Whamcloud - gitweb
1b7cb7eef1f3fd1630ae6c386bf57df33a2235af
[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/udsp.h>
45 #include <lnet/lib-lnet.h>
46
47 #define D_LNI D_CONSOLE
48
49 /*
50  * initialize ln_api_mutex statically, since it needs to be used in
51  * discovery_set callback. That module parameter callback can be called
52  * before module init completes. The mutex needs to be ready for use then.
53  */
54 struct lnet the_lnet = {
55         .ln_api_mutex = __MUTEX_INITIALIZER(the_lnet.ln_api_mutex),
56 };              /* THE state of the network */
57 EXPORT_SYMBOL(the_lnet);
58
59 static char *ip2nets = "";
60 module_param(ip2nets, charp, 0444);
61 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
62
63 static char *networks = "";
64 module_param(networks, charp, 0444);
65 MODULE_PARM_DESC(networks, "local networks");
66
67 static char *routes = "";
68 module_param(routes, charp, 0444);
69 MODULE_PARM_DESC(routes, "routes to non-local networks");
70
71 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
72 module_param(rnet_htable_size, int, 0444);
73 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
74
75 static int use_tcp_bonding = false;
76 module_param(use_tcp_bonding, int, 0444);
77 MODULE_PARM_DESC(use_tcp_bonding,
78                  "use_tcp_bonding parameter has been deprecated");
79
80 unsigned int lnet_numa_range = 0;
81 module_param(lnet_numa_range, uint, 0444);
82 MODULE_PARM_DESC(lnet_numa_range,
83                 "NUMA range to consider during Multi-Rail selection");
84
85 /*
86  * lnet_health_sensitivity determines by how much we decrement the health
87  * value on sending error. The value defaults to 100, which means health
88  * interface health is decremented by 100 points every failure.
89  */
90 unsigned int lnet_health_sensitivity = 100;
91 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
92 #ifdef HAVE_KERNEL_PARAM_OPS
93 static struct kernel_param_ops param_ops_health_sensitivity = {
94         .set = sensitivity_set,
95         .get = param_get_int,
96 };
97 #define param_check_health_sensitivity(name, p) \
98                 __param_check(name, p, int)
99 module_param(lnet_health_sensitivity, health_sensitivity, S_IRUGO|S_IWUSR);
100 #else
101 module_param_call(lnet_health_sensitivity, sensitivity_set, param_get_int,
102                   &lnet_health_sensitivity, S_IRUGO|S_IWUSR);
103 #endif
104 MODULE_PARM_DESC(lnet_health_sensitivity,
105                 "Value to decrement the health value by on error");
106
107 /*
108  * lnet_recovery_interval determines how often we should perform recovery
109  * on unhealthy interfaces.
110  */
111 unsigned int lnet_recovery_interval = 1;
112 static int recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp);
113 #ifdef HAVE_KERNEL_PARAM_OPS
114 static struct kernel_param_ops param_ops_recovery_interval = {
115         .set = recovery_interval_set,
116         .get = param_get_int,
117 };
118 #define param_check_recovery_interval(name, p) \
119                 __param_check(name, p, int)
120 module_param(lnet_recovery_interval, recovery_interval, S_IRUGO|S_IWUSR);
121 #else
122 module_param_call(lnet_recovery_interval, recovery_interval_set, param_get_int,
123                   &lnet_recovery_interval, S_IRUGO|S_IWUSR);
124 #endif
125 MODULE_PARM_DESC(lnet_recovery_interval,
126                 "Interval to recover unhealthy interfaces in seconds");
127
128 unsigned int lnet_recovery_limit;
129 module_param(lnet_recovery_limit, uint, 0644);
130 MODULE_PARM_DESC(lnet_recovery_limit,
131                  "How long to attempt recovery of unhealthy peer interfaces in seconds. Set to 0 to allow indefinite recovery");
132
133 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
134 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
135
136 static struct kernel_param_ops param_ops_interfaces_max = {
137         .set = intf_max_set,
138         .get = param_get_int,
139 };
140
141 #define param_check_interfaces_max(name, p) \
142                 __param_check(name, p, int)
143
144 #ifdef HAVE_KERNEL_PARAM_OPS
145 module_param(lnet_interfaces_max, interfaces_max, 0644);
146 #else
147 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
148                   &param_ops_interfaces_max, 0644);
149 #endif
150 MODULE_PARM_DESC(lnet_interfaces_max,
151                 "Maximum number of interfaces in a node.");
152
153 unsigned lnet_peer_discovery_disabled = 0;
154 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
155
156 static struct kernel_param_ops param_ops_discovery_disabled = {
157         .set = discovery_set,
158         .get = param_get_int,
159 };
160
161 #define param_check_discovery_disabled(name, p) \
162                 __param_check(name, p, int)
163 #ifdef HAVE_KERNEL_PARAM_OPS
164 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
165 #else
166 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
167                   &param_ops_discovery_disabled, 0644);
168 #endif
169 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
170                 "Set to 1 to disable peer discovery on this node.");
171
172 unsigned int lnet_drop_asym_route;
173 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
174
175 static struct kernel_param_ops param_ops_drop_asym_route = {
176         .set = drop_asym_route_set,
177         .get = param_get_int,
178 };
179
180 #define param_check_drop_asym_route(name, p)    \
181         __param_check(name, p, int)
182 #ifdef HAVE_KERNEL_PARAM_OPS
183 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
184 #else
185 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
186                   &param_ops_drop_asym_route, 0644);
187 #endif
188 MODULE_PARM_DESC(lnet_drop_asym_route,
189                  "Set to 1 to drop asymmetrical route messages.");
190
191 #define LNET_TRANSACTION_TIMEOUT_DEFAULT 50
192 unsigned int lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_DEFAULT;
193 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
194 #ifdef HAVE_KERNEL_PARAM_OPS
195 static struct kernel_param_ops param_ops_transaction_timeout = {
196         .set = transaction_to_set,
197         .get = param_get_int,
198 };
199
200 #define param_check_transaction_timeout(name, p) \
201                 __param_check(name, p, int)
202 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
203 #else
204 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
205                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
206 #endif
207 MODULE_PARM_DESC(lnet_transaction_timeout,
208                 "Maximum number of seconds to wait for a peer response.");
209
210 #define LNET_RETRY_COUNT_DEFAULT 2
211 unsigned int lnet_retry_count = LNET_RETRY_COUNT_DEFAULT;
212 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
213 #ifdef HAVE_KERNEL_PARAM_OPS
214 static struct kernel_param_ops param_ops_retry_count = {
215         .set = retry_count_set,
216         .get = param_get_int,
217 };
218
219 #define param_check_retry_count(name, p) \
220                 __param_check(name, p, int)
221 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
222 #else
223 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
224                   &lnet_retry_count, S_IRUGO|S_IWUSR);
225 #endif
226 MODULE_PARM_DESC(lnet_retry_count,
227                  "Maximum number of times to retry transmitting a message");
228
229 unsigned int lnet_response_tracking = 3;
230 static int response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp);
231
232 #ifdef HAVE_KERNEL_PARAM_OPS
233 static struct kernel_param_ops param_ops_response_tracking = {
234         .set = response_tracking_set,
235         .get = param_get_int,
236 };
237
238 #define param_check_response_tracking(name, p)  \
239         __param_check(name, p, int)
240 module_param(lnet_response_tracking, response_tracking, 0644);
241 #else
242 module_param_call(lnet_response_tracking, response_tracking_set, param_get_int,
243                   &lnet_response_tracking, 0644);
244 #endif
245 MODULE_PARM_DESC(lnet_response_tracking,
246                  "(0|1|2|3) LNet Internal Only|GET Reply only|PUT ACK only|Full Tracking (default)");
247
248 #define LNET_LND_TIMEOUT_DEFAULT ((LNET_TRANSACTION_TIMEOUT_DEFAULT - 1) / \
249                                   (LNET_RETRY_COUNT_DEFAULT + 1))
250 unsigned int lnet_lnd_timeout = LNET_LND_TIMEOUT_DEFAULT;
251 static void lnet_set_lnd_timeout(void)
252 {
253         lnet_lnd_timeout = (lnet_transaction_timeout - 1) /
254                            (lnet_retry_count + 1);
255 }
256
257 unsigned int lnet_current_net_count;
258
259 /*
260  * This sequence number keeps track of how many times DLC was used to
261  * update the local NIs. It is incremented when a NI is added or
262  * removed and checked when sending a message to determine if there is
263  * a need to re-run the selection algorithm. See lnet_select_pathway()
264  * for more details on its usage.
265  */
266 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
267
268 static int lnet_ping(struct lnet_process_id id, signed long timeout,
269                      struct lnet_process_id __user *ids, int n_ids);
270
271 static int lnet_discover(struct lnet_process_id id, __u32 force,
272                          struct lnet_process_id __user *ids, int n_ids);
273
274 static int
275 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
276 {
277         int rc;
278         unsigned *sensitivity = (unsigned *)kp->arg;
279         unsigned long value;
280
281         rc = kstrtoul(val, 0, &value);
282         if (rc) {
283                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
284                 return rc;
285         }
286
287         /*
288          * The purpose of locking the api_mutex here is to ensure that
289          * the correct value ends up stored properly.
290          */
291         mutex_lock(&the_lnet.ln_api_mutex);
292
293         if (value > LNET_MAX_HEALTH_VALUE) {
294                 mutex_unlock(&the_lnet.ln_api_mutex);
295                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
296                        LNET_MAX_HEALTH_VALUE, value);
297                 return -EINVAL;
298         }
299
300         if (*sensitivity != 0 && value == 0 && lnet_retry_count != 0) {
301                 lnet_retry_count = 0;
302                 lnet_set_lnd_timeout();
303         }
304
305         *sensitivity = value;
306
307         mutex_unlock(&the_lnet.ln_api_mutex);
308
309         return 0;
310 }
311
312 static int
313 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
314 {
315         int rc;
316         unsigned *interval = (unsigned *)kp->arg;
317         unsigned long value;
318
319         rc = kstrtoul(val, 0, &value);
320         if (rc) {
321                 CERROR("Invalid module parameter value for 'lnet_recovery_interval'\n");
322                 return rc;
323         }
324
325         if (value < 1) {
326                 CERROR("lnet_recovery_interval must be at least 1 second\n");
327                 return -EINVAL;
328         }
329
330         /*
331          * The purpose of locking the api_mutex here is to ensure that
332          * the correct value ends up stored properly.
333          */
334         mutex_lock(&the_lnet.ln_api_mutex);
335
336         *interval = value;
337
338         mutex_unlock(&the_lnet.ln_api_mutex);
339
340         return 0;
341 }
342
343 static int
344 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
345 {
346         int rc;
347         unsigned *discovery_off = (unsigned *)kp->arg;
348         unsigned long value;
349         struct lnet_ping_buffer *pbuf;
350
351         rc = kstrtoul(val, 0, &value);
352         if (rc) {
353                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
354                 return rc;
355         }
356
357         value = (value) ? 1 : 0;
358
359         /*
360          * The purpose of locking the api_mutex here is to ensure that
361          * the correct value ends up stored properly.
362          */
363         mutex_lock(&the_lnet.ln_api_mutex);
364
365         if (value == *discovery_off) {
366                 mutex_unlock(&the_lnet.ln_api_mutex);
367                 return 0;
368         }
369
370         /*
371          * We still want to set the discovery value even when LNet is not
372          * running. This is the case when LNet is being loaded and we want
373          * the module parameters to take effect. Otherwise if we're
374          * changing the value dynamically, we want to set it after
375          * updating the peers
376          */
377         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
378                 *discovery_off = value;
379                 mutex_unlock(&the_lnet.ln_api_mutex);
380                 return 0;
381         }
382
383         /* tell peers that discovery setting has changed */
384         lnet_net_lock(LNET_LOCK_EX);
385         pbuf = the_lnet.ln_ping_target;
386         if (value)
387                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
388         else
389                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
390         lnet_net_unlock(LNET_LOCK_EX);
391
392         /* only send a push when we're turning off discovery */
393         if (*discovery_off <= 0 && value > 0)
394                 lnet_push_update_to_peers(1);
395         *discovery_off = value;
396
397         mutex_unlock(&the_lnet.ln_api_mutex);
398
399         return 0;
400 }
401
402 static int
403 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
404 {
405         int rc;
406         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
407         unsigned long value;
408
409         rc = kstrtoul(val, 0, &value);
410         if (rc) {
411                 CERROR("Invalid module parameter value for "
412                        "'lnet_drop_asym_route'\n");
413                 return rc;
414         }
415
416         /*
417          * The purpose of locking the api_mutex here is to ensure that
418          * the correct value ends up stored properly.
419          */
420         mutex_lock(&the_lnet.ln_api_mutex);
421
422         if (value == *drop_asym_route) {
423                 mutex_unlock(&the_lnet.ln_api_mutex);
424                 return 0;
425         }
426
427         *drop_asym_route = value;
428
429         mutex_unlock(&the_lnet.ln_api_mutex);
430
431         return 0;
432 }
433
434 static int
435 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
436 {
437         int rc;
438         unsigned *transaction_to = (unsigned *)kp->arg;
439         unsigned long value;
440
441         rc = kstrtoul(val, 0, &value);
442         if (rc) {
443                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
444                 return rc;
445         }
446
447         /*
448          * The purpose of locking the api_mutex here is to ensure that
449          * the correct value ends up stored properly.
450          */
451         mutex_lock(&the_lnet.ln_api_mutex);
452
453         if (value <= lnet_retry_count || value == 0) {
454                 mutex_unlock(&the_lnet.ln_api_mutex);
455                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
456                        "Has to be greater than lnet_retry_count (%u)\n",
457                        value, lnet_retry_count);
458                 return -EINVAL;
459         }
460
461         if (value == *transaction_to) {
462                 mutex_unlock(&the_lnet.ln_api_mutex);
463                 return 0;
464         }
465
466         *transaction_to = value;
467         /* Update the lnet_lnd_timeout now that we've modified the
468          * transaction timeout
469          */
470         lnet_set_lnd_timeout();
471
472         mutex_unlock(&the_lnet.ln_api_mutex);
473
474         return 0;
475 }
476
477 static int
478 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
479 {
480         int rc;
481         unsigned *retry_count = (unsigned *)kp->arg;
482         unsigned long value;
483
484         rc = kstrtoul(val, 0, &value);
485         if (rc) {
486                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
487                 return rc;
488         }
489
490         /*
491          * The purpose of locking the api_mutex here is to ensure that
492          * the correct value ends up stored properly.
