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 <linux/random.h>
  40 #include <lnet/lib-lnet.h>
  41 #include <uapi/linux/lnet/lnetctl.h>
  42
  43 #define LNET_MSG_MASK           (LNET_PUT_BIT | LNET_ACK_BIT | \
  44                                  LNET_GET_BIT | LNET_REPLY_BIT)
  45
  46 struct lnet_drop_rule {
  47         /** link chain on the_lnet.ln_drop_rules */
  48         struct list_head        dr_link;
  49         /** attributes of this rule */
  50         struct lnet_fault_attr  dr_attr;
  51         /** lock to protect \a dr_drop_at and \a dr_stat */
  52         spinlock_t              dr_lock;
  53         /**
  54          * the message sequence to drop, which means message is dropped when
  55          * dr_stat.drs_count == dr_drop_at
  56          */
  57         unsigned long           dr_drop_at;
  58         /**
  59          * seconds to drop the next message, it's exclusive with dr_drop_at
  60          */
  61         time64_t                dr_drop_time;
  62         /** baseline to caculate dr_drop_time */
  63         time64_t                dr_time_base;
  64         /** statistic of dropped messages */
  65         struct lnet_fault_stat  dr_stat;
  66 };
  67
  68 static bool
  69 lnet_fault_nid_match(lnet_nid_t nid, lnet_nid_t msg_nid)
  70 {
  71         if (nid == msg_nid || nid == LNET_NID_ANY)
  72                 return true;
  73
  74         if (LNET_NIDNET(nid) != LNET_NIDNET(msg_nid))
  75                 return false;
  76
  77         /* 255.255.255.255@net is wildcard for all addresses in a network */
  78         return LNET_NIDADDR(nid) == LNET_NIDADDR(LNET_NID_ANY);
  79 }
  80
  81 static bool
  82 lnet_fault_attr_match(struct lnet_fault_attr *attr, lnet_nid_t src,
  83                       lnet_nid_t local_nid, lnet_nid_t dst,
  84                       unsigned int type, unsigned int portal)
  85 {
  86         if (!lnet_fault_nid_match(attr->fa_src, src) ||
  87             !lnet_fault_nid_match(attr->fa_dst, dst) ||
  88             !lnet_fault_nid_match(attr->fa_local_nid, local_nid))
  89                 return false;
  90
  91         if (!(attr->fa_msg_mask & (1 << type)))
  92                 return false;
  93
  94         /* NB: ACK and REPLY have no portal, but they should have been
  95          * rejected by message mask */
  96         if (attr->fa_ptl_mask != 0 && /* has portal filter */
  97             !(attr->fa_ptl_mask & (1ULL << portal)))
  98                 return false;
  99
 100         return true;
 101 }
 102
 103 static int
 104 lnet_fault_attr_validate(struct lnet_fault_attr *attr)
 105 {
 106         if (attr->fa_msg_mask == 0)
 107                 attr->fa_msg_mask = LNET_MSG_MASK; /* all message types */
 108
 109         if (attr->fa_ptl_mask == 0) /* no portal filter */
 110                 return 0;
 111
 112         /* NB: only PUT and GET can be filtered if portal filter has been set */
 113         attr->fa_msg_mask &= LNET_GET_BIT | LNET_PUT_BIT;
 114         if (attr->fa_msg_mask == 0) {
 115                 CDEBUG(D_NET, "can't find valid message type bits %x\n",
 116                        attr->fa_msg_mask);
 117                 return -EINVAL;
 118         }
 119         return 0;
 120 }
 121
 122 static void
 123 lnet_fault_stat_inc(struct lnet_fault_stat *stat, unsigned int type)
 124 {
 125         /* NB: fs_counter is NOT updated by this function */
 126         switch (type) {
 127         case LNET_MSG_PUT:
 128                 stat->fs_put++;
 129                 return;
 130         case LNET_MSG_ACK:
 131                 stat->fs_ack++;
 132                 return;
 133         case LNET_MSG_GET:
 134                 stat->fs_get++;
 135                 return;
 136         case LNET_MSG_REPLY:
 137                 stat->fs_reply++;
 138                 return;
 139         }
 140 }
 141
 142 /**
 143  * LNet message drop simulation
 144  */
 145
 146 /**
 147  * Add a new drop rule to LNet
 148  * There is no check for duplicated drop rule, all rules will be checked for
 149  * incoming message.
