lustre/ldlm/ldlm_test.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (c) 2002 Cluster File Systems, Inc. <info@clusterfs.com>
   5  * Copyright (c) 2002 Lawrence Livermore National Laboratory
   6  *  Author: James Newsome <newsome2@llnl.gov>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #include <asm/atomic.h>
  27 #include <linux/types.h>
  28 #include <linux/random.h>
  29
  30 #include <linux/lustre_dlm.h>
  31 #include <linux/obd.h>
  32
  33 struct ldlm_test_thread {
  34         struct obd_device *obddev;
  35         struct ldlm_namespace *t_ns;
  36         struct list_head t_link;
  37         __u32 t_flags;
  38         wait_queue_head_t t_ctl_waitq;
  39 };
  40
  41 struct ldlm_test_lock {
  42         struct list_head l_link;
  43         struct lustre_handle l_lockh;
  44 };
  45
  46 static unsigned int max_locks;
  47 static unsigned int num_resources;
  48 static unsigned int num_extents;
  49
  50 static spinlock_t ctl_lock = SPIN_LOCK_UNLOCKED;
  51 /* protect these with the ctl_lock */
  52 static LIST_HEAD(ctl_threads);
  53 static int regression_running = 0;
  54 static LIST_HEAD(lock_list);
  55 static int num_locks = 0;
  56
  57 /* cumulative stats for regression test */
  58 static atomic_t locks_requested = ATOMIC_INIT(0);
  59 static atomic_t converts_requested = ATOMIC_INIT(0);
  60 static atomic_t locks_granted = ATOMIC_INIT(0);
  61 static atomic_t locks_matched = ATOMIC_INIT(0);
  62
  63 /* making this a global avoids the problem of having pointers
  64  * to garbage after the test exits.
  65  */
  66 static struct lustre_handle regress_connh;
  67
  68 static int ldlm_do_decrement(void);
  69 static int ldlm_do_enqueue(struct ldlm_test_thread *thread);
  70 static int ldlm_do_convert(void);
  71
  72 /*
  73  * blocking ast for regression test.
  74  * Just cancels lock
  75  */
  76 static int ldlm_test_blocking_ast(struct ldlm_lock *lock,
  77                                   struct ldlm_lock_desc *new,
  78                                   void *data, __u32 data_len, int flag)
  79 {
  80         int rc;
  81         struct lustre_handle lockh;
  82         ENTRY;
  83
  84         switch (flag) {
  85         case LDLM_CB_BLOCKING:
  86                 LDLM_DEBUG(lock, "We're blocking. Cancelling lock");
  87                 ldlm_lock2handle(lock, &lockh);
  88                 rc = ldlm_cli_cancel(&lockh);
  89                 if (rc < 0) {
  90                         CERROR("ldlm_cli_cancel: %d\n", rc);
  91                         LBUG();
  92                 }
  93                 break;
  94         case LDLM_CB_CANCELING:
  95                 LDLM_DEBUG(lock, "this lock is being cancelled");
  96                 break;
  97         default:
  98                 LBUG();
  99         }
 100
 101         RETURN(0);
 102 }
 103
 104 /* blocking ast for basic tests. noop */
 105 static int ldlm_blocking_ast(struct ldlm_lock *lock,
 106                              struct ldlm_lock_desc *new,
 107                              void *data, __u32 data_len, int flag)
 108 {
 109         ENTRY;
 110         CERROR("ldlm_blocking_ast: lock=%p, new=%p, flag=%d\n", lock, new,
 111                flag);
 112         RETURN(0);
 113 }
 114
 115 /* Completion ast for regression test.
 116  * Does not sleep when blocked.
