794
795         /* Checks for struct lnet_ping_info and related constants */
796         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
797         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
798         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
799         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
800         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
801         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
802         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
803         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 31);
804
805         /* Checks for struct lnet_ping_info */
806         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
807         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
808         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
809         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
810         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
811         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
812         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
813         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
814         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
815         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
816         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) != 0);
817 }
818
819 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
820 {
821         const struct lnet_lnd *lnd;
822
823         /* holding lnd mutex */
824         if (type >= NUM_LNDS)
825                 return NULL;
826         lnd = the_lnet.ln_lnds[type];
827         LASSERT(!lnd || lnd->lnd_type == type);
828
829         return lnd;
830 }
831
832 unsigned int
833 lnet_get_lnd_timeout(void)
834 {
835         return lnet_lnd_timeout;
836 }
837 EXPORT_SYMBOL(lnet_get_lnd_timeout);
838
839 void
840 lnet_register_lnd(const struct lnet_lnd *lnd)
841 {
842         mutex_lock(&the_lnet.ln_lnd_mutex);
843
844         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
845         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
846
847         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
848
849         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
850
851         mutex_unlock(&the_lnet.ln_lnd_mutex);
852 }
853 EXPORT_SYMBOL(lnet_register_lnd);
854
855 void
856 lnet_unregister_lnd(const struct lnet_lnd *lnd)
857 {
858         mutex_lock(&the_lnet.ln_lnd_mutex);
859
860         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
861
862         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
863         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
864
865         mutex_unlock(&the_lnet.ln_lnd_mutex);
866 }
867 EXPORT_SYMBOL(lnet_unregister_lnd);
868
869 void
870 lnet_counters_get_common(struct lnet_counters_common *common)
871 {
872         struct lnet_counters *ctr;
873         int i;
874
875         memset(common, 0, sizeof(*common));
876
877         lnet_net_lock(LNET_LOCK_EX);
878
879         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
880                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
881                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
882                 common->lcc_errors       += ctr->lct_common.lcc_errors;
883                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
884                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
885                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
886                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
887                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
888                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
889                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
890                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
891         }
892         lnet_net_unlock(LNET_LOCK_EX);
893 }
894 EXPORT_SYMBOL(lnet_counters_get_common);
895
896 void
897 lnet_counters_get(struct lnet_counters *counters)
898 {
899         struct lnet_counters *ctr;
900         struct lnet_counters_health *health = &counters->lct_health;
901         int             i;
902
903         memset(counters, 0, sizeof(*counters));
904
905         lnet_counters_get_common(&counters->lct_common);
906
907         lnet_net_lock(LNET_LOCK_EX);
908
909         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
910                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
911                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
912                 health->lch_response_timeout_count +=
913                                 ctr->lct_health.lch_response_timeout_count;
914                 health->lch_local_interrupt_count +=
915                                 ctr->lct_health.lch_local_interrupt_count;
916                 health->lch_local_dropped_count +=
917                                 ctr->lct_health.lch_local_dropped_count;
918                 health->lch_local_aborted_count +=
919                                 ctr->lct_health.lch_local_aborted_count;
920                 health->lch_local_no_route_count +=
921                                 ctr->lct_health.lch_local_no_route_count;
922                 health->lch_local_timeout_count +=
923                                 ctr->lct_health.lch_local_timeout_count;
924                 health->lch_local_error_count +=
925                                 ctr->lct_health.lch_local_error_count;
926                 health->lch_remote_dropped_count +=
927                                 ctr->lct_health.lch_remote_dropped_count;
928                 health->lch_remote_error_count +=
929                                 ctr->lct_health.lch_remote_error_count;
930                 health->lch_remote_timeout_count +=
931                                 ctr->lct_health.lch_remote_timeout_count;
932                 health->lch_network_timeout_count +=
933                                 ctr->lct_health.lch_network_timeout_count;
934         }
935         lnet_net_unlock(LNET_LOCK_EX);
936 }
937 EXPORT_SYMBOL(lnet_counters_get);
938
939 void
940 lnet_counters_reset(void)
941 {
942         struct lnet_counters *counters;
943         int             i;
944
945         lnet_net_lock(LNET_LOCK_EX);
946
947         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
948                 memset(counters, 0, sizeof(struct lnet_counters));
949
950         lnet_net_unlock(LNET_LOCK_EX);
951 }
952
953 static char *
954 lnet_res_type2str(int type)
955 {
956         switch (type) {
957         default:
958                 LBUG();
959         case LNET_COOKIE_TYPE_MD:
960                 return "MD";
961         case LNET_COOKIE_TYPE_ME:
962                 return "ME";
963         case LNET_COOKIE_TYPE_EQ:
964                 return "EQ";
965         }
966 }
967
968 static void
969 lnet_res_container_cleanup(struct lnet_res_container *rec)
970 {
971         int     count = 0;
972
973         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
974                 return;
975
976         while (!list_empty(&rec->rec_active)) {
977                 struct list_head *e = rec->rec_active.next;
978
979                 list_del_init(e);
980                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
981                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
982
983                 } else { /* NB: Active MEs should be attached on portals */
984                         LBUG();
985                 }
986                 count++;
987         }
988
989         if (count > 0) {
990                 /* Found alive MD/ME/EQ, user really should unlink/free
991                  * all of them before finalize LNet, but if someone didn't,
992                  * we have to recycle garbage for him */
993                 CERROR("%d active elements on exit of %s container\n",
994                        count, lnet_res_type2str(rec->rec_type));
995         }
996
997         if (rec->rec_lh_hash != NULL) {
998                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
999                 rec->rec_lh_hash = NULL;
1000         }
1001
1002         rec->rec_type = 0; /* mark it as finalized */
1003 }
1004
1005 static int
1006 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1007 {
1008         int     rc = 0;
1009         int     i;
1010
1011         LASSERT(rec->rec_type == 0);
1012
1013         rec->rec_type = type;
1014         INIT_LIST_HEAD(&rec->rec_active);
1015
1016         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1017
1018         /* Arbitrary choice of hash table size */
1019         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1020                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1021         if (rec->rec_lh_hash == NULL) {
1022                 rc = -ENOMEM;
1023                 goto out;
1024         }
1025
1026         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1027                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1028
1029         return 0;
1030
1031 out:
1032         CERROR("Failed to setup %s resource container\n",
1033                lnet_res_type2str(type));
1034         lnet_res_container_cleanup(rec);
1035         return rc;
1036 }
1037
1038 static void
1039 lnet_res_containers_destroy(struct lnet_res_container **recs)
1040 {
1041         struct lnet_res_container       *rec;
1042         int                             i;
1043
1044         cfs_percpt_for_each(rec, i, recs)
1045                 lnet_res_container_cleanup(rec);
1046
1047         cfs_percpt_free(recs);
1048 }
1049
1050 static struct lnet_res_container **
1051 lnet_res_containers_create(int type)
1052 {
1053         struct lnet_res_container       **recs;
1054         struct lnet_res_container       *rec;
1055         int                             rc;
1056         int                             i;
1057
1058         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1059         if (recs == NULL) {
1060                 CERROR("Failed to allocate %s resource containers\n",
1061                        lnet_res_type2str(type));
1062                 return NULL;
1063         }
1064
1065         cfs_percpt_for_each(rec, i, recs) {
1066                 rc = lnet_res_container_setup(rec, i, type);
1067                 if (rc != 0) {
1068                         lnet_res_containers_destroy(recs);
1069                         return NULL;
1070                 }
1071         }
1072
1073         return recs;
1074 }
1075
1076 struct lnet_libhandle *
1077 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1078 {
1079         /* ALWAYS called with lnet_res_lock held */
1080         struct list_head        *head;
1081         struct lnet_libhandle   *lh;
1082         unsigned int            hash;
1083
1084         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1085                 return NULL;
1086
1087         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1088         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1089
1090         list_for_each_entry(lh, head, lh_hash_chain) {
1091                 if (lh->lh_cookie == cookie)
1092                         return lh;
1093         }
1094
1095         return NULL;
1096 }
1097
1098 void
1099 lnet_res_lh_initialize(struct lnet_res_container *rec,
1100                        struct lnet_libhandle *lh)
1101 {
1102         /* ALWAYS called with lnet_res_lock held */
1103         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1104         unsigned int    hash;
1105
1106         lh->lh_cookie = rec->rec_lh_cookie;
1107         rec->rec_lh_cookie += 1 << ibits;
1108
1109         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1110
1111         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1112 }
1113
1114 struct list_head **
1115 lnet_create_array_of_queues(void)
1116 {
1117         struct list_head **qs;
1118         struct list_head *q;
1119         int i;
1120
1121         qs = cfs_percpt_alloc(lnet_cpt_table(),
1122                               sizeof(struct list_head));
1123         if (!qs) {
1124                 CERROR("Failed to allocate queues\n");
1125                 return NULL;
1126         }
1127
1128         cfs_percpt_for_each(q, i, qs)
1129                 INIT_LIST_HEAD(q);
1130
1131         return qs;
1132 }
1133
1134 static int lnet_unprepare(void);
1135
1136 static int
1137 lnet_prepare(lnet_pid_t requested_pid)
1138 {
1139         /* Prepare to bring up the network */
1140         struct lnet_res_container **recs;
1141         int                       rc = 0;
1142
1143         if (requested_pid == LNET_PID_ANY) {
1144                 /* Don't instantiate LNET just for me */
1145                 return -ENETDOWN;
1146         }
1147
1148         LASSERT(the_lnet.ln_refcount == 0);
1149
1150         the_lnet.ln_routing = 0;
1151
1152         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1153         the_lnet.ln_pid = requested_pid;
1154
1155         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1156         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1157         INIT_LIST_HEAD(&the_lnet.ln_nets);
1158         INIT_LIST_HEAD(&the_lnet.ln_routers);
1159         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1160         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1161         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1162         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1163         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1164         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1165         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1166         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1167         the_lnet.ln_mt_eq = NULL;
1168         init_completion(&the_lnet.ln_started);
1169
1170         rc = lnet_slab_setup();
1171         if (rc != 0)
1172                 goto failed;
1173
1174         rc = lnet_create_remote_nets_table();
1175         if (rc != 0)
1176                 goto failed;
1177
1178         /*
1179          * NB the interface cookie in wire handles guards against delayed
1180          * replies and ACKs appearing valid after reboot.
