lustre/ldlm/ldlm_resource.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, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46
  47 int ldlm_srv_namespace_nr = 0;
  48 int ldlm_cli_namespace_nr = 0;
  49
  50 struct mutex ldlm_srv_namespace_lock;
  51 struct list_head ldlm_srv_namespace_list;
  52
  53 struct mutex ldlm_cli_namespace_lock;
  54 /* Client Namespaces that have active resources in them.
  55  * Once all resources go away, ldlm_poold moves such namespaces to the
  56  * inactive list */
  57 struct list_head ldlm_cli_active_namespace_list;
  58 /* Client namespaces that don't have any locks in them */
  59 struct list_head ldlm_cli_inactive_namespace_list;
  60
  61 static struct proc_dir_entry *ldlm_type_proc_dir;
  62 static struct proc_dir_entry *ldlm_ns_proc_dir;
  63 struct proc_dir_entry *ldlm_svc_proc_dir;
  64
  65 /* during debug dump certain amount of granted locks for one resource to avoid
  66  * DDOS. */
  67 static unsigned int ldlm_dump_granted_max = 256;
  68
  69 #ifdef CONFIG_PROC_FS
  70 static ssize_t
  71 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
  72                           size_t count, loff_t *off)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
  79
  80 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
  81 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
  82
  83 #ifdef HAVE_SERVER_SUPPORT
  84
  85 static int seq_watermark_show(struct seq_file *m, void *data)
  86 {
  87         seq_printf(m, "%llu\n", *(__u64 *)m->private);
  88         return 0;
  89 }
  90
  91 static ssize_t seq_watermark_write(struct file *file,
  92                                    const char __user *buffer, size_t count,
  93                                    loff_t *off)
  94 {
  95         __s64 value;
  96         __u64 watermark;
  97         __u64 *data = ((struct seq_file *)file->private_data)->private;
  98         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
  99         int rc;
 100
 101         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 102         if (rc) {
 103                 CERROR("Failed to set %s, rc = %d.\n",
 104                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 105                        rc);
 106                 return rc;
 107         } else if (value != 0 && value < (1 << 20)) {
 108                 CERROR("%s should be greater than 1MB.\n",
 109                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 110                 return -EINVAL;
 111         }
 112         watermark = value >> 20;
 113
 114         if (wm_low) {
 115                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 116                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 117                                "lock_limit_mb.\n");
 118                         return -EINVAL;
 119                 }
 120
 121                 *data = watermark;
 122                 if (watermark != 0) {
 123                         watermark <<= 20;
 124                         do_div(watermark, sizeof(struct ldlm_lock));
 125                 }
 126                 ldlm_reclaim_threshold = watermark;
 127         } else {
 128                 if (ldlm_reclaim_threshold_mb != 0 &&
 129                     watermark < ldlm_reclaim_threshold_mb) {
 130                         CERROR("lock_limit_mb must be greater than "
 131                                "lock_reclaim_threshold_mb.\n");
 132                         return -EINVAL;
 133                 }
 134
 135                 *data = watermark;
 136                 if (watermark != 0) {
 137                         watermark <<= 20;
 138                         do_div(watermark, sizeof(struct ldlm_lock));
 139                 }
 140                 ldlm_lock_limit = watermark;
 141         }
 142
 143         return count;
 144 }
 145
 146 static int seq_watermark_open(struct inode *inode, struct file *file)
 147 {
 148         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 149 }
 150
 151 static const struct file_operations ldlm_watermark_fops = {
 152         .owner          = THIS_MODULE,
 153         .open           = seq_watermark_open,
 154         .read           = seq_read,
 155         .write          = seq_watermark_write,
 156         .llseek         = seq_lseek,
 157         .release        = lprocfs_single_release,
 158 };
 159
 160 static int seq_granted_show(struct seq_file *m, void *data)
 161 {
 162         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 163                    (struct percpu_counter *)m->private));
 164         return 0;
 165 }
 166
 167 static int seq_granted_open(struct inode *inode, struct file *file)
 168 {
 169         return single_open(file, seq_granted_show, PDE_DATA(inode));
 170 }
 171
 172 static const struct file_operations ldlm_granted_fops = {
 173         .owner  = THIS_MODULE,
 174         .open   = seq_granted_open,
 175         .read   = seq_read,
 176         .llseek = seq_lseek,
 177         .release = seq_release,
 178 };
 179
 180 #endif /* HAVE_SERVER_SUPPORT */
 181
 182 int ldlm_proc_setup(void)
 183 {
 184         int rc;
 185         struct lprocfs_vars list[] = {
 186                 { .name =       "dump_namespaces",
 187                   .fops =       &ldlm_dump_ns_fops,
 188                   .proc_mode =  0222 },
 189                 { .name =       "dump_granted_max",
 190                   .fops =       &ldlm_rw_uint_fops,
 191                   .data =       &ldlm_dump_granted_max },
 192 #ifdef HAVE_SERVER_SUPPORT
 193                 { .name =       "lock_reclaim_threshold_mb",
 194                   .fops =       &ldlm_watermark_fops,
 195                   .data =       &ldlm_reclaim_threshold_mb },
 196                 { .name =       "lock_limit_mb",
 197                   .fops =       &ldlm_watermark_fops,
 198                   .data =       &ldlm_lock_limit_mb },
 199                 { .name =       "lock_granted_count",
 200                   .fops =       &ldlm_granted_fops,
 201                   .data =       &ldlm_granted_total },
 202 #endif
 203                 { NULL }};
 204         ENTRY;
 205         LASSERT(ldlm_ns_proc_dir == NULL);
 206
 207         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 208                                               proc_lustre_root,
 209                                               NULL, NULL);
 210         if (IS_ERR(ldlm_type_proc_dir)) {
 211                 CERROR("LProcFS failed in ldlm-init\n");
 212                 rc = PTR_ERR(ldlm_type_proc_dir);
 213                 GOTO(err, rc);
 214         }
 215
 216         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 217                                             ldlm_type_proc_dir,
 218                                             NULL, NULL);
 219         if (IS_ERR(ldlm_ns_proc_dir)) {