493          */
494         mutex_lock(&the_lnet.ln_api_mutex);
495
496         if (lnet_health_sensitivity == 0 && value > 0) {
497                 mutex_unlock(&the_lnet.ln_api_mutex);
498                 CERROR("Can not set lnet_retry_count when health feature is turned off\n");
499                 return -EINVAL;
500         }
501
502         if (value > lnet_transaction_timeout) {
503                 mutex_unlock(&the_lnet.ln_api_mutex);
504                 CERROR("Invalid value for lnet_retry_count (%lu). "
505                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
506                        value, lnet_transaction_timeout);
507                 return -EINVAL;
508         }
509
510         *retry_count = value;
511
512         /* Update the lnet_lnd_timeout now that we've modified the
513          * retry count
514          */
515         lnet_set_lnd_timeout();
516
517         mutex_unlock(&the_lnet.ln_api_mutex);
518
519         return 0;
520 }
521
522 static int
523 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
524 {
525         int value, rc;
526
527         rc = kstrtoint(val, 0, &value);
528         if (rc) {
529                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
530                 return rc;
531         }
532
533         if (value < LNET_INTERFACES_MIN) {
534                 CWARN("max interfaces provided are too small, setting to %d\n",
535                       LNET_INTERFACES_MAX_DEFAULT);
536                 value = LNET_INTERFACES_MAX_DEFAULT;
537         }
538
539         *(int *)kp->arg = value;
540
541         return 0;
542 }
543
544 static int
545 response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp)
546 {
547         int rc;
548         unsigned long new_value;
549
550         rc = kstrtoul(val, 0, &new_value);
551         if (rc) {
552                 CERROR("Invalid value for 'lnet_response_tracking'\n");
553                 return -EINVAL;
554         }
555
556         if (new_value < 0 || new_value > 3) {
557                 CWARN("Invalid value (%lu) for 'lnet_response_tracking'\n",
558                       new_value);
559                 return -EINVAL;
560         }
561
562         lnet_response_tracking = new_value;
563
564         return 0;
565 }
566
567 static const char *
568 lnet_get_routes(void)
569 {
570         return routes;
571 }
572
573 static const char *
574 lnet_get_networks(void)
575 {
576         const char *nets;
577         int rc;
578
579         if (*networks != 0 && *ip2nets != 0) {
580                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
581                                    "'ip2nets' but not both at once\n");
582                 return NULL;
583         }
584
585         if (*ip2nets != 0) {
586                 rc = lnet_parse_ip2nets(&nets, ip2nets);
587                 return (rc == 0) ? nets : NULL;
588         }
589
590         if (*networks != 0)
591                 return networks;
592
593         return "tcp";
594 }
595
596 static void
597 lnet_init_locks(void)
598 {
599         spin_lock_init(&the_lnet.ln_eq_wait_lock);
600         spin_lock_init(&the_lnet.ln_msg_resend_lock);
601         init_completion(&the_lnet.ln_mt_wait_complete);
602         mutex_init(&the_lnet.ln_lnd_mutex);
603 }
604
605 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
606 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
607                                             *  MDs kmem_cache */
608 struct kmem_cache *lnet_udsp_cachep;       /* udsp cache */
609 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
610 struct kmem_cache *lnet_msg_cachep;
611
612 static int
613 lnet_slab_setup(void)
614 {
615         /* create specific kmem_cache for MEs and small MDs (i.e., originally
616          * allocated in <size-xxx> kmem_cache).
617          */
618         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
619                                             0, 0, NULL);
620         if (!lnet_mes_cachep)
621                 return -ENOMEM;
622
623         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
624                                                   LNET_SMALL_MD_SIZE, 0, 0,
625                                                   NULL);
626         if (!lnet_small_mds_cachep)
627                 return -ENOMEM;
628
629         lnet_udsp_cachep = kmem_cache_create("lnet_udsp",
630                                              sizeof(struct lnet_udsp),
631                                              0, 0, NULL);
632         if (!lnet_udsp_cachep)
633                 return -ENOMEM;
634
635         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
636                                             0, 0, NULL);
637         if (!lnet_rspt_cachep)
638                 return -ENOMEM;
639
640         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
641                                             0, 0, NULL);
642         if (!lnet_msg_cachep)
643                 return -ENOMEM;
644
645         return 0;
646 }
647
648 static void
649 lnet_slab_cleanup(void)
650 {
651         if (lnet_msg_cachep) {
652                 kmem_cache_destroy(lnet_msg_cachep);
653                 lnet_msg_cachep = NULL;
654         }
655
656         if (lnet_rspt_cachep) {
657                 kmem_cache_destroy(lnet_rspt_cachep);
658                 lnet_rspt_cachep = NULL;
659         }
660
661         if (lnet_udsp_cachep) {
662                 kmem_cache_destroy(lnet_udsp_cachep);
663                 lnet_udsp_cachep = NULL;
664         }
665
666         if (lnet_small_mds_cachep) {
667                 kmem_cache_destroy(lnet_small_mds_cachep);
668                 lnet_small_mds_cachep = NULL;
669         }
670
671         if (lnet_mes_cachep) {
672                 kmem_cache_destroy(lnet_mes_cachep);
673                 lnet_mes_cachep = NULL;
674         }
675 }
676
677 static int
678 lnet_create_remote_nets_table(void)
679 {
680         int               i;
681         struct list_head *hash;
682
683         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
684         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
685         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
686         if (hash == NULL) {
687                 CERROR("Failed to create remote nets hash table\n");
688                 return -ENOMEM;
689         }
690
691         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
692                 INIT_LIST_HEAD(&hash[i]);
693         the_lnet.ln_remote_nets_hash = hash;
694         return 0;
695 }
696
697 static void
698 lnet_destroy_remote_nets_table(void)
699 {
700         int i;
701
702         if (the_lnet.ln_remote_nets_hash == NULL)
703                 return;
704
705         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
706                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
707
708         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
709                            LNET_REMOTE_NETS_HASH_SIZE);
710         the_lnet.ln_remote_nets_hash = NULL;
711 }
712
713 static void
714 lnet_destroy_locks(void)
715 {
716         if (the_lnet.ln_res_lock != NULL) {
717                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
718                 the_lnet.ln_res_lock = NULL;
719         }
720
721         if (the_lnet.ln_net_lock != NULL) {
722                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
723                 the_lnet.ln_net_lock = NULL;
724         }
725 }
726
727 static int
728 lnet_create_locks(void)
729 {
730         lnet_init_locks();
731
732         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
733         if (the_lnet.ln_res_lock == NULL)
734                 goto failed;
735
736         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
737         if (the_lnet.ln_net_lock == NULL)
738                 goto failed;
739
740         return 0;
741
742  failed:
743         lnet_destroy_locks();
744         return -ENOMEM;
745 }
746
747 static void lnet_assert_wire_constants(void)
748 {
749         /* Wire protocol assertions generated by 'wirecheck'
750          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
751          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
752          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7)
753          */
754
755         /* Constants... */
756         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
757         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
758         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
759         BUILD_BUG_ON(LNET_MSG_ACK != 0);
760         BUILD_BUG_ON(LNET_MSG_PUT != 1);
761         BUILD_BUG_ON(LNET_MSG_GET != 2);
762         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
763         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
764
765         BUILD_BUG_ON((int)sizeof(lnet_nid_t) != 8);
766         BUILD_BUG_ON((int)sizeof(lnet_pid_t) != 4);
767
768         /* Checks for struct lnet_process_id_packed */
769         BUILD_BUG_ON((int)sizeof(struct lnet_process_id_packed) != 12);
770         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, nid) != 0);
771         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->nid) != 8);
772         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, pid) != 8);
773         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->pid) != 4);
774
775         /* Checks for struct lnet_handle_wire */
776         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
777         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
778                                    wh_interface_cookie) != 0);
779         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
780         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
781                                    wh_object_cookie) != 8);
782         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
783
784         /* Checks for struct struct lnet_magicversion */
785         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
786         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
787         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
788         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
789         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
790         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
791                                    version_minor) != 6);
792         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
793
794         /* Checks for struct struct lnet_hdr */
795         BUILD_BUG_ON((int)sizeof(struct lnet_hdr) != 72);
796         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_nid) != 0);
797         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_nid) != 8);
798         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_nid) != 8);
799         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_nid) != 8);
800         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_pid) != 16);
801         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_pid) != 4);
802         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_pid) != 20);
803         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_pid) != 4);
804         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, type) != 24);
805         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->type) != 4);
806         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, payload_length) != 28);
807         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->payload_length) != 4);
808         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg) != 32);
809         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg) != 40);
810
811         /* Ack */
812         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) != 32);
813         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) != 16);
814         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.match_bits) != 48);
815         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) != 8);
816         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.mlength) != 56);
817         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) != 4);
818
819         /* Put */
820         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) != 32);
821         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) != 16);
822         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.match_bits) != 48);
823         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) != 8);
824         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.hdr_data) != 56);
825         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) != 8);
826         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ptl_index) != 64);
827         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) != 4);
828         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.offset) != 68);
829         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) != 4);
830
831         /* Get */
832         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.return_wmd) != 32);
833         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) != 16);
834         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.match_bits) != 48);
835         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) != 8);
836         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.ptl_index) != 56);
837         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) != 4);
838         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.src_offset) != 60);
839         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) != 4);
840         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.sink_length) != 64);
841         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) != 4);
842
843         /* Reply */
844         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) != 32);
845         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) != 16);
846
847         /* Hello */
848         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.incarnation) != 32);
849         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) != 8);
850         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.type) != 40);
851         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) != 4);
852
853         /* Checks for struct lnet_ni_status and related constants */
854         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
855         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
856         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
857
858         /* Checks for struct lnet_ni_status */
859         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
860         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
861         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
862         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
863         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
864         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_unused) != 12);
865         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) != 4);
866
867         /* Checks for struct lnet_ping_info and related constants */
868         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
869         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
870         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
871         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
872         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
873         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
874         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
875         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 31);
876
877         /* Checks for struct lnet_ping_info */
878         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
879         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
880         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
881         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
882         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
883         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
884         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
885         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
886         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
887         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
888         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) != 0);
889
890         /* Acceptor connection request */
891         BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
892
893         /* Checks for struct lnet_acceptor_connreq */
894         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq) != 16);
895         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_magic) != 0);
896         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_magic) != 4);
897         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_version) != 4);
898         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_version) != 4);
899         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_nid) != 8);
900         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_nid) != 8);
901
902         /* Checks for struct lnet_counters_common */
903         BUILD_BUG_ON((int)sizeof(struct lnet_counters_common) != 60);
904         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_alloc) != 0);
905         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_alloc) != 4);
906         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_max) != 4);
907         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_max) != 4);
908         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_errors) != 8);
909         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_errors) != 4);
910         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_count) != 12);
911         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_count) != 4);
912         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_count) != 16);
913         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_count) != 4);
914         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_count) != 20);
915         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_count) != 4);
916         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_count) != 24);
917         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_count) != 4);
918         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_length) != 28);
919         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_length) != 8);
920         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_length) != 36);
921         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_length) != 8);
922         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_length) != 44);
923         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_length) != 8);
924         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_length) != 52);
925         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_length) != 8);
926 }
927
928 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
929 {
930         const struct lnet_lnd *lnd;
931
932         /* holding lnd mutex */
933         if (type >= NUM_LNDS)
934                 return NULL;
935         lnd = the_lnet.ln_lnds[type];
936         LASSERT(!lnd || lnd->lnd_type == type);
937
938         return lnd;
939 }
940
941 unsigned int
942 lnet_get_lnd_timeout(void)
943 {
944         return lnet_lnd_timeout;
945 }
946 EXPORT_SYMBOL(lnet_get_lnd_timeout);
947
948 void
949 lnet_register_lnd(const struct lnet_lnd *lnd)
950 {
951         mutex_lock(&the_lnet.ln_lnd_mutex);
952
953         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
954         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
955
956         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
957
958         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
959
960         mutex_unlock(&the_lnet.ln_lnd_mutex);
961 }
962 EXPORT_SYMBOL(lnet_register_lnd);
963
964 void
965 lnet_unregister_lnd(const struct lnet_lnd *lnd)
966 {
967         mutex_lock(&the_lnet.ln_lnd_mutex);
968
969         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
970
971         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
972         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
973
974         mutex_unlock(&the_lnet.ln_lnd_mutex);
975 }
976 EXPORT_SYMBOL(lnet_unregister_lnd);
977
978 static void
979 lnet_counters_get_common_locked(struct lnet_counters_common *common)
980 {
981         struct lnet_counters *ctr;
982         int i;
983
984         /* FIXME !!! Their is no assert_lnet_net_locked() to ensure this
985          * actually called under the protection of the lnet_net_lock.
