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  *
  29  * lnet/lnet/net_fault.c
  30  *
  31  * Lustre network fault simulation
  32  *
  33  * Author: liang.zhen@intel.com
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <linux/random.h>
  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         time64_t                dr_drop_time;
  61         /** baseline to caculate dr_drop_time */
  62         time64_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 local_nid, lnet_nid_t dst,
  83                       unsigned int type, unsigned int portal)
  84 {
  85         if (!lnet_fault_nid_match(attr->fa_src, src) ||
  86             !lnet_fault_nid_match(attr->fa_dst, dst) ||
  87             !lnet_fault_nid_match(attr->fa_local_nid, local_nid))
  88                 return false;
  89
  90         if (!(attr->fa_msg_mask & BIT(type)))
  91                 return false;
  92
  93         /* NB: ACK and REPLY have no portal, but they should have been
  94          * rejected by message mask */
  95         if (attr->fa_ptl_mask != 0 && /* has portal filter */
  96             !(attr->fa_ptl_mask & (1ULL << portal)))
  97                 return false;
  98
  99         return true;
 100 }
 101
 102 static int
 103 lnet_fault_attr_validate(struct lnet_fault_attr *attr)
 104 {
 105         if (attr->fa_msg_mask == 0)
 106                 attr->fa_msg_mask = LNET_MSG_MASK; /* all message types */
 107
 108         if (attr->fa_ptl_mask == 0) /* no portal filter */
 109                 return 0;
 110
 111         /* NB: only PUT and GET can be filtered if portal filter has been set */
 112         attr->fa_msg_mask &= LNET_GET_BIT | LNET_PUT_BIT;
 113         if (attr->fa_msg_mask == 0) {
 114                 CDEBUG(D_NET, "can't find valid message type bits %x\n",
 115                        attr->fa_msg_mask);
 116                 return -EINVAL;
 117         }
 118         return 0;
 119 }
 120
 121 static void
 122 lnet_fault_stat_inc(struct lnet_fault_stat *stat, unsigned int type)
 123 {
 124         /* NB: fs_counter is NOT updated by this function */
 125         switch (type) {
 126         case LNET_MSG_PUT:
 127                 stat->fs_put++;
 128                 return;
 129         case LNET_MSG_ACK:
 130                 stat->fs_ack++;
 131                 return;
 132         case LNET_MSG_GET:
 133                 stat->fs_get++;
 134                 return;
 135         case LNET_MSG_REPLY:
 136                 stat->fs_reply++;
 137                 return;
 138         }
 139 }
 140
 141 /**
 142  * LNet message drop simulation
 143  */
 144
 145 /**
 146  * Add a new drop rule to LNet
 147  * There is no check for duplicated drop rule, all rules will be checked for
 148  * incoming message.
 149  */
 150 static int
 151 lnet_drop_rule_add(struct lnet_fault_attr *attr)
 152 {
 153         struct lnet_drop_rule *rule;
 154         ENTRY;
 155
 156         if (!((attr->u.drop.da_rate == 0) ^ (attr->u.drop.da_interval == 0))) {
 157                 CDEBUG(D_NET,
 158                        "please provide either drop rate or drop interval, "
 159                        "but not both at the same time %d/%d\n",
 160                        attr->u.drop.da_rate, attr->u.drop.da_interval);
 161                 RETURN(-EINVAL);
 162         }
 163
 164         if (lnet_fault_attr_validate(attr) != 0)
 165                 RETURN(-EINVAL);
 166
 167         CFS_ALLOC_PTR(rule);
 168         if (rule == NULL)
 169                 RETURN(-ENOMEM);
 170
 171         spin_lock_init(&rule->dr_lock);
 172
 173         rule->dr_attr = *attr;
 174         if (attr->u.drop.da_interval != 0) {
 175                 rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
 176                 rule->dr_drop_time = ktime_get_seconds() +
 177                                      prandom_u32_max(attr->u.drop.da_interval);
 178         } else {
 179                 rule->dr_drop_at = prandom_u32_max(attr->u.drop.da_rate);
 180         }
 181
 182         lnet_net_lock(LNET_LOCK_EX);
 183         list_add(&rule->dr_link, &the_lnet.ln_drop_rules);
 184         lnet_net_unlock(LNET_LOCK_EX);
 185
 186         CDEBUG(D_NET, "Added drop rule: src %s, dst %s, rate %d, interval %d\n",
 187                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
 188                attr->u.drop.da_rate, attr->u.drop.da_interval);
 189         RETURN(0);
 190 }
 191
 192 /**
 193  * Remove matched drop rules from lnet, all rules that can match \a src and
 194  * \a dst will be removed.