 220                 CERROR("LProcFS failed in ldlm-init\n");
 221                 rc = PTR_ERR(ldlm_ns_proc_dir);
 222                 GOTO(err_type, rc);
 223         }
 224
 225         ldlm_svc_proc_dir = lprocfs_register("services",
 226                                              ldlm_type_proc_dir,
 227                                              NULL, NULL);
 228         if (IS_ERR(ldlm_svc_proc_dir)) {
 229                 CERROR("LProcFS failed in ldlm-init\n");
 230                 rc = PTR_ERR(ldlm_svc_proc_dir);
 231                 GOTO(err_ns, rc);
 232         }
 233
 234         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 235         if (rc != 0) {
 236                 CERROR("LProcFS failed in ldlm-init\n");
 237                 GOTO(err_svc, rc);
 238         }
 239
 240         RETURN(0);
 241
 242 err_svc:
 243         lprocfs_remove(&ldlm_svc_proc_dir);
 244 err_ns:
 245         lprocfs_remove(&ldlm_ns_proc_dir);
 246 err_type:
 247         lprocfs_remove(&ldlm_type_proc_dir);
 248 err:
 249         ldlm_svc_proc_dir = NULL;
 250         RETURN(rc);
 251 }
 252
 253 void ldlm_proc_cleanup(void)
 254 {
 255         if (ldlm_svc_proc_dir)
 256                 lprocfs_remove(&ldlm_svc_proc_dir);
 257
 258         if (ldlm_ns_proc_dir)
 259                 lprocfs_remove(&ldlm_ns_proc_dir);
 260
 261         if (ldlm_type_proc_dir)
 262                 lprocfs_remove(&ldlm_type_proc_dir);
 263 }
 264
 265 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 266                                    char *buf)
 267 {
 268         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 269                                                  ns_kobj);
 270         __u64                   res = 0;
 271         struct cfs_hash_bd              bd;
 272         int                     i;
 273
 274         /* result is not strictly consistant */
 275         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 276                 res += cfs_hash_bd_count_get(&bd);
 277         return sprintf(buf, "%lld\n", res);
 278 }
 279 LUSTRE_RO_ATTR(resource_count);
 280
 281 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 282                                char *buf)
 283 {
 284         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 285                                                  ns_kobj);
 286         __u64                   locks;
 287
 288         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 289                                         LPROCFS_FIELDS_FLAGS_SUM);
 290         return sprintf(buf, "%lld\n", locks);
 291 }
 292 LUSTRE_RO_ATTR(lock_count);
 293
 294 static ssize_t lock_unused_count_show(struct kobject *kobj,
 295                                       struct attribute *attr,
 296                                       char *buf)
 297 {
 298         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 299                                                  ns_kobj);
 300
 301         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 302 }
 303 LUSTRE_RO_ATTR(lock_unused_count);
 304
 305 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 306                              char *buf)
 307 {
 308         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 309                                                  ns_kobj);
 310         __u32 *nr = &ns->ns_max_unused;
 311
 312         if (ns_connect_lru_resize(ns))
 313                 nr = &ns->ns_nr_unused;
 314         return sprintf(buf, "%u\n", *nr);
 315 }
 316
 317 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 318                               const char *buffer, size_t count)
 319 {
 320         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 321                                                  ns_kobj);
 322         unsigned long tmp;
 323         int lru_resize;
 324         int err;
 325
 326         if (strncmp(buffer, "clear", 5) == 0) {
 327                 CDEBUG(D_DLMTRACE,
 328                        "dropping all unused locks from namespace %s\n",
 329                        ldlm_ns_name(ns));
 330                 if (ns_connect_lru_resize(ns)) {
 331                         int canceled, unused  = ns->ns_nr_unused;
 332
 333                         /* Try to cancel all @ns_nr_unused locks. */
 334                         canceled = ldlm_cancel_lru(ns, unused, 0,
 335                                                    LDLM_LRU_FLAG_PASSED);
 336                         if (canceled < unused) {
 337                                 CDEBUG(D_DLMTRACE,
 338                                        "not all requested locks are canceled, requested: %d, canceled: %d\n",
 339                                        unused,
 340                                        canceled);
 341                                 return -EINVAL;
 342                         }
 343                 } else {
 344                         tmp = ns->ns_max_unused;
 345                         ns->ns_max_unused = 0;
 346                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED);
 347                         ns->ns_max_unused = tmp;
 348                 }
 349                 return count;
 350         }
 351
 352         err = kstrtoul(buffer, 10, &tmp);
 353         if (err != 0) {
 354                 CERROR("lru_size: invalid value written\n");
 355                 return -EINVAL;
 356         }
 357         lru_resize = (tmp == 0);
 358
 359         if (ns_connect_lru_resize(ns)) {
 360                 if (!lru_resize)
 361                         ns->ns_max_unused = (unsigned int)tmp;
 362
 363                 if (tmp > ns->ns_nr_unused)
 364                         tmp = ns->ns_nr_unused;
 365                 tmp = ns->ns_nr_unused - tmp;
 366
 367                 CDEBUG(D_DLMTRACE,
 368                        "changing namespace %s unused locks from %u to %u\n",
 369                        ldlm_ns_name(ns), ns->ns_nr_unused,
 370                        (unsigned int)tmp);
 371                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 372
 373                 if (!lru_resize) {
 374                         CDEBUG(D_DLMTRACE,
 375                                "disable lru_resize for namespace %s\n",
 376                                ldlm_ns_name(ns));
 377                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 378                 }
 379         } else {
 380                 CDEBUG(D_DLMTRACE,
 381                        "changing namespace %s max_unused from %u to %u\n",
 382                        ldlm_ns_name(ns), ns->ns_max_unused,
 383                        (unsigned int)tmp);
 384                 ns->ns_max_unused = (unsigned int)tmp;
 385                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 386
 387                 /* Make sure that LRU resize was originally supported before
 388                  * turning it on here.