986          */
987         memset(common, 0, sizeof(*common));
988
989         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
990                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
991                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
992                 common->lcc_errors       += ctr->lct_common.lcc_errors;
993                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
994                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
995                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
996                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
997                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
998                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
999                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
1000                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
1001         }
1002 }
1003
1004 void
1005 lnet_counters_get_common(struct lnet_counters_common *common)
1006 {
1007         lnet_net_lock(LNET_LOCK_EX);
1008         lnet_counters_get_common_locked(common);
1009         lnet_net_unlock(LNET_LOCK_EX);
1010 }
1011 EXPORT_SYMBOL(lnet_counters_get_common);
1012
1013 int
1014 lnet_counters_get(struct lnet_counters *counters)
1015 {
1016         struct lnet_counters *ctr;
1017         struct lnet_counters_health *health = &counters->lct_health;
1018         int i, rc = 0;
1019
1020         memset(counters, 0, sizeof(*counters));
1021
1022         lnet_net_lock(LNET_LOCK_EX);
1023
1024         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1025                 GOTO(out_unlock, rc = -ENODEV);
1026
1027         lnet_counters_get_common_locked(&counters->lct_common);
1028
1029         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1030                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
1031                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
1032                 health->lch_response_timeout_count +=
1033                                 ctr->lct_health.lch_response_timeout_count;
1034                 health->lch_local_interrupt_count +=
1035                                 ctr->lct_health.lch_local_interrupt_count;
1036                 health->lch_local_dropped_count +=
1037                                 ctr->lct_health.lch_local_dropped_count;
1038                 health->lch_local_aborted_count +=
1039                                 ctr->lct_health.lch_local_aborted_count;
1040                 health->lch_local_no_route_count +=
1041                                 ctr->lct_health.lch_local_no_route_count;
1042                 health->lch_local_timeout_count +=
1043                                 ctr->lct_health.lch_local_timeout_count;
1044                 health->lch_local_error_count +=
1045                                 ctr->lct_health.lch_local_error_count;
1046                 health->lch_remote_dropped_count +=
1047                                 ctr->lct_health.lch_remote_dropped_count;
1048                 health->lch_remote_error_count +=
1049                                 ctr->lct_health.lch_remote_error_count;
1050                 health->lch_remote_timeout_count +=
1051                                 ctr->lct_health.lch_remote_timeout_count;
1052                 health->lch_network_timeout_count +=
1053                                 ctr->lct_health.lch_network_timeout_count;
1054         }
1055 out_unlock:
1056         lnet_net_unlock(LNET_LOCK_EX);
1057         return rc;
1058 }
1059 EXPORT_SYMBOL(lnet_counters_get);
1060
1061 void
1062 lnet_counters_reset(void)
1063 {
1064         struct lnet_counters *counters;
1065         int             i;
1066
1067         lnet_net_lock(LNET_LOCK_EX);
1068
1069         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1070                 goto avoid_reset;
1071
1072         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
1073                 memset(counters, 0, sizeof(struct lnet_counters));
1074 avoid_reset:
1075         lnet_net_unlock(LNET_LOCK_EX);
1076 }
1077
1078 static char *
1079 lnet_res_type2str(int type)
1080 {
1081         switch (type) {
1082         default:
1083                 LBUG();
1084         case LNET_COOKIE_TYPE_MD:
1085                 return "MD";
1086         case LNET_COOKIE_TYPE_ME:
1087                 return "ME";
1088         case LNET_COOKIE_TYPE_EQ:
1089                 return "EQ";
1090         }
1091 }
1092
1093 static void
1094 lnet_res_container_cleanup(struct lnet_res_container *rec)
1095 {
1096         int     count = 0;
1097
1098         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
1099                 return;
1100
1101         while (!list_empty(&rec->rec_active)) {
1102                 struct list_head *e = rec->rec_active.next;
1103
1104                 list_del_init(e);
1105                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
1106                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
1107
1108                 } else { /* NB: Active MEs should be attached on portals */
1109                         LBUG();
1110                 }
1111                 count++;
1112         }
1113
1114         if (count > 0) {
1115                 /* Found alive MD/ME/EQ, user really should unlink/free
1116                  * all of them before finalize LNet, but if someone didn't,
1117                  * we have to recycle garbage for him */
1118                 CERROR("%d active elements on exit of %s container\n",
1119                        count, lnet_res_type2str(rec->rec_type));
1120         }
1121
1122         if (rec->rec_lh_hash != NULL) {
1123                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
1124                 rec->rec_lh_hash = NULL;
1125         }
1126
1127         rec->rec_type = 0; /* mark it as finalized */
1128 }
1129
1130 static int
1131 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1132 {
1133         int     rc = 0;
1134         int     i;
1135
1136         LASSERT(rec->rec_type == 0);
1137
1138         rec->rec_type = type;
1139         INIT_LIST_HEAD(&rec->rec_active);
1140
1141         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1142
1143         /* Arbitrary choice of hash table size */
1144         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1145                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1146         if (rec->rec_lh_hash == NULL) {
1147                 rc = -ENOMEM;
1148                 goto out;
1149         }
1150
1151         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1152                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1153
1154         return 0;
1155
1156 out:
1157         CERROR("Failed to setup %s resource container\n",
1158                lnet_res_type2str(type));
1159         lnet_res_container_cleanup(rec);
1160         return rc;
1161 }
1162
1163 static void
1164 lnet_res_containers_destroy(struct lnet_res_container **recs)
1165 {
1166         struct lnet_res_container       *rec;
1167         int                             i;
1168
1169         cfs_percpt_for_each(rec, i, recs)
1170                 lnet_res_container_cleanup(rec);
1171
1172         cfs_percpt_free(recs);
1173 }
1174
1175 static struct lnet_res_container **
1176 lnet_res_containers_create(int type)
1177 {
1178         struct lnet_res_container       **recs;
1179         struct lnet_res_container       *rec;
1180         int                             rc;
1181         int                             i;
1182
1183         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1184         if (recs == NULL) {
1185                 CERROR("Failed to allocate %s resource containers\n",
1186                        lnet_res_type2str(type));
1187                 return NULL;
1188         }
1189
1190         cfs_percpt_for_each(rec, i, recs) {
1191                 rc = lnet_res_container_setup(rec, i, type);
1192                 if (rc != 0) {
1193                         lnet_res_containers_destroy(recs);
1194                         return NULL;
1195                 }
1196         }
1197
1198         return recs;
1199 }
1200
1201 struct lnet_libhandle *
1202 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1203 {
1204         /* ALWAYS called with lnet_res_lock held */
1205         struct list_head        *head;
1206         struct lnet_libhandle   *lh;
1207         unsigned int            hash;
1208
1209         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1210                 return NULL;
1211
1212         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1213         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1214
1215         list_for_each_entry(lh, head, lh_hash_chain) {
1216                 if (lh->lh_cookie == cookie)
1217                         return lh;
1218         }
1219
1220         return NULL;
1221 }
1222
1223 void
1224 lnet_res_lh_initialize(struct lnet_res_container *rec,
1225                        struct lnet_libhandle *lh)
1226 {
1227         /* ALWAYS called with lnet_res_lock held */
1228         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1229         unsigned int    hash;
1230
1231         lh->lh_cookie = rec->rec_lh_cookie;
1232         rec->rec_lh_cookie += 1 << ibits;
1233
1234         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1235
1236         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1237 }
1238
1239 struct list_head **
1240 lnet_create_array_of_queues(void)
1241 {
1242         struct list_head **qs;
1243         struct list_head *q;
1244         int i;
1245
1246         qs = cfs_percpt_alloc(lnet_cpt_table(),
1247                               sizeof(struct list_head));
1248         if (!qs) {
1249                 CERROR("Failed to allocate queues\n");
1250                 return NULL;
1251         }
1252
1253         cfs_percpt_for_each(q, i, qs)
1254                 INIT_LIST_HEAD(q);
1255
1256         return qs;
1257 }
1258
1259 static int lnet_unprepare(void);
1260
1261 static int
1262 lnet_prepare(lnet_pid_t requested_pid)
1263 {
1264         /* Prepare to bring up the network */
1265         struct lnet_res_container **recs;
1266         int                       rc = 0;
1267
1268         if (requested_pid == LNET_PID_ANY) {
1269                 /* Don't instantiate LNET just for me */
1270                 return -ENETDOWN;
1271         }
1272
1273         LASSERT(the_lnet.ln_refcount == 0);
1274
1275         the_lnet.ln_routing = 0;
1276
1277         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1278         the_lnet.ln_pid = requested_pid;
1279
1280         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1281         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1282         INIT_LIST_HEAD(&the_lnet.ln_nets);
1283         INIT_LIST_HEAD(&the_lnet.ln_routers);
1284         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1285         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1286         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1287         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1288         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1289         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1290         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1291         INIT_LIST_HEAD(&the_lnet.ln_udsp_list);
1292         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1293         the_lnet.ln_mt_handler = NULL;
1294         init_completion(&the_lnet.ln_started);
1295
1296         rc = lnet_slab_setup();
1297         if (rc != 0)
1298                 goto failed;
1299
1300         rc = lnet_create_remote_nets_table();
1301         if (rc != 0)
1302                 goto failed;
1303
1304         /*
1305          * NB the interface cookie in wire handles guards against delayed
1306          * replies and ACKs appearing valid after reboot.
1307          */
1308         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1309
1310         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1311                                                 sizeof(struct lnet_counters));
1312         if (the_lnet.ln_counters == NULL) {
1313                 CERROR("Failed to allocate counters for LNet\n");
1314                 rc = -ENOMEM;
1315                 goto failed;
1316         }
1317
1318         rc = lnet_peer_tables_create();
1319         if (rc != 0)
1320                 goto failed;
1321
1322         rc = lnet_msg_containers_create();
1323         if (rc != 0)
1324                 goto failed;
1325
1326         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1327                                       LNET_COOKIE_TYPE_EQ);
1328         if (rc != 0)
1329                 goto failed;
1330
1331         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1332         if (recs == NULL) {
1333                 rc = -ENOMEM;
1334                 goto failed;
1335         }
1336
1337         the_lnet.ln_md_containers = recs;
1338
1339         rc = lnet_portals_create();
1340         if (rc != 0) {
1341                 CERROR("Failed to create portals for LNet: %d\n", rc);
1342                 goto failed;
1343         }
1344
1345         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1346         if (!the_lnet.ln_mt_zombie_rstqs) {
1347                 rc = -ENOMEM;
1348                 goto failed;
1349         }
1350
1351         return 0;
1352
1353  failed:
1354         lnet_unprepare();
1355         return rc;
1356 }
1357
1358 static int
1359 lnet_unprepare (void)
1360 {
1361         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1362          * have shut down already, so it is safe to unlink and free all
1363          * descriptors, even those that appear committed to a network op (eg MD
1364          * with non-zero pending count) */
1365
1366         lnet_fail_nid(LNET_NID_ANY, 0);
1367
1368         LASSERT(the_lnet.ln_refcount == 0);
1369         LASSERT(list_empty(&the_lnet.ln_test_peers));
1370         LASSERT(list_empty(&the_lnet.ln_nets));
1371
1372         if (the_lnet.ln_mt_zombie_rstqs) {
1373                 lnet_clean_zombie_rstqs();
1374                 the_lnet.ln_mt_zombie_rstqs = NULL;
1375         }
1376
1377         lnet_assert_handler_unused(the_lnet.ln_mt_handler);
1378         the_lnet.ln_mt_handler = NULL;
1379
1380         lnet_portals_destroy();
1381
1382         if (the_lnet.ln_md_containers != NULL) {
1383                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1384                 the_lnet.ln_md_containers = NULL;
1385         }
1386
1387         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1388
1389         lnet_msg_containers_destroy();
1390         lnet_peer_uninit();
1391         lnet_rtrpools_free(0);
1392
1393         if (the_lnet.ln_counters != NULL) {
1394                 cfs_percpt_free(the_lnet.ln_counters);
1395                 the_lnet.ln_counters = NULL;
1396         }
1397         lnet_destroy_remote_nets_table();
1398         lnet_udsp_destroy(true);
1399         lnet_slab_cleanup();
1400
1401         return 0;
1402 }
1403
1404 struct lnet_ni  *
1405 lnet_net2ni_locked(__u32 net_id, int cpt)
1406 {
1407         struct lnet_ni   *ni;
1408         struct lnet_net  *net;
1409
1410         LASSERT(cpt != LNET_LOCK_EX);
1411
1412         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1413                 if (net->net_id == net_id) {
1414                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1415                                         ni_netlist);
1416                         return ni;
1417                 }
1418         }
1419
1420         return NULL;
1421 }
1422
1423 struct lnet_ni *
1424 lnet_net2ni_addref(__u32 net)
1425 {
1426         struct lnet_ni *ni;
1427
1428         lnet_net_lock(0);
1429         ni = lnet_net2ni_locked(net, 0);
1430         if (ni)
1431                 lnet_ni_addref_locked(ni, 0);
1432         lnet_net_unlock(0);
1433
1434         return ni;
1435 }
1436 EXPORT_SYMBOL(lnet_net2ni_addref);
1437
1438 struct lnet_net *
1439 lnet_get_net_locked(__u32 net_id)
1440 {
1441         struct lnet_net  *net;
1442
1443         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1444                 if (net->net_id == net_id)
1445                         return net;
1446         }
1447
1448         return NULL;
1449 }
1450
1451 void
1452 lnet_net_clr_pref_rtrs(struct lnet_net *net)
1453 {
1454         struct list_head zombies;
1455         struct lnet_nid_list *ne;
1456         struct lnet_nid_list *tmp;
1457
1458         INIT_LIST_HEAD(&zombies);
1459
1460         lnet_net_lock(LNET_LOCK_EX);
1461         list_splice_init(&net->net_rtr_pref_nids, &zombies);
1462         lnet_net_unlock(LNET_LOCK_EX);
1463
1464         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1465                 list_del_init(&ne->nl_list);
1466                 LIBCFS_FREE(ne, sizeof(*ne));
1467         }
1468 }
1469
1470 int
1471 lnet_net_add_pref_rtr(struct lnet_net *net,
1472                       lnet_nid_t gw_nid)
1473 __must_hold(&the_lnet.ln_api_mutex)
1474 {
1475         struct lnet_nid_list *ne;
1476
1477         /* This function is called with api_mutex held. When the api_mutex
1478          * is held the list can not be modified, as it is only modified as
1479          * a result of applying a UDSP and that happens under api_mutex
1480          * lock.
1481          */
1482         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1483                 if (ne->nl_nid == gw_nid)
1484                         return -EEXIST;
1485         }
1486
1487         LIBCFS_ALLOC(ne, sizeof(*ne));
1488         if (!ne)
1489                 return -ENOMEM;
1490
1491         ne->nl_nid = gw_nid;
1492
1493         /* Lock the cpt to protect against addition and checks in the
1494          * selection algorithm
1495          */
1496         lnet_net_lock(LNET_LOCK_EX);
1497         list_add(&ne->nl_list, &net->net_rtr_pref_nids);
1498         lnet_net_unlock(LNET_LOCK_EX);
1499
1500         return 0;
1501 }
1502
1503 bool
1504 lnet_net_is_pref_rtr_locked(struct lnet_net *net, lnet_nid_t rtr_nid)
1505 {
1506         struct lnet_nid_list *ne;
1507
1508         CDEBUG(D_NET, "%s: rtr pref emtpy: %d\n",
1509                libcfs_net2str(net->net_id),
1510                list_empty(&net->net_rtr_pref_nids));
1511
1512         if (list_empty(&net->net_rtr_pref_nids))
1513                 return false;
1514
1515         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1516                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1517                        libcfs_nid2str(ne->nl_nid),
1518                        libcfs_nid2str(rtr_nid));
1519                 if (rtr_nid == ne->nl_nid)
1520                         return true;
1521         }
1522
1523         return false;
1524 }
1525
1526 unsigned int
1527 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
1528 {
1529         __u64           key = nid;
1530         unsigned int    val;
1531
1532         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1533
1534         if (number == 1)
1535                 return 0;
1536
1537         val = hash_long(key, LNET_CPT_BITS);
1538         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1539         if (val < number)
1540                 return val;
1541
1542         return (unsigned int)(key + val + (val >> 1)) % number;
1543 }
1544
1545 int
1546 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
1547 {
1548         struct lnet_net *net;
1549
1550         /* must called with hold of lnet_net_lock */
1551         if (LNET_CPT_NUMBER == 1)
1552                 return 0; /* the only one */
1553
1554         /*
1555          * If NI is provided then use the CPT identified in the NI cpt
1556          * list if one exists. If one doesn't exist, then that NI is
1557          * associated with all CPTs and it follows that the net it belongs
1558          * to is implicitly associated with all CPTs, so just hash the nid
1559          * and return that.
