Whamcloud - gitweb
eaf2078af8cef24ecd83d6a591e32dc4076e30dc
[fs/lustre-release.git] / lnet / lnet / net_fault.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, write to the
18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19  * Boston, MA 021110-1307, USA
20  *
21  * GPL HEADER END
22  */
23 /*
24  * Copyright (c) 2014, 2017, Intel Corporation.
25  */
26 /*
27  * This file is part of Lustre, http://www.lustre.org/
28  * Lustre is a trademark of Sun Microsystems, Inc.
29  *
30  * lnet/lnet/net_fault.c
31  *
32  * Lustre network fault simulation
33  *
34  * Author: liang.zhen@intel.com
35  */
36
37 #define DEBUG_SUBSYSTEM S_LNET
38
39 #include <lnet/lib-lnet.h>
40 #include <uapi/linux/lnet/lnetctl.h>
41
42 #define LNET_MSG_MASK           (LNET_PUT_BIT | LNET_ACK_BIT | \
43                                  LNET_GET_BIT | LNET_REPLY_BIT)
44
45 struct lnet_drop_rule {
46         /** link chain on the_lnet.ln_drop_rules */
47         struct list_head        dr_link;
48         /** attributes of this rule */
49         struct lnet_fault_attr  dr_attr;
50         /** lock to protect \a dr_drop_at and \a dr_stat */
51         spinlock_t              dr_lock;
52         /**
53          * the message sequence to drop, which means message is dropped when
54          * dr_stat.drs_count == dr_drop_at
55          */
56         unsigned long           dr_drop_at;
57         /**
58          * seconds to drop the next message, it's exclusive with dr_drop_at
59          */
60         cfs_time_t              dr_drop_time;
61         /** baseline to caculate dr_drop_time */
62         cfs_time_t              dr_time_base;
63         /** statistic of dropped messages */
64         struct lnet_fault_stat  dr_stat;
65 };
66
67 static bool
68 lnet_fault_nid_match(lnet_nid_t nid, lnet_nid_t msg_nid)
69 {
70         if (nid == msg_nid || nid == LNET_NID_ANY)
71                 return true;
72
73         if (LNET_NIDNET(nid) != LNET_NIDNET(msg_nid))
74                 return false;
75
76         /* 255.255.255.255@net is wildcard for all addresses in a network */
77         return LNET_NIDADDR(nid) == LNET_NIDADDR(LNET_NID_ANY);
78 }
79
80 static bool
81 lnet_fault_attr_match(struct lnet_fault_attr *attr, lnet_nid_t src,
82                       lnet_nid_t dst, unsigned int type, unsigned int portal)
83 {
84         if (!lnet_fault_nid_match(attr->fa_src, src) ||
85             !lnet_fault_nid_match(attr->fa_dst, dst))
86                 return false;
87
88         if (!(attr->fa_msg_mask & (1 << type)))
89                 return false;
90
91         /* NB: ACK and REPLY have no portal, but they should have been
92          * rejected by message mask */
93         if (attr->fa_ptl_mask != 0 && /* has portal filter */
94             !(attr->fa_ptl_mask & (1ULL << portal)))
95                 return false;
96
97         return true;
98 }
99
100 static int
101 lnet_fault_attr_validate(struct lnet_fault_attr *attr)
102 {
103         if (attr->fa_msg_mask == 0)
104                 attr->fa_msg_mask = LNET_MSG_MASK; /* all message types */
105
106         if (attr->fa_ptl_mask == 0) /* no portal filter */
107                 return 0;
108
109         /* NB: only PUT and GET can be filtered if portal filter has been set */
110         attr->fa_msg_mask &= LNET_GET_BIT | LNET_PUT_BIT;
111         if (attr->fa_msg_mask == 0) {
112                 CDEBUG(D_NET, "can't find valid message type bits %x\n",
113                        attr->fa_msg_mask);
114                 return -EINVAL;
115         }
116         return 0;
117 }
118
119 static void
120 lnet_fault_stat_inc(struct lnet_fault_stat *stat, unsigned int type)
121 {
122         /* NB: fs_counter is NOT updated by this function */
123         switch (type) {
124         case LNET_MSG_PUT:
125                 stat->fs_put++;
126                 return;
127         case LNET_MSG_ACK:
128                 stat->fs_ack++;
129                 return;
130         case LNET_MSG_GET:
131                 stat->fs_get++;
132                 return;
133         case LNET_MSG_REPLY:
134                 stat->fs_reply++;
135                 return;
136         }
137 }
138
139 /**
140  * LNet message drop simulation
141  */
142
143 /**
144  * Add a new drop rule to LNet
145  * There is no check for duplicated drop rule, all rules will be checked for
146  * incoming message.