 389                  */
 390                 if (lru_resize &&
 391                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 392                         CDEBUG(D_DLMTRACE,
 393                                "enable lru_resize for namespace %s\n",
 394                                ldlm_ns_name(ns));
 395                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 396                 }
 397         }
 398
 399         return count;
 400 }
 401 LUSTRE_RW_ATTR(lru_size);
 402
 403 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 404                                 char *buf)
 405 {
 406         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 407                                                  ns_kobj);
 408
 409         return sprintf(buf, "%u\n", ns->ns_max_age);
 410 }
 411
 412 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 413                                  const char *buffer, size_t count)
 414 {
 415         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 416                                                  ns_kobj);
 417         unsigned long tmp;
 418         int err;
 419
 420         err = kstrtoul(buffer, 10, &tmp);
 421         if (err != 0)
 422                 return -EINVAL;
 423
 424         ns->ns_max_age = tmp;
 425
 426         return count;
 427 }
 428 LUSTRE_RW_ATTR(lru_max_age);
 429
 430 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 431                                       struct attribute *attr,
 432                                       char *buf)
 433 {
 434         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 435                                                  ns_kobj);
 436
 437         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 438 }
 439
 440 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 441                                        struct attribute *attr,
 442                                        const char *buffer,
 443                                        size_t count)
 444 {
 445         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 446                                                  ns_kobj);
 447         unsigned long supp = -1;
 448         int rc;
 449
 450         rc = kstrtoul(buffer, 10, &supp);
 451         if (rc < 0)
 452                 return rc;
 453
 454         if (supp == 0)
 455                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 456         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 457                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 458         return count;
 459 }
 460 LUSTRE_RW_ATTR(early_lock_cancel);
 461
 462 #ifdef HAVE_SERVER_SUPPORT
 463 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 464                                     struct attribute *attr, char *buf)
 465 {
 466         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 467                                                  ns_kobj);
 468
 469         return sprintf(buf, "%u\n", ns->ns_ctime_age_limit);
 470 }
 471
 472 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 473                                      struct attribute *attr,
 474                                      const char *buffer, size_t count)
 475 {
 476         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 477                                                  ns_kobj);
 478         unsigned long tmp;
 479         int err;
 480
 481         err = kstrtoul(buffer, 10, &tmp);
 482         if (err != 0)
 483                 return -EINVAL;
 484
 485         ns->ns_ctime_age_limit = tmp;
 486
 487         return count;
 488 }
 489 LUSTRE_RW_ATTR(ctime_age_limit);
 490
 491 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 492                                   char *buf)
 493 {
 494         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 495                                                  ns_kobj);
 496
 497         return sprintf(buf, "%d\n", ns->ns_timeouts);
 498 }
 499 LUSTRE_RO_ATTR(lock_timeouts);
 500
 501 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 502                                      struct attribute *attr, char *buf)
 503 {
 504         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 505                                                  ns_kobj);
 506
 507         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 508 }
 509
 510 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 511                                       struct attribute *attr,
 512                                       const char *buffer, size_t count)
 513 {
 514         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 515                                                  ns_kobj);
 516         unsigned long tmp;
 517         int err;
 518
 519         err = kstrtoul(buffer, 10, &tmp);
 520         if (err != 0)
 521                 return -EINVAL;
 522
 523         ns->ns_max_nolock_size = tmp;
 524
 525         return count;
 526 }
 527 LUSTRE_RW_ATTR(max_nolock_bytes);
 528
 529 static ssize_t contention_seconds_show(struct kobject *kobj,
 530                                        struct attribute *attr, char *buf)
 531 {
 532         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 533                                                  ns_kobj);
 534
 535         return sprintf(buf, "%u\n", ns->ns_contention_time);
 536 }
 537
 538 static ssize_t contention_seconds_store(struct kobject *kobj,
 539                                         struct attribute *attr,
 540                                         const char *buffer, size_t count)
 541 {
 542         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 543                                                  ns_kobj);
 544         unsigned long tmp;
 545         int err;
 546
 547         err = kstrtoul(buffer, 10, &tmp);
 548         if (err != 0)
 549                 return -EINVAL;
 550
 551         ns->ns_contention_time = tmp;
 552
 553         return count;
 554 }
 555 LUSTRE_RW_ATTR(contention_seconds);
 556
 557 static ssize_t contended_locks_show(struct kobject *kobj,
 558                                     struct attribute *attr, char *buf)
 559 {
 560         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 561                                                  ns_kobj);
 562
 563         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 564 }
 565
 566 static ssize_t contended_locks_store(struct kobject *kobj,
 567                                      struct attribute *attr,
 568                                      const char *buffer, size_t count)
 569 {
 570         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 571                                                  ns_kobj);
 572         unsigned long tmp;
 573         int err;
 574
 575         err = kstrtoul(buffer, 10, &tmp);
 576         if (err != 0)
 577                 return -EINVAL;
 578
 579         ns->ns_contended_locks = tmp;
 580
 581         return count;
 582 }
 583 LUSTRE_RW_ATTR(contended_locks);
 584
 585 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 586                                      struct attribute *attr, char *buf)
 587 {
 588         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 589                                                  ns_kobj);
 590
 591         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 592 }
 593
 594 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 595                                       struct attribute *attr,
 596                                       const char *buffer, size_t count)
 597 {
 598         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 599                                                  ns_kobj);
 600         unsigned long tmp;
 601         int err;
 602
 603         err = kstrtoul(buffer, 10, &tmp);
 604         if (err != 0)
 605                 return -EINVAL;
 606
 607         ns->ns_max_parallel_ast = tmp;
 608
 609         return count;
 610 }
 611 LUSTRE_RW_ATTR(max_parallel_ast);
 612
 613 #endif /* HAVE_SERVER_SUPPORT */
 614
 615 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 616 static struct attribute *ldlm_ns_attrs[] = {
 617         &lustre_attr_resource_count.attr,
 618         &lustre_attr_lock_count.attr,
 619         &lustre_attr_lock_unused_count.attr,
 620         &lustre_attr_lru_size.attr,
 621         &lustre_attr_lru_max_age.attr,
 622         &lustre_attr_early_lock_cancel.attr,
 623 #ifdef HAVE_SERVER_SUPPORT
 624         &lustre_attr_ctime_age_limit.attr,
