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", *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", 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 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 463 static struct attribute *ldlm_ns_attrs[] = {
 464         &lustre_attr_resource_count.attr,
 465         &lustre_attr_lock_count.attr,
 466         &lustre_attr_lock_unused_count.attr,
 467         &lustre_attr_lru_size.attr,
 468         &lustre_attr_lru_max_age.attr,
 469         &lustre_attr_early_lock_cancel.attr,
 470         NULL,
 471 };
 472
 473 static void ldlm_ns_release(struct kobject *kobj)
 474 {
 475         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 476                                                  ns_kobj);
 477         complete(&ns->ns_kobj_unregister);
 478 }
 479
 480 static struct kobj_type ldlm_ns_ktype = {
 481         .default_attrs  = ldlm_ns_attrs,
 482         .sysfs_ops      = &lustre_sysfs_ops,
 483         .release        = ldlm_ns_release,
 484 };
 485
 486 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 487 {
 488         if (ns->ns_proc_dir_entry == NULL)
 489                 CERROR("dlm namespace %s has no procfs dir?\n",
 490                        ldlm_ns_name(ns));
 491         else
 492                 lprocfs_remove(&ns->ns_proc_dir_entry);
 493
 494         if (ns->ns_stats != NULL)
 495                 lprocfs_free_stats(&ns->ns_stats);
 496 }
 497
 498 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 499 {
 500         kobject_put(&ns->ns_kobj);
 501         wait_for_completion(&ns->ns_kobj_unregister);
 502 }
 503
 504 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 505 {
 506         int err;
 507
 508         ns->ns_kobj.kset = ldlm_ns_kset;
 509         init_completion(&ns->ns_kobj_unregister);
 510         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 511                                    "%s", ldlm_ns_name(ns));
 512
 513         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 514         if (!ns->ns_stats) {
 515                 kobject_put(&ns->ns_kobj);
 516                 return -ENOMEM;
 517         }
 518
 519         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 520                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 521
 522         return err;
 523 }
 524
 525 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 526 {
 527         struct lprocfs_vars lock_vars[2];
 528         char lock_name[MAX_STRING_SIZE + 1];
 529         struct proc_dir_entry *ns_pde;
 530
 531         LASSERT(ns != NULL);
 532         LASSERT(ns->ns_rs_hash != NULL);
 533
 534         if (ns->ns_proc_dir_entry != NULL) {
 535                 ns_pde = ns->ns_proc_dir_entry;
 536         } else {
 537                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 538                 if (ns_pde == NULL)
 539                         return -ENOMEM;
 540                 ns->ns_proc_dir_entry = ns_pde;
 541         }
 542
 543         lock_name[MAX_STRING_SIZE] = '\0';
 544
 545         memset(lock_vars, 0, sizeof(lock_vars));
 546         lock_vars[0].name = lock_name;
 547
 548         if (!ns_is_client(ns)) {
 549                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 550                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 551                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 552                              &ns->ns_timeouts, &ldlm_uint_fops);
 553                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 554                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 555                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 556                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 557                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 558                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 559                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 560                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 561         }
 562         return 0;
 563 }
 564 #undef MAX_STRING_SIZE
 565 #else /* CONFIG_PROC_FS */
 566
 567 #define ldlm_namespace_proc_unregister(ns)      ({;})
 568 #define ldlm_namespace_proc_register(ns)        ({0;})
 569
 570 #endif /* CONFIG_PROC_FS */
 571
 572 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 573                                   const void *key, unsigned mask)
 574 {
 575         const struct ldlm_res_id     *id  = key;
 576         unsigned                val = 0;
 577         unsigned                i;
 578
 579         for (i = 0; i < RES_NAME_SIZE; i++)
 580                 val += id->name[i];
 581         return val & mask;
 582 }
 583
 584 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 585                                       const void *key, unsigned mask)
 586 {
 587         const struct ldlm_res_id *id = key;
 588         struct lu_fid       fid;
 589         __u32               hash;