147  */
148 static int
149 lnet_drop_rule_add(struct lnet_fault_attr *attr)
150 {
151         struct lnet_drop_rule *rule;
152         ENTRY;
153
154         if (!((attr->u.drop.da_rate == 0) ^ (attr->u.drop.da_interval == 0))) {
155                 CDEBUG(D_NET,
156                        "please provide either drop rate or drop interval, "
157                        "but not both at the same time %d/%d\n",
158                        attr->u.drop.da_rate, attr->u.drop.da_interval);
159                 RETURN(-EINVAL);
160         }
161
162         if (lnet_fault_attr_validate(attr) != 0)
163                 RETURN(-EINVAL);
164
165         CFS_ALLOC_PTR(rule);
166         if (rule == NULL)
167                 RETURN(-ENOMEM);
168
169         spin_lock_init(&rule->dr_lock);
170
171         rule->dr_attr = *attr;
172         if (attr->u.drop.da_interval != 0) {
173                 rule->dr_time_base = cfs_time_shift(attr->u.drop.da_interval);
174                 rule->dr_drop_time = cfs_time_shift(cfs_rand() %
175                                                     attr->u.drop.da_interval);
176         } else {
177                 rule->dr_drop_at = cfs_rand() % attr->u.drop.da_rate;
178         }
179
180         lnet_net_lock(LNET_LOCK_EX);
181         list_add(&rule->dr_link, &the_lnet.ln_drop_rules);
182         lnet_net_unlock(LNET_LOCK_EX);
183
184         CDEBUG(D_NET, "Added drop rule: src %s, dst %s, rate %d, interval %d\n",
185                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
186                attr->u.drop.da_rate, attr->u.drop.da_interval);
187         RETURN(0);
188 }
189
190 /**
191  * Remove matched drop rules from lnet, all rules that can match \a src and
192  * \a dst will be removed.
193  * If \a src is zero, then all rules have \a dst as destination will be remove
194  * If \a dst is zero, then all rules have \a src as source will be removed
195  * If both of them are zero, all rules will be removed
196  */
197 static int
198 lnet_drop_rule_del(lnet_nid_t src, lnet_nid_t dst)
199 {
200         struct lnet_drop_rule *rule;
201         struct lnet_drop_rule *tmp;
202         struct list_head       zombies;
203         int                    n = 0;
204         ENTRY;
205
206         INIT_LIST_HEAD(&zombies);
207
208         lnet_net_lock(LNET_LOCK_EX);
209         list_for_each_entry_safe(rule, tmp, &the_lnet.ln_drop_rules, dr_link) {
210                 if (rule->dr_attr.fa_src != src && src != 0)
211                         continue;
212
213                 if (rule->dr_attr.fa_dst != dst && dst != 0)
214                         continue;
215
216                 list_move(&rule->dr_link, &zombies);
217         }
218         lnet_net_unlock(LNET_LOCK_EX);
219
220         list_for_each_entry_safe(rule, tmp, &zombies, dr_link) {
221                 CDEBUG(D_NET, "Remove drop rule: src %s->dst: %s (1/%d, %d)\n",
222                        libcfs_nid2str(rule->dr_attr.fa_src),
223                        libcfs_nid2str(rule->dr_attr.fa_dst),
224                        rule->dr_attr.u.drop.da_rate,
225                        rule->dr_attr.u.drop.da_interval);
226
227                 list_del(&rule->dr_link);
228                 CFS_FREE_PTR(rule);
229                 n++;
230         }
231
232         RETURN(n);
233 }
234
235 /**
236  * List drop rule at position of \a pos
237  */
238 static int
239 lnet_drop_rule_list(int pos, struct lnet_fault_attr *attr,
240                     struct lnet_fault_stat *stat)
241 {
242         struct lnet_drop_rule *rule;
243         int                    cpt;
244         int                    i = 0;
245         int                    rc = -ENOENT;
246         ENTRY;
247
248         cpt = lnet_net_lock_current();
249         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
250                 if (i++ < pos)
251                         continue;
252
253                 spin_lock(&rule->dr_lock);
254                 *attr = rule->dr_attr;
255                 *stat = rule->dr_stat;
256                 spin_unlock(&rule->dr_lock);
257                 rc = 0;
258                 break;
259         }
260
261         lnet_net_unlock(cpt);
262         RETURN(rc);
263 }
264
265 /**
266  * reset counters for all drop rules
267  */
268 static void
269 lnet_drop_rule_reset(void)
270 {
271         struct lnet_drop_rule *rule;
272         int                    cpt;
273         ENTRY;
274
275         cpt = lnet_net_lock_current();
276
277         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
278                 struct lnet_fault_attr *attr = &rule->dr_attr;
279
280                 spin_lock(&rule->dr_lock);
281
282                 memset(&rule->dr_stat, 0, sizeof(rule->dr_stat));
283                 if (attr->u.drop.da_rate != 0) {
284                         rule->dr_drop_at = cfs_rand() % attr->u.drop.da_rate;
285                 } else {
286                         rule->dr_drop_time = cfs_time_shift(cfs_rand() %
287                                                 attr->u.drop.da_interval);
288                         rule->dr_time_base = cfs_time_shift(attr->u.drop.