 625         &lustre_attr_lock_timeouts.attr,
 626         &lustre_attr_max_nolock_bytes.attr,
 627         &lustre_attr_contention_seconds.attr,
 628         &lustre_attr_contended_locks.attr,
 629         &lustre_attr_max_parallel_ast.attr,
 630 #endif
 631         NULL,
 632 };
 633
 634 static void ldlm_ns_release(struct kobject *kobj)
 635 {
 636         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 637                                                  ns_kobj);
 638         complete(&ns->ns_kobj_unregister);
 639 }
 640
 641 static struct kobj_type ldlm_ns_ktype = {
 642         .default_attrs  = ldlm_ns_attrs,
 643         .sysfs_ops      = &lustre_sysfs_ops,
 644         .release        = ldlm_ns_release,
 645 };
 646
 647 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 648 {
 649         if (ns->ns_proc_dir_entry == NULL)
 650                 CERROR("dlm namespace %s has no procfs dir?\n",
 651                        ldlm_ns_name(ns));
 652         else
 653                 lprocfs_remove(&ns->ns_proc_dir_entry);
 654
 655         if (ns->ns_stats != NULL)
 656                 lprocfs_free_stats(&ns->ns_stats);
 657 }
 658
 659 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 660 {
 661         kobject_put(&ns->ns_kobj);
 662         wait_for_completion(&ns->ns_kobj_unregister);
 663 }
 664
 665 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 666 {
 667         int err;
 668
 669         ns->ns_kobj.kset = ldlm_ns_kset;
 670         init_completion(&ns->ns_kobj_unregister);
 671         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 672                                    "%s", ldlm_ns_name(ns));
 673
 674         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 675         if (!ns->ns_stats) {
 676                 kobject_put(&ns->ns_kobj);
 677                 return -ENOMEM;
 678         }
 679
 680         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 681                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 682
 683         return err;
 684 }
 685
 686 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 687 {
 688         struct proc_dir_entry *ns_pde;
 689
 690         LASSERT(ns != NULL);
 691         LASSERT(ns->ns_rs_hash != NULL);
 692
 693         if (ns->ns_proc_dir_entry != NULL) {
 694                 ns_pde = ns->ns_proc_dir_entry;
 695         } else {
 696                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 697                 if (ns_pde == NULL)
 698                         return -ENOMEM;
 699                 ns->ns_proc_dir_entry = ns_pde;
 700         }
 701
 702         return 0;
 703 }
 704 #undef MAX_STRING_SIZE
 705 #else /* CONFIG_PROC_FS */
 706
 707 #define ldlm_namespace_proc_unregister(ns)      ({;})
 708 #define ldlm_namespace_proc_register(ns)        ({0;})
 709
 710 #endif /* CONFIG_PROC_FS */
 711
 712 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 713                                   const void *key, unsigned mask)
 714 {
 715         const struct ldlm_res_id     *id  = key;
 716         unsigned                val = 0;
 717         unsigned                i;
 718
 719         for (i = 0; i < RES_NAME_SIZE; i++)
 720                 val += id->name[i];
 721         return val & mask;
 722 }
 723
 724 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 725                                       const void *key, unsigned mask)
 726 {
 727         const struct ldlm_res_id *id = key;
 728         struct lu_fid       fid;
 729         __u32               hash;
 730         __u32               val;
 731
 732         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 733         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 734         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 735
 736         hash = fid_flatten32(&fid);
 737         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 738         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 739                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 740                 hash += (val >> 5) + (val << 11);
 741         } else {
 742                 val = fid_oid(&fid);
 743         }
 744         hash = hash_long(hash, hs->hs_bkt_bits);
 745         /* give me another random factor */
 746         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 747
 748         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 749         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 750
 751         return hash & mask;
 752 }
 753
 754 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 755 {
 756         struct ldlm_resource   *res;
 757
 758         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 759         return &res->lr_name;
 760 }
 761
 762 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 763 {
 764         struct ldlm_resource   *res;
 765
 766         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 767         return ldlm_res_eq((const struct ldlm_res_id *)key,
 768                            (const struct ldlm_res_id *)&res->lr_name);
 769 }
 770
 771 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 772 {
 773         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 774 }
 775
 776 static void
 777 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 778 {
 779         struct ldlm_resource *res;
 780
 781         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 782         ldlm_resource_getref(res);
 783 }
 784
 785 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 786 {
 787         struct ldlm_resource *res;
 788
 789         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 790         ldlm_resource_putref(res);
 791 }
 792
 793 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 794         .hs_hash        = ldlm_res_hop_hash,
 795         .hs_key         = ldlm_res_hop_key,
 796         .hs_keycmp      = ldlm_res_hop_keycmp,
 797         .hs_keycpy      = NULL,
 798         .hs_object      = ldlm_res_hop_object,
 799         .hs_get         = ldlm_res_hop_get_locked,
 800         .hs_put         = ldlm_res_hop_put
 801 };
 802
 803 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 804         .hs_hash        = ldlm_res_hop_fid_hash,
 805         .hs_key         = ldlm_res_hop_key,
 806         .hs_keycmp      = ldlm_res_hop_keycmp,
 807         .hs_keycpy      = NULL,
 808         .hs_object      = ldlm_res_hop_object,
 809         .hs_get         = ldlm_res_hop_get_locked,
 810         .hs_put         = ldlm_res_hop_put
 811 };
 812
 813 typedef struct ldlm_ns_hash_def {
 814         enum ldlm_ns_type       nsd_type;
 815         /** hash bucket bits */
 816         unsigned                nsd_bkt_bits;
 817         /** hash bits */
 818         unsigned                nsd_all_bits;
 819         /** hash operations */
 820         struct cfs_hash_ops *nsd_hops;
 821 } ldlm_ns_hash_def_t;
 822
 823 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 824 {
 825         {
 826                 .nsd_type       = LDLM_NS_TYPE_MDC,
 827                 .nsd_bkt_bits   = 11,
 828                 .nsd_all_bits   = 16,
 829                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 830         },
 831         {
 832                 .nsd_type       = LDLM_NS_TYPE_MDT,
 833                 .nsd_bkt_bits   = 14,
 834                 .nsd_all_bits   = 21,
 835                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 836         },
 837         {
 838                 .nsd_type       = LDLM_NS_TYPE_OSC,
 839                 .nsd_bkt_bits   = 8,
 840                 .nsd_all_bits   = 12,
 841                 .nsd_hops       = &ldlm_ns_hash_ops,
 842         },
 843         {
 844                 .nsd_type       = LDLM_NS_TYPE_OST,
 845                 .nsd_bkt_bits   = 11,
 846                 .nsd_all_bits   = 17,
 847                 .nsd_hops       = &ldlm_ns_hash_ops,
 848         },
 849         {
 850                 .nsd_type       = LDLM_NS_TYPE_MGC,
 851                 .nsd_bkt_bits   = 4,
 852                 .nsd_all_bits   = 4,
 853                 .nsd_hops       = &ldlm_ns_hash_ops,
 854         },
 855         {
 856                 .nsd_type       = LDLM_NS_TYPE_MGT,
 857                 .nsd_bkt_bits   = 4,
 858                 .nsd_all_bits   = 4,
 859                 .nsd_hops       = &ldlm_ns_hash_ops,
 860         },
 861         {
 862                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 863         },
 864 };
 865
 866 /**
 867  * Create and initialize new empty namespace.