 150  */
 151 static int
 152 lnet_drop_rule_add(struct lnet_fault_attr *attr)
 153 {
 154         struct lnet_drop_rule *rule;
 155         ENTRY;
 156
 157         if (!((attr->u.drop.da_rate == 0) ^ (attr->u.drop.da_interval == 0))) {
 158                 CDEBUG(D_NET,
 159                        "please provide either drop rate or drop interval, "
 160                        "but not both at the same time %d/%d\n",
 161                        attr->u.drop.da_rate, attr->u.drop.da_interval);
 162                 RETURN(-EINVAL);
 163         }
 164
 165         if (lnet_fault_attr_validate(attr) != 0)
 166                 RETURN(-EINVAL);
 167
 168         CFS_ALLOC_PTR(rule);
 169         if (rule == NULL)
 170                 RETURN(-ENOMEM);
 171
 172         spin_lock_init(&rule->dr_lock);
 173
 174         rule->dr_attr = *attr;
 175         if (attr->u.drop.da_interval != 0) {
 176                 rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
 177                 rule->dr_drop_time = ktime_get_seconds() +
 178                                      prandom_u32_max(attr->u.drop.da_interval);
 179         } else {
 180                 rule->dr_drop_at = prandom_u32_max(attr->u.drop.da_rate);
 181         }
 182
 183         lnet_net_lock(LNET_LOCK_EX);
 184         list_add(&rule->dr_link, &the_lnet.ln_drop_rules);
 185         lnet_net_unlock(LNET_LOCK_EX);
 186
 187         CDEBUG(D_NET, "Added drop rule: src %s, dst %s, rate %d, interval %d\n",
 188                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
 189                attr->u.drop.da_rate, attr->u.drop.da_interval);
 190         RETURN(0);
 191 }
 192
 193 /**
 194  * Remove matched drop rules from lnet, all rules that can match \a src and
 195  * \a dst will be removed.
 196  * If \a src is zero, then all rules have \a dst as destination will be remove
 197  * If \a dst is zero, then all rules have \a src as source will be removed
 198  * If both of them are zero, all rules will be removed
 199  */
 200 static int
 201 lnet_drop_rule_del(lnet_nid_t src, lnet_nid_t dst)
 202 {
 203         struct lnet_drop_rule *rule;
 204         struct lnet_drop_rule *tmp;
 205         struct list_head       zombies;
 206         int                    n = 0;
 207         ENTRY;
 208
 209         INIT_LIST_HEAD(&zombies);
 210
 211         lnet_net_lock(LNET_LOCK_EX);
 212         list_for_each_entry_safe(rule, tmp, &the_lnet.ln_drop_rules, dr_link) {
 213                 if (rule->dr_attr.fa_src != src && src != 0)
 214                         continue;
 215
 216                 if (rule->dr_attr.fa_dst != dst && dst != 0)
 217                         continue;
 218
 219                 list_move(&rule->dr_link, &zombies);
 220         }
 221         lnet_net_unlock(LNET_LOCK_EX);
 222
 223         list_for_each_entry_safe(rule, tmp, &zombies, dr_link) {
 224                 CDEBUG(D_NET, "Remove drop rule: src %s->dst: %s (1/%d, %d)\n",
 225                        libcfs_nid2str(rule->dr_attr.fa_src),
 226                        libcfs_nid2str(rule->dr_attr.fa_dst),
 227                        rule->dr_attr.u.drop.da_rate,
 228                        rule->dr_attr.u.drop.da_interval);
 229
 230                 list_del(&rule->dr_link);
 231                 CFS_FREE_PTR(rule);
 232                 n++;
 233         }
 234
 235         RETURN(n);
 236 }
 237
 238 /**
 239  * List drop rule at position of \a pos
 240  */
 241 static int
 242 lnet_drop_rule_list(int pos, struct lnet_fault_attr *attr,
 243                     struct lnet_fault_stat *stat)
 244 {
 245         struct lnet_drop_rule *rule;
 246         int                    cpt;
 247         int                    i = 0;
 248         int                    rc = -ENOENT;
 249         ENTRY;
 250
 251         cpt = lnet_net_lock_current();
 252         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
 253                 if (i++ < pos)
 254                         continue;
 255
 256                 spin_lock(&rule->dr_lock);
 257                 *attr = rule->dr_attr;
 258                 *stat = rule->dr_stat;
 259                 spin_unlock(&rule->dr_lock);
 260                 rc = 0;
 261                 break;
 262         }
 263
 264         lnet_net_unlock(cpt);
 265         RETURN(rc);
 266 }
 267
 268 /**
 269  * reset counters for all drop rules
 270  */
 271 static void
 272 lnet_drop_rule_reset(void)
 273 {
 274         struct lnet_drop_rule *rule;
 275         int                    cpt;
 276         ENTRY;
 277
 278         cpt = lnet_net_lock_current();
 279
 280         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
 281                 struct lnet_fault_attr *attr = &rule->dr_attr;
 282
 283                 spin_lock(&rule->dr_lock);
 284
 285                 memset(&rule->dr_stat, 0, sizeof(rule->dr_stat));
 286                 if (attr->u.drop.da_rate != 0) {
 287                         rule->dr_drop_at = prandom_u32_max(attr->u.drop.da_rate);
 288                 } else {
 289                         rule->dr_drop_time = ktime_get_seconds() +
 290                                              prandom_u32_max(attr->u.drop.da_interval);