 195  * If \a src is zero, then all rules have \a dst as destination will be remove
 196  * If \a dst is zero, then all rules have \a src as source will be removed
 197  * If both of them are zero, all rules will be removed
 198  */
 199 static int
 200 lnet_drop_rule_del(lnet_nid_t src, lnet_nid_t dst)
 201 {
 202         struct lnet_drop_rule *rule;
 203         struct lnet_drop_rule *tmp;
 204         LIST_HEAD(zombies);
 205         int n = 0;
 206         ENTRY;
 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 = prandom_u32_max(attr->u.drop.da_rate);
 285                 } else {
 286                         rule->dr_drop_time = ktime_get_seconds() +
 287                                              prandom_u32_max(attr->u.drop.da_interval);
 288                         rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
 289                 }
 290                 spin_unlock(&rule->dr_lock);
 291         }
 292
 293         lnet_net_unlock(cpt);
 294         EXIT;
 295 }
 296
 297 static void
 298 lnet_fault_match_health(enum lnet_msg_hstatus *hstatus, __u32 mask)
 299 {
 300         int choice;
 301         int delta;
 302         int best_delta;
 303         int i;
 304
 305         /* assign a random failure */
 306         choice = prandom_u32_max(LNET_MSG_STATUS_END - LNET_MSG_STATUS_OK);
 307         if (choice == 0)
 308                 choice++;
 309
 310         if (mask == HSTATUS_RANDOM) {
 311                 *hstatus = choice;
 312                 return;
 313         }
 314
 315         if (mask & BIT(choice)) {
 316                 *hstatus = choice;
 317                 return;
 318         }
 319
 320         /* round to the closest ON bit */
 321         i = HSTATUS_END;
 322         best_delta = HSTATUS_END;
 323         while (i > 0) {
 324                 if (mask & BIT(i)) {
 325                         delta = choice - i;
 326                         if (delta < 0)
 327                                 delta *= -1;
 328                         if (delta < best_delta) {
 329                                 best_delta = delta;
 330                                 choice = i;
 331                         }
 332                 }
 333                 i--;
 334         }
 335
 336         *hstatus = choice;
 337 }
 338
 339 /**
 340  * check source/destination NID, portal, message type and drop rate,
 341  * decide whether should drop this message or not
 342  */
 343 static bool
 344 drop_rule_match(struct lnet_drop_rule *rule, lnet_nid_t src,
 345                 lnet_nid_t local_nid, lnet_nid_t dst,
 346                 unsigned int type, unsigned int portal,
 347                 enum lnet_msg_hstatus *hstatus)
 348 {
 349         struct lnet_fault_attr  *attr = &rule->dr_attr;
 350         bool                     drop;
 351
 352         if (!lnet_fault_attr_match(attr, src, local_nid, dst, type, portal))
 353                 return false;
 354
 355         if (attr->u.drop.da_drop_all) {
 356                 CDEBUG(D_NET, "set to drop all messages\n");
 357                 drop = true;
 358                 goto drop_matched;
 359         }
 360
 361         /*
 362          * if we're trying to match a health status error but it hasn't
 363          * been set in the rule, then don't match
 364          */
 365         if ((hstatus && !attr->u.drop.da_health_error_mask) ||
 366             (!hstatus && attr->u.drop.da_health_error_mask))
 367                 return false;
 368
 369         /* match this rule, check drop rate now */
 370         spin_lock(&rule->dr_lock);
 371         if (attr->u.drop.da_random) {
 372                 int value = prandom_u32_max(attr->u.drop.da_interval);
 373                 if (value >= (attr->u.drop.da_interval / 2))
 374                         drop = true;
 375                 else
 376                         drop = false;
 377         } else if (rule->dr_drop_time != 0) { /* time based drop */
 378                 time64_t now = ktime_get_seconds();
 379
 380                 rule->dr_stat.fs_count++;
 381                 drop = now >= rule->dr_drop_time;
 382                 if (drop) {
 383                         if (now > rule->dr_time_base)
 384                                 rule->dr_time_base = now;
 385
 386                         rule->dr_drop_time = rule->dr_time_base +
 387                                              prandom_u32_max(attr->u.drop.da_interval);
 388                         rule->dr_time_base += attr->u.drop.da_interval;