 868  */
 869 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 870                                           enum ldlm_side client,
 871                                           enum ldlm_appetite apt,
 872                                           enum ldlm_ns_type ns_type)
 873 {
 874         struct ldlm_namespace *ns = NULL;
 875         struct ldlm_ns_bucket *nsb;
 876         struct ldlm_ns_hash_def *nsd;
 877         struct cfs_hash_bd bd;
 878         int idx;
 879         int rc;
 880         ENTRY;
 881
 882         LASSERT(obd != NULL);
 883
 884         rc = ldlm_get_ref();
 885         if (rc) {
 886                 CERROR("ldlm_get_ref failed: %d\n", rc);
 887                 RETURN(NULL);
 888         }
 889
 890         for (idx = 0;;idx++) {
 891                 nsd = &ldlm_ns_hash_defs[idx];
 892                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 893                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 894                         GOTO(out_ref, NULL);
 895                 }
 896
 897                 if (nsd->nsd_type == ns_type)
 898                         break;
 899         }
 900
 901         OBD_ALLOC_PTR(ns);
 902         if (!ns)
 903                 GOTO(out_ref, NULL);
 904
 905         ns->ns_rs_hash = cfs_hash_create(name,
 906                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 907                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 908                                          CFS_HASH_MIN_THETA,
 909                                          CFS_HASH_MAX_THETA,
 910                                          nsd->nsd_hops,
 911                                          CFS_HASH_DEPTH |
 912                                          CFS_HASH_BIGNAME |
 913                                          CFS_HASH_SPIN_BKTLOCK |
 914                                          CFS_HASH_NO_ITEMREF);
 915         if (ns->ns_rs_hash == NULL)
 916                 GOTO(out_ns, NULL);
 917
 918         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 919                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 920                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 921                 nsb->nsb_namespace = ns;
 922                 nsb->nsb_reclaim_start = 0;
 923         }
 924
 925         ns->ns_obd      = obd;
 926         ns->ns_appetite = apt;
 927         ns->ns_client   = client;
 928
 929         INIT_LIST_HEAD(&ns->ns_list_chain);
 930         INIT_LIST_HEAD(&ns->ns_unused_list);
 931         spin_lock_init(&ns->ns_lock);
 932         atomic_set(&ns->ns_bref, 0);
 933         init_waitqueue_head(&ns->ns_waitq);
 934
 935         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 936         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 937         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 938
 939         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 940         ns->ns_nr_unused          = 0;
 941         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 942         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 943         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 944         ns->ns_timeouts           = 0;
 945         ns->ns_orig_connect_flags = 0;
 946         ns->ns_connect_flags      = 0;
 947         ns->ns_stopping           = 0;
 948         ns->ns_reclaim_start      = 0;
 949
 950         rc = ldlm_namespace_sysfs_register(ns);
 951         if (rc) {
 952                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 953                 GOTO(out_hash, rc);
 954         }
 955
 956         rc = ldlm_namespace_proc_register(ns);
 957         if (rc) {
 958                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 959                 GOTO(out_sysfs, rc);
 960         }
 961
 962         idx = ldlm_namespace_nr_read(client);
 963         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 964         if (rc) {
 965                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 966                 GOTO(out_proc, rc);
 967         }
 968
 969         ldlm_namespace_register(ns, client);
 970         RETURN(ns);
 971 out_proc:
 972         ldlm_namespace_proc_unregister(ns);
 973 out_sysfs:
 974         ldlm_namespace_sysfs_unregister(ns);
 975         ldlm_namespace_cleanup(ns, 0);
 976 out_hash:
 977         cfs_hash_putref(ns->ns_rs_hash);
 978 out_ns:
 979         OBD_FREE_PTR(ns);
 980 out_ref:
 981         ldlm_put_ref();
 982         RETURN(NULL);
 983 }
 984 EXPORT_SYMBOL(ldlm_namespace_new);
 985
 986 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 987
 988 /**
 989  * Cancel and destroy all locks on a resource.
 990  *
 991  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 992  * clean up.  This is currently only used for recovery, and we make
 993  * certain assumptions as a result--notably, that we shouldn't cancel
 994  * locks with refs.
 995  */
 996 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 997                              __u64 flags)
 998 {
 999         struct list_head *tmp;
1000         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1001         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1002
1003         do {
1004                 struct ldlm_lock *lock = NULL;
1005
1006                 /* First, we look for non-cleaned-yet lock
1007                  * all cleaned locks are marked by CLEANED flag. */
1008                 lock_res(res);
1009                 list_for_each(tmp, q) {
1010                         lock = list_entry(tmp, struct ldlm_lock,
1011                                           l_res_link);
1012                         if (ldlm_is_cleaned(lock)) {
1013                                 lock = NULL;
1014                                 continue;
1015                         }
1016                         LDLM_LOCK_GET(lock);
1017                         ldlm_set_cleaned(lock);
1018                         break;
1019                 }
1020
1021                 if (lock == NULL) {
1022                         unlock_res(res);
1023                         break;
1024                 }
1025
1026                 /* Set CBPENDING so nothing in the cancellation path
1027                  * can match this lock. */
1028                 ldlm_set_cbpending(lock);
1029                 ldlm_set_failed(lock);
1030                 lock->l_flags |= flags;
1031
1032                 /* ... without sending a CANCEL message for local_only. */
1033                 if (local_only)
1034                         ldlm_set_local_only(lock);
1035
1036                 if (local_only && (lock->l_readers || lock->l_writers)) {
1037                         /* This is a little bit gross, but much better than the
1038                          * alternative: pretend that we got a blocking AST from
1039                          * the server, so that when the lock is decref'd, it
1040                          * will go away ... */
1041                         unlock_res(res);
1042                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1043                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1044                                 set_current_state(TASK_UNINTERRUPTIBLE);
1045                                 schedule_timeout(cfs_time_seconds(4));
1046                                 set_current_state(TASK_RUNNING);
1047                         }
1048                         if (lock->l_completion_ast)
1049                                 lock->l_completion_ast(lock,
1050                                                        LDLM_FL_FAILED, NULL);
1051                         LDLM_LOCK_RELEASE(lock);
1052                         continue;
1053                 }
1054
1055                 if (client) {
1056                         struct lustre_handle lockh;
1057
1058                         unlock_res(res);
1059                         ldlm_lock2handle(lock, &lockh);
1060                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1061                         if (rc)
1062                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1063                 } else {
1064                         unlock_res(res);
1065                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
1066                                    "client node");
1067                         ldlm_lock_cancel(lock);
1068                 }
1069                 LDLM_LOCK_RELEASE(lock);
1070         } while (1);
1071 }
1072
1073 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1074                                struct hlist_node *hnode, void *arg)
1075 {
1076         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1077         __u64 flags = *(__u64 *)arg;
1078
1079         cleanup_resource(res, &res->lr_granted, flags);
1080         cleanup_resource(res, &res->lr_converting, flags);
1081         cleanup_resource(res, &res->lr_waiting, flags);
1082
1083         return 0;
1084 }
1085
1086 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1087                                   struct hlist_node *hnode, void *arg)
1088 {
1089         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1090
1091         lock_res(res);
1092         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1093                "(%d) after lock cleanup; forcing cleanup.\n",
1094                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1095                atomic_read(&res->lr_refcount) - 1);
1096
1097         ldlm_resource_dump(D_ERROR, res);
1098         unlock_res(res);
1099         return 0;
1100 }
1101
1102 /**
1103  * Cancel and destroy all locks in the namespace.