 291                         rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
 292                 }
 293                 spin_unlock(&rule->dr_lock);
 294         }
 295
 296         lnet_net_unlock(cpt);
 297         EXIT;
 298 }
 299
 300 static void
 301 lnet_fault_match_health(enum lnet_msg_hstatus *hstatus, __u32 mask)
 302 {
 303         int choice;
 304         int delta;
 305         int best_delta;
 306         int i;
 307
 308         /* assign a random failure */
 309         choice = prandom_u32_max(LNET_MSG_STATUS_END - LNET_MSG_STATUS_OK);
 310         if (choice == 0)
 311                 choice++;
 312
 313         if (mask == HSTATUS_RANDOM) {
 314                 *hstatus = choice;
 315                 return;
 316         }
 317
 318         if (mask & (1 << choice)) {
 319                 *hstatus = choice;
 320                 return;
 321         }
 322
 323         /* round to the closest ON bit */
 324         i = HSTATUS_END;
 325         best_delta = HSTATUS_END;
 326         while (i > 0) {
 327                 if (mask & (1 << i)) {
 328                         delta = choice - i;
 329                         if (delta < 0)
 330                                 delta *= -1;
 331                         if (delta < best_delta) {
 332                                 best_delta = delta;
 333                                 choice = i;
 334                         }
 335                 }
 336                 i--;
 337         }
 338
 339         *hstatus = choice;
 340 }
 341
 342 /**
 343  * check source/destination NID, portal, message type and drop rate,
 344  * decide whether should drop this message or not
 345  */
 346 static bool
 347 drop_rule_match(struct lnet_drop_rule *rule, lnet_nid_t src,
 348                 lnet_nid_t local_nid, lnet_nid_t dst,
 349                 unsigned int type, unsigned int portal,
 350                 enum lnet_msg_hstatus *hstatus)
 351 {
 352         struct lnet_fault_attr  *attr = &rule->dr_attr;
 353         bool                     drop;
 354
 355         if (!lnet_fault_attr_match(attr, src, local_nid, dst, type, portal))
 356                 return false;
 357
 358         if (attr->u.drop.da_drop_all) {
 359                 CDEBUG(D_NET, "set to drop all messages\n");
 360                 drop = true;
 361                 goto drop_matched;
 362         }
 363
 364         /*
 365          * if we're trying to match a health status error but it hasn't
 366          * been set in the rule, then don't match
 367          */
 368         if ((hstatus && !attr->u.drop.da_health_error_mask) ||
 369             (!hstatus && attr->u.drop.da_health_error_mask))
 370                 return false;
 371
 372         /* match this rule, check drop rate now */
 373         spin_lock(&rule->dr_lock);
 374         if (attr->u.drop.da_random) {
 375                 int value = prandom_u32_max(attr->u.drop.da_interval);
 376                 if (value >= (attr->u.drop.da_interval / 2))
 377                         drop = true;
 378                 else
 379                         drop = false;
 380         } else if (rule->dr_drop_time != 0) { /* time based drop */
 381                 time64_t now = ktime_get_seconds();
 382
 383                 rule->dr_stat.fs_count++;
 384                 drop = now >= rule->dr_drop_time;
 385                 if (drop) {
 386                         if (now > rule->dr_time_base)
 387                                 rule->dr_time_base = now;
 388
 389                         rule->dr_drop_time = rule->dr_time_base +
 390                                              prandom_u32_max(attr->u.drop.da_interval);
 391                         rule->dr_time_base += attr->u.drop.da_interval;
 392
 393                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop : %lld\n",
 394                                libcfs_nid2str(attr->fa_src),
 395                                libcfs_nid2str(attr->fa_dst),
 396                                rule->dr_drop_time);
 397                 }
 398
 399         } else { /* rate based drop */
 400                 __u64 count;
 401
 402                 drop = rule->dr_stat.fs_count++ == rule->dr_drop_at;
 403                 count = rule->dr_stat.fs_count;
 404                 if (do_div(count, attr->u.drop.da_rate) == 0) {
 405                         rule->dr_drop_at = rule->dr_stat.fs_count +
 406                                            prandom_u32_max(attr->u.drop.da_rate);
 407                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop: %lu\n",
 408                                libcfs_nid2str(attr->fa_src),
 409                                libcfs_nid2str(attr->fa_dst), rule->dr_drop_at);
 410                 }
 411         }
 412
 413 drop_matched:
 414
 415         if (drop) { /* drop this message, update counters */
 416                 if (hstatus)
 417                         lnet_fault_match_health(hstatus,
 418                                 attr->u.drop.da_health_error_mask);
 419                 lnet_fault_stat_inc(&rule->dr_stat, type);