 389
 390                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop : %lld\n",
 391                                libcfs_nid2str(attr->fa_src),
 392                                libcfs_nid2str(attr->fa_dst),
 393                                rule->dr_drop_time);
 394                 }
 395
 396         } else { /* rate based drop */
 397                 __u64 count;
 398
 399                 drop = rule->dr_stat.fs_count++ == rule->dr_drop_at;
 400                 count = rule->dr_stat.fs_count;
 401                 if (do_div(count, attr->u.drop.da_rate) == 0) {
 402                         rule->dr_drop_at = rule->dr_stat.fs_count +
 403                                            prandom_u32_max(attr->u.drop.da_rate);
 404                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop: %lu\n",
 405                                libcfs_nid2str(attr->fa_src),
 406                                libcfs_nid2str(attr->fa_dst), rule->dr_drop_at);
 407                 }
 408         }
 409
 410 drop_matched:
 411
 412         if (drop) { /* drop this message, update counters */
 413                 if (hstatus)
 414                         lnet_fault_match_health(hstatus,
 415                                 attr->u.drop.da_health_error_mask);
 416                 lnet_fault_stat_inc(&rule->dr_stat, type);
 417                 rule->dr_stat.u.drop.ds_dropped++;
 418         }
 419
 420         spin_unlock(&rule->dr_lock);
 421         return drop;
 422 }
 423
 424 /**
 425  * Check if message from \a src to \a dst can match any existed drop rule
 426  */
 427 bool
 428 lnet_drop_rule_match(struct lnet_hdr *hdr,
 429                      lnet_nid_t local_nid,
 430                      enum lnet_msg_hstatus *hstatus)
 431 {
 432         lnet_nid_t src = le64_to_cpu(hdr->src_nid);
 433         lnet_nid_t dst = le64_to_cpu(hdr->dest_nid);
 434         unsigned int typ = le32_to_cpu(hdr->type);
 435         struct lnet_drop_rule *rule;
 436         unsigned int ptl = -1;
 437         bool drop = false;
 438         int cpt;
 439
 440         /* NB: if Portal is specified, then only PUT and GET will be
 441          * filtered by drop rule */
 442         if (typ == LNET_MSG_PUT)
 443                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 444         else if (typ == LNET_MSG_GET)
 445                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 446
 447         cpt = lnet_net_lock_current();
 448         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
 449                 drop = drop_rule_match(rule, src, local_nid, dst, typ, ptl,
 450                                        hstatus);
 451                 if (drop)
 452                         break;
 453         }
 454         lnet_net_unlock(cpt);
 455
 456         return drop;
 457 }
 458
 459 /**
 460  * LNet Delay Simulation
 461  */
 462 /** timestamp (second) to send delayed message */
 463 #define msg_delay_send           msg_ev.hdr_data
 464
 465 struct lnet_delay_rule {
 466         /** link chain on the_lnet.ln_delay_rules */
 467         struct list_head        dl_link;
 468         /** link chain on delay_dd.dd_sched_rules */
 469         struct list_head        dl_sched_link;
 470         /** attributes of this rule */
 471         struct lnet_fault_attr  dl_attr;
 472         /** lock to protect \a below members */
 473         spinlock_t              dl_lock;
 474         /** refcount of delay rule */
 475         atomic_t                dl_refcount;
 476         /**
 477          * the message sequence to delay, which means message is delayed when
 478          * dl_stat.fs_count == dl_delay_at
 479          */
 480         unsigned long           dl_delay_at;
 481         /**
 482          * seconds to delay the next message, it's exclusive with dl_delay_at
 483          */
 484         time64_t                dl_delay_time;
 485         /** baseline to caculate dl_delay_time */
 486         time64_t                dl_time_base;
 487         /** seconds until we send the next delayed message */
 488         time64_t                dl_msg_send;
 489         /** delayed message list */
 490         struct list_head        dl_msg_list;
 491         /** statistic of delayed messages */
 492         struct lnet_fault_stat  dl_stat;
 493         /** timer to wakeup delay_daemon */
 494         struct timer_list       dl_timer;
 495 };
 496
 497 struct delay_daemon_data {
 498         /** serialise rule add/remove */
