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, 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 <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 : "
 330                                       CFS_TIME_T"\n",
 331                                       libcfs_nid2str(attr->fa_src),
 332                                       libcfs_nid2str(attr->fa_dst),
 333                                       rule->dr_drop_time);
 334                 }
 335
 336         } else { /* rate based drop */
 337                 __u64 count;
 338
 339                 drop = rule->dr_stat.fs_count++ == rule->dr_drop_at;
 340                 count = rule->dr_stat.fs_count;
 341                 if (do_div(count, attr->u.drop.da_rate) == 0) {
 342                         rule->dr_drop_at = rule->dr_stat.fs_count +
 343                                            cfs_rand() % attr->u.drop.da_rate;
 344                         CDEBUG(D_NET, "Drop Rule %s->%s: next drop: %lu\n",
 345                                libcfs_nid2str(attr->fa_src),
 346                                libcfs_nid2str(attr->fa_dst), rule->dr_drop_at);
 347                 }
 348         }
 349
 350         if (drop) { /* drop this message, update counters */
 351                 lnet_fault_stat_inc(&rule->dr_stat, type);
 352                 rule->dr_stat.u.drop.ds_dropped++;
 353         }
 354
 355         spin_unlock(&rule->dr_lock);
 356         return drop;
 357 }
 358
 359 /**
 360  * Check if message from \a src to \a dst can match any existed drop rule
 361  */
 362 bool
 363 lnet_drop_rule_match(lnet_hdr_t *hdr)
 364 {
 365         struct lnet_drop_rule   *rule;
 366         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
 367         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
 368         unsigned int             typ = le32_to_cpu(hdr->type);
 369         unsigned int             ptl = -1;
 370         bool                     drop = false;
 371         int                      cpt;
 372
 373         /* NB: if Portal is specified, then only PUT and GET will be
 374          * filtered by drop rule */
 375         if (typ == LNET_MSG_PUT)
 376                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 377         else if (typ == LNET_MSG_GET)
 378                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 379
 380         cpt = lnet_net_lock_current();
 381         list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
 382                 drop = drop_rule_match(rule, src, dst, typ, ptl);
 383                 if (drop)
 384                         break;
 385         }
 386
 387         lnet_net_unlock(cpt);
 388         return drop;
 389 }
 390
 391 /**
 392  * LNet Delay Simulation
 393  */
 394 /** timestamp (second) to send delayed message */
 395 #define msg_delay_send           msg_ev.hdr_data
 396
 397 struct lnet_delay_rule {
 398         /** link chain on the_lnet.ln_delay_rules */
 399         struct list_head        dl_link;
 400         /** link chain on delay_dd.dd_sched_rules */
 401         struct list_head        dl_sched_link;
 402         /** attributes of this rule */
 403         struct lnet_fault_attr  dl_attr;
 404         /** lock to protect \a below members */
 405         spinlock_t              dl_lock;
 406         /** refcount of delay rule */
 407         atomic_t                dl_refcount;
 408         /**
 409          * the message sequence to delay, which means message is delayed when
 410          * dl_stat.fs_count == dl_delay_at
 411          */
 412         unsigned long           dl_delay_at;
 413         /**
 414          * seconds to delay the next message, it's exclusive with dl_delay_at
 415          */
 416         cfs_time_t              dl_delay_time;
 417         /** baseline to caculate dl_delay_time */
 418         cfs_time_t              dl_time_base;
 419         /** jiffies to send the next delayed message */
 420         unsigned long           dl_msg_send;
 421         /** delayed message list */
 422         struct list_head        dl_msg_list;
 423         /** statistic of delayed messages */
 424         struct lnet_fault_stat  dl_stat;
 425         /** timer to wakeup delay_daemon */
 426         struct timer_list       dl_timer;
 427 };
 428
 429 struct delay_daemon_data {
 430         /** serialise rule add/remove */
 431         struct mutex            dd_mutex;
 432         /** protect rules on \a dd_sched_rules */
 433         spinlock_t              dd_lock;
 434         /** scheduled delay rules (by timer) */
 435         struct list_head        dd_sched_rules;