1181          */
1182         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1183
1184         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1185                                                 sizeof(struct lnet_counters));
1186         if (the_lnet.ln_counters == NULL) {
1187                 CERROR("Failed to allocate counters for LNet\n");
1188                 rc = -ENOMEM;
1189                 goto failed;
1190         }
1191
1192         rc = lnet_peer_tables_create();
1193         if (rc != 0)
1194                 goto failed;
1195
1196         rc = lnet_msg_containers_create();
1197         if (rc != 0)
1198                 goto failed;
1199
1200         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1201                                       LNET_COOKIE_TYPE_EQ);
1202         if (rc != 0)
1203                 goto failed;
1204
1205         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1206         if (recs == NULL) {
1207                 rc = -ENOMEM;
1208                 goto failed;
1209         }
1210
1211         the_lnet.ln_md_containers = recs;
1212
1213         rc = lnet_portals_create();
1214         if (rc != 0) {
1215                 CERROR("Failed to create portals for LNet: %d\n", rc);
1216                 goto failed;
1217         }
1218
1219         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1220         if (!the_lnet.ln_mt_zombie_rstqs) {
1221                 rc = -ENOMEM;
1222                 goto failed;
1223         }
1224
1225         return 0;
1226
1227  failed:
1228         lnet_unprepare();
1229         return rc;
1230 }
1231
1232 static int
1233 lnet_unprepare (void)
1234 {
1235         int rc;
1236
1237         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1238          * have shut down already, so it is safe to unlink and free all
1239          * descriptors, even those that appear committed to a network op (eg MD
1240          * with non-zero pending count) */
1241
1242         lnet_fail_nid(LNET_NID_ANY, 0);
1243
1244         LASSERT(the_lnet.ln_refcount == 0);
1245         LASSERT(list_empty(&the_lnet.ln_test_peers));
1246         LASSERT(list_empty(&the_lnet.ln_nets));
1247
1248         if (the_lnet.ln_mt_zombie_rstqs) {
1249                 lnet_clean_zombie_rstqs();
1250                 the_lnet.ln_mt_zombie_rstqs = NULL;
1251         }
1252
1253         if (the_lnet.ln_mt_eq) {
1254                 rc = LNetEQFree(the_lnet.ln_mt_eq);
1255                 the_lnet.ln_mt_eq = NULL;
1256                 LASSERT(rc == 0);
1257         }
1258
1259         lnet_portals_destroy();
1260
1261         if (the_lnet.ln_md_containers != NULL) {
1262                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1263                 the_lnet.ln_md_containers = NULL;
1264         }
1265
1266         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1267
1268         lnet_msg_containers_destroy();
1269         lnet_peer_uninit();
1270         lnet_rtrpools_free(0);
1271
1272         if (the_lnet.ln_counters != NULL) {
1273                 cfs_percpt_free(the_lnet.ln_counters);
1274                 the_lnet.ln_counters = NULL;
1275         }
1276         lnet_destroy_remote_nets_table();
1277         lnet_slab_cleanup();
1278
1279         return 0;
1280 }
1281
1282 struct lnet_ni  *
1283 lnet_net2ni_locked(__u32 net_id, int cpt)
1284 {
1285         struct lnet_ni   *ni;
1286         struct lnet_net  *net;
1287
1288         LASSERT(cpt != LNET_LOCK_EX);
1289
1290         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1291                 if (net->net_id == net_id) {
1292                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1293                                         ni_netlist);
1294                         return ni;
1295                 }
1296         }
1297
1298         return NULL;
1299 }
1300
1301 struct lnet_ni *
1302 lnet_net2ni_addref(__u32 net)
1303 {
1304         struct lnet_ni *ni;
1305
1306         lnet_net_lock(0);
1307         ni = lnet_net2ni_locked(net, 0);
1308         if (ni)
1309                 lnet_ni_addref_locked(ni, 0);
1310         lnet_net_unlock(0);
1311
1312         return ni;
1313 }
1314 EXPORT_SYMBOL(lnet_net2ni_addref);
1315
1316 struct lnet_net *
1317 lnet_get_net_locked(__u32 net_id)
1318 {
1319         struct lnet_net  *net;
1320
1321         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1322                 if (net->net_id == net_id)
1323                         return net;
1324         }
1325
1326         return NULL;
1327 }
1328
1329 unsigned int
1330 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
1331 {
1332         __u64           key = nid;
1333         unsigned int    val;
1334
1335         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1336
1337         if (number == 1)
1338                 return 0;
1339
1340         val = hash_long(key, LNET_CPT_BITS);
1341         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1342         if (val < number)
1343                 return val;
1344
1345         return (unsigned int)(key + val + (val >> 1)) % number;
1346 }
1347
1348 int
1349 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
1350 {
1351         struct lnet_net *net;
1352
1353         /* must called with hold of lnet_net_lock */
1354         if (LNET_CPT_NUMBER == 1)
1355                 return 0; /* the only one */
1356
1357         /*
1358          * If NI is provided then use the CPT identified in the NI cpt
1359          * list if one exists. If one doesn't exist, then that NI is
1360          * associated with all CPTs and it follows that the net it belongs
1361          * to is implicitly associated with all CPTs, so just hash the nid
1362          * and return that.
1363          */
1364         if (ni != NULL) {
1365                 if (ni->ni_cpts != NULL)
1366                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1367                                                              ni->ni_ncpts)];
1368                 else
1369                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1370         }
1371
1372         /* no NI provided so look at the net */
1373         net = lnet_get_net_locked(LNET_NIDNET(nid));
1374
1375         if (net != NULL && net->net_cpts != NULL) {
1376                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1377         }
1378
1379         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1380 }
1381
1382 int
1383 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
1384 {
1385         int     cpt;
1386         int     cpt2;
1387
1388         if (LNET_CPT_NUMBER == 1)
1389                 return 0; /* the only one */
1390
1391         cpt = lnet_net_lock_current();
1392
1393         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1394
1395         lnet_net_unlock(cpt);
1396
1397         return cpt2;
1398 }
1399 EXPORT_SYMBOL(lnet_cpt_of_nid);
1400
1401 int
1402 lnet_islocalnet_locked(__u32 net_id)
1403 {
1404         struct lnet_net *net;
1405         bool local;
1406
1407         net = lnet_get_net_locked(net_id);
1408
1409         local = net != NULL;
1410
1411         return local;
1412 }
1413
1414 int
1415 lnet_islocalnet(__u32 net_id)
1416 {
1417         int cpt;
1418         bool local;
1419
1420         cpt = lnet_net_lock_current();
1421
1422         local = lnet_islocalnet_locked(net_id);
1423
1424         lnet_net_unlock(cpt);
1425
1426         return local;
1427 }
1428
1429 struct lnet_ni  *
1430 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
1431 {
1432         struct lnet_net  *net;
1433         struct lnet_ni   *ni;
1434
1435         LASSERT(cpt != LNET_LOCK_EX);
1436
1437         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1438                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1439                         if (ni->ni_nid == nid)
1440                                 return ni;
1441                 }
1442         }
1443
1444         return NULL;
1445 }
1446
1447 struct lnet_ni *
1448 lnet_nid2ni_addref(lnet_nid_t nid)
1449 {
1450         struct lnet_ni *ni;
1451
1452         lnet_net_lock(0);
1453         ni = lnet_nid2ni_locked(nid, 0);
1454         if (ni)
1455                 lnet_ni_addref_locked(ni, 0);
1456         lnet_net_unlock(0);
1457
1458         return ni;
1459 }
1460 EXPORT_SYMBOL(lnet_nid2ni_addref);
1461
1462 int
1463 lnet_islocalnid(lnet_nid_t nid)
1464 {
1465         struct lnet_ni  *ni;
1466         int             cpt;
1467
1468         cpt = lnet_net_lock_current();
1469         ni = lnet_nid2ni_locked(nid, cpt);
1470         lnet_net_unlock(cpt);
1471
1472         return ni != NULL;
1473 }
1474
1475 int
1476 lnet_count_acceptor_nets(void)
1477 {
1478         /* Return the # of NIs that need the acceptor. */
1479         int              count = 0;
1480         struct lnet_net  *net;
1481         int              cpt;
1482
1483         cpt = lnet_net_lock_current();
1484         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1485                 /* all socklnd type networks should have the acceptor
1486                  * thread started */
1487                 if (net->net_lnd->lnd_accept != NULL)
1488                         count++;
1489         }
1490
1491         lnet_net_unlock(cpt);
1492
1493         return count;
1494 }
1495
1496 struct lnet_ping_buffer *
1497 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1498 {
1499         struct lnet_ping_buffer *pbuf;
1500
1501         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1502         if (pbuf) {
1503                 pbuf->pb_nnis = nnis;
1504                 pbuf->pb_needs_post = false;
1505                 atomic_set(&pbuf->pb_refcnt, 1);
1506         }
1507
1508         return pbuf;
1509 }
1510
1511 void
1512 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1513 {
1514         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1515         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1516 }
1517
1518 static struct lnet_ping_buffer *
1519 lnet_ping_target_create(int nnis)
1520 {
1521         struct lnet_ping_buffer *pbuf;
1522
1523         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1524         if (pbuf == NULL) {
1525                 CERROR("Can't allocate ping source [%d]\n", nnis);
1526                 return NULL;
1527         }
1528
1529         pbuf->pb_info.pi_nnis = nnis;
1530         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1531         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1532         pbuf->pb_info.pi_features =
1533                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1534
1535         return pbuf;
1536 }
1537
1538 static inline int
1539 lnet_get_net_ni_count_locked(struct lnet_net *net)
1540 {
1541         struct lnet_ni  *ni;
1542         int             count = 0;
1543
1544         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1545                 count++;
1546
1547         return count;
1548 }
1549
1550 static inline int
1551 lnet_get_net_ni_count_pre(struct lnet_net *net)
1552 {
1553         struct lnet_ni  *ni;
1554         int             count = 0;
1555
1556         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1557                 count++;
1558
1559         return count;
1560 }
1561
1562 static inline int
1563 lnet_get_ni_count(void)
1564 {
1565         struct lnet_ni  *ni;
1566         struct lnet_net *net;
1567         int             count = 0;
1568
1569         lnet_net_lock(0);
1570
1571         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1572                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1573                         count++;
1574         }
1575
1576         lnet_net_unlock(0);