 117  */
 118 static int ldlm_test_completion_ast(struct ldlm_lock *lock, int flags)
 119 {
 120         struct ldlm_test_lock *lock_info;
 121         ENTRY;
 122
 123         if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 124                      LDLM_FL_BLOCK_CONV)) {
 125                 LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock");
 126                 RETURN(0);
 127         }
 128
 129         if (lock->l_granted_mode != lock->l_req_mode)
 130                 CERROR("completion ast called with non-granted lock\n");
 131
 132         /* add to list of granted locks */
 133
 134         if (flags & LDLM_FL_WAIT_NOREPROC) {
 135                 atomic_inc(&locks_matched);
 136                 LDLM_DEBUG(lock, "lock matched");
 137         } else {
 138                 atomic_inc(&locks_granted);
 139                 LDLM_DEBUG(lock, "lock granted");
 140         }
 141
 142         OBD_ALLOC(lock_info, sizeof(*lock_info));
 143         if (lock_info == NULL) {
 144                 LBUG();
 145                 RETURN(-ENOMEM);
 146         }
 147
 148         ldlm_lock2handle(lock, &lock_info->l_lockh);
 149
 150         spin_lock(&ctl_lock);
 151         list_add_tail(&lock_info->l_link, &lock_list);
 152         num_locks++;
 153         spin_unlock(&ctl_lock);
 154
 155         RETURN(0);
 156 }
 157
 158 int ldlm_test_basics(struct obd_device *obddev)
 159 {
 160         struct ldlm_namespace *ns;
 161         struct ldlm_resource *res;
 162         __u64 res_id[RES_NAME_SIZE] = {1, 2, 3};
 163         ldlm_error_t err;
 164         struct ldlm_lock *lock1, *lock;
 165         int flags;
 166         ENTRY;
 167
 168         ns = ldlm_namespace_new("test_server", LDLM_NAMESPACE_SERVER);
 169         if (ns == NULL)
 170                 LBUG();
 171
 172         lock1 = ldlm_lock_create(ns, NULL, res_id, LDLM_PLAIN, LCK_CR, NULL, 0);
 173         if (lock1 == NULL)
 174                 LBUG();
 175         err = ldlm_lock_enqueue(lock1, NULL, 0, &flags,
 176                                 ldlm_completion_ast, ldlm_blocking_ast);
 177         if (err != ELDLM_OK)
 178                 LBUG();
 179
 180         lock = ldlm_lock_create(ns, NULL, res_id, LDLM_PLAIN, LCK_EX, NULL, 0);
 181         if (lock == NULL)
 182                 LBUG();
 183         err = ldlm_lock_enqueue(lock, NULL, 0, &flags,
 184                                 ldlm_completion_ast, ldlm_blocking_ast);
 185         if (err != ELDLM_OK)
 186                 LBUG();
 187         if (!(flags & LDLM_FL_BLOCK_GRANTED))
 188                 LBUG();
 189
 190         res = ldlm_resource_get(ns, NULL, res_id, LDLM_PLAIN, 1);
 191         if (res == NULL)
 192                 LBUG();
 193         ldlm_resource_dump(res);
 194
 195         res = ldlm_lock_convert(lock1, LCK_NL, &flags);
 196         if (res != NULL)
 197                 ldlm_reprocess_all(res);
 198
 199         ldlm_resource_dump(res);
 200         ldlm_namespace_free(ns);
 201
 202         RETURN(0);
 203 }
 204
 205 int ldlm_test_extents(struct obd_device *obddev)
 206 {
 207         struct ldlm_namespace *ns;
 208         struct ldlm_resource *res;
 209         struct ldlm_lock *lock, *lock1, *lock2;
 210         __u64 res_id[RES_NAME_SIZE] = {0, 0, 0};