1560          */
1561         if (ni != NULL) {
1562                 if (ni->ni_cpts != NULL)
1563                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1564                                                              ni->ni_ncpts)];
1565                 else
1566                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1567         }
1568
1569         /* no NI provided so look at the net */
1570         net = lnet_get_net_locked(LNET_NIDNET(nid));
1571
1572         if (net != NULL && net->net_cpts != NULL) {
1573                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1574         }
1575
1576         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1577 }
1578
1579 int
1580 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
1581 {
1582         int     cpt;
1583         int     cpt2;
1584
1585         if (LNET_CPT_NUMBER == 1)
1586                 return 0; /* the only one */
1587
1588         cpt = lnet_net_lock_current();
1589
1590         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1591
1592         lnet_net_unlock(cpt);
1593
1594         return cpt2;
1595 }
1596 EXPORT_SYMBOL(lnet_cpt_of_nid);
1597
1598 int
1599 lnet_islocalnet_locked(__u32 net_id)
1600 {
1601         struct lnet_net *net;
1602         bool local;
1603
1604         net = lnet_get_net_locked(net_id);
1605
1606         local = net != NULL;
1607
1608         return local;
1609 }
1610
1611 int
1612 lnet_islocalnet(__u32 net_id)
1613 {
1614         int cpt;
1615         bool local;
1616
1617         cpt = lnet_net_lock_current();
1618
1619         local = lnet_islocalnet_locked(net_id);
1620
1621         lnet_net_unlock(cpt);
1622
1623         return local;
1624 }
1625
1626 struct lnet_ni  *
1627 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
1628 {
1629         struct lnet_net  *net;
1630         struct lnet_ni   *ni;
1631
1632         LASSERT(cpt != LNET_LOCK_EX);
1633
1634         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1635                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1636                         if (ni->ni_nid == nid)
1637                                 return ni;
1638                 }
1639         }
1640
1641         return NULL;
1642 }
1643
1644 struct lnet_ni *
1645 lnet_nid2ni_addref(lnet_nid_t nid)
1646 {
1647         struct lnet_ni *ni;
1648
1649         lnet_net_lock(0);
1650         ni = lnet_nid2ni_locked(nid, 0);
1651         if (ni)
1652                 lnet_ni_addref_locked(ni, 0);
1653         lnet_net_unlock(0);
1654
1655         return ni;
1656 }
1657 EXPORT_SYMBOL(lnet_nid2ni_addref);
1658
1659 int
1660 lnet_islocalnid(lnet_nid_t nid)
1661 {
1662         struct lnet_ni  *ni;
1663         int             cpt;
1664
1665         cpt = lnet_net_lock_current();
1666         ni = lnet_nid2ni_locked(nid, cpt);
1667         lnet_net_unlock(cpt);
1668
1669         return ni != NULL;
1670 }
1671
1672 int
1673 lnet_count_acceptor_nets(void)
1674 {
1675         /* Return the # of NIs that need the acceptor. */
1676         int              count = 0;
1677         struct lnet_net  *net;
1678         int              cpt;
1679
1680         cpt = lnet_net_lock_current();
1681         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1682                 /* all socklnd type networks should have the acceptor
1683                  * thread started */
1684                 if (net->net_lnd->lnd_accept != NULL)
1685                         count++;
1686         }
1687
1688         lnet_net_unlock(cpt);
1689
1690         return count;
1691 }
1692
1693 struct lnet_ping_buffer *
1694 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1695 {
1696         struct lnet_ping_buffer *pbuf;
1697
1698         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1699         if (pbuf) {
1700                 pbuf->pb_nnis = nnis;
1701                 pbuf->pb_needs_post = false;
1702                 atomic_set(&pbuf->pb_refcnt, 1);
1703         }
1704
1705         return pbuf;
1706 }
1707
1708 void
1709 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1710 {
1711         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1712         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1713 }
1714
1715 static struct lnet_ping_buffer *
1716 lnet_ping_target_create(int nnis)
1717 {
1718         struct lnet_ping_buffer *pbuf;
1719
1720         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1721         if (pbuf == NULL) {
1722                 CERROR("Can't allocate ping source [%d]\n", nnis);
1723                 return NULL;
1724         }
1725
1726         pbuf->pb_info.pi_nnis = nnis;
1727         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1728         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1729         pbuf->pb_info.pi_features =
1730                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1731
1732         return pbuf;
1733 }
1734
1735 static inline int
1736 lnet_get_net_ni_count_locked(struct lnet_net *net)
1737 {
1738         struct lnet_ni  *ni;
1739         int             count = 0;
1740
1741         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1742                 count++;
1743
1744         return count;
1745 }
1746
1747 static inline int
1748 lnet_get_net_ni_count_pre(struct lnet_net *net)
1749 {
1750         struct lnet_ni  *ni;
1751         int             count = 0;
1752
1753         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1754                 count++;
1755
1756         return count;
1757 }
1758
1759 static inline int
1760 lnet_get_ni_count(void)
1761 {
1762         struct lnet_ni  *ni;
1763         struct lnet_net *net;
1764         int             count = 0;
1765
1766         lnet_net_lock(0);
1767
1768         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1769                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1770                         count++;
1771         }
1772
1773         lnet_net_unlock(0);
1774
1775         return count;
1776 }
1777
1778 int
1779 lnet_get_net_count(void)
1780 {
1781         struct lnet_net *net;
1782         int count = 0;
1783
1784         lnet_net_lock(0);
1785
1786         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1787                 count++;
1788         }
1789
1790         lnet_net_unlock(0);
1791
1792         return count;
1793 }
1794
1795 void
1796 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1797 {
1798         struct lnet_ni_status *stat;
1799         int nnis;
1800         int i;
1801
1802         __swab32s(&pbuf->pb_info.pi_magic);
1803         __swab32s(&pbuf->pb_info.pi_features);
1804         __swab32s(&pbuf->pb_info.pi_pid);
1805         __swab32s(&pbuf->pb_info.pi_nnis);
1806         nnis = pbuf->pb_info.pi_nnis;
1807         if (nnis > pbuf->pb_nnis)
1808                 nnis = pbuf->pb_nnis;
1809         for (i = 0; i < nnis; i++) {
1810                 stat = &pbuf->pb_info.pi_ni[i];
1811                 __swab64s(&stat->ns_nid);
1812                 __swab32s(&stat->ns_status);
1813         }
1814 }
1815
1816 int
1817 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1818 {
1819         if (!pinfo)
1820                 return -EINVAL;
1821         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1822                 return -EPROTO;
1823         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1824                 return -EPROTO;
1825         /* Loopback is guaranteed to be present */
1826         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1827                 return -ERANGE;
1828         if (LNET_PING_INFO_LONI(pinfo) != LNET_NID_LO_0)
1829                 return -EPROTO;
1830         return 0;
1831 }
1832
1833 static void
1834 lnet_ping_target_destroy(void)
1835 {
1836         struct lnet_net *net;
1837         struct lnet_ni  *ni;
1838
1839         lnet_net_lock(LNET_LOCK_EX);
1840
1841         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1842                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1843                         lnet_ni_lock(ni);
1844                         ni->ni_status = NULL;
1845                         lnet_ni_unlock(ni);
1846                 }
1847         }
1848
1849         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1850         the_lnet.ln_ping_target = NULL;
1851
1852         lnet_net_unlock(LNET_LOCK_EX);
1853 }
1854
1855 static void
1856 lnet_ping_target_event_handler(struct lnet_event *event)
1857 {
1858         struct lnet_ping_buffer *pbuf = event->md_user_ptr;
1859
1860         if (event->unlinked)
1861                 lnet_ping_buffer_decref(pbuf);
1862 }
1863
1864 static int
1865 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1866                        struct lnet_handle_md *ping_mdh,
1867                        int ni_count, bool set_eq)
1868 {
1869         struct lnet_process_id id = {
1870                 .nid = LNET_NID_ANY,
1871                 .pid = LNET_PID_ANY
1872         };
1873         struct lnet_me *me;
1874         struct lnet_md md = { NULL };
1875         int rc;
1876
1877         if (set_eq)
1878                 the_lnet.ln_ping_target_handler =
1879                         lnet_ping_target_event_handler;
1880
1881         *ppbuf = lnet_ping_target_create(ni_count);
1882         if (*ppbuf == NULL) {
1883                 rc = -ENOMEM;
1884                 goto fail_free_eq;
1885         }
1886
1887         /* Ping target ME/MD */
1888         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1889                           LNET_PROTO_PING_MATCHBITS, 0,
1890                           LNET_UNLINK, LNET_INS_AFTER);
1891         if (IS_ERR(me)) {
1892                 rc = PTR_ERR(me);
1893                 CERROR("Can't create ping target ME: %d\n", rc);
1894                 goto fail_decref_ping_buffer;
1895         }
1896
1897         /* initialize md content */
1898         md.start     = &(*ppbuf)->pb_info;
1899         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1900         md.threshold = LNET_MD_THRESH_INF;
1901         md.max_size  = 0;
1902         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1903                        LNET_MD_MANAGE_REMOTE;
1904         md.handler   = the_lnet.ln_ping_target_handler;
1905         md.user_ptr  = *ppbuf;
1906
1907         rc = LNetMDAttach(me, &md, LNET_RETAIN, ping_mdh);
1908         if (rc != 0) {
1909                 CERROR("Can't attach ping target MD: %d\n", rc);
1910                 goto fail_decref_ping_buffer;
1911         }
1912         lnet_ping_buffer_addref(*ppbuf);
1913
1914         return 0;
1915
1916 fail_decref_ping_buffer:
1917         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
1918         lnet_ping_buffer_decref(*ppbuf);
1919         *ppbuf = NULL;
1920 fail_free_eq:
1921         return rc;
1922 }
1923
1924 static void
1925 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1926                     struct lnet_handle_md *ping_mdh)
1927 {
1928         LNetMDUnlink(*ping_mdh);
1929         LNetInvalidateMDHandle(ping_mdh);
1930
1931         /* NB the MD could be busy; this just starts the unlink */
1932         wait_var_event_warning(&pbuf->pb_refcnt,
1933                                atomic_read(&pbuf->pb_refcnt) <= 1,
1934                                "Still waiting for ping data MD to unlink\n");
1935 }
1936
1937 static void
1938 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1939 {
1940         struct lnet_ni          *ni;
1941         struct lnet_net         *net;
1942         struct lnet_ni_status *ns;
1943         int                     i;
1944         int                     rc;
1945
1946         i = 0;
1947         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1948                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1949                         LASSERT(i < pbuf->pb_nnis);
1950
1951                         ns = &pbuf->pb_info.pi_ni[i];
1952
1953                         ns->ns_nid = ni->ni_nid;
1954
1955                         lnet_ni_lock(ni);
1956                         ns->ns_status = (ni->ni_status != NULL) ?
1957                                          ni->ni_status->ns_status :
1958                                                 LNET_NI_STATUS_UP;
1959                         ni->ni_status = ns;
1960                         lnet_ni_unlock(ni);
1961
1962                         i++;
1963                 }
1964         }
1965         /*
1966          * We (ab)use the ns_status of the loopback interface to
1967          * transmit the sequence number. The first interface listed
1968          * must be the loopback interface.
1969          */
1970         rc = lnet_ping_info_validate(&pbuf->pb_info);
1971         if (rc) {
1972                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1973                 LBUG();
1974         }
1975         LNET_PING_BUFFER_SEQNO(pbuf) =
1976                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1977 }
1978
1979 static void
1980 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1981                         struct lnet_handle_md ping_mdh)
1982 {
1983         struct lnet_ping_buffer *old_pbuf = NULL;
1984         struct lnet_handle_md old_ping_md;
1985
1986         /* switch the NIs to point to the new ping info created */
1987         lnet_net_lock(LNET_LOCK_EX);
1988
1989         if (!the_lnet.ln_routing)
1990                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1991         if (!lnet_peer_discovery_disabled)
1992                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1993
1994         /* Ensure only known feature bits have been set. */
1995         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1996         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1997
1998         lnet_ping_target_install_locked(pbuf);
1999
2000         if (the_lnet.ln_ping_target) {
2001                 old_pbuf = the_lnet.ln_ping_target;
2002                 old_ping_md = the_lnet.ln_ping_target_md;
2003         }
2004         the_lnet.ln_ping_target_md = ping_mdh;
2005         the_lnet.ln_ping_target = pbuf;
2006
2007         lnet_net_unlock(LNET_LOCK_EX);
2008
2009         if (old_pbuf) {
2010                 /* unlink and free the old ping info */
2011                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
2012                 lnet_ping_buffer_decref(old_pbuf);
2013         }
2014
2015         lnet_push_update_to_peers(0);
2016 }
2017
2018 static void
2019 lnet_ping_target_fini(void)
2020 {
2021         lnet_ping_md_unlink(the_lnet.ln_ping_target,
2022                             &the_lnet.ln_ping_target_md);
2023
2024         lnet_assert_handler_unused(the_lnet.ln_ping_target_handler);
2025         lnet_ping_target_destroy();
2026 }
2027
2028 /* Resize the push target. */
2029 int lnet_push_target_resize(void)
2030 {
2031         struct lnet_handle_md mdh;
2032         struct lnet_handle_md old_mdh;
2033         struct lnet_ping_buffer *pbuf;
2034         struct lnet_ping_buffer *old_pbuf;
2035         int nnis;
2036         int rc;
2037
2038 again:
2039         nnis = the_lnet.ln_push_target_nnis;
2040         if (nnis <= 0) {
2041                 CDEBUG(D_NET, "Invalid nnis %d\n", nnis);
2042                 return -EINVAL;
2043         }
2044
2045         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
2046          * dropped when we need to resize again (see "old_pbuf" below) or when
2047          * LNet is shutdown (see lnet_push_target_fini())
2048          */
2049         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
2050         if (!pbuf) {
2051                 CDEBUG(D_NET, "Can't allocate pbuf for nnis %d\n", nnis);
2052                 return -ENOMEM;
2053         }
2054
2055         rc = lnet_push_target_post(pbuf, &mdh);
2056         if (rc) {
2057                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
2058                 lnet_ping_buffer_decref(pbuf);
2059                 return rc;
2060         }
2061
2062         lnet_net_lock(LNET_LOCK_EX);
2063         old_pbuf = the_lnet.ln_push_target;
2064         old_mdh = the_lnet.ln_push_target_md;
2065         the_lnet.ln_push_target = pbuf;
2066         the_lnet.ln_push_target_md = mdh;
2067         lnet_net_unlock(LNET_LOCK_EX);
2068
2069         if (old_pbuf) {
2070                 LNetMDUnlink(old_mdh);
2071                 /* Drop ref set by lnet_ping_buffer_alloc() */
2072                 lnet_ping_buffer_decref(old_pbuf);
2073         }
2074
2075         /* Received another push or reply that requires a larger buffer */
2076         if (nnis < the_lnet.ln_push_target_nnis)
2077                 goto again;
2078
2079         CDEBUG(D_NET, "nnis %d success\n", nnis);
2080         return 0;
2081 }
2082
2083 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
2084                           struct lnet_handle_md *mdhp)
2085 {
2086         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
2087         struct lnet_md md = { NULL };
2088         struct lnet_me *me;
2089         int rc;
2090
2091         me = LNetMEAttach(LNET_RESERVED_PORTAL, id,
2092                           LNET_PROTO_PING_MATCHBITS, 0,
2093                           LNET_UNLINK, LNET_INS_AFTER);
2094         if (IS_ERR(me)) {
2095                 rc = PTR_ERR(me);
2096                 CERROR("Can't create push target ME: %d\n", rc);
2097                 return rc;
2098         }
2099
2100         pbuf->pb_needs_post = false;
2101
2102         /* This reference is dropped by lnet_push_target_event_handler() */
2103         lnet_ping_buffer_addref(pbuf);
2104
2105         /* initialize md content */
2106         md.start     = &pbuf->pb_info;
2107         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
2108         md.threshold = 1;
2109         md.max_size  = 0;
2110         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
2111         md.user_ptr  = pbuf;
2112         md.handler   = the_lnet.ln_push_target_handler;
2113
2114         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdhp);
2115         if (rc) {
2116                 CERROR("Can't attach push MD: %d\n", rc);
2117                 lnet_ping_buffer_decref(pbuf);
2118                 pbuf->pb_needs_post = true;
2119                 return rc;
2120         }
2121
2122         CDEBUG(D_NET, "posted push target %p\n", pbuf);
2123
2124         return 0;
2125 }
2126
2127 static void lnet_push_target_event_handler(struct lnet_event *ev)
2128 {
2129         struct lnet_ping_buffer *pbuf = ev->md_user_ptr;
2130
2131         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
2132                ev->unlinked);
2133
2134         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2135                 lnet_swap_pinginfo(pbuf);
2136
2137         if (ev->type == LNET_EVENT_UNLINK) {
2138                 /* Drop ref added by lnet_push_target_post() */
2139                 lnet_ping_buffer_decref(pbuf);
2140                 return;
2141         }
2142
2143         lnet_peer_push_event(ev);
2144         if (ev->unlinked)
2145                 /* Drop ref added by lnet_push_target_post */
2146                 lnet_ping_buffer_decref(pbuf);
2147 }
2148
2149 /* Initialize the push target. */
2150 static int lnet_push_target_init(void)
2151 {
2152         int rc;
2153
2154         if (the_lnet.ln_push_target)
2155                 return -EALREADY;
2156
2157         the_lnet.ln_push_target_handler =
2158                 lnet_push_target_event_handler;
2159
2160         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
2161         LASSERT(rc == 0);
2162
2163         /* Start at the required minimum, we'll enlarge if required. */
2164         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
2165
2166         rc = lnet_push_target_resize();
2167
2168         if (rc) {
2169                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2170                 the_lnet.ln_push_target_handler = NULL;
2171         }
2172
2173         return rc;
2174 }
2175
2176 /* Clean up the push target. */
2177 static void lnet_push_target_fini(void)
2178 {
2179         if (!the_lnet.ln_push_target)
2180                 return;
2181
2182         /* Unlink and invalidate to prevent new references. */
2183         LNetMDUnlink(the_lnet.ln_push_target_md);
2184         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2185
2186         /* Wait for the unlink to complete. */
2187         wait_var_event_warning(&the_lnet.ln_push_target->pb_refcnt,
2188                                atomic_read(&the_lnet.ln_push_target->pb_refcnt) <= 1,
2189                                "Still waiting for ping data MD to unlink\n");
2190
2191         /* Drop ref set by lnet_ping_buffer_alloc() */
2192         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2193         the_lnet.ln_push_target = NULL;
2194         the_lnet.ln_push_target_nnis = 0;
2195
2196         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2197         lnet_assert_handler_unused(the_lnet.ln_push_target_handler);
2198         the_lnet.ln_push_target_handler = NULL;
2199 }
2200
2201 static int
2202 lnet_ni_tq_credits(struct lnet_ni *ni)
2203 {
2204         int     credits;
2205
2206         LASSERT(ni->ni_ncpts >= 1);
2207
2208         if (ni->ni_ncpts == 1)
2209                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2210
2211         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2212         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2213         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2214
2215         return credits;
2216 }
2217
2218 static void
2219 lnet_ni_unlink_locked(struct lnet_ni *ni)
2220 {
2221         /* move it to zombie list and nobody can find it anymore */
2222         LASSERT(!list_empty(&ni->ni_netlist));
2223         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2224         lnet_ni_decref_locked(ni, 0);
2225 }
2226
2227 static void
2228 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2229 {
2230         int             i;
2231         int             islo;
2232         struct lnet_ni  *ni;
2233         struct list_head *zombie_list = &net->net_ni_zombie;
2234
2235         /*
2236          * Now wait for the NIs I just nuked to show up on the zombie
2237          * list and shut them down in guaranteed thread context
2238          */
2239         i = 2;
2240         while (!list_empty(zombie_list)) {
2241                 int     *ref;
2242                 int     j;
2243
2244                 ni = list_entry(zombie_list->next,
2245                                 struct lnet_ni, ni_netlist);
2246                 list_del_init(&ni->ni_netlist);
2247                 /* the ni should be in deleting state. If it's not it's
2248                  * a bug */
2249                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2250                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2251                         if (*ref == 0)
2252                                 continue;
2253                         /* still busy, add it back to zombie list */
2254                         list_add(&ni->ni_netlist, zombie_list);
2255                         break;
2256                 }
2257
2258                 if (!list_empty(&ni->ni_netlist)) {
2259                         /* Unlock mutex while waiting to allow other
2260                          * threads to read the LNet state and fall through
2261                          * to avoid deadlock
2262                          */
2263                         lnet_net_unlock(LNET_LOCK_EX);
2264                         mutex_unlock(&the_lnet.ln_api_mutex);
2265
2266                         ++i;
2267                         if ((i & (-i)) == i) {
2268                                 CDEBUG(D_WARNING,
2269                                        "Waiting for zombie LNI %s\n",
2270                                        libcfs_nid2str(ni->ni_nid));
2271                         }
2272                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2273
2274                         mutex_lock(&the_lnet.ln_api_mutex);
2275                         lnet_net_lock(LNET_LOCK_EX);
2276                         continue;
2277                 }
2278
2279                 lnet_net_unlock(LNET_LOCK_EX);
2280
2281                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2282
2283                 LASSERT(!in_interrupt());
2284                 /* Holding the mutex makes it safe for lnd_shutdown
2285                  * to call module_put(). Module unload cannot finish
2286                  * until lnet_unregister_lnd() completes, and that
2287                  * requires the mutex.