1577
1578         return count;
1579 }
1580
1581 int
1582 lnet_get_net_count(void)
1583 {
1584         struct lnet_net *net;
1585         int count = 0;
1586
1587         lnet_net_lock(0);
1588
1589         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1590                 count++;
1591         }
1592
1593         lnet_net_unlock(0);
1594
1595         return count;
1596 }
1597
1598 void
1599 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1600 {
1601         struct lnet_ni_status *stat;
1602         int nnis;
1603         int i;
1604
1605         __swab32s(&pbuf->pb_info.pi_magic);
1606         __swab32s(&pbuf->pb_info.pi_features);
1607         __swab32s(&pbuf->pb_info.pi_pid);
1608         __swab32s(&pbuf->pb_info.pi_nnis);
1609         nnis = pbuf->pb_info.pi_nnis;
1610         if (nnis > pbuf->pb_nnis)
1611                 nnis = pbuf->pb_nnis;
1612         for (i = 0; i < nnis; i++) {
1613                 stat = &pbuf->pb_info.pi_ni[i];
1614                 __swab64s(&stat->ns_nid);
1615                 __swab32s(&stat->ns_status);
1616         }
1617 }
1618
1619 int
1620 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1621 {
1622         if (!pinfo)
1623                 return -EINVAL;
1624         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1625                 return -EPROTO;
1626         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1627                 return -EPROTO;
1628         /* Loopback is guaranteed to be present */
1629         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1630                 return -ERANGE;
1631         if (LNET_NETTYP(LNET_NIDNET(LNET_PING_INFO_LONI(pinfo))) != LOLND)
1632                 return -EPROTO;
1633         return 0;
1634 }
1635
1636 static void
1637 lnet_ping_target_destroy(void)
1638 {
1639         struct lnet_net *net;
1640         struct lnet_ni  *ni;
1641
1642         lnet_net_lock(LNET_LOCK_EX);
1643
1644         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1645                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1646                         lnet_ni_lock(ni);
1647                         ni->ni_status = NULL;
1648                         lnet_ni_unlock(ni);
1649                 }
1650         }
1651
1652         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1653         the_lnet.ln_ping_target = NULL;
1654
1655         lnet_net_unlock(LNET_LOCK_EX);
1656 }
1657
1658 static void
1659 lnet_ping_target_event_handler(struct lnet_event *event)
1660 {
1661         struct lnet_ping_buffer *pbuf = event->md.user_ptr;
1662
1663         if (event->unlinked)
1664                 lnet_ping_buffer_decref(pbuf);
1665 }
1666
1667 static int
1668 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1669                        struct lnet_handle_md *ping_mdh,
1670                        int ni_count, bool set_eq)
1671 {
1672         struct lnet_process_id id = {
1673                 .nid = LNET_NID_ANY,
1674                 .pid = LNET_PID_ANY
1675         };
1676         struct lnet_me *me;
1677         struct lnet_md md = { NULL };
1678         int rc, rc2;
1679
1680         if (set_eq) {
1681                 the_lnet.ln_ping_target_eq =
1682                         LNetEQAlloc(lnet_ping_target_event_handler);
1683                 if (IS_ERR(the_lnet.ln_ping_target_eq)) {
1684                         rc = PTR_ERR(the_lnet.ln_ping_target_eq);
1685                         CERROR("Can't allocate ping buffer EQ: %d\n", rc);
1686                         return rc;
1687                 }
1688         }
1689
1690         *ppbuf = lnet_ping_target_create(ni_count);
1691         if (*ppbuf == NULL) {
1692                 rc = -ENOMEM;
1693                 goto fail_free_eq;
1694         }
1695
1696         /* Ping target ME/MD */
1697         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1698                           LNET_PROTO_PING_MATCHBITS, 0,
1699                           LNET_UNLINK, LNET_INS_AFTER);
1700         if (IS_ERR(me)) {
1701                 rc = PTR_ERR(me);
1702                 CERROR("Can't create ping target ME: %d\n", rc);
1703                 goto fail_decref_ping_buffer;
1704         }
1705
1706         /* initialize md content */
1707         md.start     = &(*ppbuf)->pb_info;
1708         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1709         md.threshold = LNET_MD_THRESH_INF;
1710         md.max_size  = 0;
1711         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1712                        LNET_MD_MANAGE_REMOTE;
1713         md.eq_handle = the_lnet.ln_ping_target_eq;
1714         md.user_ptr  = *ppbuf;
1715
1716         rc = LNetMDAttach(me, md, LNET_RETAIN, ping_mdh);
1717         if (rc != 0) {
1718                 CERROR("Can't attach ping target MD: %d\n", rc);
1719                 goto fail_unlink_ping_me;
1720         }
1721         lnet_ping_buffer_addref(*ppbuf);
1722
1723         return 0;
1724
1725 fail_unlink_ping_me:
1726         LNetMEUnlink(me);
1727 fail_decref_ping_buffer:
1728         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
1729         lnet_ping_buffer_decref(*ppbuf);
1730         *ppbuf = NULL;
1731 fail_free_eq:
1732         if (set_eq) {
1733                 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1734                 LASSERT(rc2 == 0);
1735         }
1736         return rc;
1737 }
1738
1739 static void
1740 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1741                     struct lnet_handle_md *ping_mdh)
1742 {
1743         LNetMDUnlink(*ping_mdh);
1744         LNetInvalidateMDHandle(ping_mdh);
1745
1746         /* NB the MD could be busy; this just starts the unlink */
1747         while (atomic_read(&pbuf->pb_refcnt) > 1) {
1748                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1749                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
1750         }
1751 }
1752
1753 static void
1754 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1755 {
1756         struct lnet_ni          *ni;
1757         struct lnet_net         *net;
1758         struct lnet_ni_status *ns;
1759         int                     i;
1760         int                     rc;
1761
1762         i = 0;
1763         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1764                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1765                         LASSERT(i < pbuf->pb_nnis);
1766
1767                         ns = &pbuf->pb_info.pi_ni[i];
1768
1769                         ns->ns_nid = ni->ni_nid;
1770
1771                         lnet_ni_lock(ni);
1772                         ns->ns_status = (ni->ni_status != NULL) ?
1773                                          ni->ni_status->ns_status :
1774                                                 LNET_NI_STATUS_UP;
1775                         ni->ni_status = ns;
1776                         lnet_ni_unlock(ni);
1777
1778                         i++;
1779                 }
1780         }
1781         /*
1782          * We (ab)use the ns_status of the loopback interface to
1783          * transmit the sequence number. The first interface listed
1784          * must be the loopback interface.
1785          */
1786         rc = lnet_ping_info_validate(&pbuf->pb_info);
1787         if (rc) {
1788                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1789                 LBUG();
1790         }
1791         LNET_PING_BUFFER_SEQNO(pbuf) =
1792                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1793 }
1794
1795 static void
1796 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1797                         struct lnet_handle_md ping_mdh)
1798 {
1799         struct lnet_ping_buffer *old_pbuf = NULL;
1800         struct lnet_handle_md old_ping_md;
1801
1802         /* switch the NIs to point to the new ping info created */
1803         lnet_net_lock(LNET_LOCK_EX);
1804
1805         if (!the_lnet.ln_routing)
1806                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1807         if (!lnet_peer_discovery_disabled)
1808                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1809
1810         /* Ensure only known feature bits have been set. */
1811         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1812         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1813
1814         lnet_ping_target_install_locked(pbuf);
1815
1816         if (the_lnet.ln_ping_target) {
1817                 old_pbuf = the_lnet.ln_ping_target;
1818                 old_ping_md = the_lnet.ln_ping_target_md;
1819         }
1820         the_lnet.ln_ping_target_md = ping_mdh;
1821         the_lnet.ln_ping_target = pbuf;
1822
1823         lnet_net_unlock(LNET_LOCK_EX);
1824
1825         if (old_pbuf) {
1826                 /* unlink and free the old ping info */
1827                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1828                 lnet_ping_buffer_decref(old_pbuf);
1829         }
1830
1831         lnet_push_update_to_peers(0);
1832 }
1833
1834 static void
1835 lnet_ping_target_fini(void)
1836 {
1837         int             rc;
1838
1839         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1840                             &the_lnet.ln_ping_target_md);
1841
1842         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1843         LASSERT(rc == 0);
1844
1845         lnet_ping_target_destroy();
1846 }
1847
1848 /* Resize the push target. */
1849 int lnet_push_target_resize(void)
1850 {
1851         struct lnet_handle_md mdh;
1852         struct lnet_handle_md old_mdh;
1853         struct lnet_ping_buffer *pbuf;
1854         struct lnet_ping_buffer *old_pbuf;
1855         int nnis;
1856         int rc;
1857
1858 again:
1859         nnis = the_lnet.ln_push_target_nnis;
1860         if (nnis <= 0) {
1861                 CDEBUG(D_NET, "Invalid nnis %d\n", nnis);
1862                 return -EINVAL;
1863         }
1864
1865         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
1866          * dropped when we need to resize again (see "old_pbuf" below) or when
1867          * LNet is shutdown (see lnet_push_target_fini())
1868          */
1869         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1870         if (!pbuf) {
1871                 CDEBUG(D_NET, "Can't allocate pbuf for nnis %d\n", nnis);
1872                 return -ENOMEM;
1873         }
1874
1875         rc = lnet_push_target_post(pbuf, &mdh);
1876         if (rc) {
1877                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
1878                 lnet_ping_buffer_decref(pbuf);
1879                 return rc;
1880         }
1881
1882         lnet_net_lock(LNET_LOCK_EX);
1883         old_pbuf = the_lnet.ln_push_target;
1884         old_mdh = the_lnet.ln_push_target_md;
1885         the_lnet.ln_push_target = pbuf;
1886         the_lnet.ln_push_target_md = mdh;
1887         lnet_net_unlock(LNET_LOCK_EX);
1888
1889         if (old_pbuf) {
1890                 LNetMDUnlink(old_mdh);
1891                 /* Drop ref set by lnet_ping_buffer_alloc() */
1892                 lnet_ping_buffer_decref(old_pbuf);
1893         }
1894
1895         /* Received another push or reply that requires a larger buffer */
1896         if (nnis < the_lnet.ln_push_target_nnis)
1897                 goto again;
1898
1899         CDEBUG(D_NET, "nnis %d success\n", nnis);
1900         return 0;
1901 }
1902
1903 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
1904                           struct lnet_handle_md *mdhp)
1905 {
1906         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
1907         struct lnet_md md = { NULL };
1908         struct lnet_me *me;
1909         int rc;