 420                 rule->dr_stat.u.drop.ds_dropped++;
 421         }
 422
 423         spin_unlock(&rule->dr_lock);
 424         return drop;
 425 }
 426
 427 /**
 428  * Check if message from \a src to \a dst can match any existed drop rule
 429  */
 430 bool
 431 lnet_drop_rule_match(struct lnet_hdr *hdr,
 432                      lnet_nid_t local_nid,
 433                      enum lnet_msg_hstatus *hstatus)
 434 {
 435         lnet_nid_t src = le64_to_cpu(hdr->src_nid);
 436         lnet_nid_t dst = le64_to_cpu(hdr->dest_nid);
 437         unsigned int typ = le32_to_cpu(hdr->type);
 438         struct lnet_drop_rule *rule;
 439         unsigned int ptl = -1;
 440         bool drop = false;
 441         int cpt;
 442
 443         /* NB: if Portal is specified, then only PUT and GET will be
 444          * filtered by drop rule */
 445         if (typ == LNET_MSG_PUT)
 446                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 447         else if (typ == LNET_MSG_GET)
 448                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 449
 450         cpt = lnet_net_lock_current();
 451         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
 452                 drop = drop_rule_match(rule, src, local_nid, dst, typ, ptl,
 453                                        hstatus);
 454                 if (drop)
 455                         break;
 456         }
 457         lnet_net_unlock(cpt);
 458
 459         return drop;
 460 }
 461
 462 /**
 463  * LNet Delay Simulation
 464  */
 465 /** timestamp (second) to send delayed message */
 466 #define msg_delay_send           msg_ev.hdr_data
 467
 468 struct lnet_delay_rule {
 469         /** link chain on the_lnet.ln_delay_rules */
 470         struct list_head        dl_link;
 471         /** link chain on delay_dd.dd_sched_rules */
 472         struct list_head        dl_sched_link;
 473         /** attributes of this rule */
 474         struct lnet_fault_attr  dl_attr;
 475         /** lock to protect \a below members */
 476         spinlock_t              dl_lock;
 477         /** refcount of delay rule */
 478         atomic_t                dl_refcount;
 479         /**
 480          * the message sequence to delay, which means message is delayed when
 481          * dl_stat.fs_count == dl_delay_at
 482          */
 483         unsigned long           dl_delay_at;
 484         /**
 485          * seconds to delay the next message, it's exclusive with dl_delay_at
 486          */
 487         time64_t                dl_delay_time;
 488         /** baseline to caculate dl_delay_time */
 489         time64_t                dl_time_base;
 490         /** jiffies to send the next delayed message */
 491         unsigned long           dl_msg_send;
 492         /** delayed message list */
 493         struct list_head        dl_msg_list;
 494         /** statistic of delayed messages */
 495         struct lnet_fault_stat  dl_stat;
 496         /** timer to wakeup delay_daemon */
 497         struct timer_list       dl_timer;
 498 };
 499
 500 struct delay_daemon_data {
 501         /** serialise rule add/remove */
 502         struct mutex            dd_mutex;
 503         /** protect rules on \a dd_sched_rules */
 504         spinlock_t              dd_lock;
 505         /** scheduled delay rules (by timer) */
 506         struct list_head        dd_sched_rules;
 507         /** deamon thread sleeps at here */
 508         wait_queue_head_t       dd_waitq;
 509         /** controler (lctl command) wait at here */
 510         wait_queue_head_t       dd_ctl_waitq;
 511         /** deamon is running */
 512         unsigned int            dd_running;
 513         /** deamon stopped */
 514         unsigned int            dd_stopped;
 515 };
 516
 517 static struct delay_daemon_data delay_dd;
 518
 519 static void
 520 delay_rule_decref(struct lnet_delay_rule *rule)
 521 {
 522         if (atomic_dec_and_test(&rule->dl_refcount)) {
 523                 LASSERT(list_empty(&rule->dl_sched_link));
 524                 LASSERT(list_empty(&rule->dl_msg_list));
 525                 LASSERT(list_empty(&rule->dl_link));
 526
 527                 CFS_FREE_PTR(rule);
 528         }
 529 }
 530
 531 /**
 532  * check source/destination NID, portal, message type and delay rate,
 533  * decide whether should delay this message or not
 534  */
 535 static bool
 536 delay_rule_match(struct lnet_delay_rule *rule, lnet_nid_t src,
 537                 lnet_nid_t dst, unsigned int type, unsigned int portal,
 538                 struct lnet_msg *msg)
 539 {
 540         struct lnet_fault_attr  *attr = &rule->dl_attr;
 541         bool                     delay;
 542
 543         if (!lnet_fault_attr_match(attr, src, LNET_NID_ANY,
 544                                    dst, type, portal))
 545                 return false;
 546
 547         /* match this rule, check delay rate now */
 548         spin_lock(&rule->dl_lock);
 549         if (rule->dl_delay_time != 0) { /* time based delay */
 550                 time64_t now = ktime_get_seconds();
 551
 552                 rule->dl_stat.fs_count++;
 553                 delay = now >= rule->dl_delay_time;