 211         struct ldlm_extent ext1 = {4, 6}, ext2 = {6, 9}, ext3 = {10, 11};
 212         ldlm_error_t err;
 213         int flags;
 214         ENTRY;
 215
 216         ns = ldlm_namespace_new("test_server", LDLM_NAMESPACE_SERVER);
 217         if (ns == NULL)
 218                 LBUG();
 219
 220         flags = 0;
 221         lock1 = ldlm_lock_create(ns, NULL, res_id, LDLM_EXTENT, LCK_PR, NULL,
 222                                  0);
 223         if (lock1 == NULL)
 224                 LBUG();
 225         err = ldlm_lock_enqueue(lock1, &ext1, sizeof(ext1), &flags, NULL, NULL);
 226         if (err != ELDLM_OK)
 227                 LBUG();
 228         if (!(flags & LDLM_FL_LOCK_CHANGED))
 229                 LBUG();
 230
 231         flags = 0;
 232         lock2 = ldlm_lock_create(ns, NULL, res_id, LDLM_EXTENT, LCK_PR,
 233                                 NULL, 0);
 234         err = ldlm_lock_enqueue(lock2, &ext2, sizeof(ext2), &flags, NULL, NULL);
 235         if (err != ELDLM_OK)
 236                 LBUG();
 237         if (!(flags & LDLM_FL_LOCK_CHANGED))
 238                 LBUG();
 239
 240         flags = 0;
 241         lock = ldlm_lock_create(ns, NULL, res_id, LDLM_EXTENT, LCK_EX, NULL, 0);
 242         if (lock == NULL)
 243                 LBUG();
 244         err = ldlm_lock_enqueue(lock, &ext3, sizeof(ext3), &flags,
 245                                 NULL, NULL);
 246         if (err != ELDLM_OK)
 247                 LBUG();
 248         if (!(flags & LDLM_FL_BLOCK_GRANTED))
 249                 LBUG();
 250         if (flags & LDLM_FL_LOCK_CHANGED)
 251                 LBUG();
 252
 253         /* Convert/cancel blocking locks */
 254         flags = 0;
 255         res = ldlm_lock_convert(lock1, LCK_NL, &flags);
 256         if (res != NULL)
 257                 ldlm_reprocess_all(res);
 258
 259         ldlm_lock_cancel(lock2);
 260         if (res != NULL)
 261                 ldlm_reprocess_all(res);
 262
 263         /* Dump the results */
 264         res = ldlm_resource_get(ns, NULL, res_id, LDLM_EXTENT, 0);
 265         if (res == NULL)
 266                 LBUG();
 267         ldlm_resource_dump(res);
 268         ldlm_namespace_free(ns);
 269
 270         RETURN(0);
 271 }
 272
 273 static int ldlm_test_network(struct obd_device *obddev,
 274                              struct lustre_handle *connh)
 275 {
 276
 277         __u64 res_id[RES_NAME_SIZE] = {1, 2, 3};
 278         struct ldlm_extent ext = {4, 6};
 279         struct lustre_handle lockh1;
 280         struct ldlm_lock *lock;
 281         int flags = 0;
 282         ldlm_error_t err;
 283         ENTRY;
 284
 285         err = ldlm_cli_enqueue(connh, NULL, obddev->obd_namespace, NULL, res_id,
 286                                LDLM_EXTENT, &ext, sizeof(ext), LCK_PR, &flags,
 287                                ldlm_completion_ast, NULL, NULL, 0, &lockh1);
 288
 289         CERROR("ldlm_cli_enqueue: %d\n", err);
 290
 291         flags = 0;
 292         err = ldlm_cli_convert(&lockh1, LCK_EX, &flags);
 293         CERROR("ldlm_cli_convert: %d\n", err);
 294
 295         lock = ldlm_handle2lock(&lockh1);
 296         ldlm_lock_dump(lock);
 297         ldlm_lock_put(lock);
 298
 299         /* Need to decrement old mode. Don't bother incrementing new
 300          * mode since the test is done.