1910
1911         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1912                           LNET_PROTO_PING_MATCHBITS, 0,
1913                           LNET_UNLINK, LNET_INS_AFTER);
1914         if (IS_ERR(me)) {
1915                 rc = PTR_ERR(me);
1916                 CERROR("Can't create push target ME: %d\n", rc);
1917                 return rc;
1918         }
1919
1920         pbuf->pb_needs_post = false;
1921
1922         /* This reference is dropped by lnet_push_target_event_handler() */
1923         lnet_ping_buffer_addref(pbuf);
1924
1925         /* initialize md content */
1926         md.start     = &pbuf->pb_info;
1927         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
1928         md.threshold = 1;
1929         md.max_size  = 0;
1930         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
1931         md.user_ptr  = pbuf;
1932         md.eq_handle = the_lnet.ln_push_target_eq;
1933
1934         rc = LNetMDAttach(me, md, LNET_UNLINK, mdhp);
1935         if (rc) {
1936                 CERROR("Can't attach push MD: %d\n", rc);
1937                 LNetMEUnlink(me);
1938                 lnet_ping_buffer_decref(pbuf);
1939                 pbuf->pb_needs_post = true;
1940                 return rc;
1941         }
1942
1943         CDEBUG(D_NET, "posted push target %p\n", pbuf);
1944
1945         return 0;
1946 }
1947
1948 static void lnet_push_target_event_handler(struct lnet_event *ev)
1949 {
1950         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1951
1952         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
1953                ev->unlinked);
1954
1955         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
1956                 lnet_swap_pinginfo(pbuf);
1957
1958         if (ev->type == LNET_EVENT_UNLINK) {
1959                 /* Drop ref added by lnet_push_target_post() */
1960                 lnet_ping_buffer_decref(pbuf);
1961                 return;
1962         }
1963
1964         lnet_peer_push_event(ev);
1965         if (ev->unlinked)
1966                 /* Drop ref added by lnet_push_target_post */
1967                 lnet_ping_buffer_decref(pbuf);
1968 }
1969
1970 /* Initialize the push target. */
1971 static int lnet_push_target_init(void)
1972 {
1973         int rc;
1974
1975         if (the_lnet.ln_push_target)
1976                 return -EALREADY;
1977
1978         the_lnet.ln_push_target_eq =
1979                 LNetEQAlloc(lnet_push_target_event_handler);
1980         if (IS_ERR(the_lnet.ln_push_target_eq)) {
1981                 rc = PTR_ERR(the_lnet.ln_push_target_eq);
1982                 CERROR("Can't allocated push target EQ: %d\n", rc);
1983                 return rc;
1984         }
1985
1986         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
1987         LASSERT(rc == 0);
1988
1989         /* Start at the required minimum, we'll enlarge if required. */
1990         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
1991
1992         rc = lnet_push_target_resize();
1993
1994         if (rc) {
1995                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
1996                 LNetEQFree(the_lnet.ln_push_target_eq);
1997                 the_lnet.ln_push_target_eq = NULL;
1998         }
1999
2000         return rc;
2001 }
2002
2003 /* Clean up the push target. */
2004 static void lnet_push_target_fini(void)
2005 {
2006         if (!the_lnet.ln_push_target)
2007                 return;
2008
2009         /* Unlink and invalidate to prevent new references. */
2010         LNetMDUnlink(the_lnet.ln_push_target_md);
2011         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2012
2013         /* Wait for the unlink to complete. */
2014         while (atomic_read(&the_lnet.ln_push_target->pb_refcnt) > 1) {
2015                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
2016                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2017         }
2018
2019         /* Drop ref set by lnet_ping_buffer_alloc() */
2020         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2021         the_lnet.ln_push_target = NULL;
2022         the_lnet.ln_push_target_nnis = 0;
2023
2024         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2025         LNetEQFree(the_lnet.ln_push_target_eq);
2026         the_lnet.ln_push_target_eq = NULL;
2027 }
2028
2029 static int
2030 lnet_ni_tq_credits(struct lnet_ni *ni)
2031 {
2032         int     credits;
2033
2034         LASSERT(ni->ni_ncpts >= 1);
2035
2036         if (ni->ni_ncpts == 1)
2037                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2038
2039         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2040         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2041         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2042
2043         return credits;
2044 }
2045
2046 static void
2047 lnet_ni_unlink_locked(struct lnet_ni *ni)
2048 {
2049         /* move it to zombie list and nobody can find it anymore */
2050         LASSERT(!list_empty(&ni->ni_netlist));
2051         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2052         lnet_ni_decref_locked(ni, 0);
2053 }
2054
2055 static void
2056 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2057 {
2058         int             i;
2059         int             islo;
2060         struct lnet_ni  *ni;
2061         struct list_head *zombie_list = &net->net_ni_zombie;
2062
2063         /*
2064          * Now wait for the NIs I just nuked to show up on the zombie
2065          * list and shut them down in guaranteed thread context
2066          */
2067         i = 2;
2068         while (!list_empty(zombie_list)) {
2069                 int     *ref;
2070                 int     j;
2071
2072                 ni = list_entry(zombie_list->next,
2073                                 struct lnet_ni, ni_netlist);
2074                 list_del_init(&ni->ni_netlist);
2075                 /* the ni should be in deleting state. If it's not it's
2076                  * a bug */
2077                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2078                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2079                         if (*ref == 0)
2080                                 continue;
2081                         /* still busy, add it back to zombie list */
2082                         list_add(&ni->ni_netlist, zombie_list);
2083                         break;
2084                 }
2085
2086                 if (!list_empty(&ni->ni_netlist)) {
2087                         lnet_net_unlock(LNET_LOCK_EX);
2088                         ++i;
2089                         if ((i & (-i)) == i) {
2090                                 CDEBUG(D_WARNING,
2091                                        "Waiting for zombie LNI %s\n",
2092                                        libcfs_nid2str(ni->ni_nid));
2093                         }
2094                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2095                         lnet_net_lock(LNET_LOCK_EX);
2096                         continue;
2097                 }
2098
2099                 lnet_net_unlock(LNET_LOCK_EX);
2100
2101                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2102
2103                 LASSERT(!in_interrupt());
2104                 /* Holding the mutex makes it safe for lnd_shutdown
2105                  * to call module_put(). Module unload cannot finish
2106                  * until lnet_unregister_lnd() completes, and that
2107                  * requires the mutex.
2108                  */
2109                 mutex_lock(&the_lnet.ln_lnd_mutex);
2110                 (net->net_lnd->lnd_shutdown)(ni);
2111                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2112
2113                 if (!islo)
2114                         CDEBUG(D_LNI, "Removed LNI %s\n",
2115                               libcfs_nid2str(ni->ni_nid));
2116
2117                 lnet_ni_free(ni);
2118                 i = 2;
2119                 lnet_net_lock(LNET_LOCK_EX);
2120         }
2121 }
2122
2123 /* shutdown down the NI and release refcount */
2124 static void
2125 lnet_shutdown_lndni(struct lnet_ni *ni)
2126 {
2127         int i;
2128         struct lnet_net *net = ni->ni_net;
2129
2130         lnet_net_lock(LNET_LOCK_EX);
2131         lnet_ni_lock(ni);
2132         ni->ni_state = LNET_NI_STATE_DELETING;
2133         lnet_ni_unlock(ni);
2134         lnet_ni_unlink_locked(ni);
2135         lnet_incr_dlc_seq();
2136         lnet_net_unlock(LNET_LOCK_EX);
2137
2138         /* clear messages for this NI on the lazy portal */
2139         for (i = 0; i < the_lnet.ln_nportals; i++)
2140                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2141
2142         lnet_net_lock(LNET_LOCK_EX);
2143         lnet_clear_zombies_nis_locked(net);
2144         lnet_net_unlock(LNET_LOCK_EX);
2145 }
2146
2147 static void
2148 lnet_shutdown_lndnet(struct lnet_net *net)
2149 {
2150         struct lnet_ni *ni;
2151
2152         lnet_net_lock(LNET_LOCK_EX);
2153
2154         list_del_init(&net->net_list);
2155
2156         while (!list_empty(&net->net_ni_list)) {
2157                 ni = list_entry(net->net_ni_list.next,
2158                                 struct lnet_ni, ni_netlist);
2159                 lnet_net_unlock(LNET_LOCK_EX);
2160                 lnet_shutdown_lndni(ni);
2161                 lnet_net_lock(LNET_LOCK_EX);
2162         }
2163
2164         lnet_net_unlock(LNET_LOCK_EX);
2165
2166         /* Do peer table cleanup for this net */
2167         lnet_peer_tables_cleanup(net);
2168
2169         lnet_net_free(net);
2170 }
2171
2172 static void
2173 lnet_shutdown_lndnets(void)
2174 {
2175         struct lnet_net *net;
2176         LIST_HEAD(resend);
2177         struct lnet_msg *msg, *tmp;
2178
2179         /* NB called holding the global mutex */
2180
2181         /* All quiet on the API front */
2182         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2183         LASSERT(the_lnet.ln_refcount == 0);
2184
2185         lnet_net_lock(LNET_LOCK_EX);
2186         the_lnet.ln_state = LNET_STATE_STOPPING;
2187
2188         /*
2189          * move the nets to the zombie list to avoid them being
2190          * picked up for new work. LONET is also included in the
2191          * Nets that will be moved to the zombie list
2192          */
2193         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2194
2195         /* Drop the cached loopback Net. */
2196         if (the_lnet.ln_loni != NULL) {
2197                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2198                 the_lnet.ln_loni = NULL;
2199         }
2200         lnet_net_unlock(LNET_LOCK_EX);
2201
2202         /* iterate through the net zombie list and delete each net */
2203         while (!list_empty(&the_lnet.ln_net_zombie)) {
2204                 net = list_entry(the_lnet.ln_net_zombie.next,
2205                                  struct lnet_net, net_list);
2206                 lnet_shutdown_lndnet(net);
2207         }
2208
2209         spin_lock(&the_lnet.ln_msg_resend_lock);
2210         list_splice(&the_lnet.ln_msg_resend, &resend);
2211         spin_unlock(&the_lnet.ln_msg_resend_lock);
2212
2213         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2214                 list_del_init(&msg->msg_list);
2215                 msg->msg_no_resend = true;
2216                 lnet_finalize(msg, -ECANCELED);
2217         }
2218
2219         lnet_net_lock(LNET_LOCK_EX);
2220         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2221         lnet_net_unlock(LNET_LOCK_EX);
2222 }
2223
2224 static int
2225 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2226 {