 436         /** deamon thread sleeps at here */
 437         wait_queue_head_t       dd_waitq;
 438         /** controler (lctl command) wait at here */
 439         wait_queue_head_t       dd_ctl_waitq;
 440         /** deamon is running */
 441         unsigned int            dd_running;
 442         /** deamon stopped */
 443         unsigned int            dd_stopped;
 444 };
 445
 446 static struct delay_daemon_data delay_dd;
 447
 448 static cfs_time_t
 449 round_timeout(cfs_time_t timeout)
 450 {
 451         return cfs_time_seconds((unsigned int)
 452                         cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
 453 }
 454
 455 static void
 456 delay_rule_decref(struct lnet_delay_rule *rule)
 457 {
 458         if (atomic_dec_and_test(&rule->dl_refcount)) {
 459                 LASSERT(list_empty(&rule->dl_sched_link));
 460                 LASSERT(list_empty(&rule->dl_msg_list));
 461                 LASSERT(list_empty(&rule->dl_link));
 462
 463                 CFS_FREE_PTR(rule);
 464         }
 465 }
 466
 467 /**
 468  * check source/destination NID, portal, message type and delay rate,
 469  * decide whether should delay this message or not
 470  */
 471 static bool
 472 delay_rule_match(struct lnet_delay_rule *rule, lnet_nid_t src,
 473                 lnet_nid_t dst, unsigned int type, unsigned int portal,
 474                 struct lnet_msg *msg)
 475 {
 476         struct lnet_fault_attr  *attr = &rule->dl_attr;
 477         bool                     delay;
 478
 479         if (!lnet_fault_attr_match(attr, src, dst, type, portal))
 480                 return false;
 481
 482         /* match this rule, check delay rate now */
 483         spin_lock(&rule->dl_lock);
 484         if (rule->dl_delay_time != 0) { /* time based delay */
 485                 cfs_time_t now = cfs_time_current();
 486
 487                 rule->dl_stat.fs_count++;
 488                 delay = cfs_time_aftereq(now, rule->dl_delay_time);
 489                 if (delay) {
 490                         if (cfs_time_after(now, rule->dl_time_base))
 491                                 rule->dl_time_base = now;
 492
 493                         rule->dl_delay_time = rule->dl_time_base +
 494                                              cfs_time_seconds(cfs_rand() %
 495                                                 attr->u.delay.la_interval);
 496                         rule->dl_time_base += cfs_time_seconds(attr->u.delay.
 497                                                                la_interval);
 498
 499                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay : "
 500                                       CFS_TIME_T"\n",
 501                                       libcfs_nid2str(attr->fa_src),
 502                                       libcfs_nid2str(attr->fa_dst),
 503                                       rule->dl_delay_time);
 504                 }
 505
 506         } else { /* rate based delay */
 507                 __u64 count;
 508
 509                 delay = rule->dl_stat.fs_count++ == rule->dl_delay_at;
 510                 /* generate the next random rate sequence */
 511                 count = rule->dl_stat.fs_count;
 512                 if (do_div(count, attr->u.delay.la_rate) == 0) {
 513                         rule->dl_delay_at = rule->dl_stat.fs_count +
 514                                             cfs_rand() % attr->u.delay.la_rate;
 515                         CDEBUG(D_NET, "Delay Rule %s->%s: next delay: %lu\n",
 516                                libcfs_nid2str(attr->fa_src),
 517                                libcfs_nid2str(attr->fa_dst), rule->dl_delay_at);
 518                 }
 519         }
 520
 521         if (!delay) {
 522                 spin_unlock(&rule->dl_lock);
 523                 return false;
 524         }
 525
 526         /* delay this message, update counters */
 527         lnet_fault_stat_inc(&rule->dl_stat, type);
 528         rule->dl_stat.u.delay.ls_delayed++;
 529
 530         list_add_tail(&msg->msg_list, &rule->dl_msg_list);
 531         msg->msg_delay_send = round_timeout(
 532                         cfs_time_shift(attr->u.delay.la_latency));
 533         if (rule->dl_msg_send == -1) {
 534                 rule->dl_msg_send = msg->msg_delay_send;
 535                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
 536         }
 537
 538         spin_unlock(&rule->dl_lock);
 539         return true;
 540 }
 541
 542 /**
 543  * check if \a msg can match any Delay Rule, receiving of this message
 544  * will be delayed if there is a match.