 301          */
 302         if (err == ELDLM_OK)
 303                 ldlm_lock_decref(&lockh1, LCK_PR);
 304
 305         RETURN(err);
 306 }
 307
 308 static int ldlm_do_decrement(void)
 309 {
 310         struct ldlm_test_lock *lock_info;
 311         struct ldlm_lock *lock;
 312         int rc = 0;
 313         ENTRY;
 314
 315         spin_lock(&ctl_lock);
 316         if(list_empty(&lock_list)) {
 317                 CERROR("lock_list is empty\n");
 318                 spin_unlock(&ctl_lock);
 319                 RETURN(0);
 320         }
 321
 322         /* delete from list */
 323         lock_info = list_entry(lock_list.next,
 324                         struct ldlm_test_lock, l_link);
 325         list_del(lock_list.next);
 326         num_locks--;
 327         spin_unlock(&ctl_lock);
 328
 329         /* decrement and free the info */
 330         lock = ldlm_handle2lock(&lock_info->l_lockh);
 331         ldlm_lock_decref(&lock_info->l_lockh, lock->l_granted_mode);
 332         ldlm_lock_put(lock);
 333
 334         OBD_FREE(lock_info, sizeof(*lock_info));
 335
 336         RETURN(rc);
 337 }
 338
 339 static int ldlm_do_enqueue(struct ldlm_test_thread *thread)
 340 {
 341         struct lustre_handle lockh;
 342         __u64 res_id[3] = {0};
 343         __u32 lock_mode;
 344         struct ldlm_extent ext;
 345         unsigned char random;
 346         int flags = 0, rc = 0;
 347         ENTRY;
 348
 349         /* Pick a random resource from 1 to num_resources */
 350         get_random_bytes(&random, sizeof(random));
 351         res_id[0] = random % num_resources;
 352
 353         /* Pick a random lock mode */
 354         get_random_bytes(&random, sizeof(random));
 355         lock_mode = random % LCK_NL + 1;
 356
 357         /* Pick a random extent */
 358         get_random_bytes(&random, sizeof(random));
 359         ext.start = random % num_extents;
 360         get_random_bytes(&random, sizeof(random));
 361         ext.end = random %
 362                 (num_extents - (int)ext.start) + ext.start;
 363
 364         LDLM_DEBUG_NOLOCK("about to enqueue with resource "LPX64", mode %d,"
 365                           " extent "LPX64" -> "LPX64, res_id[0], lock_mode,
 366                           ext.start, ext.end);
 367
 368         rc = ldlm_match_or_enqueue(&regress_connh, NULL,
 369                                    thread->obddev->obd_namespace,
 370                                    NULL, res_id, LDLM_EXTENT, &ext,
 371                                    sizeof(ext), lock_mode, &flags,
 372                                    ldlm_test_completion_ast,
 373                                    ldlm_test_blocking_ast,
 374                                    NULL, 0, &lockh);
 375
 376         atomic_inc(&locks_requested);
 377
 378         if (rc < 0) {
 379                 CERROR("ldlm_cli_enqueue: %d\n", rc);
 380                 LBUG();
 381         }
 382
 383         RETURN(rc);
 384 }
 385
 386 static int ldlm_do_convert(void)
 387 {
 388         __u32 lock_mode;
 389         unsigned char random;
 390         int flags = 0, rc = 0;
 391         struct ldlm_test_lock *lock_info;
 392         struct ldlm_lock *lock;
 393         ENTRY;
 394
 395         /* delete from list */
 396         spin_lock(&ctl_lock);
 397         lock_info = list_entry(lock_list.next, struct ldlm_test_lock, l_link);
 398         list_del(lock_list.next);
 399         num_locks--;
 400         spin_unlock(&ctl_lock);
 401
 402         /* Pick a random lock mode */
 403         get_random_bytes(&random, sizeof(random));
 404         lock_mode = random % LCK_NL + 1;
 405
 406         /* do the conversion */
 407         rc = ldlm_cli_convert(&lock_info->l_lockh , lock_mode, &flags);
 408         atomic_inc(&converts_requested);
 409
 410         if (rc < 0) {
 411                 CERROR("ldlm_cli_convert: %d\n", rc);
 412                 LBUG();
 413         }
 414
 415         /*
 416          *  Adjust reference counts.