2288                  */
2289                 mutex_lock(&the_lnet.ln_lnd_mutex);
2290                 (net->net_lnd->lnd_shutdown)(ni);
2291                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2292
2293                 if (!islo)
2294                         CDEBUG(D_LNI, "Removed LNI %s\n",
2295                               libcfs_nid2str(ni->ni_nid));
2296
2297                 lnet_ni_free(ni);
2298                 i = 2;
2299                 lnet_net_lock(LNET_LOCK_EX);
2300         }
2301 }
2302
2303 /* shutdown down the NI and release refcount */
2304 static void
2305 lnet_shutdown_lndni(struct lnet_ni *ni)
2306 {
2307         int i;
2308         struct lnet_net *net = ni->ni_net;
2309
2310         lnet_net_lock(LNET_LOCK_EX);
2311         lnet_ni_lock(ni);
2312         ni->ni_state = LNET_NI_STATE_DELETING;
2313         lnet_ni_unlock(ni);
2314         lnet_ni_unlink_locked(ni);
2315         lnet_incr_dlc_seq();
2316         lnet_net_unlock(LNET_LOCK_EX);
2317
2318         /* clear messages for this NI on the lazy portal */
2319         for (i = 0; i < the_lnet.ln_nportals; i++)
2320                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2321
2322         lnet_net_lock(LNET_LOCK_EX);
2323         lnet_clear_zombies_nis_locked(net);
2324         lnet_net_unlock(LNET_LOCK_EX);
2325 }
2326
2327 static void
2328 lnet_shutdown_lndnet(struct lnet_net *net)
2329 {
2330         struct lnet_ni *ni;
2331
2332         lnet_net_lock(LNET_LOCK_EX);
2333
2334         list_del_init(&net->net_list);
2335
2336         while (!list_empty(&net->net_ni_list)) {
2337                 ni = list_entry(net->net_ni_list.next,
2338                                 struct lnet_ni, ni_netlist);
2339                 lnet_net_unlock(LNET_LOCK_EX);
2340                 lnet_shutdown_lndni(ni);
2341                 lnet_net_lock(LNET_LOCK_EX);
2342         }
2343
2344         lnet_net_unlock(LNET_LOCK_EX);
2345
2346         /* Do peer table cleanup for this net */
2347         lnet_peer_tables_cleanup(net);
2348
2349         lnet_net_free(net);
2350 }
2351
2352 static void
2353 lnet_shutdown_lndnets(void)
2354 {
2355         struct lnet_net *net;
2356         LIST_HEAD(resend);
2357         struct lnet_msg *msg, *tmp;
2358
2359         /* NB called holding the global mutex */
2360
2361         /* All quiet on the API front */
2362         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2363         LASSERT(the_lnet.ln_refcount == 0);
2364
2365         lnet_net_lock(LNET_LOCK_EX);
2366         the_lnet.ln_state = LNET_STATE_STOPPING;
2367
2368         /*
2369          * move the nets to the zombie list to avoid them being
2370          * picked up for new work. LONET is also included in the
2371          * Nets that will be moved to the zombie list
2372          */
2373         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2374
2375         /* Drop the cached loopback Net. */
2376         if (the_lnet.ln_loni != NULL) {
2377                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2378                 the_lnet.ln_loni = NULL;
2379         }
2380         lnet_net_unlock(LNET_LOCK_EX);
2381
2382         /* iterate through the net zombie list and delete each net */
2383         while (!list_empty(&the_lnet.ln_net_zombie)) {
2384                 net = list_entry(the_lnet.ln_net_zombie.next,
2385                                  struct lnet_net, net_list);
2386                 lnet_shutdown_lndnet(net);
2387         }
2388
2389         spin_lock(&the_lnet.ln_msg_resend_lock);
2390         list_splice(&the_lnet.ln_msg_resend, &resend);
2391         spin_unlock(&the_lnet.ln_msg_resend_lock);
2392
2393         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2394                 list_del_init(&msg->msg_list);
2395                 msg->msg_no_resend = true;
2396                 lnet_finalize(msg, -ECANCELED);
2397         }
2398
2399         lnet_net_lock(LNET_LOCK_EX);
2400         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2401         lnet_net_unlock(LNET_LOCK_EX);
2402 }
2403
2404 static int
2405 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2406 {
2407         int                     rc = -EINVAL;
2408         struct lnet_tx_queue    *tq;
2409         int                     i;
2410         struct lnet_net         *net = ni->ni_net;
2411
2412         mutex_lock(&the_lnet.ln_lnd_mutex);
2413
2414         if (tun) {
2415                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2416                 ni->ni_lnd_tunables_set = true;
2417         }
2418
2419         rc = (net->net_lnd->lnd_startup)(ni);
2420
2421         mutex_unlock(&the_lnet.ln_lnd_mutex);
2422
2423         if (rc != 0) {
2424                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2425                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2426                 goto failed0;
2427         }
2428
2429         lnet_ni_lock(ni);
2430         ni->ni_state = LNET_NI_STATE_ACTIVE;
2431         lnet_ni_unlock(ni);
2432
2433         /* We keep a reference on the loopback net through the loopback NI */
2434         if (net->net_lnd->lnd_type == LOLND) {
2435                 lnet_ni_addref(ni);
2436                 LASSERT(the_lnet.ln_loni == NULL);
2437                 the_lnet.ln_loni = ni;
2438                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2439                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2440                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2441                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2442                 return 0;
2443         }
2444
2445         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2446             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2447                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2448                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2449                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2450                                         "" : "per-peer ");
2451                 /* shutdown the NI since if we get here then it must've already
2452                  * been started
2453                  */
2454                 lnet_shutdown_lndni(ni);
2455                 return -EINVAL;
2456         }
2457
2458         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2459                 tq->tq_credits_min =
2460                 tq->tq_credits_max =
2461                 tq->tq_credits = lnet_ni_tq_credits(ni);
2462         }
2463
2464         atomic_set(&ni->ni_tx_credits,
2465                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2466         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2467
2468         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2469                 libcfs_nid2str(ni->ni_nid),
2470                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2471                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2472                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2473                 ni->ni_net->net_tunables.lct_peer_timeout);
2474
2475         return 0;
2476 failed0:
2477         lnet_ni_free(ni);
2478         return rc;
2479 }
2480
2481 static int
2482 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2483 {
2484         struct lnet_ni *ni;
2485         struct lnet_net *net_l = NULL;
2486         LIST_HEAD(local_ni_list);
2487         int rc;
2488         int ni_count = 0;
2489         __u32 lnd_type;
2490         const struct lnet_lnd  *lnd;
2491         int peer_timeout =
2492                 net->net_tunables.lct_peer_timeout;
2493         int maxtxcredits =
2494                 net->net_tunables.lct_max_tx_credits;
2495         int peerrtrcredits =
2496                 net->net_tunables.lct_peer_rtr_credits;
2497
2498         /*
2499          * make sure that this net is unique. If it isn't then
2500          * we are adding interfaces to an already existing network, and
2501          * 'net' is just a convenient way to pass in the list.
2502          * if it is unique we need to find the LND and load it if
2503          * necessary.
2504          */
2505         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2506                 lnd_type = LNET_NETTYP(net->net_id);
2507
2508                 mutex_lock(&the_lnet.ln_lnd_mutex);
2509                 lnd = lnet_find_lnd_by_type(lnd_type);
2510
2511                 if (lnd == NULL) {
2512                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2513                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2514                         mutex_lock(&the_lnet.ln_lnd_mutex);
2515
2516                         lnd = lnet_find_lnd_by_type(lnd_type);
2517                         if (lnd == NULL) {
2518                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2519                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2520                                 libcfs_lnd2str(lnd_type),
2521                                 libcfs_lnd2modname(lnd_type), rc);
2522 #ifndef HAVE_MODULE_LOADING_SUPPORT
2523                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2524                                                 "compiled with kernel module "
2525                                                 "loading support.");
2526 #endif
2527                                 rc = -EINVAL;
2528                                 goto failed0;
2529                         }
2530                 }
2531
2532                 net->net_lnd = lnd;
2533
2534                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2535
2536                 net_l = net;
2537         }
2538
2539         /*
2540          * net_l: if the network being added is unique then net_l
2541          *        will point to that network
2542          *        if the network being added is not unique then
2543          *        net_l points to the existing network.
2544          *
2545          * When we enter the loop below, we'll pick NIs off he
2546          * network beign added and start them up, then add them to
2547          * a local ni list. Once we've successfully started all
2548          * the NIs then we join the local NI list (of started up
2549          * networks) with the net_l->net_ni_list, which should
2550          * point to the correct network to add the new ni list to
2551          *
2552          * If any of the new NIs fail to start up, then we want to
2553          * iterate through the local ni list, which should include
2554          * any NIs which were successfully started up, and shut
2555          * them down.
2556          *
2557          * After than we want to delete the network being added,
2558          * to avoid a memory leak.
2559          */
2560
2561         /*
2562          * When a network uses TCP bonding then all its interfaces
2563          * must be specified when the network is first defined: the
2564          * TCP bonding code doesn't allow for interfaces to be added
2565          * or removed.
2566          */
2567         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2568             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2569                 rc = -EINVAL;
2570                 goto failed0;
2571         }
2572
2573         while (!list_empty(&net->net_ni_added)) {
2574                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2575                                 ni_netlist);
2576                 list_del_init(&ni->ni_netlist);
2577
2578                 /* make sure that the the NI we're about to start
2579                  * up is actually unique. if it's not fail. */
2580                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2581                                         ni->ni_interfaces[0])) {
2582                         rc = -EEXIST;
2583                         goto failed1;
2584                 }
2585
2586                 /* adjust the pointer the parent network, just in case it
2587                  * the net is a duplicate */
2588                 ni->ni_net = net_l;
2589
2590                 rc = lnet_startup_lndni(ni, tun);
2591
2592                 if (rc < 0)
2593                         goto failed1;
2594
2595                 lnet_ni_addref(ni);
2596                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2597
2598                 ni_count++;
2599         }
2600
2601         lnet_net_lock(LNET_LOCK_EX);
2602         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2603         lnet_incr_dlc_seq();
2604         lnet_net_unlock(LNET_LOCK_EX);
2605
2606         /* if the network is not unique then we don't want to keep
2607          * it around after we're done. Free it. Otherwise add that
2608          * net to the global the_lnet.ln_nets */
2609         if (net_l != net && net_l != NULL) {
2610                 /*
2611                  * TODO - note. currently the tunables can not be updated
2612                  * once added
2613                  */
2614                 lnet_net_free(net);
2615         } else {
2616                 /*
2617                  * restore tunables after it has been overwitten by the
2618                  * lnd
2619                  */
2620                 if (peer_timeout != -1)
2621                         net->net_tunables.lct_peer_timeout = peer_timeout;
2622                 if (maxtxcredits != -1)
2623                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2624                 if (peerrtrcredits != -1)
2625                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2626
2627                 lnet_net_lock(LNET_LOCK_EX);
2628                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2629                 lnet_net_unlock(LNET_LOCK_EX);
2630         }
2631
2632         /* update net count */
2633         lnet_current_net_count = lnet_get_net_count();
2634
2635         return ni_count;
2636
2637 failed1:
2638         /*
2639          * shutdown the new NIs that are being started up
2640          * free the NET being started
2641          */
2642         while (!list_empty(&local_ni_list)) {
2643                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2644                                 ni_netlist);
2645
2646                 lnet_shutdown_lndni(ni);
2647         }
2648
2649 failed0:
2650         lnet_net_free(net);
2651
2652         return rc;
2653 }
2654
2655 static int
2656 lnet_startup_lndnets(struct list_head *netlist)
2657 {
2658         struct lnet_net         *net;
2659         int                     rc;
2660         int                     ni_count = 0;
2661
2662         /*
2663          * Change to running state before bringing up the LNDs. This
2664          * allows lnet_shutdown_lndnets() to assert that we've passed
2665          * through here.