289                                                                   da_interval);
290                 }
291                 spin_unlock(&rule->dr_lock);
292         }
293
294         lnet_net_unlock(cpt);
295         EXIT;
296 }
297
298 /**
299  * check source/destination NID, portal, message type and drop rate,
300  * decide whether should drop this message or not
301  */
302 static bool
303 drop_rule_match(struct lnet_drop_rule *rule, lnet_nid_t src,
304                 lnet_nid_t dst, unsigned int type, unsigned int portal)
305 {
306         struct lnet_fault_attr  *attr = &rule->dr_attr;
307         bool                     drop;
308
309         if (!lnet_fault_attr_match(attr, src, dst, type, portal))
310                 return false;
311
312         /* match this rule, check drop rate now */
313         spin_lock(&rule->dr_lock);
314         if (rule->dr_drop_time != 0) { /* time based drop */
315                 cfs_time_t now = cfs_time_current();
316
317                 rule->dr_stat.fs_count++;
318                 drop = cfs_time_aftereq(now, rule->dr_drop_time);
319                 if (drop) {
320                         if (cfs_time_after(now, rule->dr_time_base))
321                                 rule->dr_time_base = now;
322
323                         rule->dr_drop_time = rule->dr_time_base +
324                                              cfs_time_seconds(cfs_rand() %
325                                                 attr->u.drop.da_interval);
326                         rule->dr_time_base += cfs_time_seconds(attr->u.drop.
327                                                                da_interval);
328
329                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop : %ld\n",
330                                libcfs_nid2str(attr->fa_src),
331                                libcfs_nid2str(attr->fa_dst),
332                                rule->dr_drop_time);
333                 }
334
335         } else { /* rate based drop */
336                 __u64 count;
337
338                 drop = rule->dr_stat.fs_count++ == rule->dr_drop_at;
339                 count = rule->dr_stat.fs_count;
340                 if (do_div(count, attr->u.drop.da_rate) == 0) {
341                         rule->dr_drop_at = rule->dr_stat.fs_count +
342                                            cfs_rand() % attr->u.drop.da_rate;
343                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop: %lu\n",
344                                libcfs_nid2str(attr->fa_src),
345                                libcfs_nid2str(attr->fa_dst), rule->dr_drop_at);
346                 }
347         }
348
349         if (drop) { /* drop this message, update counters */
350                 lnet_fault_stat_inc(&rule->dr_stat, type);
351                 rule->dr_stat.u.drop.ds_dropped++;
352         }
353
354         spin_unlock(&rule->dr_lock);
355         return drop;
356 }
357
358 /**
359  * Check if message from \a src to \a dst can match any existed drop rule
360  */
361 bool
362 lnet_drop_rule_match(struct lnet_hdr *hdr)
363 {
364         struct lnet_drop_rule   *rule;
365         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
366         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
367         unsigned int             typ = le32_to_cpu(hdr->type);
368         unsigned int             ptl = -1;
369         bool                     drop = false;
370         int                      cpt;
371
372         /* NB: if Portal is specified, then only PUT and GET will be
373          * filtered by drop rule */
374         if (typ == LNET_MSG_PUT)
375                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
376         else if (typ == LNET_MSG_GET)
377                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
378
379         cpt = lnet_net_lock_current();
380         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
381                 drop = drop_rule_match(rule, src, dst, typ, ptl);
382                 if (drop)
383                         break;
384         }
385
386         lnet_net_unlock(cpt);
387         return drop;
388 }
389
390 /**
391  * LNet Delay Simulation
392  */
393 /** timestamp (second) to send delayed message */
394 #define msg_delay_send           msg_ev.hdr_data
395
396 struct lnet_delay_rule {
397         /** link chain on the_lnet.ln_delay_rules */
398         struct list_head        dl_link;
399         /** link chain on delay_dd.dd_sched_rules */
400         struct list_head        dl_sched_link;
401         /** attributes of this rule */
402         struct lnet_fault_attr  dl_attr;
403         /** lock to protect \a below members */
404         spinlock_t              dl_lock;
405         /** refcount of delay rule */
406         atomic_t                dl_refcount;
407         /**
408          * the message sequence to delay, which means message is delayed when
409          * dl_stat.fs_count == dl_delay_at
410          */
411         unsigned long           dl_delay_at;
412         /**
413          * seconds to delay the next message, it's exclusive with dl_delay_at
414          */
415         cfs_time_t              dl_delay_time;
416         /** baseline to caculate dl_delay_time */