 590         __u32               val;
 591
 592         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 593         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 594         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 595
 596         hash = fid_flatten32(&fid);
 597         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 598         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 599                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 600                 hash += (val >> 5) + (val << 11);
 601         } else {
 602                 val = fid_oid(&fid);
 603         }
 604         hash = hash_long(hash, hs->hs_bkt_bits);
 605         /* give me another random factor */
 606         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 607
 608         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 609         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 610
 611         return hash & mask;
 612 }
 613
 614 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 615 {
 616         struct ldlm_resource   *res;
 617
 618         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 619         return &res->lr_name;
 620 }
 621
 622 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 623 {
 624         struct ldlm_resource   *res;
 625
 626         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 627         return ldlm_res_eq((const struct ldlm_res_id *)key,
 628                            (const struct ldlm_res_id *)&res->lr_name);
 629 }
 630
 631 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 632 {
 633         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 634 }
 635
 636 static void
 637 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 638 {
 639         struct ldlm_resource *res;
 640
 641         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 642         ldlm_resource_getref(res);
 643 }
 644
 645 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 646 {
 647         struct ldlm_resource *res;
 648
 649         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 650         ldlm_resource_putref(res);
 651 }
 652
 653 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 654         .hs_hash        = ldlm_res_hop_hash,
 655         .hs_key         = ldlm_res_hop_key,
 656         .hs_keycmp      = ldlm_res_hop_keycmp,
 657         .hs_keycpy      = NULL,
 658         .hs_object      = ldlm_res_hop_object,
 659         .hs_get         = ldlm_res_hop_get_locked,
 660         .hs_put         = ldlm_res_hop_put
 661 };
 662
 663 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 664         .hs_hash        = ldlm_res_hop_fid_hash,
 665         .hs_key         = ldlm_res_hop_key,
 666         .hs_keycmp      = ldlm_res_hop_keycmp,
 667         .hs_keycpy      = NULL,
 668         .hs_object      = ldlm_res_hop_object,
 669         .hs_get         = ldlm_res_hop_get_locked,
 670         .hs_put         = ldlm_res_hop_put
 671 };
 672
 673 typedef struct ldlm_ns_hash_def {
 674         enum ldlm_ns_type       nsd_type;
 675         /** hash bucket bits */
 676         unsigned                nsd_bkt_bits;
 677         /** hash bits */
 678         unsigned                nsd_all_bits;
 679         /** hash operations */
 680         struct cfs_hash_ops *nsd_hops;
 681 } ldlm_ns_hash_def_t;
 682
 683 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 684 {
 685         {
 686                 .nsd_type       = LDLM_NS_TYPE_MDC,
 687                 .nsd_bkt_bits   = 11,
 688                 .nsd_all_bits   = 16,
 689                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 690         },
 691         {
 692                 .nsd_type       = LDLM_NS_TYPE_MDT,
 693                 .nsd_bkt_bits   = 14,
 694                 .nsd_all_bits   = 21,
 695                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 696         },
 697         {
 698                 .nsd_type       = LDLM_NS_TYPE_OSC,
 699                 .nsd_bkt_bits   = 8,
 700                 .nsd_all_bits   = 12,
 701                 .nsd_hops       = &ldlm_ns_hash_ops,
 702         },
 703         {
 704                 .nsd_type       = LDLM_NS_TYPE_OST,
 705                 .nsd_bkt_bits   = 11,
 706                 .nsd_all_bits   = 17,
 707                 .nsd_hops       = &ldlm_ns_hash_ops,
 708         },
 709         {
 710                 .nsd_type       = LDLM_NS_TYPE_MGC,
 711                 .nsd_bkt_bits   = 4,
 712                 .nsd_all_bits   = 4,
 713                 .nsd_hops       = &ldlm_ns_hash_ops,
 714         },
 715         {
 716                 .nsd_type       = LDLM_NS_TYPE_MGT,
 717                 .nsd_bkt_bits   = 4,
 718                 .nsd_all_bits   = 4,
 719                 .nsd_hops       = &ldlm_ns_hash_ops,
 720         },
 721         {
 722                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 723         },
 724 };
 725
 726 /**
 727  * Create and initialize new empty namespace.