 417          *  FIXME: This is technically a bit... wrong,
 418          *  since we don't know when/if the convert succeeded
 419          */
 420         ldlm_lock_addref(&lock_info->l_lockh, lock_mode);
 421         lock = ldlm_handle2lock(&lock_info->l_lockh);
 422         ldlm_lock_decref(&lock_info->l_lockh, lock->l_granted_mode);
 423         ldlm_lock_put(lock);
 424
 425         OBD_FREE(lock_info, sizeof(*lock_info));
 426
 427         RETURN(rc);
 428 }
 429
 430
 431
 432 static int ldlm_test_main(void *data)
 433 {
 434         struct ldlm_test_thread *thread = data;
 435         ENTRY;
 436
 437         lock_kernel();
 438         daemonize();
 439         spin_lock_irq(&current->sigmask_lock);
 440         sigfillset(&current->blocked);
 441 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
 442         recalc_sigpending();
 443 #else
 444         recalc_sigpending(current);
 445 #endif
 446         spin_unlock_irq(&current->sigmask_lock);
 447
 448         sprintf(current->comm, "ldlm_test");
 449         unlock_kernel();
 450
 451         /* Record that the thread is running */
 452         thread->t_flags |= SVC_RUNNING;
 453         wake_up(&thread->t_ctl_waitq);
 454
 455         while (!(thread->t_flags & SVC_STOPPING)) {
 456                 unsigned char random;
 457                 unsigned char dec_chance, con_chance;
 458                 unsigned char chance_left = 100;
 459
 460                 spin_lock(&ctl_lock);
 461                 /* probability of decrementing increases linearly
 462                  * as more locks are held.
 463                  */
 464                 dec_chance = chance_left * num_locks / max_locks;
 465                 chance_left -= dec_chance;
 466
 467                 /* FIXME: conversions temporarily disabled
 468                  * until they are working correctly.
 469                  */
 470                 /* con_chance = chance_left * num_locks / max_locks; */
 471                 con_chance = 0;
 472                 chance_left -= con_chance;
 473                 spin_unlock(&ctl_lock);
 474
 475                 get_random_bytes(&random, sizeof(random));
 476
 477                 random = random % 100;
 478                 if (random < dec_chance)
 479                         ldlm_do_decrement();
 480                 else if (random < (dec_chance + con_chance))
 481                         ldlm_do_convert();
 482                 else
 483                         ldlm_do_enqueue(thread);
 484
 485                 LDLM_DEBUG_NOLOCK("locks requested: %d, "
 486                                   "conversions requested %d",
 487                                   atomic_read(&locks_requested),
 488                                   atomic_read(&converts_requested));
 489                 LDLM_DEBUG_NOLOCK("locks granted: %d, "
 490                                   "locks matched: %d",
 491                                   atomic_read(&locks_granted),
 492                                   atomic_read(&locks_matched));
 493
 494                 spin_lock(&ctl_lock);
 495                 LDLM_DEBUG_NOLOCK("lock references currently held: %d, ",
 496                                   num_locks);
 497                 spin_unlock(&ctl_lock);
 498
 499                 /*
 500                  * We don't sleep after a lock being blocked, so let's
 501                  * make sure other things can run.