 554                 if (delay) {
 555                         if (now > rule->dl_time_base)
 556                                 rule->dl_time_base = now;
 557
 558                         rule->dl_delay_time = rule->dl_time_base +
 559                                               prandom_u32_max(attr->u.delay.la_interval);
 560                         rule->dl_time_base += attr->u.delay.la_interval;
 561
 562                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay : %lld\n",
 563                                libcfs_nid2str(attr->fa_src),
 564                                libcfs_nid2str(attr->fa_dst),
 565                                rule->dl_delay_time);
 566                 }
 567
 568         } else { /* rate based delay */
 569                 __u64 count;
 570
 571                 delay = rule->dl_stat.fs_count++ == rule->dl_delay_at;
 572                 /* generate the next random rate sequence */
 573                 count = rule->dl_stat.fs_count;
 574                 if (do_div(count, attr->u.delay.la_rate) == 0) {
 575                         rule->dl_delay_at = rule->dl_stat.fs_count +
 576                                             prandom_u32_max(attr->u.delay.la_rate);
 577                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay: %lu\n",
 578                                libcfs_nid2str(attr->fa_src),
 579                                libcfs_nid2str(attr->fa_dst), rule->dl_delay_at);
 580                 }
 581         }
 582
 583         if (!delay) {
 584                 spin_unlock(&rule->dl_lock);
 585                 return false;
 586         }
 587
 588         /* delay this message, update counters */
 589         lnet_fault_stat_inc(&rule->dl_stat, type);
 590         rule->dl_stat.u.delay.ls_delayed++;
 591
 592         list_add_tail(&msg->msg_list, &rule->dl_msg_list);
 593         msg->msg_delay_send = ktime_get_seconds() + attr->u.delay.la_latency;
 594         if (rule->dl_msg_send == -1) {
 595                 rule->dl_msg_send = msg->msg_delay_send;
 596                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
 597         }
 598
 599         spin_unlock(&rule->dl_lock);
 600         return true;
 601 }
 602
 603 /**
 604  * check if \a msg can match any Delay Rule, receiving of this message
 605  * will be delayed if there is a match.
 606  */
 607 bool
 608 lnet_delay_rule_match_locked(struct lnet_hdr *hdr, struct lnet_msg *msg)
 609 {
 610         struct lnet_delay_rule  *rule;
 611         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
 612         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
 613         unsigned int             typ = le32_to_cpu(hdr->type);
 614         unsigned int             ptl = -1;
 615
 616         /* NB: called with hold of lnet_net_lock */
 617
 618         /* NB: if Portal is specified, then only PUT and GET will be
 619          * filtered by delay rule */
 620         if (typ == LNET_MSG_PUT)
 621                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 622         else if (typ == LNET_MSG_GET)
 623                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 624
 625         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 626                 if (delay_rule_match(rule, src, dst, typ, ptl, msg))
 627                         return true;
 628         }
 629
 630         return false;
 631 }
 632
 633 /** check out delayed messages for send */
 634 static void
 635 delayed_msg_check(struct lnet_delay_rule *rule, bool all,
 636                   struct list_head *msg_list)
 637 {
 638         struct lnet_msg *msg;
 639         struct lnet_msg *tmp;
 640         time64_t now = ktime_get_seconds();
 641
 642         if (!all && cfs_time_seconds(rule->dl_msg_send) > now)
 643                 return;
 644
 645         spin_lock(&rule->dl_lock);
 646         list_for_each_entry_safe(msg, tmp, &rule->dl_msg_list, msg_list) {
 647                 if (!all && msg->msg_delay_send > now)
 648                         break;
 649
 650                 msg->msg_delay_send = 0;
 651                 list_move_tail(&msg->msg_list, msg_list);
 652         }
 653
 654         if (list_empty(&rule->dl_msg_list)) {
 655                 del_timer(&rule->dl_timer);
 656                 rule->dl_msg_send = -1;
 657
 658         } else if (!list_empty(msg_list)) {
 659                 /* dequeued some timedout messages, update timer for the
 660                  * next delayed message on rule */
 661                 msg = list_entry(rule->dl_msg_list.next,
 662                                  struct lnet_msg, msg_list);
 663                 rule->dl_msg_send = msg->msg_delay_send;
 664                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
 665         }
 666         spin_unlock(&rule->dl_lock);
 667 }
 668
 669 static void
 670 delayed_msg_process(struct list_head *msg_list, bool drop)
 671 {
 672         struct lnet_msg *msg;
 673
 674         while (!list_empty(msg_list)) {
 675                 struct lnet_ni *ni;
 676                 int             cpt;
 677                 int             rc;