2227         int                     rc = -EINVAL;
2228         struct lnet_tx_queue    *tq;
2229         int                     i;
2230         struct lnet_net         *net = ni->ni_net;
2231
2232         mutex_lock(&the_lnet.ln_lnd_mutex);
2233
2234         if (tun) {
2235                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2236                 ni->ni_lnd_tunables_set = true;
2237         }
2238
2239         rc = (net->net_lnd->lnd_startup)(ni);
2240
2241         mutex_unlock(&the_lnet.ln_lnd_mutex);
2242
2243         if (rc != 0) {
2244                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2245                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2246                 goto failed0;
2247         }
2248
2249         lnet_ni_lock(ni);
2250         ni->ni_state = LNET_NI_STATE_ACTIVE;
2251         lnet_ni_unlock(ni);
2252
2253         /* We keep a reference on the loopback net through the loopback NI */
2254         if (net->net_lnd->lnd_type == LOLND) {
2255                 lnet_ni_addref(ni);
2256                 LASSERT(the_lnet.ln_loni == NULL);
2257                 the_lnet.ln_loni = ni;
2258                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2259                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2260                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2261                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2262                 return 0;
2263         }
2264
2265         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2266             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2267                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2268                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2269                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2270                                         "" : "per-peer ");
2271                 /* shutdown the NI since if we get here then it must've already
2272                  * been started
2273                  */
2274                 lnet_shutdown_lndni(ni);
2275                 return -EINVAL;
2276         }
2277
2278         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2279                 tq->tq_credits_min =
2280                 tq->tq_credits_max =
2281                 tq->tq_credits = lnet_ni_tq_credits(ni);
2282         }
2283
2284         atomic_set(&ni->ni_tx_credits,
2285                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2286         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2287
2288         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2289                 libcfs_nid2str(ni->ni_nid),
2290                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2291                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2292                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2293                 ni->ni_net->net_tunables.lct_peer_timeout);
2294
2295         return 0;
2296 failed0:
2297         lnet_ni_free(ni);
2298         return rc;
2299 }
2300
2301 static int
2302 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2303 {
2304         struct lnet_ni *ni;
2305         struct lnet_net *net_l = NULL;
2306         LIST_HEAD(local_ni_list);
2307         int rc;
2308         int ni_count = 0;
2309         __u32 lnd_type;
2310         const struct lnet_lnd  *lnd;
2311         int peer_timeout =
2312                 net->net_tunables.lct_peer_timeout;
2313         int maxtxcredits =
2314                 net->net_tunables.lct_max_tx_credits;
2315         int peerrtrcredits =
2316                 net->net_tunables.lct_peer_rtr_credits;
2317
2318         /*
2319          * make sure that this net is unique. If it isn't then
2320          * we are adding interfaces to an already existing network, and
2321          * 'net' is just a convenient way to pass in the list.
2322          * if it is unique we need to find the LND and load it if
2323          * necessary.
2324          */
2325         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2326                 lnd_type = LNET_NETTYP(net->net_id);
2327
2328                 mutex_lock(&the_lnet.ln_lnd_mutex);
2329                 lnd = lnet_find_lnd_by_type(lnd_type);
2330
2331                 if (lnd == NULL) {
2332                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2333                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2334                         mutex_lock(&the_lnet.ln_lnd_mutex);
2335
2336                         lnd = lnet_find_lnd_by_type(lnd_type);
2337                         if (lnd == NULL) {
2338                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2339                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2340                                 libcfs_lnd2str(lnd_type),
2341                                 libcfs_lnd2modname(lnd_type), rc);
2342 #ifndef HAVE_MODULE_LOADING_SUPPORT
2343                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2344                                                 "compiled with kernel module "
2345                                                 "loading support.");
2346 #endif
2347                                 rc = -EINVAL;
2348                                 goto failed0;
2349                         }
2350                 }
2351
2352                 net->net_lnd = lnd;
2353
2354                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2355
2356                 net_l = net;
2357         }
2358
2359         /*
2360          * net_l: if the network being added is unique then net_l
2361          *        will point to that network
2362          *        if the network being added is not unique then
2363          *        net_l points to the existing network.
2364          *
2365          * When we enter the loop below, we'll pick NIs off he
2366          * network beign added and start them up, then add them to
2367          * a local ni list. Once we've successfully started all
2368          * the NIs then we join the local NI list (of started up
2369          * networks) with the net_l->net_ni_list, which should
2370          * point to the correct network to add the new ni list to
2371          *
2372          * If any of the new NIs fail to start up, then we want to
2373          * iterate through the local ni list, which should include
2374          * any NIs which were successfully started up, and shut
2375          * them down.
2376          *
2377          * After than we want to delete the network being added,
2378          * to avoid a memory leak.
2379          */
2380
2381         /*
2382          * When a network uses TCP bonding then all its interfaces
2383          * must be specified when the network is first defined: the
2384          * TCP bonding code doesn't allow for interfaces to be added
2385          * or removed.
2386          */
2387         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2388             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2389                 rc = -EINVAL;
2390                 goto failed0;
2391         }
2392
2393         while (!list_empty(&net->net_ni_added)) {
2394                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2395                                 ni_netlist);
2396                 list_del_init(&ni->ni_netlist);
2397
2398                 /* make sure that the the NI we're about to start
2399                  * up is actually unique. if it's not fail. */
2400                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2401                                         ni->ni_interfaces[0])) {
2402                         rc = -EEXIST;
2403                         goto failed1;
2404                 }
2405
2406                 /* adjust the pointer the parent network, just in case it
2407                  * the net is a duplicate */
2408                 ni->ni_net = net_l;
2409
2410                 rc = lnet_startup_lndni(ni, tun);
2411
2412                 if (rc < 0)
2413                         goto failed1;
2414
2415                 lnet_ni_addref(ni);
2416                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2417
2418                 ni_count++;
2419         }
2420
2421         lnet_net_lock(LNET_LOCK_EX);
2422         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2423         lnet_incr_dlc_seq();
2424         lnet_net_unlock(LNET_LOCK_EX);
2425
2426         /* if the network is not unique then we don't want to keep
2427          * it around after we're done. Free it. Otherwise add that
2428          * net to the global the_lnet.ln_nets */
2429         if (net_l != net && net_l != NULL) {
2430                 /*
2431                  * TODO - note. currently the tunables can not be updated
2432                  * once added
2433                  */
2434                 lnet_net_free(net);
2435         } else {
2436                 /*
2437                  * restore tunables after it has been overwitten by the
2438                  * lnd
2439                  */
2440                 if (peer_timeout != -1)
2441                         net->net_tunables.lct_peer_timeout = peer_timeout;
2442                 if (maxtxcredits != -1)
2443                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2444                 if (peerrtrcredits != -1)
2445                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2446
2447                 lnet_net_lock(LNET_LOCK_EX);
2448                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2449                 lnet_net_unlock(LNET_LOCK_EX);
2450         }
2451
2452         /* update net count */
2453         lnet_current_net_count = lnet_get_net_count();
2454
2455         return ni_count;
2456
2457 failed1:
2458         /*
2459          * shutdown the new NIs that are being started up
2460          * free the NET being started
2461          */
2462         while (!list_empty(&local_ni_list)) {
2463                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2464                                 ni_netlist);
2465
2466                 lnet_shutdown_lndni(ni);
2467         }
2468
2469 failed0:
2470         lnet_net_free(net);
2471
2472         return rc;
2473 }
2474
2475 static int
2476 lnet_startup_lndnets(struct list_head *netlist)
2477 {
2478         struct lnet_net         *net;
2479         int                     rc;
2480         int                     ni_count = 0;
2481
2482         /*
2483          * Change to running state before bringing up the LNDs. This
2484          * allows lnet_shutdown_lndnets() to assert that we've passed
2485          * through here.
2486          */
2487         lnet_net_lock(LNET_LOCK_EX);
2488         the_lnet.ln_state = LNET_STATE_RUNNING;
2489         lnet_net_unlock(LNET_LOCK_EX);
2490
2491         while (!list_empty(netlist)) {
2492                 net = list_entry(netlist->next, struct lnet_net, net_list);
2493                 list_del_init(&net->net_list);
2494
2495                 rc = lnet_startup_lndnet(net, NULL);
2496
2497                 if (rc < 0)
2498                         goto failed;
2499
2500                 ni_count += rc;
2501         }
2502
2503         return ni_count;
2504 failed:
2505         lnet_shutdown_lndnets();
2506
2507         return rc;
2508 }
2509
2510 /**
2511  * Initialize LNet library.
2512  *
2513  * Automatically called at module loading time. Caller has to call
2514  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2515  * latter returned 0. It must be called exactly once.
2516  *
2517  * \retval 0 on success
2518  * \retval -ve on failures.