 499         struct mutex            dd_mutex;
 500         /** protect rules on \a dd_sched_rules */
 501         spinlock_t              dd_lock;
 502         /** scheduled delay rules (by timer) */
 503         struct list_head        dd_sched_rules;
 504         /** deamon thread sleeps at here */
 505         wait_queue_head_t       dd_waitq;
 506         /** controler (lctl command) wait at here */
 507         wait_queue_head_t       dd_ctl_waitq;
 508         /** deamon is running */
 509         unsigned int            dd_running;
 510         /** deamon stopped */
 511         unsigned int            dd_stopped;
 512 };
 513
 514 static struct delay_daemon_data delay_dd;
 515
 516 static void
 517 delay_rule_decref(struct lnet_delay_rule *rule)
 518 {
 519         if (atomic_dec_and_test(&rule->dl_refcount)) {
 520                 LASSERT(list_empty(&rule->dl_sched_link));
 521                 LASSERT(list_empty(&rule->dl_msg_list));
 522                 LASSERT(list_empty(&rule->dl_link));
 523
 524                 CFS_FREE_PTR(rule);
 525         }
 526 }
 527
 528 /**
 529  * check source/destination NID, portal, message type and delay rate,
 530  * decide whether should delay this message or not
 531  */
 532 static bool
 533 delay_rule_match(struct lnet_delay_rule *rule, lnet_nid_t src,
 534                 lnet_nid_t dst, unsigned int type, unsigned int portal,
 535                 struct lnet_msg *msg)
 536 {
 537         struct lnet_fault_attr  *attr = &rule->dl_attr;
 538         bool                     delay;
 539
 540         if (!lnet_fault_attr_match(attr, src, LNET_NID_ANY,
 541                                    dst, type, portal))
 542                 return false;
 543
 544         /* match this rule, check delay rate now */
 545         spin_lock(&rule->dl_lock);
 546         if (rule->dl_delay_time != 0) { /* time based delay */
 547                 time64_t now = ktime_get_seconds();
 548
 549                 rule->dl_stat.fs_count++;
 550                 delay = now >= rule->dl_delay_time;
 551                 if (delay) {
 552                         if (now > rule->dl_time_base)
 553                                 rule->dl_time_base = now;
 554
 555                         rule->dl_delay_time = rule->dl_time_base +
 556                                               prandom_u32_max(attr->u.delay.la_interval);
 557                         rule->dl_time_base += attr->u.delay.la_interval;
 558
 559                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay : %lld\n",
 560                                libcfs_nid2str(attr->fa_src),
 561                                libcfs_nid2str(attr->fa_dst),
 562                                rule->dl_delay_time);
 563                 }
 564
 565         } else { /* rate based delay */
 566                 __u64 count;
 567
 568                 delay = rule->dl_stat.fs_count++ == rule->dl_delay_at;
 569                 /* generate the next random rate sequence */
 570                 count = rule->dl_stat.fs_count;
 571                 if (do_div(count, attr->u.delay.la_rate) == 0) {
 572                         rule->dl_delay_at = rule->dl_stat.fs_count +
 573                                             prandom_u32_max(attr->u.delay.la_rate);
 574                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay: %lu\n",
 575                                libcfs_nid2str(attr->fa_src),
 576                                libcfs_nid2str(attr->fa_dst), rule->dl_delay_at);
 577                 }
 578         }
 579
 580         if (!delay) {
 581                 spin_unlock(&rule->dl_lock);
 582                 return false;
 583         }
 584
 585         /* delay this message, update counters */
 586         lnet_fault_stat_inc(&rule->dl_stat, type);
 587         rule->dl_stat.u.delay.ls_delayed++;
 588
 589         list_add_tail(&msg->msg_list, &rule->dl_msg_list);
 590         msg->msg_delay_send = ktime_get_seconds() + attr->u.delay.la_latency;
 591         if (rule->dl_msg_send == -1) {
 592                 rule->dl_msg_send = msg->msg_delay_send;
 593                 mod_timer(&rule->dl_timer,
 594                           jiffies + cfs_time_seconds(rule->dl_msg_send));
 595         }
 596
 597         spin_unlock(&rule->dl_lock);
 598         return true;
 599 }
 600
 601 /**
 602  * check if \a msg can match any Delay Rule, receiving of this message
 603  * will be delayed if there is a match.