417         cfs_time_t              dl_time_base;
418         /** jiffies to send the next delayed message */
419         unsigned long           dl_msg_send;
420         /** delayed message list */
421         struct list_head        dl_msg_list;
422         /** statistic of delayed messages */
423         struct lnet_fault_stat  dl_stat;
424         /** timer to wakeup delay_daemon */
425         struct timer_list       dl_timer;
426 };
427
428 struct delay_daemon_data {
429         /** serialise rule add/remove */
430         struct mutex            dd_mutex;
431         /** protect rules on \a dd_sched_rules */
432         spinlock_t              dd_lock;
433         /** scheduled delay rules (by timer) */
434         struct list_head        dd_sched_rules;
435         /** deamon thread sleeps at here */
436         wait_queue_head_t       dd_waitq;
437         /** controler (lctl command) wait at here */
438         wait_queue_head_t       dd_ctl_waitq;
439         /** deamon is running */
440         unsigned int            dd_running;
441         /** deamon stopped */
442         unsigned int            dd_stopped;
443 };
444
445 static struct delay_daemon_data delay_dd;
446
447 static cfs_time_t
448 round_timeout(cfs_time_t timeout)
449 {
450         return cfs_time_seconds((unsigned int)
451                         cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
452 }
453
454 static void
455 delay_rule_decref(struct lnet_delay_rule *rule)
456 {
457         if (atomic_dec_and_test(&rule->dl_refcount)) {
458                 LASSERT(list_empty(&rule->dl_sched_link));
459                 LASSERT(list_empty(&rule->dl_msg_list));
460                 LASSERT(list_empty(&rule->dl_link));
461
462                 CFS_FREE_PTR(rule);
463         }
464 }
465
466 /**
467  * check source/destination NID, portal, message type and delay rate,
468  * decide whether should delay this message or not
469  */
470 static bool
471 delay_rule_match(struct lnet_delay_rule *rule, lnet_nid_t src,
472                 lnet_nid_t dst, unsigned int type, unsigned int portal,
473                 struct lnet_msg *msg)
474 {
475         struct lnet_fault_attr  *attr = &rule->dl_attr;
476         bool                     delay;
477
478         if (!lnet_fault_attr_match(attr, src, dst, type, portal))
479                 return false;
480
481         /* match this rule, check delay rate now */
482         spin_lock(&rule->dl_lock);
483         if (rule->dl_delay_time != 0) { /* time based delay */
484                 cfs_time_t now = cfs_time_current();
485
486                 rule->dl_stat.fs_count++;
487                 delay = cfs_time_aftereq(now, rule->dl_delay_time);
488                 if (delay) {
489                         if (cfs_time_after(now, rule->dl_time_base))
490                                 rule->dl_time_base = now;
491
492                         rule->dl_delay_time = rule->dl_time_base +
493                                              cfs_time_seconds(cfs_rand() %
494                                                 attr->u.delay.la_interval);
495                         rule->dl_time_base += cfs_time_seconds(attr->u.delay.
496                                                                la_interval);
497
498                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay : %ld\n",
499                                libcfs_nid2str(attr->fa_src),
500                                libcfs_nid2str(attr->fa_dst),
501                                rule->dl_delay_time);
502                 }
503
504         } else { /* rate based delay */
505                 __u64 count;
506
507                 delay = rule->dl_stat.fs_count++ == rule->dl_delay_at;
508                 /* generate the next random rate sequence */
509                 count = rule->dl_stat.fs_count;
510                 if (do_div(count, attr->u.delay.la_rate) == 0) {
511                         rule->dl_delay_at = rule->dl_stat.fs_count +
512                                             cfs_rand() % attr->u.delay.la_rate;
513                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay: %lu\n",
514                                libcfs_nid2str(attr->fa_src),
515                                libcfs_nid2str(attr->fa_dst), rule->dl_delay_at);
516                 }
517         }
518
519         if (!delay) {
520                 spin_unlock(&rule->dl_lock);
521                 return false;
522         }
523
524         /* delay this message, update counters */
525         lnet_fault_stat_inc(&rule->dl_stat, type);
526         rule->dl_stat.u.delay.ls_delayed++;
527
528         list_add_tail(&msg->msg_list, &rule->dl_msg_list);
529         msg->msg_delay_send = round_timeout(
530                         cfs_time_shift(attr->u.delay.la_latency));
531         if (rule->dl_msg_send == -1) {
532                 rule->dl_msg_send = msg->msg_delay_send;
533                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
534         }
535
536         spin_unlock(&rule->dl_lock);
537         return true;
538 }
539
540 /**
541  * check if \a msg can match any Delay Rule, receiving of this message
542  * will be delayed if there is a match.