2519  */
2520 int lnet_lib_init(void)
2521 {
2522         int rc;
2523
2524         lnet_assert_wire_constants();
2525
2526         /* refer to global cfs_cpt_table for now */
2527         the_lnet.ln_cpt_table = cfs_cpt_tab;
2528         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2529
2530         LASSERT(the_lnet.ln_cpt_number > 0);
2531         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2532                 /* we are under risk of consuming all lh_cookie */
2533                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2534                        "please change setting of CPT-table and retry\n",
2535                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2536                 return -E2BIG;
2537         }
2538
2539         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2540                 the_lnet.ln_cpt_bits++;
2541
2542         rc = lnet_create_locks();
2543         if (rc != 0) {
2544                 CERROR("Can't create LNet global locks: %d\n", rc);
2545                 return rc;
2546         }
2547
2548         the_lnet.ln_refcount = 0;
2549         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2550         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2551
2552         /* The hash table size is the number of bits it takes to express the set
2553          * ln_num_routes, minus 1 (better to under estimate than over so we
2554          * don't waste memory). */
2555         if (rnet_htable_size <= 0)
2556                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2557         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2558                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2559         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2560                                            order_base_2(rnet_htable_size) - 1);
2561
2562         /* All LNDs apart from the LOLND are in separate modules.  They
2563          * register themselves when their module loads, and unregister
2564          * themselves when their module is unloaded. */
2565         lnet_register_lnd(&the_lolnd);
2566         return 0;
2567 }
2568
2569 /**
2570  * Finalize LNet library.
2571  *
2572  * \pre lnet_lib_init() called with success.
2573  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2574  *
2575  * As this happens at module-unload, all lnds must already be unloaded,
2576  * so they must already be unregistered.
2577  */
2578 void lnet_lib_exit(void)
2579 {
2580         int i;
2581
2582         LASSERT(the_lnet.ln_refcount == 0);
2583         lnet_unregister_lnd(&the_lolnd);
2584         for (i = 0; i < NUM_LNDS; i++)
2585                 LASSERT(!the_lnet.ln_lnds[i]);
2586         lnet_destroy_locks();
2587 }
2588
2589 /**
2590  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2591  *
2592  * Users must call this function at least once before any other functions.
2593  * For each successful call there must be a corresponding call to
2594  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2595  * ignored.
2596  *
2597  * The PID used by LNet may be different from the one requested.
2598  * See LNetGetId().
2599  *
2600  * \param requested_pid PID requested by the caller.
2601  *
2602  * \return >= 0 on success, and < 0 error code on failures.
2603  */
2604 int
2605 LNetNIInit(lnet_pid_t requested_pid)
2606 {
2607         int                     im_a_router = 0;
2608         int                     rc;
2609         int                     ni_count;
2610         struct lnet_ping_buffer *pbuf;
2611         struct lnet_handle_md   ping_mdh;
2612         LIST_HEAD(net_head);
2613         struct lnet_net         *net;
2614
2615         mutex_lock(&the_lnet.ln_api_mutex);
2616
2617         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2618
2619         if (the_lnet.ln_refcount > 0) {
2620                 rc = the_lnet.ln_refcount++;
2621                 mutex_unlock(&the_lnet.ln_api_mutex);
2622                 return rc;
2623         }
2624
2625         rc = lnet_prepare(requested_pid);
2626         if (rc != 0) {
2627                 mutex_unlock(&the_lnet.ln_api_mutex);
2628                 return rc;
2629         }
2630
2631         /* create a network for Loopback network */
2632         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2633         if (net == NULL) {
2634                 rc = -ENOMEM;
2635                 goto err_empty_list;
2636         }
2637
2638         /* Add in the loopback NI */
2639         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2640                 rc = -ENOMEM;
2641                 goto err_empty_list;
2642         }
2643
2644         /* If LNet is being initialized via DLC it is possible
2645          * that the user requests not to load module parameters (ones which
2646          * are supported by DLC) on initialization.  Therefore, make sure not
2647          * to load networks, routes and forwarding from module parameters
2648          * in this case.  On cleanup in case of failure only clean up
2649          * routes if it has been loaded */
2650         if (!the_lnet.ln_nis_from_mod_params) {
2651                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2652                                          use_tcp_bonding);
2653                 if (rc < 0)
2654                         goto err_empty_list;
2655         }
2656
2657         ni_count = lnet_startup_lndnets(&net_head);
2658         if (ni_count < 0) {
2659                 rc = ni_count;
2660                 goto err_empty_list;
2661         }
2662
2663         if (!the_lnet.ln_nis_from_mod_params) {
2664                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2665                 if (rc != 0)
2666                         goto err_shutdown_lndnis;
2667
2668                 rc = lnet_rtrpools_alloc(im_a_router);
2669                 if (rc != 0)
2670                         goto err_destroy_routes;
2671         }
2672
2673         rc = lnet_acceptor_start();
2674         if (rc != 0)
2675                 goto err_destroy_routes;
2676
2677         the_lnet.ln_refcount = 1;
2678         /* Now I may use my own API functions... */
2679
2680         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2681         if (rc != 0)
2682                 goto err_acceptor_stop;
2683
2684         lnet_ping_target_update(pbuf, ping_mdh);
2685
2686         the_lnet.ln_mt_eq = LNetEQAlloc(lnet_mt_event_handler);
2687         if (IS_ERR(the_lnet.ln_mt_eq)) {
2688                 rc = PTR_ERR(the_lnet.ln_mt_eq);
2689                 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
2690                 goto err_stop_ping;
2691         }
2692
2693         rc = lnet_push_target_init();
2694         if (rc != 0)
2695                 goto err_stop_ping;
2696
2697         rc = lnet_peer_discovery_start();
2698         if (rc != 0)
2699                 goto err_destroy_push_target;
2700
2701         rc = lnet_monitor_thr_start();
2702         if (rc != 0)
2703                 goto err_stop_discovery_thr;
2704
2705         lnet_fault_init();
2706         lnet_router_debugfs_init();
2707
2708         mutex_unlock(&the_lnet.ln_api_mutex);
2709
2710         complete_all(&the_lnet.ln_started);
2711
2712         /* wait for all routers to start */
2713         lnet_wait_router_start();
2714
2715         return 0;
2716
2717 err_stop_discovery_thr:
2718         lnet_peer_discovery_stop();
2719 err_destroy_push_target:
2720         lnet_push_target_fini();
2721 err_stop_ping:
2722         lnet_ping_target_fini();
2723 err_acceptor_stop:
2724         the_lnet.ln_refcount = 0;
2725         lnet_acceptor_stop();
2726 err_destroy_routes:
2727         if (!the_lnet.ln_nis_from_mod_params)
2728                 lnet_destroy_routes();
2729 err_shutdown_lndnis:
2730         lnet_shutdown_lndnets();
2731 err_empty_list:
2732         lnet_unprepare();
2733         LASSERT(rc < 0);
2734         mutex_unlock(&the_lnet.ln_api_mutex);
2735         while (!list_empty(&net_head)) {
2736                 struct lnet_net *net;
2737
2738                 net = list_entry(net_head.next, struct lnet_net, net_list);
2739                 list_del_init(&net->net_list);
2740                 lnet_net_free(net);
2741         }
2742         return rc;
2743 }
2744 EXPORT_SYMBOL(LNetNIInit);
2745
2746 /**
2747  * Stop LNet interfaces, routing, and forwarding.
2748  *
2749  * Users must call this function once for each successful call to LNetNIInit().
2750  * Once the LNetNIFini() operation has been started, the results of pending
2751  * API operations are undefined.
2752  *
2753  * \return always 0 for current implementation.
2754  */
2755 int
2756 LNetNIFini(void)
2757 {
2758         mutex_lock(&the_lnet.ln_api_mutex);
2759
2760         LASSERT(the_lnet.ln_refcount > 0);
2761
2762         if (the_lnet.ln_refcount != 1) {
2763                 the_lnet.ln_refcount--;
2764         } else {
2765                 LASSERT(!the_lnet.ln_niinit_self);
2766
2767                 lnet_fault_fini();
2768
2769                 lnet_router_debugfs_fini();
2770                 lnet_monitor_thr_stop();
2771                 lnet_peer_discovery_stop();
2772                 lnet_push_target_fini();
2773                 lnet_ping_target_fini();
2774
2775                 /* Teardown fns that use my own API functions BEFORE here */
2776                 the_lnet.ln_refcount = 0;
2777
2778                 lnet_acceptor_stop();
2779                 lnet_destroy_routes();
2780                 lnet_shutdown_lndnets();
2781                 lnet_unprepare();
2782         }
2783
2784         mutex_unlock(&the_lnet.ln_api_mutex);
2785         return 0;
2786 }
2787 EXPORT_SYMBOL(LNetNIFini);
2788
2789 /**
2790  * Grabs the ni data from the ni structure and fills the out
2791  * parameters
2792  *
2793  * \param[in] ni network        interface structure
2794  * \param[out] cfg_ni           NI config information
2795  * \param[out] tun              network and LND tunables
2796  */
2797 static void
2798 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2799                    struct lnet_ioctl_config_lnd_tunables *tun,
2800                    struct lnet_ioctl_element_stats *stats,
2801                    __u32 tun_size)
2802 {
2803         size_t min_size = 0;
2804         int i;
2805
2806         if (!ni || !cfg_ni || !tun)
2807                 return;
2808
2809         if (ni->ni_interfaces[0] != NULL) {
2810                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2811                         if (ni->ni_interfaces[i] != NULL) {
2812                                 strncpy(cfg_ni->lic_ni_intf[i],
2813                                         ni->ni_interfaces[i],
2814                                         sizeof(cfg_ni->lic_ni_intf[i]));
2815                         }
2816                 }
2817         }
2818
2819         cfg_ni->lic_nid = ni->ni_nid;
2820         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2821                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2822         else
2823                 cfg_ni->lic_status = ni->ni_status->ns_status;
2824         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2825         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2826
2827         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2828
2829         if (stats) {
2830                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
2831                                                        LNET_STATS_TYPE_SEND);
2832                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
2833                                                        LNET_STATS_TYPE_RECV);
2834                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
2835                                                        LNET_STATS_TYPE_DROP);
2836         }
2837
2838         /*
2839          * tun->lt_tun will always be present, but in order to be
2840          * backwards compatible, we need to deal with the cases when
2841          * tun->lt_tun is smaller than what the kernel has, because it
2842          * comes from an older version of a userspace program, then we'll
2843          * need to copy as much information as we have available space.