 604  */
 605 bool
 606 lnet_delay_rule_match_locked(struct lnet_hdr *hdr, struct lnet_msg *msg)
 607 {
 608         struct lnet_delay_rule  *rule;
 609         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
 610         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
 611         unsigned int             typ = le32_to_cpu(hdr->type);
 612         unsigned int             ptl = -1;
 613
 614         /* NB: called with hold of lnet_net_lock */
 615
 616         /* NB: if Portal is specified, then only PUT and GET will be
 617          * filtered by delay rule */
 618         if (typ == LNET_MSG_PUT)
 619                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 620         else if (typ == LNET_MSG_GET)
 621                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 622
 623         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 624                 if (delay_rule_match(rule, src, dst, typ, ptl, msg))
 625                         return true;
 626         }
 627
 628         return false;
 629 }
 630
 631 /** check out delayed messages for send */
 632 static void
 633 delayed_msg_check(struct lnet_delay_rule *rule, bool all,
 634                   struct list_head *msg_list)
 635 {
 636         struct lnet_msg *msg;
 637         struct lnet_msg *tmp;
 638         time64_t now = ktime_get_seconds();
 639
 640         if (!all && cfs_time_seconds(rule->dl_msg_send) > now)
 641                 return;
 642
 643         spin_lock(&rule->dl_lock);
 644         list_for_each_entry_safe(msg, tmp, &rule->dl_msg_list, msg_list) {
 645                 if (!all && msg->msg_delay_send > now)
 646                         break;
 647
 648                 msg->msg_delay_send = 0;
 649                 list_move_tail(&msg->msg_list, msg_list);
 650         }
 651
 652         if (list_empty(&rule->dl_msg_list)) {
 653                 del_timer(&rule->dl_timer);
 654                 rule->dl_msg_send = -1;
 655
 656         } else if (!list_empty(msg_list)) {
 657                 /* dequeued some timedout messages, update timer for the
 658                  * next delayed message on rule */
 659                 msg = list_entry(rule->dl_msg_list.next,
 660                                  struct lnet_msg, msg_list);
 661                 rule->dl_msg_send = msg->msg_delay_send;
 662                 mod_timer(&rule->dl_timer,
 663                           jiffies + cfs_time_seconds(rule->dl_msg_send));
 664         }
 665         spin_unlock(&rule->dl_lock);
 666 }
 667
 668 static void
 669 delayed_msg_process(struct list_head *msg_list, bool drop)
 670 {
 671         struct lnet_msg *msg;
 672
 673         while (!list_empty(msg_list)) {
 674                 struct lnet_ni *ni;
 675                 int             cpt;
 676                 int             rc;
 677
 678                 msg = list_entry(msg_list->next, struct lnet_msg, msg_list);
 679                 LASSERT(msg->msg_rxpeer != NULL);
 680                 LASSERT(msg->msg_rxni != NULL);
 681
 682                 ni = msg->msg_rxni;
 683                 cpt = msg->msg_rx_cpt;
 684
 685                 list_del_init(&msg->msg_list);
 686                 if (drop) {
 687                         rc = -ECANCELED;
 688
 689                 } else if (!msg->msg_routing) {
 690                         rc = lnet_parse_local(ni, msg);
 691                         if (rc == 0)
 692                                 continue;
 693
 694                 } else {
 695                         lnet_net_lock(cpt);
 696                         rc = lnet_parse_forward_locked(ni, msg);
 697                         lnet_net_unlock(cpt);
 698
 699                         switch (rc) {
 700                         case LNET_CREDIT_OK:
 701                                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
 702                                              0, msg->msg_len, msg->msg_len);
 703                                 /* fallthrough */
 704                         case LNET_CREDIT_WAIT:
 705                                 continue;
 706                         default: /* failures */
 707                                 break;
 708                         }
 709                 }
 710
 711                 lnet_drop_message(ni, cpt, msg->msg_private, msg->msg_len,
 712                                   msg->msg_type);
 713                 lnet_finalize(msg, rc);
 714         }
 715 }
 716
 717 /**
 718  * Process delayed messages for scheduled rules
 719  * This function can either be called by delay_rule_daemon, or by lnet_finalise
 720  */
 721 void
 722 lnet_delay_rule_check(void)
 723 {
 724         struct lnet_delay_rule *rule;
 725         LIST_HEAD(msgs);
 726
 727         while (1) {
 728                 if (list_empty(&delay_dd.dd_sched_rules))
 729                         break;
 730
 731                 spin_lock_bh(&delay_dd.dd_lock);
 732                 if (list_empty(&delay_dd.dd_sched_rules)) {
 733                         spin_unlock_bh(&delay_dd.dd_lock);
 734                         break;
 735                 }
 736
 737                 rule = list_entry(delay_dd.dd_sched_rules.next,
 738                                   struct lnet_delay_rule, dl_sched_link);
 739                 list_del_init(&rule->dl_sched_link);