543  */
544 bool
545 lnet_delay_rule_match_locked(struct lnet_hdr *hdr, struct lnet_msg *msg)
546 {
547         struct lnet_delay_rule  *rule;
548         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
549         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
550         unsigned int             typ = le32_to_cpu(hdr->type);
551         unsigned int             ptl = -1;
552
553         /* NB: called with hold of lnet_net_lock */
554
555         /* NB: if Portal is specified, then only PUT and GET will be
556          * filtered by delay rule */
557         if (typ == LNET_MSG_PUT)
558                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
559         else if (typ == LNET_MSG_GET)
560                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
561
562         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
563                 if (delay_rule_match(rule, src, dst, typ, ptl, msg))
564                         return true;
565         }
566
567         return false;
568 }
569
570 /** check out delayed messages for send */
571 static void
572 delayed_msg_check(struct lnet_delay_rule *rule, bool all,
573                   struct list_head *msg_list)
574 {
575         struct lnet_msg *msg;
576         struct lnet_msg *tmp;
577         unsigned long    now = cfs_time_current();
578
579         if (!all && rule->dl_msg_send > now)
580                 return;
581
582         spin_lock(&rule->dl_lock);
583         list_for_each_entry_safe(msg, tmp, &rule->dl_msg_list, msg_list) {
584                 if (!all && msg->msg_delay_send > now)
585                         break;
586
587                 msg->msg_delay_send = 0;
588                 list_move_tail(&msg->msg_list, msg_list);
589         }
590
591         if (list_empty(&rule->dl_msg_list)) {
592                 del_timer(&rule->dl_timer);
593                 rule->dl_msg_send = -1;
594
595         } else if (!list_empty(msg_list)) {
596                 /* dequeued some timedout messages, update timer for the
597                  * next delayed message on rule */
598                 msg = list_entry(rule->dl_msg_list.next,
599                                  struct lnet_msg, msg_list);
600                 rule->dl_msg_send = msg->msg_delay_send;
601                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
602         }
603         spin_unlock(&rule->dl_lock);
604 }
605
606 static void
607 delayed_msg_process(struct list_head *msg_list, bool drop)
608 {
609         struct lnet_msg *msg;
610
611         while (!list_empty(msg_list)) {
612                 struct lnet_ni *ni;
613                 int             cpt;
614                 int             rc;
615
616                 msg = list_entry(msg_list->next, struct lnet_msg, msg_list);
617                 LASSERT(msg->msg_rxpeer != NULL);
618                 LASSERT(msg->msg_rxni != NULL);
619
620                 ni = msg->msg_rxni;
621                 cpt = msg->msg_rx_cpt;
622
623                 list_del_init(&msg->msg_list);
624                 if (drop) {
625                         rc = -ECANCELED;
626
627                 } else if (!msg->msg_routing) {
628                         rc = lnet_parse_local(ni, msg);
629                         if (rc == 0)
630                                 continue;
631
632                 } else {
633                         lnet_net_lock(cpt);
634                         rc = lnet_parse_forward_locked(ni, msg);
635                         lnet_net_unlock(cpt);
636
637                         switch (rc) {
638                         case LNET_CREDIT_OK:
639                                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
640                                              0, msg->msg_len, msg->msg_len);
641                         case LNET_CREDIT_WAIT:
642                                 continue;
643                         default: /* failures */
644                                 break;
645                         }
646                 }
647
648                 lnet_drop_message(ni, cpt, msg->msg_private, msg->msg_len,
649                                   msg->msg_type);