2844          */
2845         min_size = tun_size - sizeof(tun->lt_cmn);
2846         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2847
2848         /* copy over the cpts */
2849         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2850             ni->ni_cpts == NULL)  {
2851                 for (i = 0; i < ni->ni_ncpts; i++)
2852                         cfg_ni->lic_cpts[i] = i;
2853         } else {
2854                 for (i = 0;
2855                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2856                      i < LNET_MAX_SHOW_NUM_CPT;
2857                      i++)
2858                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2859         }
2860         cfg_ni->lic_ncpts = ni->ni_ncpts;
2861 }
2862
2863 /**
2864  * NOTE: This is a legacy function left in the code to be backwards
2865  * compatible with older userspace programs. It should eventually be
2866  * removed.
2867  *
2868  * Grabs the ni data from the ni structure and fills the out
2869  * parameters
2870  *
2871  * \param[in] ni network        interface structure
2872  * \param[out] config           config information
2873  */
2874 static void
2875 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2876                          struct lnet_ioctl_config_data *config)
2877 {
2878         struct lnet_ioctl_net_config *net_config;
2879         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2880         size_t min_size, tunable_size = 0;
2881         int i;
2882
2883         if (!ni || !config)
2884                 return;
2885
2886         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2887         if (!net_config)
2888                 return;
2889
2890         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2891                      ARRAY_SIZE(net_config->ni_interfaces));
2892
2893         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2894                 if (!ni->ni_interfaces[i])
2895                         break;
2896
2897                 strncpy(net_config->ni_interfaces[i],
2898                         ni->ni_interfaces[i],
2899                         sizeof(net_config->ni_interfaces[i]));
2900         }
2901
2902         config->cfg_nid = ni->ni_nid;
2903         config->cfg_config_u.cfg_net.net_peer_timeout =
2904                 ni->ni_net->net_tunables.lct_peer_timeout;
2905         config->cfg_config_u.cfg_net.net_max_tx_credits =
2906                 ni->ni_net->net_tunables.lct_max_tx_credits;
2907         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2908                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2909         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2910                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2911
2912         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2913                 net_config->ni_status = LNET_NI_STATUS_UP;
2914         else
2915                 net_config->ni_status = ni->ni_status->ns_status;
2916
2917         if (ni->ni_cpts) {
2918                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2919
2920                 for (i = 0; i < num_cpts; i++)
2921                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2922
2923                 config->cfg_ncpts = num_cpts;
2924         }
2925
2926         /*
2927          * See if user land tools sent in a newer and larger version
2928          * of struct lnet_tunables than what the kernel uses.
2929          */
2930         min_size = sizeof(*config) + sizeof(*net_config);
2931
2932         if (config->cfg_hdr.ioc_len > min_size)
2933                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2934
2935         /* Don't copy too much data to user space */
2936         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2937         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2938
2939         if (lnd_cfg && min_size) {
2940                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2941                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2942
2943                 /* Tell user land that kernel side has less data */
2944                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2945                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2946                         config->cfg_hdr.ioc_len -= min_size;
2947                 }
2948         }
2949 }
2950
2951 struct lnet_ni *
2952 lnet_get_ni_idx_locked(int idx)
2953 {
2954         struct lnet_ni          *ni;
2955         struct lnet_net         *net;
2956
2957         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2958                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2959                         if (idx-- == 0)
2960                                 return ni;
2961                 }
2962         }
2963
2964         return NULL;
2965 }
2966
2967 struct lnet_ni *
2968 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2969 {
2970         struct lnet_ni          *ni;
2971         struct lnet_net         *net = mynet;
2972
2973         /*
2974          * It is possible that the net has been cleaned out while there is
2975          * a message being sent. This function accessed the net without
2976          * checking if the list is empty
2977          */
2978         if (prev == NULL) {
2979                 if (net == NULL)
2980                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2981                                         net_list);
2982                 if (list_empty(&net->net_ni_list))
2983                         return NULL;
2984                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2985                                 ni_netlist);
2986
2987                 return ni;
2988         }
2989
2990         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2991                 /* if you reached the end of the ni list and the net is
2992                  * specified, then there are no more nis in that net */
2993                 if (net != NULL)
2994                         return NULL;
2995
2996                 /* we reached the end of this net ni list. move to the
2997                  * next net */
2998                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2999                         /* no more nets and no more NIs. */
3000                         return NULL;
3001
3002                 /* get the next net */
3003                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
3004                                  net_list);
3005                 if (list_empty(&net->net_ni_list))
3006                         return NULL;
3007                 /* get the ni on it */
3008                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3009                                 ni_netlist);
3010
3011                 return ni;
3012         }
3013
3014         if (list_empty(&prev->ni_netlist))
3015                 return NULL;
3016
3017         /* there are more nis left */
3018         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
3019
3020         return ni;
3021 }
3022
3023 int
3024 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3025 {
3026         struct lnet_ni *ni;
3027         int cpt;
3028         int rc = -ENOENT;
3029         int idx = config->cfg_count;
3030
3031         cpt = lnet_net_lock_current();
3032
3033         ni = lnet_get_ni_idx_locked(idx);
3034
3035         if (ni != NULL) {
3036                 rc = 0;
3037                 lnet_ni_lock(ni);
3038                 lnet_fill_ni_info_legacy(ni, config);
3039                 lnet_ni_unlock(ni);
3040         }
3041
3042         lnet_net_unlock(cpt);
3043         return rc;
3044 }
3045
3046 int
3047 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3048                    struct lnet_ioctl_config_lnd_tunables *tun,
3049                    struct lnet_ioctl_element_stats *stats,
3050                    __u32 tun_size)
3051 {
3052         struct lnet_ni          *ni;
3053         int                     cpt;
3054         int                     rc = -ENOENT;
3055
3056         if (!cfg_ni || !tun || !stats)
3057                 return -EINVAL;
3058
3059         cpt = lnet_net_lock_current();
3060
3061         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3062
3063         if (ni) {
3064                 rc = 0;
3065                 lnet_ni_lock(ni);
3066                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3067                 lnet_ni_unlock(ni);
3068         }
3069
3070         lnet_net_unlock(cpt);
3071         return rc;
3072 }
3073
3074 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3075 {
3076         struct lnet_ni *ni;
3077         int cpt;
3078         int rc = -ENOENT;
3079
3080         if (!msg_stats)
3081                 return -EINVAL;
3082
3083         cpt = lnet_net_lock_current();
3084
3085         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3086
3087         if (ni) {
3088                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3089                 rc = 0;
3090         }
3091
3092         lnet_net_unlock(cpt);
3093
3094         return rc;
3095 }
3096
3097 static int lnet_add_net_common(struct lnet_net *net,
3098                                struct lnet_ioctl_config_lnd_tunables *tun)
3099 {
3100         __u32                   net_id;
3101         struct lnet_ping_buffer *pbuf;
3102         struct lnet_handle_md   ping_mdh;
3103         int                     rc;
3104         struct lnet_remotenet *rnet;
3105         int                     net_ni_count;
3106         int                     num_acceptor_nets;
3107
3108         lnet_net_lock(LNET_LOCK_EX);
3109         rnet = lnet_find_rnet_locked(net->net_id);
3110         lnet_net_unlock(LNET_LOCK_EX);
3111         /*
3112          * make sure that the net added doesn't invalidate the current
3113          * configuration LNet is keeping
3114          */
3115         if (rnet) {
3116                 CERROR("Adding net %s will invalidate routing configuration\n",
3117                        libcfs_net2str(net->net_id));
3118                 lnet_net_free(net);
3119                 return -EUSERS;
3120         }
3121
3122         /*
3123          * make sure you calculate the correct number of slots in the ping
3124          * buffer. Since the ping info is a flattened list of all the NIs,
3125          * we should allocate enough slots to accomodate the number of NIs
3126          * which will be added.
3127          *
3128          * since ni hasn't been configured yet, use
3129          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3130          */
3131         net_ni_count = lnet_get_net_ni_count_pre(net);
3132
3133         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3134                                     net_ni_count + lnet_get_ni_count(),
3135                                     false);
3136         if (rc < 0) {
3137                 lnet_net_free(net);
3138                 return rc;
3139         }
3140
3141         if (tun)
3142                 memcpy(&net->net_tunables,
3143                        &tun->lt_cmn, sizeof(net->net_tunables));
3144         else
3145                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3146
3147         /*
3148          * before starting this network get a count of the current TCP
3149          * networks which require the acceptor thread running. If that
3150          * count is == 0 before we start up this network, then we'd want to
3151          * start up the acceptor thread after starting up this network
3152          */
3153         num_acceptor_nets = lnet_count_acceptor_nets();
3154
3155         net_id = net->net_id;
3156
3157         rc = lnet_startup_lndnet(net,
3158                                  (tun) ? &tun->lt_tun : NULL);
3159         if (rc < 0)
3160                 goto failed;
3161
3162         lnet_net_lock(LNET_LOCK_EX);
3163         net = lnet_get_net_locked(net_id);
3164         lnet_net_unlock(LNET_LOCK_EX);
3165
3166         LASSERT(net);
3167
3168         /*
3169          * Start the acceptor thread if this is the first network
3170          * being added that requires the thread.