1104  *
1105  * Typically used during evictions when server notified client that it was
1106  * evicted and all of its state needs to be destroyed.
1107  * Also used during shutdown.
1108  */
1109 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1110 {
1111         if (ns == NULL) {
1112                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1113                 return ELDLM_OK;
1114         }
1115
1116         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1117                                  &flags, 0);
1118         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1119                                  NULL, 0);
1120         return ELDLM_OK;
1121 }
1122 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1123
1124 /**
1125  * Attempts to free namespace.
1126  *
1127  * Only used when namespace goes away, like during an unmount.
1128  */
1129 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1130 {
1131         ENTRY;
1132
1133         /* At shutdown time, don't call the cancellation callback */
1134         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1135
1136         if (atomic_read(&ns->ns_bref) > 0) {
1137                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1138                 int rc;
1139                 CDEBUG(D_DLMTRACE,
1140                        "dlm namespace %s free waiting on refcount %d\n",
1141                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1142 force_wait:
1143                 if (force)
1144                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
1145                                           MSEC_PER_SEC) / 4, NULL, NULL);
1146
1147                 rc = l_wait_event(ns->ns_waitq,
1148                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1149
1150                 /* Forced cleanups should be able to reclaim all references,
1151                  * so it's safe to wait forever... we can't leak locks... */
1152                 if (force && rc == -ETIMEDOUT) {
1153                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1154                                        "namespace with %d resources in use, "
1155                                        "(rc=%d)\n", ldlm_ns_name(ns),
1156                                        atomic_read(&ns->ns_bref), rc);
1157                         GOTO(force_wait, rc);
1158                 }
1159
1160                 if (atomic_read(&ns->ns_bref)) {
1161                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1162                                        "with %d resources in use, (rc=%d)\n",
1163                                        ldlm_ns_name(ns),
1164                                        atomic_read(&ns->ns_bref), rc);
1165                         RETURN(ELDLM_NAMESPACE_EXISTS);
1166                 }
1167                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1168                        ldlm_ns_name(ns));
1169         }
1170
1171         RETURN(ELDLM_OK);
1172 }
1173
1174 /**
1175  * Performs various cleanups for passed \a ns to make it drop refc and be
1176  * ready for freeing. Waits for refc == 0.
1177  *
1178  * The following is done:
1179  * (0) Unregister \a ns from its list to make inaccessible for potential
1180  * users like pools thread and others;
1181  * (1) Clear all locks in \a ns.
1182  */
1183 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1184                                struct obd_import *imp,
1185                                int force)
1186 {
1187         int rc;
1188         ENTRY;
1189         if (!ns) {
1190                 EXIT;
1191                 return;
1192         }
1193
1194         spin_lock(&ns->ns_lock);
1195         ns->ns_stopping = 1;
1196         spin_unlock(&ns->ns_lock);
1197
1198         /*
1199          * Can fail with -EINTR when force == 0 in which case try harder.
1200          */
1201         rc = __ldlm_namespace_free(ns, force);
1202         if (rc != ELDLM_OK) {
1203                 if (imp) {
1204                         ptlrpc_disconnect_import(imp, 0);
1205                         ptlrpc_invalidate_import(imp);
1206                 }
1207
1208                 /*
1209                  * With all requests dropped and the import inactive
1210                  * we are gaurenteed all reference will be dropped.
1211                  */
1212                 rc = __ldlm_namespace_free(ns, 1);
1213                 LASSERT(rc == 0);
1214         }
1215         EXIT;
1216 }
1217 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1218
1219 /**
1220  * Performs freeing memory structures related to \a ns. This is only done
1221  * when ldlm_namespce_free_prior() successfully removed all resources
1222  * referencing \a ns and its refc == 0.
1223  */
1224 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1225 {
1226         ENTRY;
1227         if (!ns) {
1228                 EXIT;
1229                 return;
1230         }
1231
1232         /* Make sure that nobody can find this ns in its list. */
1233         ldlm_namespace_unregister(ns, ns->ns_client);
1234         /* Fini pool _before_ parent proc dir is removed. This is important as
1235          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1236          * Removing it after @dir may cause oops. */
1237         ldlm_pool_fini(&ns->ns_pool);
1238
1239         ldlm_namespace_proc_unregister(ns);
1240         ldlm_namespace_sysfs_unregister(ns);
1241         cfs_hash_putref(ns->ns_rs_hash);
1242         /* Namespace \a ns should be not on list at this time, otherwise
1243          * this will cause issues related to using freed \a ns in poold
1244          * thread. */
1245         LASSERT(list_empty(&ns->ns_list_chain));
1246         OBD_FREE_PTR(ns);
1247         ldlm_put_ref();
1248         EXIT;
1249 }
1250 EXPORT_SYMBOL(ldlm_namespace_free_post);
1251
1252 /**
1253  * Cleanup the resource, and free namespace.
1254  * bug 12864:
1255  * Deadlock issue:
1256  * proc1: destroy import
1257  *        class_disconnect_export(grab cl_sem) ->
1258  *              -> ldlm_namespace_free ->
1259  *              -> lprocfs_remove(grab _lprocfs_lock).
1260  * proc2: read proc info
1261  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1262  *              -> osc_rd_active, etc(grab cl_sem).
1263  *
1264  * So that I have to split the ldlm_namespace_free into two parts - the first
1265  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1266  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1267  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1268  * held.