 740                 spin_unlock_bh(&delay_dd.dd_lock);
 741
 742                 delayed_msg_check(rule, false, &msgs);
 743                 delay_rule_decref(rule); /* -1 for delay_dd.dd_sched_rules */
 744         }
 745
 746         if (!list_empty(&msgs))
 747                 delayed_msg_process(&msgs, false);
 748 }
 749
 750 /** deamon thread to handle delayed messages */
 751 static int
 752 lnet_delay_rule_daemon(void *arg)
 753 {
 754         delay_dd.dd_running = 1;
 755         wake_up(&delay_dd.dd_ctl_waitq);
 756
 757         while (delay_dd.dd_running) {
 758                 wait_event_interruptible(delay_dd.dd_waitq,
 759                                          !delay_dd.dd_running ||
 760                                          !list_empty(&delay_dd.dd_sched_rules));
 761                 lnet_delay_rule_check();
 762         }
 763
 764         /* in case more rules have been enqueued after my last check */
 765         lnet_delay_rule_check();
 766         delay_dd.dd_stopped = 1;
 767         wake_up(&delay_dd.dd_ctl_waitq);
 768
 769         return 0;
 770 }
 771
 772 static void
 773 delay_timer_cb(cfs_timer_cb_arg_t data)
 774 {
 775         struct lnet_delay_rule *rule = cfs_from_timer(rule, data, dl_timer);
 776
 777         spin_lock_bh(&delay_dd.dd_lock);
 778         if (list_empty(&rule->dl_sched_link) && delay_dd.dd_running) {
 779                 atomic_inc(&rule->dl_refcount);
 780                 list_add_tail(&rule->dl_sched_link, &delay_dd.dd_sched_rules);
 781                 wake_up(&delay_dd.dd_waitq);
 782         }
 783         spin_unlock_bh(&delay_dd.dd_lock);
 784 }
 785
 786 /**
 787  * Add a new delay rule to LNet
 788  * There is no check for duplicated delay rule, all rules will be checked for
 789  * incoming message.
 790  */
 791 int
 792 lnet_delay_rule_add(struct lnet_fault_attr *attr)
 793 {
 794         struct lnet_delay_rule *rule;
 795         int                     rc = 0;
 796         ENTRY;
 797
 798         if (!((attr->u.delay.la_rate == 0) ^
 799               (attr->u.delay.la_interval == 0))) {
 800                 CDEBUG(D_NET,
 801                        "please provide either delay rate or delay interval, "
 802                        "but not both at the same time %d/%d\n",
 803                        attr->u.delay.la_rate, attr->u.delay.la_interval);
 804                 RETURN(-EINVAL);
 805         }
 806
 807         if (attr->u.delay.la_latency == 0) {
 808                 CDEBUG(D_NET, "delay latency cannot be zero\n");
 809                 RETURN(-EINVAL);
 810         }
 811
 812         if (lnet_fault_attr_validate(attr) != 0)
 813                 RETURN(-EINVAL);
 814
 815         CFS_ALLOC_PTR(rule);
 816         if (rule == NULL)
 817                 RETURN(-ENOMEM);
 818
 819         mutex_lock(&delay_dd.dd_mutex);
 820         if (!delay_dd.dd_running) {
 821                 struct task_struct *task;
 822
 823                 /* NB: although LND threads will process delayed message
 824                  * in lnet_finalize, but there is no guarantee that LND
 825                  * threads will be waken up if no other message needs to
 826                  * be handled.
 827                  * Only one daemon thread, performance is not the concern
 828                  * of this simualation module.
 829                  */
 830                 task = kthread_run(lnet_delay_rule_daemon, NULL, "lnet_dd");
 831                 if (IS_ERR(task)) {
 832                         rc = PTR_ERR(task);
 833                         GOTO(failed, rc);
 834                 }
 835                 wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_running);
 836         }
 837
 838         cfs_timer_setup(&rule->dl_timer, delay_timer_cb,
 839                         (unsigned long)rule, 0);
 840
 841         spin_lock_init(&rule->dl_lock);
 842         INIT_LIST_HEAD(&rule->dl_msg_list);
 843         INIT_LIST_HEAD(&rule->dl_sched_link);
 844
 845         rule->dl_attr = *attr;
 846         if (attr->u.delay.la_interval != 0) {
 847                 rule->dl_time_base = ktime_get_seconds() +
 848                                      attr->u.delay.la_interval;
 849                 rule->dl_delay_time = ktime_get_seconds() +
 850                                       prandom_u32_max(attr->u.delay.la_interval);
 851         } else {
 852                 rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
 853         }
 854
 855         rule->dl_msg_send = -1;
 856
 857         lnet_net_lock(LNET_LOCK_EX);
 858         atomic_set(&rule->dl_refcount, 1);
 859         list_add(&rule->dl_link, &the_lnet.ln_delay_rules);
 860         lnet_net_unlock(LNET_LOCK_EX);
 861
 862         CDEBUG(D_NET, "Added delay rule: src %s, dst %s, rate %d\n",
 863                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
 864                attr->u.delay.la_rate);
 865
 866         mutex_unlock(&delay_dd.dd_mutex);
 867         RETURN(0);
 868  failed:
 869         mutex_unlock(&delay_dd.dd_mutex);
 870         CFS_FREE_PTR(rule);
 871         return rc;
 872 }
 873
 874 /**
 875  * Remove matched Delay Rules from lnet, if \a shutdown is true or both \a src
 876  * and \a dst are zero, all rules will be removed, otherwise only matched rules
 877  * will be removed.