 502                  */
 503                 schedule();
 504         }
 505
 506         thread->t_flags |= SVC_STOPPED;
 507         wake_up(&thread->t_ctl_waitq);
 508
 509         RETURN(0);
 510 }
 511
 512 static int ldlm_start_thread(struct obd_device *obddev,
 513                              struct lustre_handle *connh)
 514 {
 515         struct ldlm_test_thread *test;
 516         int rc;
 517         ENTRY;
 518
 519         OBD_ALLOC(test, sizeof(*test));
 520         if (test == NULL) {
 521                 LBUG();
 522                 RETURN(-ENOMEM);
 523         }
 524         init_waitqueue_head(&test->t_ctl_waitq);
 525
 526         test->obddev = obddev;
 527
 528         spin_lock(&ctl_lock);
 529         list_add(&test->t_link, &ctl_threads);
 530         spin_unlock(&ctl_lock);
 531
 532         rc = kernel_thread(ldlm_test_main, (void *)test,
 533                            CLONE_VM | CLONE_FS | CLONE_FILES);
 534         if (rc < 0) {
 535                 CERROR("cannot start thread\n");
 536                 RETURN(-EINVAL);
 537         }
 538         wait_event(test->t_ctl_waitq, test->t_flags & SVC_RUNNING);
 539
 540         RETURN(0);
 541 }
 542
 543 int ldlm_regression_start(struct obd_device *obddev,
 544                           struct lustre_handle *connh,
 545                           unsigned int threads, unsigned int max_locks_in,
 546                           unsigned int num_resources_in,
 547                           unsigned int num_extents_in)
 548 {
 549         int i, rc = 0;
 550         ENTRY;
 551
 552         spin_lock(&ctl_lock);
 553         if (regression_running) {
 554                 CERROR("You can't start the ldlm regression twice.\n");
 555                 spin_unlock(&ctl_lock);
 556                 RETURN(-EINVAL);
 557         }
 558         regression_running = 1;
 559         spin_unlock(&ctl_lock);
 560
 561         regress_connh = *connh;
 562         max_locks = max_locks_in;
 563         num_resources = num_resources_in;
 564         num_extents = num_extents_in;
 565
 566         LDLM_DEBUG_NOLOCK("regression test started: threads: %d, max_locks: "
 567                           "%d, num_res: %d, num_ext: %d\n",
 568                           threads, max_locks_in, num_resources_in,
 569                           num_extents_in);
 570
 571         for (i = 0; i < threads; i++) {
 572                 rc = ldlm_start_thread(obddev, connh);
 573                 if (rc < 0)
 574                         GOTO(cleanup, rc);
 575         }
 576
 577  cleanup:
 578         if (rc < 0)
 579                 ldlm_regression_stop();
 580         RETURN(rc);
 581 }
 582
 583 int ldlm_regression_stop(void)
 584 {
 585         ENTRY;
 586
 587         spin_lock(&ctl_lock);
 588         if (!regression_running) {
 589                 CERROR("The ldlm regression isn't started.\n");
 590                 spin_unlock(&ctl_lock);
 591                 RETURN(-EINVAL);
 592         }
 593
 594         while (!list_empty(&ctl_threads)) {
 595                 struct ldlm_test_thread *thread;
 596                 thread = list_entry(ctl_threads.next, struct ldlm_test_thread,
 597                                     t_link);
 598
 599                 thread->t_flags |= SVC_STOPPING;
 600
 601                 spin_unlock(&ctl_lock);
 602                 wake_up(&thread->t_ctl_waitq);
 603                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 604                 spin_lock(&ctl_lock);
 605
 606                 list_del(&thread->t_link);
 607                 OBD_FREE(thread, sizeof(*thread));
 608         }
 609
 610         /* decrement all held locks */
 611         while (!list_empty(&lock_list)) {
 612                 struct ldlm_lock *lock;
 613                 struct ldlm_test_lock *lock_info =
 614                        list_entry(lock_list.next, struct ldlm_test_lock,
 615                                    l_link);
 616                 list_del(lock_list.next);
 617                 num_locks--;
 618
 619                 lock = ldlm_handle2lock(&lock_info->l_lockh);
 620                 ldlm_lock_decref(&lock_info->l_lockh, lock->l_granted_mode);
 621                 ldlm_lock_put(lock);
 622
 623                 OBD_FREE(lock_info, sizeof(*lock_info));
 624         }
 625
 626         regression_running = 0;
 627         spin_unlock(&ctl_lock);
 628
 629         RETURN(0);
 630 }
 631
 632 int ldlm_test(struct obd_device *obddev, struct lustre_handle *connh)
 633 {
 634         int rc;
 635         rc = ldlm_test_basics(obddev);
 636         if (rc)
 637                 RETURN(rc);
 638
 639         rc = ldlm_test_extents(obddev);
 640         if (rc)
 641                 RETURN(rc);
 642
 643         rc = ldlm_test_network(obddev, connh);
 644         RETURN(rc);
 645 }