650                 lnet_finalize(msg, rc);
651         }
652 }
653
654 /**
655  * Process delayed messages for scheduled rules
656  * This function can either be called by delay_rule_daemon, or by lnet_finalise
657  */
658 void
659 lnet_delay_rule_check(void)
660 {
661         struct lnet_delay_rule  *rule;
662         struct list_head         msgs;
663
664         INIT_LIST_HEAD(&msgs);
665         while (1) {
666                 if (list_empty(&delay_dd.dd_sched_rules))
667                         break;
668
669                 spin_lock_bh(&delay_dd.dd_lock);
670                 if (list_empty(&delay_dd.dd_sched_rules)) {
671                         spin_unlock_bh(&delay_dd.dd_lock);
672                         break;
673                 }
674
675                 rule = list_entry(delay_dd.dd_sched_rules.next,
676                                   struct lnet_delay_rule, dl_sched_link);
677                 list_del_init(&rule->dl_sched_link);
678                 spin_unlock_bh(&delay_dd.dd_lock);
679
680                 delayed_msg_check(rule, false, &msgs);
681                 delay_rule_decref(rule); /* -1 for delay_dd.dd_sched_rules */
682         }
683
684         if (!list_empty(&msgs))
685                 delayed_msg_process(&msgs, false);
686 }
687
688 /** deamon thread to handle delayed messages */
689 static int
690 lnet_delay_rule_daemon(void *arg)
691 {
692         delay_dd.dd_running = 1;
693         wake_up(&delay_dd.dd_ctl_waitq);
694
695         while (delay_dd.dd_running) {
696                 wait_event_interruptible(delay_dd.dd_waitq,
697                                          !delay_dd.dd_running ||
698                                          !list_empty(&delay_dd.dd_sched_rules));
699                 lnet_delay_rule_check();
700         }
701
702         /* in case more rules have been enqueued after my last check */
703         lnet_delay_rule_check();
704         delay_dd.dd_stopped = 1;
705         wake_up(&delay_dd.dd_ctl_waitq);
706
707         return 0;
708 }
709
710 static void
711 delay_timer_cb(unsigned long arg)
712 {
713         struct lnet_delay_rule *rule = (struct lnet_delay_rule *)arg;
714
715         spin_lock_bh(&delay_dd.dd_lock);
716         if (list_empty(&rule->dl_sched_link) && delay_dd.dd_running) {
717                 atomic_inc(&rule->dl_refcount);
718                 list_add_tail(&rule->dl_sched_link, &delay_dd.dd_sched_rules);
719                 wake_up(&delay_dd.dd_waitq);
720         }
721         spin_unlock_bh(&delay_dd.dd_lock);
722 }
723
724 /**
725  * Add a new delay rule to LNet
726  * There is no check for duplicated delay rule, all rules will be checked for
727  * incoming message.
728  */
729 int
730 lnet_delay_rule_add(struct lnet_fault_attr *attr)
731 {
732         struct lnet_delay_rule *rule;
733         int                     rc = 0;
734         ENTRY;
735
736         if (!((attr->u.delay.la_rate == 0) ^
737               (attr->u.delay.la_interval == 0))) {
738                 CDEBUG(D_NET,
739                        "please provide either delay rate or delay interval, "
740                        "but not both at the same time %d/%d\n",
741                        attr->u.delay.la_rate, attr->u.delay.la_interval);
742                 RETURN(-EINVAL);
743         }
744
745         if (attr->u.delay.la_latency == 0) {
746                 CDEBUG(D_NET, "delay latency cannot be zero\n");
747                 RETURN(-EINVAL);
748         }
749
750         if (lnet_fault_attr_validate(attr) != 0)
751                 RETURN(-EINVAL);
752
753         CFS_ALLOC_PTR(rule);
754         if (rule == NULL)
755                 RETURN(-ENOMEM);
756
757         mutex_lock(&delay_dd.dd_mutex);
758         if (!delay_dd.dd_running) {
759                 struct task_struct *task;
760
761                 /* NB: although LND threads will process delayed message
762                  * in lnet_finalize, but there is no guarantee that LND
763                  * threads will be waken up if no other message needs to
764                  * be handled.
765                  * Only one daemon thread, performance is not the concern
766                  * of this simualation module.