 678
 679                 msg = list_entry(msg_list->next, struct lnet_msg, msg_list);
 680                 LASSERT(msg->msg_rxpeer != NULL);
 681                 LASSERT(msg->msg_rxni != NULL);
 682
 683                 ni = msg->msg_rxni;
 684                 cpt = msg->msg_rx_cpt;
 685
 686                 list_del_init(&msg->msg_list);
 687                 if (drop) {
 688                         rc = -ECANCELED;
 689
 690                 } else if (!msg->msg_routing) {
 691                         rc = lnet_parse_local(ni, msg);
 692                         if (rc == 0)
 693                                 continue;
 694
 695                 } else {
 696                         lnet_net_lock(cpt);
 697                         rc = lnet_parse_forward_locked(ni, msg);
 698                         lnet_net_unlock(cpt);
 699
 700                         switch (rc) {
 701                         case LNET_CREDIT_OK:
 702                                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
 703                                              0, msg->msg_len, msg->msg_len);
 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         struct list_head         msgs;
 726
 727         INIT_LIST_HEAD(&msgs);
 728         while (1) {
 729                 if (list_empty(&delay_dd.dd_sched_rules))
 730                         break;
 731
 732                 spin_lock_bh(&delay_dd.dd_lock);
 733                 if (list_empty(&delay_dd.dd_sched_rules)) {
 734                         spin_unlock_bh(&delay_dd.dd_lock);
 735                         break;
 736                 }
 737
 738                 rule = list_entry(delay_dd.dd_sched_rules.next,
 739                                   struct lnet_delay_rule, dl_sched_link);
 740                 list_del_init(&rule->dl_sched_link);
 741                 spin_unlock_bh(&delay_dd.dd_lock);
 742
 743                 delayed_msg_check(rule, false, &msgs);
 744                 delay_rule_decref(rule); /* -1 for delay_dd.dd_sched_rules */
 745         }
 746
 747         if (!list_empty(&msgs))
 748                 delayed_msg_process(&msgs, false);
 749 }
 750
 751 /** deamon thread to handle delayed messages */
 752 static int
 753 lnet_delay_rule_daemon(void *arg)
 754 {
 755         delay_dd.dd_running = 1;
 756         wake_up(&delay_dd.dd_ctl_waitq);
 757
 758         while (delay_dd.dd_running) {
 759                 wait_event_interruptible(delay_dd.dd_waitq,
 760                                          !delay_dd.dd_running ||
 761                                          !list_empty(&delay_dd.dd_sched_rules));
 762                 lnet_delay_rule_check();
 763         }
 764
 765         /* in case more rules have been enqueued after my last check */
 766         lnet_delay_rule_check();
 767         delay_dd.dd_stopped = 1;
 768         wake_up(&delay_dd.dd_ctl_waitq);
 769
 770         return 0;
 771 }
 772
 773 static void
 774 delay_timer_cb(cfs_timer_cb_arg_t data)
 775 {
 776         struct lnet_delay_rule *rule = cfs_from_timer(rule, data, dl_timer);
 777
 778         spin_lock_bh(&delay_dd.dd_lock);
 779         if (list_empty(&rule->dl_sched_link) && delay_dd.dd_running) {
 780                 atomic_inc(&rule->dl_refcount);
 781                 list_add_tail(&rule->dl_sched_link, &delay_dd.dd_sched_rules);
 782                 wake_up(&delay_dd.dd_waitq);
 783         }
 784         spin_unlock_bh(&delay_dd.dd_lock);
 785 }
 786
 787 /**
 788  * Add a new delay rule to LNet
 789  * There is no check for duplicated delay rule, all rules will be checked for
 790  * incoming message.
 791  */
 792 int
 793 lnet_delay_rule_add(struct lnet_fault_attr *attr)
 794 {
 795         struct lnet_delay_rule *rule;
 796         int                     rc = 0;
 797         ENTRY;
 798
 799         if (!((attr->u.delay.la_rate == 0) ^
 800               (attr->u.delay.la_interval == 0))) {
 801                 CDEBUG(D_NET,
 802                        "please provide either delay rate or delay interval, "
 803                        "but not both at the same time %d/%d\n",
 804                        attr->u.delay.la_rate, attr->u.delay.la_interval);
 805                 RETURN(-EINVAL);
 806         }
 807
 808         if (attr->u.delay.la_latency == 0) {
 809                 CDEBUG(D_NET, "delay latency cannot be zero\n");
 810                 RETURN(-EINVAL);
 811         }
 812
 813         if (lnet_fault_attr_validate(attr) != 0)
 814                 RETURN(-EINVAL);
 815
 816         CFS_ALLOC_PTR(rule);
 817         if (rule == NULL)
 818                 RETURN(-ENOMEM);
 819
 820         mutex_lock(&delay_dd.dd_mutex);
 821         if (!delay_dd.dd_running) {
 822                 struct task_struct *task;
 823
 824                 /* NB: although LND threads will process delayed message
 825                  * in lnet_finalize, but there is no guarantee that LND
 826                  * threads will be waken up if no other message needs to
 827                  * be handled.
 828                  * Only one daemon thread, performance is not the concern
 829                  * of this simualation module.