 728  */
 729 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 730                                           enum ldlm_side client,
 731                                           enum ldlm_appetite apt,
 732                                           enum ldlm_ns_type ns_type)
 733 {
 734         struct ldlm_namespace *ns = NULL;
 735         struct ldlm_ns_bucket *nsb;
 736         struct ldlm_ns_hash_def *nsd;
 737         struct cfs_hash_bd bd;
 738         int idx;
 739         int rc;
 740         ENTRY;
 741
 742         LASSERT(obd != NULL);
 743
 744         rc = ldlm_get_ref();
 745         if (rc) {
 746                 CERROR("ldlm_get_ref failed: %d\n", rc);
 747                 RETURN(NULL);
 748         }
 749
 750         for (idx = 0;;idx++) {
 751                 nsd = &ldlm_ns_hash_defs[idx];
 752                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 753                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 754                         GOTO(out_ref, NULL);
 755                 }
 756
 757                 if (nsd->nsd_type == ns_type)
 758                         break;
 759         }
 760
 761         OBD_ALLOC_PTR(ns);
 762         if (!ns)
 763                 GOTO(out_ref, NULL);
 764
 765         ns->ns_rs_hash = cfs_hash_create(name,
 766                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 767                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 768                                          CFS_HASH_MIN_THETA,
 769                                          CFS_HASH_MAX_THETA,
 770                                          nsd->nsd_hops,
 771                                          CFS_HASH_DEPTH |
 772                                          CFS_HASH_BIGNAME |
 773                                          CFS_HASH_SPIN_BKTLOCK |
 774                                          CFS_HASH_NO_ITEMREF);
 775         if (ns->ns_rs_hash == NULL)
 776                 GOTO(out_ns, NULL);
 777
 778         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 779                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 780                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 781                 nsb->nsb_namespace = ns;
 782                 nsb->nsb_reclaim_start = 0;
 783         }
 784
 785         ns->ns_obd      = obd;
 786         ns->ns_appetite = apt;
 787         ns->ns_client   = client;
 788
 789         INIT_LIST_HEAD(&ns->ns_list_chain);
 790         INIT_LIST_HEAD(&ns->ns_unused_list);
 791         spin_lock_init(&ns->ns_lock);
 792         atomic_set(&ns->ns_bref, 0);
 793         init_waitqueue_head(&ns->ns_waitq);
 794
 795         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 796         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 797         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 798
 799         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 800         ns->ns_nr_unused          = 0;
 801         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 802         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 803         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 804         ns->ns_timeouts           = 0;
 805         ns->ns_orig_connect_flags = 0;
 806         ns->ns_connect_flags      = 0;
 807         ns->ns_stopping           = 0;
 808         ns->ns_reclaim_start      = 0;
 809
 810         rc = ldlm_namespace_sysfs_register(ns);
 811         if (rc) {
 812                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 813                 GOTO(out_hash, rc);
 814         }
 815
 816         rc = ldlm_namespace_proc_register(ns);
 817         if (rc) {
 818                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 819                 GOTO(out_sysfs, rc);
 820         }
 821
 822         idx = ldlm_namespace_nr_read(client);
 823         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 824         if (rc) {
 825                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 826                 GOTO(out_proc, rc);
 827         }
 828
 829         ldlm_namespace_register(ns, client);
 830         RETURN(ns);
 831 out_proc:
 832         ldlm_namespace_proc_unregister(ns);
 833 out_sysfs:
 834         ldlm_namespace_sysfs_unregister(ns);
 835         ldlm_namespace_cleanup(ns, 0);
 836 out_hash:
 837         cfs_hash_putref(ns->ns_rs_hash);
 838 out_ns:
 839         OBD_FREE_PTR(ns);
 840 out_ref:
 841         ldlm_put_ref();
 842         RETURN(NULL);
 843 }
 844 EXPORT_SYMBOL(ldlm_namespace_new);
 845
 846 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 847
 848 /**
 849  * Cancel and destroy all locks on a resource.
 850  *
 851  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 852  * clean up.  This is currently only used for recovery, and we make
 853  * certain assumptions as a result--notably, that we shouldn't cancel
 854  * locks with refs.