3171          */
3172         if (net->net_lnd->lnd_accept && num_acceptor_nets == 0) {
3173                 rc = lnet_acceptor_start();
3174                 if (rc < 0) {
3175                         /* shutdown the net that we just started */
3176                         CERROR("Failed to start up acceptor thread\n");
3177                         lnet_shutdown_lndnet(net);
3178                         goto failed;
3179                 }
3180         }
3181
3182         lnet_net_lock(LNET_LOCK_EX);
3183         lnet_peer_net_added(net);
3184         lnet_net_unlock(LNET_LOCK_EX);
3185
3186         lnet_ping_target_update(pbuf, ping_mdh);
3187
3188         return 0;
3189
3190 failed:
3191         lnet_ping_md_unlink(pbuf, &ping_mdh);
3192         lnet_ping_buffer_decref(pbuf);
3193         return rc;
3194 }
3195
3196 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3197                                       struct lnet_ioctl_config_lnd_tunables *tun)
3198 {
3199         struct lnet_net *net;
3200         char *nets;
3201         int rc;
3202         LIST_HEAD(net_head);
3203
3204         rc = lnet_parse_ip2nets(&nets, ip2nets);
3205         if (rc < 0)
3206                 return rc;
3207
3208         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3209         if (rc < 0)
3210                 return rc;
3211
3212         mutex_lock(&the_lnet.ln_api_mutex);
3213         while (!list_empty(&net_head)) {
3214                 net = list_entry(net_head.next, struct lnet_net, net_list);
3215                 list_del_init(&net->net_list);
3216                 rc = lnet_add_net_common(net, tun);
3217                 if (rc < 0)
3218                         goto out;
3219         }
3220
3221 out:
3222         mutex_unlock(&the_lnet.ln_api_mutex);
3223
3224         while (!list_empty(&net_head)) {
3225                 net = list_entry(net_head.next, struct lnet_net, net_list);
3226                 list_del_init(&net->net_list);
3227                 lnet_net_free(net);
3228         }
3229         return rc;
3230 }
3231
3232 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3233 {
3234         struct lnet_net *net;
3235         struct lnet_ni *ni;
3236         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3237         int rc, i;
3238         __u32 net_id, lnd_type;
3239
3240         /* get the tunables if they are available */
3241         if (conf->lic_cfg_hdr.ioc_len >=
3242             sizeof(*conf) + sizeof(*tun))
3243                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3244                         conf->lic_bulk;
3245
3246         /* handle legacy ip2nets from DLC */
3247         if (conf->lic_legacy_ip2nets[0] != '\0')
3248                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3249                                                   tun);
3250
3251         net_id = LNET_NIDNET(conf->lic_nid);
3252         lnd_type = LNET_NETTYP(net_id);
3253
3254         if (!libcfs_isknown_lnd(lnd_type)) {
3255                 CERROR("No valid net and lnd information provided\n");
3256                 return -EINVAL;
3257         }
3258
3259         net = lnet_net_alloc(net_id, NULL);
3260         if (!net)
3261                 return -ENOMEM;
3262
3263         for (i = 0; i < conf->lic_ncpts; i++) {
3264                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3265                         return -EINVAL;
3266         }
3267
3268         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3269                                        conf->lic_ni_intf[0]);
3270         if (!ni)
3271                 return -ENOMEM;
3272
3273         mutex_lock(&the_lnet.ln_api_mutex);
3274
3275         rc = lnet_add_net_common(net, tun);
3276
3277         mutex_unlock(&the_lnet.ln_api_mutex);
3278
3279         return rc;
3280 }
3281
3282 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3283 {
3284         struct lnet_net  *net;
3285         struct lnet_ni *ni;
3286         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3287         struct lnet_ping_buffer *pbuf;
3288         struct lnet_handle_md  ping_mdh;
3289         int               rc;
3290         int               net_count;
3291         __u32             addr;
3292
3293         /* don't allow userspace to shutdown the LOLND */
3294         if (LNET_NETTYP(net_id) == LOLND)
3295                 return -EINVAL;
3296
3297         mutex_lock(&the_lnet.ln_api_mutex);
3298
3299         lnet_net_lock(0);
3300
3301         net = lnet_get_net_locked(net_id);
3302         if (!net) {
3303                 CERROR("net %s not found\n",
3304                        libcfs_net2str(net_id));
3305                 rc = -ENOENT;
3306                 goto unlock_net;
3307         }
3308
3309         addr = LNET_NIDADDR(conf->lic_nid);
3310         if (addr == 0) {
3311                 /* remove the entire net */
3312                 net_count = lnet_get_net_ni_count_locked(net);
3313
3314                 lnet_net_unlock(0);
3315
3316                 /* create and link a new ping info, before removing the old one */
3317                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3318                                         lnet_get_ni_count() - net_count,
3319                                         false);
3320                 if (rc != 0)
3321                         goto unlock_api_mutex;
3322
3323                 lnet_shutdown_lndnet(net);
3324
3325                 if (lnet_count_acceptor_nets() == 0)
3326                         lnet_acceptor_stop();
3327
3328                 lnet_ping_target_update(pbuf, ping_mdh);
3329
3330                 goto unlock_api_mutex;
3331         }
3332
3333         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3334         if (!ni) {
3335                 CERROR("nid %s not found\n",
3336                        libcfs_nid2str(conf->lic_nid));
3337                 rc = -ENOENT;
3338                 goto unlock_net;
3339         }
3340
3341         net_count = lnet_get_net_ni_count_locked(net);
3342
3343         lnet_net_unlock(0);
3344
3345         /* create and link a new ping info, before removing the old one */
3346         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3347                                   lnet_get_ni_count() - 1, false);
3348         if (rc != 0)
3349                 goto unlock_api_mutex;
3350
3351         lnet_shutdown_lndni(ni);
3352
3353         if (lnet_count_acceptor_nets() == 0)
3354                 lnet_acceptor_stop();
3355
3356         lnet_ping_target_update(pbuf, ping_mdh);
3357
3358         /* check if the net is empty and remove it if it is */
3359         if (net_count == 1)
3360                 lnet_shutdown_lndnet(net);
3361
3362         goto unlock_api_mutex;
3363
3364 unlock_net:
3365         lnet_net_unlock(0);
3366 unlock_api_mutex:
3367         mutex_unlock(&the_lnet.ln_api_mutex);
3368
3369         return rc;
3370 }
3371
3372 /*
3373  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3374  * They are only expected to be called for unique networks.
3375  * That can be as a result of older DLC library
3376  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3377  */
3378 int
3379 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3380 {
3381         struct lnet_net *net;
3382         LIST_HEAD(net_head);
3383         int rc;
3384         struct lnet_ioctl_config_lnd_tunables tun;
3385         char *nets = conf->cfg_config_u.cfg_net.net_intf;
3386
3387         /* Create a net/ni structures for the network string */
3388         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3389         if (rc <= 0)
3390                 return rc == 0 ? -EINVAL : rc;
3391
3392         mutex_lock(&the_lnet.ln_api_mutex);
3393
3394         if (rc > 1) {
3395                 rc = -EINVAL; /* only add one network per call */
3396                 goto out_unlock_clean;
3397         }
3398
3399         net = list_entry(net_head.next, struct lnet_net, net_list);
3400         list_del_init(&net->net_list);
3401
3402         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3403
3404         memset(&tun, 0, sizeof(tun));
3405
3406         tun.lt_cmn.lct_peer_timeout =
3407           conf->cfg_config_u.cfg_net.net_peer_timeout;
3408         tun.lt_cmn.lct_peer_tx_credits =
3409           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3410         tun.lt_cmn.lct_peer_rtr_credits =
3411           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3412         tun.lt_cmn.lct_max_tx_credits =
3413           conf->cfg_config_u.cfg_net.net_max_tx_credits;
3414
3415         rc = lnet_add_net_common(net, &tun);
3416
3417 out_unlock_clean:
3418         mutex_unlock(&the_lnet.ln_api_mutex);
3419         while (!list_empty(&net_head)) {
3420                 /* net_head list is empty in success case */
3421                 net = list_entry(net_head.next, struct lnet_net, net_list);
3422                 list_del_init(&net->net_list);
3423                 lnet_net_free(net);
3424         }
3425         return rc;
3426 }
3427
3428 int
3429 lnet_dyn_del_net(__u32 net_id)
3430 {
3431         struct lnet_net  *net;
3432         struct lnet_ping_buffer *pbuf;
3433         struct lnet_handle_md ping_mdh;
3434         int               rc;
3435         int               net_ni_count;
3436
3437         /* don't allow userspace to shutdown the LOLND */
3438         if (LNET_NETTYP(net_id) == LOLND)
3439                 return -EINVAL;
3440
3441         mutex_lock(&the_lnet.ln_api_mutex);
3442
3443         lnet_net_lock(0);
3444
3445         net = lnet_get_net_locked(net_id);
3446         if (net == NULL) {
3447                 lnet_net_unlock(0);
3448                 rc = -EINVAL;
3449                 goto out;
3450         }
3451
3452         net_ni_count = lnet_get_net_ni_count_locked(net);
3453
3454         lnet_net_unlock(0);
3455
3456         /* create and link a new ping info, before removing the old one */
3457         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3458                                     lnet_get_ni_count() - net_ni_count, false);
3459         if (rc != 0)
3460                 goto out;
3461
3462         lnet_shutdown_lndnet(net);
3463
3464         if (lnet_count_acceptor_nets() == 0)
3465                 lnet_acceptor_stop();
3466
3467         lnet_ping_target_update(pbuf, ping_mdh);
3468
3469 out:
3470         mutex_unlock(&the_lnet.ln_api_mutex);
3471
3472         return rc;
3473 }
3474
3475 void lnet_incr_dlc_seq(void)
3476 {
3477         atomic_inc(&lnet_dlc_seq_no);
3478 }
3479
3480 __u32 lnet_get_dlc_seq_locked(void)
3481 {
3482         return atomic_read(&lnet_dlc_seq_no);
3483 }
3484
3485 static void
3486 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3487 {
3488         struct lnet_net *net;
3489         struct lnet_ni *ni;
3490
3491         lnet_net_lock(LNET_LOCK_EX);
3492         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3493                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3494                         if (ni->ni_nid == nid || all) {
3495                                 atomic_set(&ni->ni_healthv, value);
3496                                 if (list_empty(&ni->ni_recovery) &&
3497                                     value < LNET_MAX_HEALTH_VALUE) {
3498                                         CERROR("manually adding local NI %s to recovery\n",
3499                                                libcfs_nid2str(ni->ni_nid));
3500                                         list_add_tail(&ni->ni_recovery,
3501                                                       &the_lnet.ln_mt_localNIRecovq);
3502                                         lnet_ni_addref_locked(ni, 0);
3503                                 }
3504                                 if (!all) {
3505                                         lnet_net_unlock(LNET_LOCK_EX);
3506                                         return;
3507                                 }
3508                         }
3509                 }
3510         }
3511         lnet_net_unlock(LNET_LOCK_EX);
3512 }
3513
3514 static int
3515 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3516 {