2666          */
2667         lnet_net_lock(LNET_LOCK_EX);
2668         the_lnet.ln_state = LNET_STATE_RUNNING;
2669         lnet_net_unlock(LNET_LOCK_EX);
2670
2671         while (!list_empty(netlist)) {
2672                 net = list_entry(netlist->next, struct lnet_net, net_list);
2673                 list_del_init(&net->net_list);
2674
2675                 rc = lnet_startup_lndnet(net, NULL);
2676
2677                 if (rc < 0)
2678                         goto failed;
2679
2680                 ni_count += rc;
2681         }
2682
2683         return ni_count;
2684 failed:
2685         lnet_shutdown_lndnets();
2686
2687         return rc;
2688 }
2689
2690 /**
2691  * Initialize LNet library.
2692  *
2693  * Automatically called at module loading time. Caller has to call
2694  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2695  * latter returned 0. It must be called exactly once.
2696  *
2697  * \retval 0 on success
2698  * \retval -ve on failures.
2699  */
2700 int lnet_lib_init(void)
2701 {
2702         int rc;
2703
2704         lnet_assert_wire_constants();
2705
2706         /* refer to global cfs_cpt_table for now */
2707         the_lnet.ln_cpt_table = cfs_cpt_tab;
2708         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2709
2710         LASSERT(the_lnet.ln_cpt_number > 0);
2711         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2712                 /* we are under risk of consuming all lh_cookie */
2713                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2714                        "please change setting of CPT-table and retry\n",
2715                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2716                 return -E2BIG;
2717         }
2718
2719         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2720                 the_lnet.ln_cpt_bits++;
2721
2722         rc = lnet_create_locks();
2723         if (rc != 0) {
2724                 CERROR("Can't create LNet global locks: %d\n", rc);
2725                 return rc;
2726         }
2727
2728         the_lnet.ln_refcount = 0;
2729         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2730         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2731
2732         /* The hash table size is the number of bits it takes to express the set
2733          * ln_num_routes, minus 1 (better to under estimate than over so we
2734          * don't waste memory). */
2735         if (rnet_htable_size <= 0)
2736                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2737         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2738                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2739         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2740                                            order_base_2(rnet_htable_size) - 1);
2741
2742         /* All LNDs apart from the LOLND are in separate modules.  They
2743          * register themselves when their module loads, and unregister
2744          * themselves when their module is unloaded. */
2745         lnet_register_lnd(&the_lolnd);
2746         return 0;
2747 }
2748
2749 /**
2750  * Finalize LNet library.
2751  *
2752  * \pre lnet_lib_init() called with success.
2753  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2754  *
2755  * As this happens at module-unload, all lnds must already be unloaded,
2756  * so they must already be unregistered.
2757  */
2758 void lnet_lib_exit(void)
2759 {
2760         int i;
2761
2762         LASSERT(the_lnet.ln_refcount == 0);
2763         lnet_unregister_lnd(&the_lolnd);
2764         for (i = 0; i < NUM_LNDS; i++)
2765                 LASSERT(!the_lnet.ln_lnds[i]);
2766         lnet_destroy_locks();
2767 }
2768
2769 /**
2770  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2771  *
2772  * Users must call this function at least once before any other functions.
2773  * For each successful call there must be a corresponding call to
2774  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2775  * ignored.
2776  *
2777  * The PID used by LNet may be different from the one requested.
2778  * See LNetGetId().
2779  *
2780  * \param requested_pid PID requested by the caller.
2781  *
2782  * \return >= 0 on success, and < 0 error code on failures.
2783  */
2784 int
2785 LNetNIInit(lnet_pid_t requested_pid)
2786 {
2787         int                     im_a_router = 0;
2788         int                     rc;
2789         int                     ni_count;
2790         struct lnet_ping_buffer *pbuf;
2791         struct lnet_handle_md   ping_mdh;
2792         LIST_HEAD(net_head);
2793         struct lnet_net         *net;
2794
2795         mutex_lock(&the_lnet.ln_api_mutex);
2796
2797         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2798
2799         if (the_lnet.ln_refcount > 0) {
2800                 rc = the_lnet.ln_refcount++;
2801                 mutex_unlock(&the_lnet.ln_api_mutex);
2802                 return rc;
2803         }
2804
2805         rc = lnet_prepare(requested_pid);
2806         if (rc != 0) {
2807                 mutex_unlock(&the_lnet.ln_api_mutex);
2808                 return rc;
2809         }
2810
2811         /* create a network for Loopback network */
2812         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2813         if (net == NULL) {
2814                 rc = -ENOMEM;
2815                 goto err_empty_list;
2816         }
2817
2818         /* Add in the loopback NI */
2819         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2820                 rc = -ENOMEM;
2821                 goto err_empty_list;
2822         }
2823
2824         if (use_tcp_bonding)
2825                 CWARN("'use_tcp_bonding' option has been deprecated. See LU-13641\n");
2826
2827         /* If LNet is being initialized via DLC it is possible
2828          * that the user requests not to load module parameters (ones which
2829          * are supported by DLC) on initialization.  Therefore, make sure not
2830          * to load networks, routes and forwarding from module parameters
2831          * in this case.  On cleanup in case of failure only clean up
2832          * routes if it has been loaded */
2833         if (!the_lnet.ln_nis_from_mod_params) {
2834                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2835                                          use_tcp_bonding);
2836                 if (rc < 0)
2837                         goto err_empty_list;
2838         }
2839
2840         ni_count = lnet_startup_lndnets(&net_head);
2841         if (ni_count < 0) {
2842                 rc = ni_count;
2843                 goto err_empty_list;
2844         }
2845
2846         if (!the_lnet.ln_nis_from_mod_params) {
2847                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2848                 if (rc != 0)
2849                         goto err_shutdown_lndnis;
2850
2851                 rc = lnet_rtrpools_alloc(im_a_router);
2852                 if (rc != 0)
2853                         goto err_destroy_routes;
2854         }
2855
2856         rc = lnet_acceptor_start();
2857         if (rc != 0)
2858                 goto err_destroy_routes;
2859
2860         the_lnet.ln_refcount = 1;
2861         /* Now I may use my own API functions... */
2862
2863         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2864         if (rc != 0)
2865                 goto err_acceptor_stop;
2866
2867         lnet_ping_target_update(pbuf, ping_mdh);
2868
2869         the_lnet.ln_mt_handler = lnet_mt_event_handler;
2870
2871         rc = lnet_push_target_init();
2872         if (rc != 0)
2873                 goto err_stop_ping;
2874
2875         rc = lnet_peer_discovery_start();
2876         if (rc != 0)
2877                 goto err_destroy_push_target;
2878
2879         rc = lnet_monitor_thr_start();
2880         if (rc != 0)
2881                 goto err_stop_discovery_thr;
2882
2883         lnet_fault_init();
2884         lnet_router_debugfs_init();
2885
2886         mutex_unlock(&the_lnet.ln_api_mutex);
2887
2888         complete_all(&the_lnet.ln_started);
2889
2890         /* wait for all routers to start */
2891         lnet_wait_router_start();
2892
2893         return 0;
2894
2895 err_stop_discovery_thr:
2896         lnet_peer_discovery_stop();
2897 err_destroy_push_target:
2898         lnet_push_target_fini();
2899 err_stop_ping:
2900         lnet_ping_target_fini();
2901 err_acceptor_stop:
2902         the_lnet.ln_refcount = 0;
2903         lnet_acceptor_stop();
2904 err_destroy_routes:
2905         if (!the_lnet.ln_nis_from_mod_params)
2906                 lnet_destroy_routes();
2907 err_shutdown_lndnis:
2908         lnet_shutdown_lndnets();
2909 err_empty_list:
2910         lnet_unprepare();
2911         LASSERT(rc < 0);
2912         mutex_unlock(&the_lnet.ln_api_mutex);
2913         while (!list_empty(&net_head)) {
2914                 struct lnet_net *net;
2915
2916                 net = list_entry(net_head.next, struct lnet_net, net_list);
2917                 list_del_init(&net->net_list);
2918                 lnet_net_free(net);
2919         }
2920         return rc;
2921 }
2922 EXPORT_SYMBOL(LNetNIInit);
2923
2924 /**
2925  * Stop LNet interfaces, routing, and forwarding.
2926  *
2927  * Users must call this function once for each successful call to LNetNIInit().
2928  * Once the LNetNIFini() operation has been started, the results of pending
2929  * API operations are undefined.
2930  *
2931  * \return always 0 for current implementation.
2932  */
2933 int
2934 LNetNIFini(void)
2935 {
2936         mutex_lock(&the_lnet.ln_api_mutex);
2937
2938         LASSERT(the_lnet.ln_refcount > 0);
2939
2940         if (the_lnet.ln_refcount != 1) {
2941                 the_lnet.ln_refcount--;
2942         } else {
2943                 LASSERT(!the_lnet.ln_niinit_self);
2944
2945                 lnet_fault_fini();
2946
2947                 lnet_router_debugfs_fini();
2948                 lnet_monitor_thr_stop();
2949                 lnet_peer_discovery_stop();
2950                 lnet_push_target_fini();
2951                 lnet_ping_target_fini();
2952
2953                 /* Teardown fns that use my own API functions BEFORE here */
2954                 the_lnet.ln_refcount = 0;
2955
2956                 lnet_acceptor_stop();
2957                 lnet_destroy_routes();
2958                 lnet_shutdown_lndnets();
2959                 lnet_unprepare();
2960         }
2961
2962         mutex_unlock(&the_lnet.ln_api_mutex);
2963         return 0;
2964 }
2965 EXPORT_SYMBOL(LNetNIFini);
2966
2967 /**
2968  * Grabs the ni data from the ni structure and fills the out
2969  * parameters
2970  *
2971  * \param[in] ni network        interface structure
2972  * \param[out] cfg_ni           NI config information
2973  * \param[out] tun              network and LND tunables
2974  */
2975 static void
2976 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2977                    struct lnet_ioctl_config_lnd_tunables *tun,
2978                    struct lnet_ioctl_element_stats *stats,
2979                    __u32 tun_size)
2980 {
2981         size_t min_size = 0;
2982         int i;
2983
2984         if (!ni || !cfg_ni || !tun)
2985                 return;
2986
2987         if (ni->ni_interfaces[0] != NULL) {
2988                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2989                         if (ni->ni_interfaces[i] != NULL) {
2990                                 strncpy(cfg_ni->lic_ni_intf[i],
2991                                         ni->ni_interfaces[i],
2992                                         sizeof(cfg_ni->lic_ni_intf[i]));
2993                         }
2994                 }
2995         }
2996
2997         cfg_ni->lic_nid = ni->ni_nid;
2998         if (ni->ni_nid == LNET_NID_LO_0)
2999                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
3000         else
3001                 cfg_ni->lic_status = ni->ni_status->ns_status;
3002         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
3003         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
3004
3005         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
3006
3007         if (stats) {
3008                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
3009                                                        LNET_STATS_TYPE_SEND);
3010                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
3011                                                        LNET_STATS_TYPE_RECV);
3012                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
3013                                                        LNET_STATS_TYPE_DROP);
3014         }
3015
3016         /*
3017          * tun->lt_tun will always be present, but in order to be
3018          * backwards compatible, we need to deal with the cases when
3019          * tun->lt_tun is smaller than what the kernel has, because it
3020          * comes from an older version of a userspace program, then we'll
3021          * need to copy as much information as we have available space.
3022          */
3023         min_size = tun_size - sizeof(tun->lt_cmn);
3024         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
3025
3026         /* copy over the cpts */
3027         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
3028             ni->ni_cpts == NULL)  {
3029                 for (i = 0; i < ni->ni_ncpts; i++)
3030                         cfg_ni->lic_cpts[i] = i;
3031         } else {
3032                 for (i = 0;
3033                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
3034                      i < LNET_MAX_SHOW_NUM_CPT;
3035                      i++)
3036                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
3037         }
3038         cfg_ni->lic_ncpts = ni->ni_ncpts;
3039 }
3040
3041 /**
3042  * NOTE: This is a legacy function left in the code to be backwards
3043  * compatible with older userspace programs. It should eventually be
3044  * removed.
3045  *
3046  * Grabs the ni data from the ni structure and fills the out
3047  * parameters
3048  *
3049  * \param[in] ni network        interface structure
3050  * \param[out] config           config information
3051  */
3052 static void
3053 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
3054                          struct lnet_ioctl_config_data *config)
3055 {
3056         struct lnet_ioctl_net_config *net_config;
3057         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
3058         size_t min_size, tunable_size = 0;
3059         int i;
3060
3061         if (!ni || !config)
3062                 return;
3063
3064         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
3065         if (!net_config)
3066                 return;
3067
3068         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
3069                      ARRAY_SIZE(net_config->ni_interfaces));
3070
3071         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
3072                 if (!ni->ni_interfaces[i])
3073                         break;
3074
3075                 strncpy(net_config->ni_interfaces[i],
3076                         ni->ni_interfaces[i],
3077                         sizeof(net_config->ni_interfaces[i]));
3078         }
3079
3080         config->cfg_nid = ni->ni_nid;
3081         config->cfg_config_u.cfg_net.net_peer_timeout =
3082                 ni->ni_net->net_tunables.lct_peer_timeout;
3083         config->cfg_config_u.cfg_net.net_max_tx_credits =
3084                 ni->ni_net->net_tunables.lct_max_tx_credits;
3085         config->cfg_config_u.cfg_net.net_peer_tx_credits =
3086                 ni->ni_net->net_tunables.lct_peer_tx_credits;
3087         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
3088                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
3089
3090         if (ni->ni_nid == LNET_NID_LO_0)
3091                 net_config->ni_status = LNET_NI_STATUS_UP;
3092         else
3093                 net_config->ni_status = ni->ni_status->ns_status;
3094
3095         if (ni->ni_cpts) {
3096                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
3097
3098                 for (i = 0; i < num_cpts; i++)
3099                         net_config->ni_cpts[i] = ni->ni_cpts[i];
3100
3101                 config->cfg_ncpts = num_cpts;
3102         }
3103
3104         /*
3105          * See if user land tools sent in a newer and larger version
3106          * of struct lnet_tunables than what the kernel uses.