767                  */
768                 task = kthread_run(lnet_delay_rule_daemon, NULL, "lnet_dd");
769                 if (IS_ERR(task)) {
770                         rc = PTR_ERR(task);
771                         GOTO(failed, rc);
772                 }
773                 wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_running);
774         }
775
776         init_timer(&rule->dl_timer);
777         rule->dl_timer.function = delay_timer_cb;
778         rule->dl_timer.data = (unsigned long)rule;
779
780         spin_lock_init(&rule->dl_lock);
781         INIT_LIST_HEAD(&rule->dl_msg_list);
782         INIT_LIST_HEAD(&rule->dl_sched_link);
783
784         rule->dl_attr = *attr;
785         if (attr->u.delay.la_interval != 0) {
786                 rule->dl_time_base = cfs_time_shift(attr->u.delay.la_interval);
787                 rule->dl_delay_time = cfs_time_shift(cfs_rand() %
788                                                      attr->u.delay.la_interval);
789         } else {
790                 rule->dl_delay_at = cfs_rand() % attr->u.delay.la_rate;
791         }
792
793         rule->dl_msg_send = -1;
794
795         lnet_net_lock(LNET_LOCK_EX);
796         atomic_set(&rule->dl_refcount, 1);
797         list_add(&rule->dl_link, &the_lnet.ln_delay_rules);
798         lnet_net_unlock(LNET_LOCK_EX);
799
800         CDEBUG(D_NET, "Added delay rule: src %s, dst %s, rate %d\n",
801                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
802                attr->u.delay.la_rate);
803
804         mutex_unlock(&delay_dd.dd_mutex);
805         RETURN(0);
806  failed:
807         mutex_unlock(&delay_dd.dd_mutex);
808         CFS_FREE_PTR(rule);
809         return rc;
810 }
811
812 /**
813  * Remove matched Delay Rules from lnet, if \a shutdown is true or both \a src
814  * and \a dst are zero, all rules will be removed, otherwise only matched rules
815  * will be removed.
816  * If \a src is zero, then all rules have \a dst as destination will be remove
817  * If \a dst is zero, then all rules have \a src as source will be removed
818  *
819  * When a delay rule is removed, all delayed messages of this rule will be
820  * processed immediately.
821  */
822 int
823 lnet_delay_rule_del(lnet_nid_t src, lnet_nid_t dst, bool shutdown)
824 {
825         struct lnet_delay_rule *rule;
826         struct lnet_delay_rule  *tmp;
827         struct list_head        rule_list;
828         struct list_head        msg_list;
829         int                     n = 0;
830         bool                    cleanup;
831         ENTRY;
832
833         INIT_LIST_HEAD(&rule_list);
834         INIT_LIST_HEAD(&msg_list);
835
836         if (shutdown)
837                 src = dst = 0;
838
839         mutex_lock(&delay_dd.dd_mutex);
840         lnet_net_lock(LNET_LOCK_EX);
841
842         list_for_each_entry_safe(rule, tmp, &the_lnet.ln_delay_rules, dl_link) {
843                 if (rule->dl_attr.fa_src != src && src != 0)
844                         continue;
845
846                 if (rule->dl_attr.fa_dst != dst && dst != 0)
847                         continue;
848
849                 CDEBUG(D_NET, "Remove delay rule: src %s->dst: %s (1/%d, %d)\n",
850                        libcfs_nid2str(rule->dl_attr.fa_src),
851                        libcfs_nid2str(rule->dl_attr.fa_dst),
852                        rule->dl_attr.u.delay.la_rate,
853                        rule->dl_attr.u.delay.la_interval);
854                 /* refcount is taken over by rule_list */
855                 list_move(&rule->dl_link, &rule_list);
856         }
857
858         /* check if we need to shutdown delay_daemon */
859         cleanup = list_empty(&the_lnet.ln_delay_rules) &&
860                   !list_empty(&rule_list);
861         lnet_net_unlock(LNET_LOCK_EX);
862
863         list_for_each_entry_safe(rule, tmp, &rule_list, dl_link) {
864                 list_del_init(&rule->dl_link);
865
866                 del_timer_sync(&rule->dl_timer);
867                 delayed_msg_check(rule, true, &msg_list);
868                 delay_rule_decref(rule); /* -1 for the_lnet.ln_delay_rules */
869                 n++;
870         }
871
872         if (cleanup) { /* no more delay rule, shutdown delay_daemon */
873                 LASSERT(delay_dd.dd_running);
874                 delay_dd.dd_running = 0;
875                 wake_up(&delay_dd.dd_waitq);
876
877                 while (!delay_dd.dd_stopped)
878                         wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_stopped);
879         }
880         mutex_unlock(&delay_dd.dd_mutex);
881
882         if (!list_empty(&msg_list))
883                 delayed_msg_process(&msg_list, shutdown);
884
885         RETURN(n);
886 }
887
888 /**
889  * List Delay Rule at position of \a pos
890  */
891 int
892 lnet_delay_rule_list(int pos, struct lnet_fault_attr *attr,
893                     struct lnet_fault_stat *stat)
894 {
895         struct lnet_delay_rule *rule;
896         int                     cpt;
897         int                     i = 0;
898         int                     rc = -ENOENT;
899         ENTRY;
900
901         cpt = lnet_net_lock_current();
902         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
903                 if (i++ < pos)
904                         continue;
905
906                 spin_lock(&rule->dl_lock);
907                 *attr = rule->dl_attr;
908                 *stat = rule->dl_stat;
909                 spin_unlock(&rule->dl_lock);
910                 rc = 0;
911                 break;
912         }
913
914         lnet_net_unlock(cpt);
915         RETURN(rc);
916 }
917
918 /**
919  * reset counters for all Delay Rules
920  */
921 void
922 lnet_delay_rule_reset(void)
923 {
924         struct lnet_delay_rule *rule;
925         int                     cpt;
926         ENTRY;
927
928         cpt = lnet_net_lock_current();
929
930         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
931                 struct lnet_fault_attr *attr = &rule->dl_attr;
932
933                 spin_lock(&rule->dl_lock);
934
935                 memset(&rule->dl_stat, 0, sizeof(rule->dl_stat));
936                 if (attr->u.delay.la_rate != 0) {
937                         rule->dl_delay_at = cfs_rand() % attr->u.delay.la_rate;
938                 } else {
939                         rule->dl_delay_time = cfs_time_shift(cfs_rand() %
940                                                 attr->u.delay.la_interval);
941                         rule->dl_time_base = cfs_time_shift(attr->u.delay.