1269  */
1270 void ldlm_namespace_free(struct ldlm_namespace *ns,
1271                          struct obd_import *imp,
1272                          int force)
1273 {
1274         ldlm_namespace_free_prior(ns, imp, force);
1275         ldlm_namespace_free_post(ns);
1276 }
1277 EXPORT_SYMBOL(ldlm_namespace_free);
1278
1279 void ldlm_namespace_get(struct ldlm_namespace *ns)
1280 {
1281         atomic_inc(&ns->ns_bref);
1282 }
1283
1284 /* This is only for callers that care about refcount */
1285 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1286 {
1287         return atomic_inc_return(&ns->ns_bref);
1288 }
1289
1290 void ldlm_namespace_put(struct ldlm_namespace *ns)
1291 {
1292         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1293                 wake_up(&ns->ns_waitq);
1294                 spin_unlock(&ns->ns_lock);
1295         }
1296 }
1297
1298 /** Register \a ns in the list of namespaces */
1299 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1300 {
1301         mutex_lock(ldlm_namespace_lock(client));
1302         LASSERT(list_empty(&ns->ns_list_chain));
1303         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1304         ldlm_namespace_nr_inc(client);
1305         mutex_unlock(ldlm_namespace_lock(client));
1306 }
1307
1308 /** Unregister \a ns from the list of namespaces. */
1309 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1310 {
1311         mutex_lock(ldlm_namespace_lock(client));
1312         LASSERT(!list_empty(&ns->ns_list_chain));
1313         /* Some asserts and possibly other parts of the code are still
1314          * using list_empty(&ns->ns_list_chain). This is why it is
1315          * important to use list_del_init() here. */
1316         list_del_init(&ns->ns_list_chain);
1317         ldlm_namespace_nr_dec(client);
1318         mutex_unlock(ldlm_namespace_lock(client));
1319 }
1320
1321 /** Should be called with ldlm_namespace_lock(client) taken. */
1322 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1323                                           enum ldlm_side client)
1324 {
1325         LASSERT(!list_empty(&ns->ns_list_chain));
1326         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1327         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1328 }
1329
1330 /** Should be called with ldlm_namespace_lock(client) taken. */
1331 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1332                                             enum ldlm_side client)
1333 {
1334         LASSERT(!list_empty(&ns->ns_list_chain));
1335         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1336         list_move_tail(&ns->ns_list_chain,
1337                        ldlm_namespace_inactive_list(client));
1338 }
1339
1340 /** Should be called with ldlm_namespace_lock(client) taken. */
1341 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1342 {
1343         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1344         LASSERT(!list_empty(ldlm_namespace_list(client)));
1345         return container_of(ldlm_namespace_list(client)->next,
1346                             struct ldlm_namespace, ns_list_chain);
1347 }
1348
1349 /** Create and initialize new resource. */
1350 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1351 {
1352         struct ldlm_resource *res;
1353         int idx;
1354
1355         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1356         if (res == NULL)
1357                 return NULL;
1358
1359         if (ldlm_type == LDLM_EXTENT) {
1360                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1361                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1362                 if (res->lr_itree == NULL) {
1363                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1364                         return NULL;
1365                 }
1366                 /* Initialize interval trees for each lock mode. */
1367                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1368                         res->lr_itree[idx].lit_size = 0;
1369                         res->lr_itree[idx].lit_mode = 1 << idx;
1370                         res->lr_itree[idx].lit_root = NULL;
1371                 }
1372         }
1373
1374         INIT_LIST_HEAD(&res->lr_granted);
1375         INIT_LIST_HEAD(&res->lr_converting);
1376         INIT_LIST_HEAD(&res->lr_waiting);
1377
1378         atomic_set(&res->lr_refcount, 1);
1379         spin_lock_init(&res->lr_lock);
1380         lu_ref_init(&res->lr_reference);
1381
1382         /* Since LVB init can be delayed now, there is no longer need to
1383          * immediatelly acquire mutex here. */
1384         mutex_init(&res->lr_lvb_mutex);
1385         res->lr_lvb_initialized = false;
1386
1387         return res;
1388 }
1389
1390 /**
1391  * Return a reference to resource with given name, creating it if necessary.
1392  * Args: namespace with ns_lock unlocked
1393  * Locks: takes and releases NS hash-lock and res->lr_lock
1394  * Returns: referenced, unlocked ldlm_resource or NULL
1395  */
1396 struct ldlm_resource *
1397 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1398                   const struct ldlm_res_id *name, enum ldlm_type type,
1399                   int create)
1400 {
1401         struct hlist_node       *hnode;
1402         struct ldlm_resource    *res = NULL;
1403         struct cfs_hash_bd              bd;
1404         __u64                   version;
1405         int                     ns_refcount = 0;
1406
1407         LASSERT(ns != NULL);
1408         LASSERT(parent == NULL);
1409         LASSERT(ns->ns_rs_hash != NULL);
1410         LASSERT(name->name[0] != 0);
1411
1412         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1413         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1414         if (hnode != NULL) {
1415                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1416                 GOTO(found, res);
1417         }
1418
1419         version = cfs_hash_bd_version_get(&bd);
1420         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1421
1422         if (create == 0)
1423                 return ERR_PTR(-ENOENT);
1424
1425         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1426                  "type: %d\n", type);
1427         res = ldlm_resource_new(type);
1428         if (res == NULL)
1429                 return ERR_PTR(-ENOMEM);
1430
1431         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1432         res->lr_name = *name;
1433         res->lr_type = type;
1434
1435         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1436         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1437                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1438
1439         if (hnode != NULL) {
1440                 /* Someone won the race and already added the resource. */
1441                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1442                 /* Clean lu_ref for failed resource. */
1443                 lu_ref_fini(&res->lr_reference);
1444                 if (res->lr_itree != NULL)
1445                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1446                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1447                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1448 found:
1449                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1450                 return res;
1451         }
1452         /* We won! Let's add the resource. */
1453         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1454         if (cfs_hash_bd_count_get(&bd) == 1)
1455                 ns_refcount = ldlm_namespace_get_return(ns);
1456
1457         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1458
1459         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1460
1461         /* Let's see if we happened to be the very first resource in this
1462          * namespace. If so, and this is a client namespace, we need to move
1463          * the namespace into the active namespaces list to be patrolled by
1464          * the ldlm_poold. */
1465         if (ns_is_client(ns) && ns_refcount == 1) {
1466                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1467                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1468                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1469         }
1470
1471         return res;
1472 }
1473 EXPORT_SYMBOL(ldlm_resource_get);
1474
1475 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1476 {
1477         LASSERT(res != NULL);
1478         LASSERT(res != LP_POISON);
1479         atomic_inc(&res->lr_refcount);
1480         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1481                atomic_read(&res->lr_refcount));
1482         return res;
1483 }
1484
1485 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1486                                          struct ldlm_resource *res)
1487 {
1488         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1489
1490         if (!list_empty(&res->lr_granted)) {
1491                 ldlm_resource_dump(D_ERROR, res);
1492                 LBUG();
1493         }
1494
1495         if (!list_empty(&res->lr_converting)) {
1496                 ldlm_resource_dump(D_ERROR, res);
1497                 LBUG();
1498         }
1499
1500         if (!list_empty(&res->lr_waiting)) {
1501                 ldlm_resource_dump(D_ERROR, res);
1502                 LBUG();
1503         }
1504
1505         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1506                                bd, &res->lr_hash);
1507         lu_ref_fini(&res->lr_reference);
1508         if (cfs_hash_bd_count_get(bd) == 0)
1509                 ldlm_namespace_put(nsb->nsb_namespace);
1510 }
1511
1512 /* Returns 1 if the resource was freed, 0 if it remains. */
1513 int ldlm_resource_putref(struct ldlm_resource *res)
1514 {
1515         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1516         struct cfs_hash_bd   bd;
1517
1518         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1519         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1520                res, atomic_read(&res->lr_refcount) - 1);
1521
1522         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1523         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1524                 __ldlm_resource_putref_final(&bd, res);
1525                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1526                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1527                         ns->ns_lvbo->lvbo_free(res);
1528                 if (res->lr_itree != NULL)
1529                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1530                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1531                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1532                 return 1;
1533         }
1534         return 0;
1535 }
1536 EXPORT_SYMBOL(ldlm_resource_putref);
1537
1538 /**
1539  * Add a lock into a given resource into specified lock list.