 878  * If \a src is zero, then all rules have \a dst as destination will be remove
 879  * If \a dst is zero, then all rules have \a src as source will be removed
 880  *
 881  * When a delay rule is removed, all delayed messages of this rule will be
 882  * processed immediately.
 883  */
 884 int
 885 lnet_delay_rule_del(lnet_nid_t src, lnet_nid_t dst, bool shutdown)
 886 {
 887         struct lnet_delay_rule *rule;
 888         struct lnet_delay_rule *tmp;
 889         LIST_HEAD(rule_list);
 890         LIST_HEAD(msg_list);
 891         int n = 0;
 892         bool cleanup;
 893         ENTRY;
 894
 895         if (shutdown)
 896                 src = dst = 0;
 897
 898         mutex_lock(&delay_dd.dd_mutex);
 899         lnet_net_lock(LNET_LOCK_EX);
 900
 901         list_for_each_entry_safe(rule, tmp, &the_lnet.ln_delay_rules, dl_link) {
 902                 if (rule->dl_attr.fa_src != src && src != 0)
 903                         continue;
 904
 905                 if (rule->dl_attr.fa_dst != dst && dst != 0)
 906                         continue;
 907
 908                 CDEBUG(D_NET, "Remove delay rule: src %s->dst: %s (1/%d, %d)\n",
 909                        libcfs_nid2str(rule->dl_attr.fa_src),
 910                        libcfs_nid2str(rule->dl_attr.fa_dst),
 911                        rule->dl_attr.u.delay.la_rate,
 912                        rule->dl_attr.u.delay.la_interval);
 913                 /* refcount is taken over by rule_list */
 914                 list_move(&rule->dl_link, &rule_list);
 915         }
 916
 917         /* check if we need to shutdown delay_daemon */
 918         cleanup = list_empty(&the_lnet.ln_delay_rules) &&
 919                   !list_empty(&rule_list);
 920         lnet_net_unlock(LNET_LOCK_EX);
 921
 922         list_for_each_entry_safe(rule, tmp, &rule_list, dl_link) {
 923                 list_del_init(&rule->dl_link);
 924
 925                 del_timer_sync(&rule->dl_timer);
 926                 delayed_msg_check(rule, true, &msg_list);
 927                 delay_rule_decref(rule); /* -1 for the_lnet.ln_delay_rules */
 928                 n++;
 929         }
 930
 931         if (cleanup) { /* no more delay rule, shutdown delay_daemon */
 932                 LASSERT(delay_dd.dd_running);
 933                 delay_dd.dd_running = 0;
 934                 wake_up(&delay_dd.dd_waitq);
 935
 936                 while (!delay_dd.dd_stopped)
 937                         wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_stopped);
 938         }
 939         mutex_unlock(&delay_dd.dd_mutex);
 940
 941         if (!list_empty(&msg_list))
 942                 delayed_msg_process(&msg_list, shutdown);
 943
 944         RETURN(n);
 945 }
 946
 947 /**
 948  * List Delay Rule at position of \a pos
 949  */
 950 int
 951 lnet_delay_rule_list(int pos, struct lnet_fault_attr *attr,
 952                     struct lnet_fault_stat *stat)
 953 {
 954         struct lnet_delay_rule *rule;
 955         int                     cpt;
 956         int                     i = 0;
 957         int                     rc = -ENOENT;
 958         ENTRY;
 959
 960         cpt = lnet_net_lock_current();
 961         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 962                 if (i++ < pos)
 963                         continue;
 964
 965                 spin_lock(&rule->dl_lock);
 966                 *attr = rule->dl_attr;
 967                 *stat = rule->dl_stat;
 968                 spin_unlock(&rule->dl_lock);
 969                 rc = 0;
 970                 break;
 971         }
 972
 973         lnet_net_unlock(cpt);
 974         RETURN(rc);
 975 }
 976
 977 /**
 978  * reset counters for all Delay Rules
 979  */
 980 void
 981 lnet_delay_rule_reset(void)
 982 {
 983         struct lnet_delay_rule *rule;
 984         int                     cpt;
 985         ENTRY;
 986
 987         cpt = lnet_net_lock_current();
 988