3107          */
3108         min_size = sizeof(*config) + sizeof(*net_config);
3109
3110         if (config->cfg_hdr.ioc_len > min_size)
3111                 tunable_size = config->cfg_hdr.ioc_len - min_size;
3112
3113         /* Don't copy too much data to user space */
3114         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
3115         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
3116
3117         if (lnd_cfg && min_size) {
3118                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
3119                 config->cfg_config_u.cfg_net.net_interface_count = 1;
3120
3121                 /* Tell user land that kernel side has less data */
3122                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
3123                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
3124                         config->cfg_hdr.ioc_len -= min_size;
3125                 }
3126         }
3127 }
3128
3129 struct lnet_ni *
3130 lnet_get_ni_idx_locked(int idx)
3131 {
3132         struct lnet_ni          *ni;
3133         struct lnet_net         *net;
3134
3135         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3136                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3137                         if (idx-- == 0)
3138                                 return ni;
3139                 }
3140         }
3141
3142         return NULL;
3143 }
3144
3145 int lnet_get_net_healthv_locked(struct lnet_net *net)
3146 {
3147         struct lnet_ni *ni;
3148         int best_healthv = 0;
3149         int healthv;
3150
3151         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3152                 healthv = atomic_read(&ni->ni_healthv);
3153                 if (healthv > best_healthv)
3154                         best_healthv = healthv;
3155         }
3156
3157         return best_healthv;
3158 }
3159
3160 struct lnet_ni *
3161 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
3162 {
3163         struct lnet_ni          *ni;
3164         struct lnet_net         *net = mynet;
3165
3166         /*
3167          * It is possible that the net has been cleaned out while there is
3168          * a message being sent. This function accessed the net without
3169          * checking if the list is empty
3170          */
3171         if (prev == NULL) {
3172                 if (net == NULL)
3173                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
3174                                         net_list);
3175                 if (list_empty(&net->net_ni_list))
3176                         return NULL;
3177                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3178                                 ni_netlist);
3179
3180                 return ni;
3181         }
3182
3183         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
3184                 /* if you reached the end of the ni list and the net is
3185                  * specified, then there are no more nis in that net */
3186                 if (net != NULL)
3187                         return NULL;
3188
3189                 /* we reached the end of this net ni list. move to the
3190                  * next net */
3191                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
3192                         /* no more nets and no more NIs. */
3193                         return NULL;
3194
3195                 /* get the next net */
3196                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
3197                                  net_list);
3198                 if (list_empty(&net->net_ni_list))
3199                         return NULL;
3200                 /* get the ni on it */
3201                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3202                                 ni_netlist);
3203
3204                 return ni;
3205         }
3206
3207         if (list_empty(&prev->ni_netlist))
3208                 return NULL;
3209
3210         /* there are more nis left */
3211         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
3212
3213         return ni;
3214 }
3215
3216 int
3217 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3218 {
3219         struct lnet_ni *ni;
3220         int cpt;
3221         int rc = -ENOENT;
3222         int idx = config->cfg_count;
3223
3224         cpt = lnet_net_lock_current();
3225
3226         ni = lnet_get_ni_idx_locked(idx);
3227
3228         if (ni != NULL) {
3229                 rc = 0;
3230                 lnet_ni_lock(ni);
3231                 lnet_fill_ni_info_legacy(ni, config);
3232                 lnet_ni_unlock(ni);
3233         }
3234
3235         lnet_net_unlock(cpt);
3236         return rc;
3237 }
3238
3239 int
3240 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3241                    struct lnet_ioctl_config_lnd_tunables *tun,
3242                    struct lnet_ioctl_element_stats *stats,
3243                    __u32 tun_size)
3244 {
3245         struct lnet_ni          *ni;
3246         int                     cpt;
3247         int                     rc = -ENOENT;
3248
3249         if (!cfg_ni || !tun || !stats)
3250                 return -EINVAL;
3251
3252         cpt = lnet_net_lock_current();
3253
3254         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3255
3256         if (ni) {
3257                 rc = 0;
3258                 lnet_ni_lock(ni);
3259                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3260                 lnet_ni_unlock(ni);
3261         }
3262
3263         lnet_net_unlock(cpt);
3264         return rc;
3265 }
3266
3267 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3268 {
3269         struct lnet_ni *ni;
3270         int cpt;
3271         int rc = -ENOENT;
3272
3273         if (!msg_stats)
3274                 return -EINVAL;
3275
3276         cpt = lnet_net_lock_current();
3277
3278         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3279
3280         if (ni) {
3281                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3282                 rc = 0;
3283         }
3284
3285         lnet_net_unlock(cpt);
3286
3287         return rc;
3288 }
3289
3290 static int lnet_add_net_common(struct lnet_net *net,
3291                                struct lnet_ioctl_config_lnd_tunables *tun)
3292 {
3293         struct lnet_handle_md ping_mdh;
3294         struct lnet_ping_buffer *pbuf;
3295         struct lnet_remotenet *rnet;
3296         struct lnet_ni *ni;
3297         int net_ni_count;
3298         __u32 net_id;
3299         int rc;
3300
3301         lnet_net_lock(LNET_LOCK_EX);
3302         rnet = lnet_find_rnet_locked(net->net_id);
3303         lnet_net_unlock(LNET_LOCK_EX);
3304         /*
3305          * make sure that the net added doesn't invalidate the current
3306          * configuration LNet is keeping
3307          */
3308         if (rnet) {
3309                 CERROR("Adding net %s will invalidate routing configuration\n",
3310                        libcfs_net2str(net->net_id));
3311                 lnet_net_free(net);
3312                 return -EUSERS;
3313         }
3314
3315         /*
3316          * make sure you calculate the correct number of slots in the ping
3317          * buffer. Since the ping info is a flattened list of all the NIs,
3318          * we should allocate enough slots to accomodate the number of NIs
3319          * which will be added.
3320          *
3321          * since ni hasn't been configured yet, use
3322          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3323          */
3324         net_ni_count = lnet_get_net_ni_count_pre(net);
3325
3326         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3327                                     net_ni_count + lnet_get_ni_count(),
3328                                     false);
3329         if (rc < 0) {
3330                 lnet_net_free(net);
3331                 return rc;
3332         }
3333
3334         if (tun)
3335                 memcpy(&net->net_tunables,
3336                        &tun->lt_cmn, sizeof(net->net_tunables));
3337         else
3338                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3339
3340         net_id = net->net_id;
3341
3342         rc = lnet_startup_lndnet(net,
3343                                  (tun) ? &tun->lt_tun : NULL);
3344         if (rc < 0)
3345                 goto failed;
3346
3347         lnet_net_lock(LNET_LOCK_EX);
3348         net = lnet_get_net_locked(net_id);
3349         LASSERT(net);
3350
3351         /* apply the UDSPs */
3352         rc = lnet_udsp_apply_policies_on_net(net);
3353         if (rc)
3354                 CERROR("Failed to apply UDSPs on local net %s\n",
3355                        libcfs_net2str(net->net_id));
3356
3357         /* At this point we lost track of which NI was just added, so we
3358          * just re-apply the policies on all of the NIs on this net
3359          */
3360         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3361                 rc = lnet_udsp_apply_policies_on_ni(ni);
3362                 if (rc)
3363                         CERROR("Failed to apply UDSPs on ni %s\n",
3364                                libcfs_nid2str(ni->ni_nid));
3365         }
3366         lnet_net_unlock(LNET_LOCK_EX);
3367
3368         /*
3369          * Start the acceptor thread if this is the first network
3370          * being added that requires the thread.
3371          */
3372         if (net->net_lnd->lnd_accept) {
3373                 rc = lnet_acceptor_start();
3374                 if (rc < 0) {
3375                         /* shutdown the net that we just started */
3376                         CERROR("Failed to start up acceptor thread\n");
3377                         lnet_shutdown_lndnet(net);
3378                         goto failed;
3379                 }
3380         }
3381
3382         lnet_net_lock(LNET_LOCK_EX);
3383         lnet_peer_net_added(net);
3384         lnet_net_unlock(LNET_LOCK_EX);
3385
3386         lnet_ping_target_update(pbuf, ping_mdh);
3387
3388         return 0;
3389
3390 failed:
3391         lnet_ping_md_unlink(pbuf, &ping_mdh);
3392         lnet_ping_buffer_decref(pbuf);
3393         return rc;
3394 }
3395
3396 static void
3397 lnet_set_tune_defaults(struct lnet_ioctl_config_lnd_tunables *tun)
3398 {
3399         if (tun) {
3400                 if (!tun->lt_cmn.lct_peer_timeout)
3401                         tun->lt_cmn.lct_peer_timeout = DEFAULT_PEER_TIMEOUT;
3402                 if (!tun->lt_cmn.lct_peer_tx_credits)
3403                         tun->lt_cmn.lct_peer_tx_credits = DEFAULT_PEER_CREDITS;
3404                 if (!tun->lt_cmn.lct_max_tx_credits)
3405                         tun->lt_cmn.lct_max_tx_credits = DEFAULT_CREDITS;
3406         }
3407 }
3408
3409 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3410                                       struct lnet_ioctl_config_lnd_tunables *tun)
3411 {
3412         struct lnet_net *net;
3413         const char *nets;
3414         int rc;
3415         LIST_HEAD(net_head);
3416
3417         rc = lnet_parse_ip2nets(&nets, ip2nets);
3418         if (rc < 0)
3419                 return rc;
3420
3421         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3422         if (rc < 0)
3423                 return rc;
3424
3425         lnet_set_tune_defaults(tun);
3426
3427         mutex_lock(&the_lnet.ln_api_mutex);
3428         while (!list_empty(&net_head)) {
3429                 net = list_entry(net_head.next, struct lnet_net, net_list);
3430                 list_del_init(&net->net_list);
3431                 rc = lnet_add_net_common(net, tun);
3432                 if (rc < 0)
3433                         goto out;
3434         }
3435
3436 out:
3437         mutex_unlock(&the_lnet.ln_api_mutex);
3438
3439         while (!list_empty(&net_head)) {
3440                 net = list_entry(net_head.next, struct lnet_net, net_list);
3441                 list_del_init(&net->net_list);
3442                 lnet_net_free(net);
3443         }
3444         return rc;
3445 }
3446
3447 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3448 {
3449         struct lnet_net *net;
3450         struct lnet_ni *ni;
3451         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3452         int rc, i;
3453         __u32 net_id, lnd_type;
3454
3455         /* get the tunables if they are available */
3456         if (conf->lic_cfg_hdr.ioc_len >=
3457             sizeof(*conf) + sizeof(*tun))
3458                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3459                         conf->lic_bulk;
3460
3461         /* handle legacy ip2nets from DLC */
3462         if (conf->lic_legacy_ip2nets[0] != '\0')
3463                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3464                                                   tun);
3465
3466         net_id = LNET_NIDNET(conf->lic_nid);
3467         lnd_type = LNET_NETTYP(net_id);
3468
3469         if (!libcfs_isknown_lnd(lnd_type)) {
3470                 CERROR("No valid net and lnd information provided\n");
3471                 return -EINVAL;
3472         }
3473
3474         net = lnet_net_alloc(net_id, NULL);
3475         if (!net)
3476                 return -ENOMEM;
3477
3478         for (i = 0; i < conf->lic_ncpts; i++) {
3479                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3480                         return -EINVAL;
3481         }
3482
3483         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3484                                        conf->lic_ni_intf[0]);
3485         if (!ni)
3486                 return -ENOMEM;
3487
3488         lnet_set_tune_defaults(tun);
3489
3490         mutex_lock(&the_lnet.ln_api_mutex);
3491
3492         rc = lnet_add_net_common(net, tun);
3493
3494         mutex_unlock(&the_lnet.ln_api_mutex);
3495
3496         return rc;
3497 }
3498
3499 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3500 {
3501         struct lnet_net  *net;
3502         struct lnet_ni *ni;
3503         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3504         struct lnet_ping_buffer *pbuf;
3505         struct lnet_handle_md  ping_mdh;
3506         int               rc;
3507         int               net_count;
3508         __u32             addr;
3509
3510         /* don't allow userspace to shutdown the LOLND */
3511         if (LNET_NETTYP(net_id) == LOLND)
3512                 return -EINVAL;
3513
3514         mutex_lock(&the_lnet.ln_api_mutex);
3515
3516         lnet_net_lock(0);
3517
3518         net = lnet_get_net_locked(net_id);
3519         if (!net) {
3520                 CERROR("net %s not found\n",
3521                        libcfs_net2str(net_id));
3522                 rc = -ENOENT;
3523                 goto unlock_net;
3524         }
3525
3526         addr = LNET_NIDADDR(conf->lic_nid);
3527         if (addr == 0) {
3528                 /* remove the entire net */
3529                 net_count = lnet_get_net_ni_count_locked(net);
3530
3531                 lnet_net_unlock(0);
3532
3533                 /* create and link a new ping info, before removing the old one */
3534                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3535                                         lnet_get_ni_count() - net_count,
3536                                         false);
3537                 if (rc != 0)
3538                         goto unlock_api_mutex;
3539
3540                 lnet_shutdown_lndnet(net);
3541
3542                 lnet_acceptor_stop();
3543
3544                 lnet_ping_target_update(pbuf, ping_mdh);
3545
3546                 goto unlock_api_mutex;
3547         }
3548
3549         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3550         if (!ni) {
3551                 CERROR("nid %s not found\n",
3552                        libcfs_nid2str(conf->lic_nid));
3553                 rc = -ENOENT;
3554                 goto unlock_net;
3555         }
3556
3557         net_count = lnet_get_net_ni_count_locked(net);
3558
3559         lnet_net_unlock(0);
3560
3561         /* create and link a new ping info, before removing the old one */
3562         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3563                                   lnet_get_ni_count() - 1, false);
3564         if (rc != 0)
3565                 goto unlock_api_mutex;
3566
3567         lnet_shutdown_lndni(ni);
3568
3569         lnet_acceptor_stop();
3570
3571         lnet_ping_target_update(pbuf, ping_mdh);
3572
3573         /* check if the net is empty and remove it if it is */
3574         if (net_count == 1)
3575                 lnet_shutdown_lndnet(net);
3576
3577         goto unlock_api_mutex;
3578
3579 unlock_net:
3580         lnet_net_unlock(0);
3581 unlock_api_mutex:
3582         mutex_unlock(&the_lnet.ln_api_mutex);
3583
3584         return rc;
3585 }
3586
3587 /*
3588  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3589  * They are only expected to be called for unique networks.