942                                                                   la_interval);
943                 }
944                 spin_unlock(&rule->dl_lock);
945         }
946
947         lnet_net_unlock(cpt);
948         EXIT;
949 }
950
951 int
952 lnet_fault_ctl(int opc, struct libcfs_ioctl_data *data)
953 {
954         struct lnet_fault_attr *attr;
955         struct lnet_fault_stat *stat;
956
957         attr = (struct lnet_fault_attr *)data->ioc_inlbuf1;
958
959         switch (opc) {
960         default:
961                 return -EINVAL;
962
963         case LNET_CTL_DROP_ADD:
964                 if (attr == NULL)
965                         return -EINVAL;
966
967                 return lnet_drop_rule_add(attr);
968
969         case LNET_CTL_DROP_DEL:
970                 if (attr == NULL)
971                         return -EINVAL;
972
973                 data->ioc_count = lnet_drop_rule_del(attr->fa_src,
974                                                      attr->fa_dst);
975                 return 0;
976
977         case LNET_CTL_DROP_RESET:
978                 lnet_drop_rule_reset();
979                 return 0;
980
981         case LNET_CTL_DROP_LIST:
982                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
983                 if (attr == NULL || stat == NULL)
984                         return -EINVAL;
985
986                 return lnet_drop_rule_list(data->ioc_count, attr, stat);
987
988         case LNET_CTL_DELAY_ADD:
989                 if (attr == NULL)
990                         return -EINVAL;
991
992                 return lnet_delay_rule_add(attr);
993
994         case LNET_CTL_DELAY_DEL:
995                 if (attr == NULL)
996                         return -EINVAL;
997
998                 data->ioc_count = lnet_delay_rule_del(attr->fa_src,
999                                                       attr->fa_dst, false);
1000                 return 0;
1001
1002         case LNET_CTL_DELAY_RESET:
1003                 lnet_delay_rule_reset();
1004                 return 0;
1005
1006         case LNET_CTL_DELAY_LIST:
1007                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1008                 if (attr == NULL || stat == NULL)
1009                         return -EINVAL;
1010
1011                 return lnet_delay_rule_list(data->ioc_count, attr, stat);
1012         }
1013 }
1014
1015 int
1016 lnet_fault_init(void)
1017 {
1018         CLASSERT(LNET_PUT_BIT == 1 << LNET_MSG_PUT);
1019         CLASSERT(LNET_ACK_BIT == 1 << LNET_MSG_ACK);
1020         CLASSERT(LNET_GET_BIT == 1 << LNET_MSG_GET);
1021         CLASSERT(LNET_REPLY_BIT == 1 << LNET_MSG_REPLY);
1022
1023         mutex_init(&delay_dd.dd_mutex);
1024         spin_lock_init(&delay_dd.dd_lock);
1025         init_waitqueue_head(&delay_dd.dd_waitq);
1026         init_waitqueue_head(&delay_dd.dd_ctl_waitq);
1027         INIT_LIST_HEAD(&delay_dd.dd_sched_rules);
1028
1029         return 0;
1030 }
1031
1032 void
1033 lnet_fault_fini(void)
1034 {
1035         lnet_drop_rule_del(0, 0);
1036         lnet_delay_rule_del(0, 0, true);
1037
1038         LASSERT(list_empty(&the_lnet.ln_drop_rules));
1039         LASSERT(list_empty(&the_lnet.ln_delay_rules));
1040         LASSERT(list_empty(&delay_dd.dd_sched_rules));
1041 }