1540  */
1541 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1542                             struct ldlm_lock *lock)
1543 {
1544         check_res_locked(res);
1545
1546         LDLM_DEBUG(lock, "About to add this lock");
1547
1548         if (ldlm_is_destroyed(lock)) {
1549                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1550                 return;
1551         }
1552
1553         LASSERT(list_empty(&lock->l_res_link));
1554
1555         list_add_tail(&lock->l_res_link, head);
1556 }
1557
1558 /**
1559  * Insert a lock into resource after specified lock.
1560  *
1561  * Obtain resource description from the lock we are inserting after.
1562  */
1563 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1564                                      struct ldlm_lock *new)
1565 {
1566         struct ldlm_resource *res = original->l_resource;
1567
1568         check_res_locked(res);
1569
1570         ldlm_resource_dump(D_INFO, res);
1571         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1572
1573         if (ldlm_is_destroyed(new)) {
1574                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1575                 goto out;
1576         }
1577
1578         LASSERT(list_empty(&new->l_res_link));
1579
1580         list_add(&new->l_res_link, &original->l_res_link);
1581  out:;
1582 }
1583
1584 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1585 {
1586         int type = lock->l_resource->lr_type;
1587
1588         check_res_locked(lock->l_resource);
1589         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1590                 ldlm_unlink_lock_skiplist(lock);
1591         else if (type == LDLM_EXTENT)
1592                 ldlm_extent_unlink_lock(lock);
1593         list_del_init(&lock->l_res_link);
1594 }
1595 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1596
1597 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1598 {
1599         desc->lr_type = res->lr_type;
1600         desc->lr_name = res->lr_name;
1601 }
1602
1603 /**
1604  * Print information about all locks in all namespaces on this node to debug
1605  * log.
1606  */
1607 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1608 {
1609         struct list_head *tmp;
1610
1611         if (!((libcfs_debug | D_ERROR) & level))
1612                 return;
1613
1614         mutex_lock(ldlm_namespace_lock(client));
1615
1616         list_for_each(tmp, ldlm_namespace_list(client)) {
1617                 struct ldlm_namespace *ns;
1618
1619                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1620                 ldlm_namespace_dump(level, ns);
1621         }
1622
1623         mutex_unlock(ldlm_namespace_lock(client));
1624 }
1625
1626 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1627                               struct hlist_node *hnode, void *arg)
1628 {
1629         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1630         int    level = (int)(unsigned long)arg;
1631
1632         lock_res(res);
1633         ldlm_resource_dump(level, res);
1634         unlock_res(res);
1635
1636         return 0;
1637 }
1638
1639 /**
1640  * Print information about all locks in this namespace on this node to debug
1641  * log.
1642  */
1643 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1644 {
1645         if (!((libcfs_debug | D_ERROR) & level))
1646                 return;
1647
1648         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1649                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1650                ns_is_client(ns) ? "client" : "server");
1651
1652         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1653                 return;
1654
1655         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1656                                  ldlm_res_hash_dump,
1657                                  (void *)(unsigned long)level, 0);
1658         spin_lock(&ns->ns_lock);
1659         ns->ns_next_dump = cfs_time_shift(10);
1660         spin_unlock(&ns->ns_lock);
1661 }
1662
1663 /**
1664  * Print information about all locks in this resource to debug log.
1665  */
1666 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1667 {
1668         struct ldlm_lock *lock;
1669         unsigned int granted = 0;
1670
1671         CLASSERT(RES_NAME_SIZE == 4);
1672
1673         if (!((libcfs_debug | D_ERROR) & level))
1674                 return;
1675
1676         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1677                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1678
1679         if (!list_empty(&res->lr_granted)) {
1680                 CDEBUG(level, "Granted locks (in reverse order):\n");
1681                 list_for_each_entry_reverse(lock, &res->lr_granted,
1682                                                 l_res_link) {
1683                         LDLM_DEBUG_LIMIT(level, lock, "###");
1684                         if (!(level & D_CANTMASK) &&
1685                             ++granted > ldlm_dump_granted_max) {
1686                                 CDEBUG(level, "only dump %d granted locks to "
1687                                        "avoid DDOS.\n", granted);
1688                                 break;
1689                         }
1690                 }
1691         }
1692         if (!list_empty(&res->lr_converting)) {
1693                 CDEBUG(level, "Converting locks:\n");
1694                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1695                         LDLM_DEBUG_LIMIT(level, lock, "###");
1696         }
1697         if (!list_empty(&res->lr_waiting)) {
1698                 CDEBUG(level, "Waiting locks:\n");
1699                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1700                         LDLM_DEBUG_LIMIT(level, lock, "###");
1701         }
1702 }
1703 EXPORT_SYMBOL(ldlm_resource_dump);