3590  * That can be as a result of older DLC library
3591  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3592  */
3593 int
3594 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3595 {
3596         struct lnet_net *net;
3597         LIST_HEAD(net_head);
3598         int rc;
3599         struct lnet_ioctl_config_lnd_tunables tun;
3600         const char *nets = conf->cfg_config_u.cfg_net.net_intf;
3601
3602         /* Create a net/ni structures for the network string */
3603         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3604         if (rc <= 0)
3605                 return rc == 0 ? -EINVAL : rc;
3606
3607         mutex_lock(&the_lnet.ln_api_mutex);
3608
3609         if (rc > 1) {
3610                 rc = -EINVAL; /* only add one network per call */
3611                 goto out_unlock_clean;
3612         }
3613
3614         net = list_entry(net_head.next, struct lnet_net, net_list);
3615         list_del_init(&net->net_list);
3616
3617         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3618
3619         memset(&tun, 0, sizeof(tun));
3620
3621         tun.lt_cmn.lct_peer_timeout =
3622           (!conf->cfg_config_u.cfg_net.net_peer_timeout) ? DEFAULT_PEER_TIMEOUT :
3623                 conf->cfg_config_u.cfg_net.net_peer_timeout;
3624         tun.lt_cmn.lct_peer_tx_credits =
3625           (!conf->cfg_config_u.cfg_net.net_peer_tx_credits) ? DEFAULT_PEER_CREDITS :
3626                 conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3627         tun.lt_cmn.lct_peer_rtr_credits =
3628           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3629         tun.lt_cmn.lct_max_tx_credits =
3630           (!conf->cfg_config_u.cfg_net.net_max_tx_credits) ? DEFAULT_CREDITS :
3631                 conf->cfg_config_u.cfg_net.net_max_tx_credits;
3632
3633         rc = lnet_add_net_common(net, &tun);
3634
3635 out_unlock_clean:
3636         mutex_unlock(&the_lnet.ln_api_mutex);
3637         while (!list_empty(&net_head)) {
3638                 /* net_head list is empty in success case */
3639                 net = list_entry(net_head.next, struct lnet_net, net_list);
3640                 list_del_init(&net->net_list);
3641                 lnet_net_free(net);
3642         }
3643         return rc;
3644 }
3645
3646 int
3647 lnet_dyn_del_net(__u32 net_id)
3648 {
3649         struct lnet_net  *net;
3650         struct lnet_ping_buffer *pbuf;
3651         struct lnet_handle_md ping_mdh;
3652         int               rc;
3653         int               net_ni_count;
3654
3655         /* don't allow userspace to shutdown the LOLND */
3656         if (LNET_NETTYP(net_id) == LOLND)
3657                 return -EINVAL;
3658
3659         mutex_lock(&the_lnet.ln_api_mutex);
3660
3661         lnet_net_lock(0);
3662
3663         net = lnet_get_net_locked(net_id);
3664         if (net == NULL) {
3665                 lnet_net_unlock(0);
3666                 rc = -EINVAL;
3667                 goto out;
3668         }
3669
3670         net_ni_count = lnet_get_net_ni_count_locked(net);
3671
3672         lnet_net_unlock(0);
3673
3674         /* create and link a new ping info, before removing the old one */
3675         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3676                                     lnet_get_ni_count() - net_ni_count, false);
3677         if (rc != 0)
3678                 goto out;
3679
3680         lnet_shutdown_lndnet(net);
3681
3682         lnet_acceptor_stop();
3683
3684         lnet_ping_target_update(pbuf, ping_mdh);
3685
3686 out:
3687         mutex_unlock(&the_lnet.ln_api_mutex);
3688
3689         return rc;
3690 }
3691
3692 void lnet_incr_dlc_seq(void)
3693 {
3694         atomic_inc(&lnet_dlc_seq_no);
3695 }
3696
3697 __u32 lnet_get_dlc_seq_locked(void)
3698 {
3699         return atomic_read(&lnet_dlc_seq_no);
3700 }
3701
3702 static void
3703 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3704 {
3705         struct lnet_net *net;
3706         struct lnet_ni *ni;
3707
3708         lnet_net_lock(LNET_LOCK_EX);
3709         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3710                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3711                         if (ni->ni_nid == nid || all) {
3712                                 atomic_set(&ni->ni_healthv, value);
3713                                 if (list_empty(&ni->ni_recovery) &&
3714                                     value < LNET_MAX_HEALTH_VALUE) {
3715                                         CERROR("manually adding local NI %s to recovery\n",
3716                                                libcfs_nid2str(ni->ni_nid));
3717                                         list_add_tail(&ni->ni_recovery,
3718                                                       &the_lnet.ln_mt_localNIRecovq);
3719                                         lnet_ni_addref_locked(ni, 0);
3720                                 }
3721                                 if (!all) {
3722                                         lnet_net_unlock(LNET_LOCK_EX);
3723                                         return;
3724                                 }
3725                         }
3726                 }
3727         }
3728         lnet_net_unlock(LNET_LOCK_EX);
3729 }
3730
3731 static int
3732 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3733 {
3734         int cpt, rc = 0;
3735         struct lnet_ni *ni;
3736         lnet_nid_t nid = stats->hlni_nid;
3737
3738         cpt = lnet_net_lock_current();
3739         ni = lnet_nid2ni_locked(nid, cpt);
3740
3741         if (!ni) {
3742                 rc = -ENOENT;
3743                 goto unlock;
3744         }
3745
3746         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3747         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3748         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3749         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3750         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3751         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3752         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3753
3754 unlock:
3755         lnet_net_unlock(cpt);
3756
3757         return rc;
3758 }
3759
3760 static int
3761 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3762 {
3763         struct lnet_ni *ni;
3764         int i = 0;
3765
3766         lnet_net_lock(LNET_LOCK_EX);
3767         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
3768                 list->rlst_nid_array[i] = ni->ni_nid;
3769                 i++;
3770                 if (i >= LNET_MAX_SHOW_NUM_NID)
3771                         break;
3772         }
3773         lnet_net_unlock(LNET_LOCK_EX);
3774         list->rlst_num_nids = i;
3775
3776         return 0;
3777 }
3778
3779 static int
3780 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3781 {
3782         struct lnet_peer_ni *lpni;
3783         int i = 0;
3784
3785         lnet_net_lock(LNET_LOCK_EX);
3786         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
3787                 list->rlst_nid_array[i] = lpni->lpni_nid;
3788                 i++;
3789                 if (i >= LNET_MAX_SHOW_NUM_NID)
3790                         break;
3791         }
3792         lnet_net_unlock(LNET_LOCK_EX);
3793         list->rlst_num_nids = i;
3794
3795         return 0;
3796 }
3797
3798 /**
3799  * LNet ioctl handler.
3800  *
3801  */
3802 int
3803 LNetCtl(unsigned int cmd, void *arg)
3804 {
3805         struct libcfs_ioctl_data *data = arg;
3806         struct lnet_ioctl_config_data *config;
3807         struct lnet_process_id    id = {0};
3808         struct lnet_ni           *ni;
3809         int                       rc;
3810
3811         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
3812                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
3813
3814         switch (cmd) {
3815         case IOC_LIBCFS_GET_NI:
3816                 rc = LNetGetId(data->ioc_count, &id);
3817                 data->ioc_nid = id.nid;
3818                 return rc;
3819
3820         case IOC_LIBCFS_FAIL_NID:
3821                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
3822
3823         case IOC_LIBCFS_ADD_ROUTE: {
3824                 /* default router sensitivity to 1 */
3825                 unsigned int sensitivity = 1;
3826                 config = arg;
3827
3828                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3829                         return -EINVAL;
3830
3831                 if (config->cfg_config_u.cfg_route.rtr_sensitivity) {
3832                         sensitivity =
3833                           config->cfg_config_u.cfg_route.rtr_sensitivity;
3834                 }
3835
3836                 mutex_lock(&the_lnet.ln_api_mutex);
3837                 rc = lnet_add_route(config->cfg_net,
3838                                     config->cfg_config_u.cfg_route.rtr_hop,
3839                                     config->cfg_nid,
3840                                     config->cfg_config_u.cfg_route.
3841                                         rtr_priority, sensitivity);
3842                 mutex_unlock(&the_lnet.ln_api_mutex);
3843                 return rc;
3844         }
3845
3846         case IOC_LIBCFS_DEL_ROUTE:
3847                 config = arg;
3848
3849                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3850                         return -EINVAL;
3851
3852                 mutex_lock(&the_lnet.ln_api_mutex);
3853                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
3854                 mutex_unlock(&the_lnet.ln_api_mutex);
3855                 return rc;
3856
3857         case IOC_LIBCFS_GET_ROUTE:
3858                 config = arg;
3859
3860                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3861                         return -EINVAL;
3862
3863                 mutex_lock(&the_lnet.ln_api_mutex);
3864                 rc = lnet_get_route(config->cfg_count,
3865                                     &config->cfg_net,
3866                                     &config->cfg_config_u.cfg_route.rtr_hop,
3867                                     &config->cfg_nid,
3868                                     &config->cfg_config_u.cfg_route.rtr_flags,
3869                                     &config->cfg_config_u.cfg_route.
3870                                         rtr_priority,
3871                                     &config->cfg_config_u.cfg_route.
3872                                         rtr_sensitivity);
3873                 mutex_unlock(&the_lnet.ln_api_mutex);
3874                 return rc;
3875
3876         case IOC_LIBCFS_GET_LOCAL_NI: {
3877                 struct lnet_ioctl_config_ni *cfg_ni;
3878                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3879                 struct lnet_ioctl_element_stats *stats;
3880                 __u32 tun_size;
3881
3882                 cfg_ni = arg;
3883
3884                 /* get the tunables if they are available */
3885                 if (cfg_ni->lic_cfg_hdr.ioc_len <
3886                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
3887                         return -EINVAL;
3888
3889                 stats = (struct lnet_ioctl_element_stats *)
3890                         cfg_ni->lic_bulk;
3891                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3892                                 (cfg_ni->lic_bulk + sizeof(*stats));
3893
3894                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
3895                         sizeof(*stats);
3896
3897                 mutex_lock(&the_lnet.ln_api_mutex);
3898                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
3899                 mutex_unlock(&the_lnet.ln_api_mutex);
3900                 return rc;
3901         }
3902
3903         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
3904                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
3905
3906                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
3907                         return -EINVAL;
3908
3909                 mutex_lock(&the_lnet.ln_api_mutex);
3910                 rc = lnet_get_ni_stats(msg_stats);
3911                 mutex_unlock(&the_lnet.ln_api_mutex);
3912
3913                 return rc;
3914         }
3915
3916         case IOC_LIBCFS_GET_NET: {
3917                 size_t total = sizeof(*config) +
3918                                sizeof(struct lnet_ioctl_net_config);
3919                 config = arg;
3920
3921                 if (config->cfg_hdr.ioc_len < total)
3922                         return -EINVAL;
3923
3924                 mutex_lock(&the_lnet.ln_api_mutex);
3925                 rc = lnet_get_net_config(config);
3926                 mutex_unlock(&the_lnet.ln_api_mutex);
3927                 return rc;
3928         }
3929
3930         case IOC_LIBCFS_GET_LNET_STATS:
3931         {
3932                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
3933
3934                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
3935                         return -EINVAL;
3936
3937                 mutex_lock(&the_lnet.ln_api_mutex);
3938                 rc = lnet_counters_get(&lnet_stats->st_cntrs);
3939                 mutex_unlock(&the_lnet.ln_api_mutex);
3940                 return rc;
3941         }
3942
3943         case IOC_LIBCFS_CONFIG_RTR:
3944                 config = arg;
3945
3946                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3947                         return -EINVAL;
3948
3949                 mutex_lock(&the_lnet.ln_api_mutex);
3950                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
3951                         rc = lnet_rtrpools_enable();
3952                         mutex_unlock(&the_lnet.ln_api_mutex);
3953                         return rc;
3954                 }
3955                 lnet_rtrpools_disable();
3956                 mutex_unlock(&the_lnet.ln_api_mutex);
3957                 return 0;
3958
3959         case IOC_LIBCFS_ADD_BUF:
3960                 config = arg;
3961
3962                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3963                         return -EINVAL;
3964
3965                 mutex_lock(&the_lnet.ln_api_mutex);
3966                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
3967                                                 buf_tiny,
3968                                           config->cfg_config_u.cfg_buffers.
3969                                                 buf_small,
3970                                           config->cfg_config_u.cfg_buffers.
3971                                                 buf_large);
3972                 mutex_unlock(&the_lnet.ln_api_mutex);
3973                 return rc;
3974
3975         case IOC_LIBCFS_SET_NUMA_RANGE: {
3976                 struct lnet_ioctl_set_value *numa;
3977                 numa = arg;
3978                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3979                         return -EINVAL;
3980                 lnet_net_lock(LNET_LOCK_EX);
3981                 lnet_numa_range = numa->sv_value;
3982                 lnet_net_unlock(LNET_LOCK_EX);
3983                 return 0;
3984         }
3985
3986         case IOC_LIBCFS_GET_NUMA_RANGE: {
3987                 struct lnet_ioctl_set_value *numa;
3988                 numa = arg;
3989                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3990                         return -EINVAL;
3991                 numa->sv_value = lnet_numa_range;
3992                 return 0;
3993         }
3994
3995         case IOC_LIBCFS_GET_BUF: {
3996                 struct lnet_ioctl_pool_cfg *pool_cfg;
3997                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
3998
3999                 config = arg;
4000
4001                 if (config->cfg_hdr.ioc_len < total)
4002                         return -EINVAL;
4003
4004                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
4005
4006                 mutex_lock(&the_lnet.ln_api_mutex);
4007                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
4008                 mutex_unlock(&the_lnet.ln_api_mutex);
4009                 return rc;
4010         }
4011
4012         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
4013                 struct lnet_ioctl_local_ni_hstats *stats = arg;
4014
4015                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
4016                         return -EINVAL;
4017
4018                 mutex_lock(&the_lnet.ln_api_mutex);
4019                 rc = lnet_get_local_ni_hstats(stats);
4020                 mutex_unlock(&the_lnet.ln_api_mutex);
4021
4022                 return rc;
4023         }
4024
4025         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
4026                 struct lnet_ioctl_recovery_list *list = arg;
4027                 if (list->rlst_hdr.ioc_len < sizeof(*list))
4028                         return -EINVAL;
4029
4030                 mutex_lock(&the_lnet.ln_api_mutex);
4031                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
4032                         rc = lnet_get_local_ni_recovery_list(list);
4033                 else
4034                         rc = lnet_get_peer_ni_recovery_list(list);
4035                 mutex_unlock(&the_lnet.ln_api_mutex);
4036                 return rc;
4037         }
4038
4039         case IOC_LIBCFS_ADD_PEER_NI: {
4040                 struct lnet_ioctl_peer_cfg *cfg = arg;
4041
4042                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4043                         return -EINVAL;
4044
4045                 mutex_lock(&the_lnet.ln_api_mutex);
4046                 rc = lnet_add_peer_ni(cfg->prcfg_prim_nid,
4047                                       cfg->prcfg_cfg_nid,
4048                                       cfg->prcfg_mr);
4049                 mutex_unlock(&the_lnet.ln_api_mutex);
4050                 return rc;
4051         }
4052
4053         case IOC_LIBCFS_DEL_PEER_NI: {
4054                 struct lnet_ioctl_peer_cfg *cfg = arg;
4055
4056                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4057                         return -EINVAL;
4058
4059                 mutex_lock(&the_lnet.ln_api_mutex);
4060                 rc = lnet_del_peer_ni(cfg->prcfg_prim_nid,
4061                                       cfg->prcfg_cfg_nid);
4062                 mutex_unlock(&the_lnet.ln_api_mutex);
4063                 return rc;
4064         }
4065
4066         case IOC_LIBCFS_GET_PEER_INFO: {
4067                 struct lnet_ioctl_peer *peer_info = arg;
4068
4069                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
4070                         return -EINVAL;
4071
4072                 mutex_lock(&the_lnet.ln_api_mutex);
4073                 rc = lnet_get_peer_ni_info(
4074                    peer_info->pr_count,
4075                    &peer_info->pr_nid,
4076                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
4077                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
4078                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
4079                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
4080                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
4081                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
4082                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
4083                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
4084                 mutex_unlock(&the_lnet.ln_api_mutex);
4085                 return rc;
4086         }
4087
4088         case IOC_LIBCFS_GET_PEER_NI: {
4089                 struct lnet_ioctl_peer_cfg *cfg = arg;
4090
4091                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4092                         return -EINVAL;
4093 </