 989         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 990                 struct lnet_fault_attr *attr = &rule->dl_attr;
 991
 992                 spin_lock(&rule->dl_lock);
 993
 994                 memset(&rule->dl_stat, 0, sizeof(rule->dl_stat));
 995                 if (attr->u.delay.la_rate != 0) {
 996                         rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
 997                 } else {
 998                         rule->dl_delay_time = ktime_get_seconds() +
 999                                               prandom_u32_max(attr->u.delay.la_interval);
1000                         rule->dl_time_base = ktime_get_seconds() +
1001                                              attr->u.delay.la_interval;
1002                 }
1003                 spin_unlock(&rule->dl_lock);
1004         }
1005
1006         lnet_net_unlock(cpt);
1007         EXIT;
1008 }
1009
1010 int
1011 lnet_fault_ctl(int opc, struct libcfs_ioctl_data *data)
1012 {
1013         struct lnet_fault_attr *attr;
1014         struct lnet_fault_stat *stat;
1015
1016         attr = (struct lnet_fault_attr *)data->ioc_inlbuf1;
1017
1018         switch (opc) {
1019         default:
1020                 return -EINVAL;
1021
1022         case LNET_CTL_DROP_ADD:
1023                 if (attr == NULL)
1024                         return -EINVAL;
1025
1026                 return lnet_drop_rule_add(attr);
1027
1028         case LNET_CTL_DROP_DEL:
1029                 if (attr == NULL)
1030                         return -EINVAL;
1031
1032                 data->ioc_count = lnet_drop_rule_del(attr->fa_src,
1033                                                      attr->fa_dst);
1034                 return 0;
1035
1036         case LNET_CTL_DROP_RESET:
1037                 lnet_drop_rule_reset();
1038                 return 0;
1039
1040         case LNET_CTL_DROP_LIST:
1041                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1042                 if (attr == NULL || stat == NULL)
1043                         return -EINVAL;
1044
1045                 return lnet_drop_rule_list(data->ioc_count, attr, stat);
1046
1047         case LNET_CTL_DELAY_ADD:
1048                 if (attr == NULL)
1049                         return -EINVAL;
1050
1051                 return lnet_delay_rule_add(attr);
1052
1053         case LNET_CTL_DELAY_DEL:
1054                 if (attr == NULL)
1055                         return -EINVAL;
1056
1057                 data->ioc_count = lnet_delay_rule_del(attr->fa_src,
1058                                                       attr->fa_dst, false);
1059                 return 0;
1060
1061         case LNET_CTL_DELAY_RESET:
1062                 lnet_delay_rule_reset();
1063                 return 0;
1064
1065         case LNET_CTL_DELAY_LIST:
1066                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1067                 if (attr == NULL || stat == NULL)
1068                         return -EINVAL;
1069
1070                 return lnet_delay_rule_list(data->ioc_count, attr, stat);
1071         }
1072 }
1073
1074 int
1075 lnet_fault_init(void)
1076 {
1077         BUILD_BUG_ON(LNET_PUT_BIT != BIT(LNET_MSG_PUT));
1078         BUILD_BUG_ON(LNET_ACK_BIT != BIT(LNET_MSG_ACK));
1079         BUILD_BUG_ON(LNET_GET_BIT != BIT(LNET_MSG_GET));
1080         BUILD_BUG_ON(LNET_REPLY_BIT != BIT(LNET_MSG_REPLY));
1081
1082         mutex_init(&delay_dd.dd_mutex);
1083         spin_lock_init(&delay_dd.dd_lock);
1084         init_waitqueue_head(&delay_dd.dd_waitq);
1085         init_waitqueue_head(&delay_dd.dd_ctl_waitq);
1086         INIT_LIST_HEAD(&delay_dd.dd_sched_rules);
1087
1088         return 0;
1089 }
1090
1091 void
1092 lnet_fault_fini(void)
1093 {
1094         lnet_drop_rule_del(0, 0);
1095         lnet_delay_rule_del(0, 0, true);
1096
1097         LASSERT(list_empty(&the_lnet.ln_drop_rules));
1098         LASSERT(list_empty(&the_lnet.ln_delay_rules));
1099         LASSERT(list_empty(&delay_dd.dd_sched_rules));
1100 }