 545  */
 546 bool
 547 lnet_delay_rule_match_locked(lnet_hdr_t *hdr, struct lnet_msg *msg)
 548 {
 549         struct lnet_delay_rule  *rule;
 550         lnet_nid_t               src = le64_to_cpu(hdr->src_nid);
 551         lnet_nid_t               dst = le64_to_cpu(hdr->dest_nid);
 552         unsigned int             typ = le32_to_cpu(hdr->type);
 553         unsigned int             ptl = -1;
 554
 555         /* NB: called with hold of lnet_net_lock */
 556
 557         /* NB: if Portal is specified, then only PUT and GET will be
 558          * filtered by delay rule */
 559         if (typ == LNET_MSG_PUT)
 560                 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
 561         else if (typ == LNET_MSG_GET)
 562                 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
 563
 564         list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
 565                 if (delay_rule_match(rule, src, dst, typ, ptl, msg))
 566                         return true;
 567         }
 568
 569         return false;
 570 }
 571
 572 /** check out delayed messages for send */
 573 static void
 574 delayed_msg_check(struct lnet_delay_rule *rule, bool all,
 575                   struct list_head *msg_list)
 576 {
 577         struct lnet_msg *msg;
 578         struct lnet_msg *tmp;
 579         unsigned long    now = cfs_time_current();
 580
 581         if (!all && rule->dl_msg_send > now)
 582                 return;
 583
 584         spin_lock(&rule->dl_lock);
 585         list_for_each_entry_safe(msg, tmp, &rule->dl_msg_list, msg_list) {
 586                 if (!all && msg->msg_delay_send > now)
 587                         break;
 588
 589                 msg->msg_delay_send = 0;
 590                 list_move_tail(&msg->msg_list, msg_list);
 591         }
 592
 593         if (list_empty(&rule->dl_msg_list)) {
 594                 del_timer(&rule->dl_timer);
 595                 rule->dl_msg_send = -1;
 596
 597         } else if (!list_empty(msg_list)) {
 598                 /* dequeued some timedout messages, update timer for the
 599                  * next delayed message on rule */
 600                 msg = list_entry(rule->dl_msg_list.next,
 601                                  struct lnet_msg, msg_list);
 602                 rule->dl_msg_send = msg->msg_delay_send;
 603                 mod_timer(&rule->dl_timer, rule->dl_msg_send);
 604         }
 605         spin_unlock(&rule->dl_lock);
 606 }
 607
 608 static void
 609 delayed_msg_process(struct list_head *msg_list, bool drop)
 610 {
 611         struct lnet_msg *msg;
 612
 613         while (!list_empty(msg_list)) {
 614                 struct lnet_ni *ni;
 615                 int             cpt;
 616                 int             rc;
 617
 618                 msg = list_entry(msg_list->next, struct lnet_msg, msg_list);
 619                 LASSERT(msg->msg_rxpeer != NULL);
 620
 621                 ni = msg->msg_rxpeer->lp_ni;
 622                 cpt = msg->msg_rx_cpt;
 623
 624                 list_del_init(&msg->msg_list);
 625                 if (drop) {
 626                         rc = -ECANCELED;
 627
 628                 } else if (!msg->msg_routing) {
 629                         rc = lnet_parse_local(ni, msg);
 630                         if (rc == 0)
 631                                 continue;
 632
 633                 } else {
 634                         lnet_net_lock(cpt);
 635                         rc = lnet_parse_forward_locked(ni, msg);
 636                         lnet_net_unlock(cpt);
 637
 638                         switch (rc) {
 639                         case LNET_CREDIT_OK:
 640                                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
 641                                              0, msg->msg_len, msg->msg_len);
 642                         case LNET_CREDIT_WAIT:
 643                                 continue;
 644                         default: /* failures */
 645                                 break;
 646                         }
 647                 }
 648
 649                 lnet_drop_message(ni, cpt, msg->msg_private, msg->msg_len);
 650                 lnet_finalize(ni, 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 }