 830                  */
 831                 task = kthread_run(lnet_delay_rule_daemon, NULL, "lnet_dd");
 832                 if (IS_ERR(task)) {
 833                         rc = PTR_ERR(task);
 834                         GOTO(failed, rc);
 835                 }
 836                 wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_running);
 837         }
 838
 839         cfs_timer_setup(&rule->dl_timer, delay_timer_cb,
 840                         (unsigned long)rule, 0);
 841
 842         spin_lock_init(&rule->dl_lock);
 843         INIT_LIST_HEAD(&rule->dl_msg_list);
 844         INIT_LIST_HEAD(&rule->dl_sched_link);
 845
 846         rule->dl_attr = *attr;
 847         if (attr->u.delay.la_interval != 0) {
 848                 rule->dl_time_base = ktime_get_seconds() +
 849                                      attr->u.delay.la_interval;
 850                 rule->dl_delay_time = ktime_get_seconds() +
 851                                       prandom_u32_max(attr->u.delay.la_interval);
 852         } else {
 853                 rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
 854         }
 855
 856         rule->dl_msg_send = -1;
 857
 858         lnet_net_lock(LNET_LOCK_EX);
 859         atomic_set(&rule->dl_refcount, 1);
 860         list_add(&rule->dl_link, &the_lnet.ln_delay_rules);
 861         lnet_net_unlock(LNET_LOCK_EX);
 862
 863         CDEBUG(D_NET, "Added delay rule: src %s, dst %s, rate %d\n",
 864                libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
 865                attr->u.delay.la_rate);
 866
 867         mutex_unlock(&delay_dd.dd_mutex);
 868         RETURN(0);
 869  failed:
 870         mutex_unlock(&delay_dd.dd_mutex);
 871         CFS_FREE_PTR(rule);
 872         return rc;
 873 }
 874
 875 /**
 876  * Remove matched Delay Rules from lnet, if \a shutdown is true or both \a src
 877  * and \a dst are zero, all rules will be removed, otherwise only matched rules
 878  * will be removed.
 879  * If \a src is zero, then all rules have \a dst as destination will be remove
 880  * If \a dst is zero, then all rules have \a src as source will be removed
 881  *
 882  * When a delay rule is removed, all delayed messages of this rule will be
 883  * processed immediately.
 884  */
 885 int
 886 lnet_delay_rule_del(lnet_nid_t src, lnet_nid_t dst, bool shutdown)
 887 {
 888         struct lnet_delay_rule *rule;
 889         struct lnet_delay_rule  *tmp;
 890         struct list_head        rule_list;
 891         struct list_head        msg_list;
 892         int                     n = 0;
 893         bool                    cleanup;
 894         ENTRY;
 895
 896         INIT_LIST_HEAD(&rule_list);
 897         INIT_LIST_HEAD(&msg_list);
 898
 899         if (shutdown)
 900                 src = dst = 0;
 901
 902         mutex_lock(&delay_dd.dd_mutex);
 903         lnet_net_lock(LNET_LOCK_EX);
 904
 905         list_for_each_entry_safe(rule, tmp, &the_lnet.ln_delay_rules, dl_link) {
 906                 if (rule->dl_attr.fa_src != src && src != 0)
 907                         continue;
 908
 909                 if (rule->dl_attr.fa_dst != dst && dst != 0)
 910                         continue;
 911
 912                 CDEBUG(D_NET, "Remove delay rule: src %s->dst: %s (1/%d, %d)\n",
 913                        libcfs_nid2str(rule->dl_attr.fa_src),
 914                        libcfs_nid2str(rule->dl_attr.fa_dst),
 915                        rule->dl_attr.u.delay.la_rate,
 916                        rule->dl_attr.u.delay.la_interval);
 917                 /* refcount is taken over by rule_list */
 918                 list_move(&rule->dl_link, &rule_list);
 919         }
 920
 921         /* check if we need to shutdown delay_daemon */
 922         cleanup = list_empty(&the_lnet.ln_delay_rules) &&
 923                   !list_empty(&rule_list);
 924         lnet_net_unlock(LNET_LOCK_EX);
 925
 926         list_for_each_entry_safe(rule, tmp, &rule_list, dl_link) {
 927                 list_del_init(&rule->dl_link);
 928
 929                 del_timer_sync(&rule->dl_timer);
 930                 delayed_msg_check(rule, true, &msg_list);
 931                 delay_rule_decref(rule); /* -1 for the_lnet.ln_delay_rules */
 932                 n++;
 933         }
 934
 935         if (cleanup) { /* no more delay rule, shutdown delay_daemon */
 936                 LASSERT(delay_dd.dd_running);
 937                 delay_dd.dd_running = 0;
 938                 wake_up(&delay_dd.dd_waitq);
 939
 940                 while (!delay_dd.dd_stopped)
 941                         wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_stopped);
 942         }
 943         mutex_unlock(&delay_dd.dd_mutex);
 944
 945         if (!list_empty(&msg_list))
 946                 delayed_msg_process(&msg_list, shutdown);
 947
 948         RETURN(n);
 949 }
 950
 951 /**
 952  * List Delay Rule at position of \a pos
 953  */
 954 int
 955 lnet_delay_rule_list(int pos, struct lnet_fault_attr *attr,