 855  */
 856 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 857                              __u64 flags)
 858 {
 859         struct list_head *tmp;
 860         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 861         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 862
 863         do {
 864                 struct ldlm_lock *lock = NULL;
 865
 866                 /* First, we look for non-cleaned-yet lock
 867                  * all cleaned locks are marked by CLEANED flag. */
 868                 lock_res(res);
 869                 list_for_each(tmp, q) {
 870                         lock = list_entry(tmp, struct ldlm_lock,
 871                                           l_res_link);
 872                         if (ldlm_is_cleaned(lock)) {
 873                                 lock = NULL;
 874                                 continue;
 875                         }
 876                         LDLM_LOCK_GET(lock);
 877                         ldlm_set_cleaned(lock);
 878                         break;
 879                 }
 880
 881                 if (lock == NULL) {
 882                         unlock_res(res);
 883                         break;
 884                 }
 885
 886                 /* Set CBPENDING so nothing in the cancellation path
 887                  * can match this lock. */
 888                 ldlm_set_cbpending(lock);
 889                 ldlm_set_failed(lock);
 890                 lock->l_flags |= flags;
 891
 892                 /* ... without sending a CANCEL message for local_only. */
 893                 if (local_only)
 894                         ldlm_set_local_only(lock);
 895
 896                 if (local_only && (lock->l_readers || lock->l_writers)) {
 897                         /* This is a little bit gross, but much better than the
 898                          * alternative: pretend that we got a blocking AST from
 899                          * the server, so that when the lock is decref'd, it
 900                          * will go away ... */
 901                         unlock_res(res);
 902                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 903                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 904                                 set_current_state(TASK_UNINTERRUPTIBLE);
 905                                 schedule_timeout(cfs_time_seconds(4));
 906                                 set_current_state(TASK_RUNNING);
 907                         }
 908                         if (lock->l_completion_ast)
 909                                 lock->l_completion_ast(lock,
 910                                                        LDLM_FL_FAILED, NULL);
 911                         LDLM_LOCK_RELEASE(lock);
 912                         continue;
 913                 }
 914
 915                 if (client) {
 916                         struct lustre_handle lockh;
 917
 918                         unlock_res(res);
 919                         ldlm_lock2handle(lock, &lockh);
 920                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
 921                         if (rc)
 922                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 923                 } else {
 924                         unlock_res(res);
 925                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 926                                    "client node");
 927                         ldlm_lock_cancel(lock);
 928                 }
 929                 LDLM_LOCK_RELEASE(lock);
 930         } while (1);
 931 }
 932
 933 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 934                                struct hlist_node *hnode, void *arg)
 935 {
 936         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 937         __u64 flags = *(__u64 *)arg;
 938
 939         cleanup_resource(res, &res->lr_granted, flags);
 940         cleanup_resource(res, &res->lr_converting, flags);
 941         cleanup_resource(res, &res->lr_waiting, flags);
 942
 943         return 0;
 944 }
 945
 946 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 947                                   struct hlist_node *hnode, void *arg)
 948 {
 949         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 950
 951         lock_res(res);
 952         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 953                "(%d) after lock cleanup; forcing cleanup.\n",
 954                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 955                atomic_read(&res->lr_refcount) - 1);
 956
 957         ldlm_resource_dump(D_ERROR, res);
 958         unlock_res(res);
 959         return 0;
 960 }
 961
 962 /**
 963  * Cancel and destroy all locks in the namespace.
 964  *
 965  * Typically used during evictions when server notified client that it was
 966  * evicted and all of its state needs to be destroyed.
 967  * Also used during shutdown.
 968  */
 969 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 970 {
 971         if (ns == NULL) {
 972                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 973                 return ELDLM_OK;
 974         }
 975
 976         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 977                                  &flags, 0);
 978         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
 979                                  NULL, 0);
 980         return ELDLM_OK;
 981 }
 982 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 983
 984 /**
 985  * Attempts to free namespace.
 986  *
 987  * Only used when namespace goes away, like during an unmount.