 956                     struct lnet_fault_stat *stat)
 957 {
 958         struct lnet_delay_rule *rule;
 959         int                     cpt;
 960         int                     i = 0;
 961         int                     rc = -ENOENT;
 962         ENTRY;
 963
 964         cpt = lnet_net_lock_current();
 965         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 966                 if (i++ < pos)
 967                         continue;
 968
 969                 spin_lock(&rule->dl_lock);
 970                 *attr = rule->dl_attr;
 971                 *stat = rule->dl_stat;
 972                 spin_unlock(&rule->dl_lock);
 973                 rc = 0;
 974                 break;
 975         }
 976
 977         lnet_net_unlock(cpt);
 978         RETURN(rc);
 979 }
 980
 981 /**
 982  * reset counters for all Delay Rules
 983  */
 984 void
 985 lnet_delay_rule_reset(void)
 986 {
 987         struct lnet_delay_rule *rule;
 988         int                     cpt;
 989         ENTRY;
 990
 991         cpt = lnet_net_lock_current();
 992
 993         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 994                 struct lnet_fault_attr *attr = &rule->dl_attr;
 995
 996                 spin_lock(&rule->dl_lock);
 997
 998                 memset(&rule->dl_stat, 0, sizeof(rule->dl_stat));
 999                 if (attr->u.delay.la_rate != 0) {
1000                         rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
1001                 } else {
1002                         rule->dl_delay_time = ktime_get_seconds() +
1003                                               prandom_u32_max(attr->u.delay.la_interval);
1004                         rule->dl_time_base = ktime_get_seconds() +
1005                                              attr->u.delay.la_interval;
1006                 }
1007                 spin_unlock(&rule->dl_lock);
1008         }
1009
1010         lnet_net_unlock(cpt);
1011         EXIT;
1012 }
1013
1014 int
1015 lnet_fault_ctl(int opc, struct libcfs_ioctl_data *data)
1016 {
1017         struct lnet_fault_attr *attr;
1018         struct lnet_fault_stat *stat;
1019
1020         attr = (struct lnet_fault_attr *)data->ioc_inlbuf1;
1021
1022         switch (opc) {
1023         default:
1024                 return -EINVAL;
1025
1026         case LNET_CTL_DROP_ADD:
1027                 if (attr == NULL)
1028                         return -EINVAL;
1029
1030                 return lnet_drop_rule_add(attr);
1031
1032         case LNET_CTL_DROP_DEL:
1033                 if (attr == NULL)
1034                         return -EINVAL;
1035
1036                 data->ioc_count = lnet_drop_rule_del(attr->fa_src,
1037                                                      attr->fa_dst);
1038                 return 0;
1039
1040         case LNET_CTL_DROP_RESET:
1041                 lnet_drop_rule_reset();
1042                 return 0;
1043
1044         case LNET_CTL_DROP_LIST:
1045                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1046                 if (attr == NULL || stat == NULL)
1047                         return -EINVAL;
1048
1049                 return lnet_drop_rule_list(data->ioc_count, attr, stat);
1050
1051         case LNET_CTL_DELAY_ADD:
1052                 if (attr == NULL)
1053                         return -EINVAL;
1054
1055                 return lnet_delay_rule_add(attr);
1056
1057         case LNET_CTL_DELAY_DEL:
1058                 if (attr == NULL)
1059                         return -EINVAL;
1060
1061                 data->ioc_count = lnet_delay_rule_del(attr->fa_src,
1062                                                       attr->fa_dst, false);
1063                 return 0;
1064
1065         case LNET_CTL_DELAY_RESET:
1066                 lnet_delay_rule_reset();
1067                 return 0;
1068
1069         case LNET_CTL_DELAY_LIST:
1070                 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1071                 if (attr == NULL || stat == NULL)
1072                         return -EINVAL;
1073
1074                 return lnet_delay_rule_list(data->ioc_count, attr, stat);
1075         }
1076 }
1077
1078 int
1079 lnet_fault_init(void)
1080 {
1081         CLASSERT(LNET_PUT_BIT == 1 << LNET_MSG_PUT);
1082         CLASSERT(LNET_ACK_BIT == 1 << LNET_MSG_ACK);
1083         CLASSERT(LNET_GET_BIT == 1 << LNET_MSG_GET);
1084         CLASSERT(LNET_REPLY_BIT == 1 << LNET_MSG_REPLY);
1085
1086         mutex_init(&delay_dd.dd_mutex);
1087         spin_lock_init(&delay_dd.dd_lock);
1088         init_waitqueue_head(&delay_dd.dd_waitq);
1089         init_waitqueue_head(&delay_dd.dd_ctl_waitq);
1090         INIT_LIST_HEAD(&delay_dd.dd_sched_rules);
1091
1092         return 0;
1093 }
1094
1095 void
1096 lnet_fault_fini(void)
1097 {
1098         lnet_drop_rule_del(0, 0);
1099         lnet_delay_rule_del(0, 0, true);
1100
1101         LASSERT(list_empty(&the_lnet.ln_drop_rules));
1102         LASSERT(list_empty(&the_lnet.ln_delay_rules));
1103         LASSERT(list_empty(&delay_dd.dd_sched_rules));
1104 }