 988  */
 989 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 990 {
 991         ENTRY;
 992
 993         /* At shutdown time, don't call the cancellation callback */
 994         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 995
 996         if (atomic_read(&ns->ns_bref) > 0) {
 997                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 998                 int rc;
 999                 CDEBUG(D_DLMTRACE,
1000                        "dlm namespace %s free waiting on refcount %d\n",
1001                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1002 force_wait:
1003                 if (force)
1004                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
1005                                           MSEC_PER_SEC) / 4, NULL, NULL);
1006
1007                 rc = l_wait_event(ns->ns_waitq,
1008                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1009
1010                 /* Forced cleanups should be able to reclaim all references,
1011                  * so it's safe to wait forever... we can't leak locks... */
1012                 if (force && rc == -ETIMEDOUT) {
1013                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1014                                        "namespace with %d resources in use, "
1015                                        "(rc=%d)\n", ldlm_ns_name(ns),
1016                                        atomic_read(&ns->ns_bref), rc);
1017                         GOTO(force_wait, rc);
1018                 }
1019
1020                 if (atomic_read(&ns->ns_bref)) {
1021                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1022                                        "with %d resources in use, (rc=%d)\n",
1023                                        ldlm_ns_name(ns),
1024                                        atomic_read(&ns->ns_bref), rc);
1025                         RETURN(ELDLM_NAMESPACE_EXISTS);
1026                 }
1027                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1028                        ldlm_ns_name(ns));
1029         }
1030
1031         RETURN(ELDLM_OK);
1032 }
1033
1034 /**
1035  * Performs various cleanups for passed \a ns to make it drop refc and be
1036  * ready for freeing. Waits for refc == 0.
1037  *
1038  * The following is done:
1039  * (0) Unregister \a ns from its list to make inaccessible for potential
1040  * users like pools thread and others;
1041  * (1) Clear all locks in \a ns.
1042  */
1043 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1044                                struct obd_import *imp,
1045                                int force)
1046 {
1047         int rc;
1048         ENTRY;
1049         if (!ns) {
1050                 EXIT;
1051                 return;
1052         }
1053
1054         spin_lock(&ns->ns_lock);
1055         ns->ns_stopping = 1;
1056         spin_unlock(&ns->ns_lock);
1057
1058         /*
1059          * Can fail with -EINTR when force == 0 in which case try harder.
1060          */
1061         rc = __ldlm_namespace_free(ns, force);
1062         if (rc != ELDLM_OK) {
1063                 if (imp) {
1064                         ptlrpc_disconnect_import(imp, 0);
1065                         ptlrpc_invalidate_import(imp);
1066                 }
1067
1068                 /*
1069                  * With all requests dropped and the import inactive
1070                  * we are gaurenteed all reference will be dropped.
1071                  */
1072                 rc = __ldlm_namespace_free(ns, 1);
1073                 LASSERT(rc == 0);
1074         }
1075         EXIT;
1076 }
1077 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1078
1079 /**
1080  * Performs freeing memory structures related to \a ns. This is only done
1081  * when ldlm_namespce_free_prior() successfully removed all resources
1082  * referencing \a ns and its refc == 0.
1083  */
1084 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1085 {
1086         ENTRY;
1087         if (!ns) {
1088                 EXIT;
1089                 return;
1090         }
1091
1092         /* Make sure that nobody can find this ns in its list. */
1093         ldlm_namespace_unregister(ns, ns->ns_client);
1094         /* Fini pool _before_ parent proc dir is removed. This is important as
1095          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1096          * Removing it after @dir may cause oops. */
1097         ldlm_pool_fini(&ns->ns_pool);
1098
1099         ldlm_namespace_proc_unregister(ns);
1100         ldlm_namespace_sysfs_unregister(ns);
1101         cfs_hash_putref(ns->ns_rs_hash);
1102         /* Namespace \a ns should be not on list at this time, otherwise
1103          * this will cause issues related to using freed \a ns in poold
1104          * thread. */
1105         LASSERT(list_empty(&ns->ns_list_chain));
1106         OBD_FREE_PTR(ns);
1107         ldlm_put_ref();
1108         EXIT;
1109 }
1110 EXPORT_SYMBOL(ldlm_namespace_free_post);
1111
1112 /**
1113  * Cleanup the resource, and free namespace.
1114  * bug 12864:
1115  * Deadlock issue:
1116  * proc1: destroy import
1117  *        class_disconnect_export(grab cl_sem) ->
1118  *              -> ldlm_namespace_free ->
1119  *              -> lprocfs_remove(grab _lprocfs_lock).
1120  * proc2: read proc info
1121  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1122  *              -> osc_rd_active, etc(grab cl_sem).
1123  *
1124  * So that I have to split the ldlm_namespace_free into two parts - the first
1125  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1126  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1127  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1128  * held.
1129  */
1130 void ldlm_namespace_free(struct ldlm_namespace *ns,
1131                          struct obd_import *imp,
1132                          int force)
1133 {
1134         ldlm_namespace_free_prior(ns, imp, force);
1135         ldlm_namespace_free_post(ns);
1136 }
1137 EXPORT_SYMBOL(ldlm_namespace_free);
1138
1139 void ldlm_namespace_get(struct ldlm_namespace *ns)
1140 {
1141         atomic_inc(&ns->ns_bref);
1142 }
1143
1144 /* This is only for callers that care about refcount */
1145 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1146 {
1147         return atomic_inc_return(&ns->ns_bref);
1148 }
1149
1150 void ldlm_namespace_put(struct ldlm_namespace *ns)
1151 {
1152         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1153                 wake_up(&ns->ns_waitq);
1154                 spin_unlock(&ns->ns_lock);
1155         }
1156 }
1157
1158 /** Register \a ns in the list of namespaces */
1159 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1160 {
1161         mutex_lock(ldlm_namespace_lock(client));
1162         LASSERT(list_empty(&ns->ns_list_chain));
1163         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1164         ldlm_namespace_nr_inc(client);
1165         mutex_unlock(ldlm_namespace_lock(client));
1166 }
1167
1168 /** Unregister \a ns from the list of namespaces. */
1169 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1170 {
1171         mutex_lock(ldlm_namespace_lock(client));
1172         LASSERT(!list_empty(&ns->ns_list_chain));
1173         /* Some asserts and possibly other parts of the code are still
1174          * using list_empty(&ns->ns_list_chain). This is why it is
1175          * important to use list_del_init() here. */
1176         list_del_init(&ns->ns_list_chain);
1177         ldlm_namespace_nr_dec(client);
1178         mutex_unlock(ldlm_namespace_lock(client));
1179 }
1180
1181 /** Should be called with ldlm_namespace_lock(client) taken. */
1182 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1183                                           enum ldlm_side client)
1184 {
1185         LASSERT(!list_empty(&ns->ns_list_chain));
1186         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1187         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1188 }
1189
1190 /** Should be called with ldlm_namespace_lock(client) taken. */
1191 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1192                                             enum ldlm_side client)
1193 {
1194         LASSERT(!list_empty(&ns->ns_list_chain));
1195         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1196         list_move_tail(&ns->ns_list_chain,
1197                        ldlm_namespace_inactive_list(client));
1198 }
1199
1200 /** Should be called with ldlm_namespace_lock(client) taken. */
1201 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1202 {
1203         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1204         LASSERT(!list_empty(ldlm_namespace_list(client)));
1205         return container_of(ldlm_namespace_list(client)->next,
1206                             struct ldlm_namespace, ns_list_chain);
1207 }
1208
1209 /** Create and initialize new resource. */
1210 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1211 {
1212         struct ldlm_resource *res;
1213         int idx;
1214
1215         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1216         if (res == NULL)
1217                 return NULL;
1218
1219         if (ldlm_type == LDLM_EXTENT) {
1220                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1221                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1222                 if (res->lr_itree == NULL) {
1223                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1224                         return NULL;
1225                 }
1226                 /* Initialize interval trees for each lock mode. */
1227                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1228                         res->lr_itree[idx].lit_size = 0;
1229                         res->lr_itree[idx].lit_mode = 1 << idx;
1230                         res->lr_itree[idx].lit_root = NULL;
1231                 }
1232         }
1233
1234         INIT_LIST_HEAD(&res->lr_granted);
1235         INIT_LIST_HEAD(&res->lr_converting);
1236         INIT_LIST_HEAD(&res->lr_waiting);
1237
1238         atomic_set(&res->lr_refcount, 1);
1239         spin_lock_init(&res->lr_lock);
1240         lu_ref_init(&res->lr_reference);
1241
1242         /* Since LVB init can be delayed now, there is no longer need to
1243          * immediatelly acquire mutex here. */
1244         mutex_init(&res->lr_lvb_mutex);
1245         res->lr_lvb_initialized = false;
1246
1247         return res;
1248 }
1249
1250 /**
1251  * Return a reference to resource with given name, creating it if necessary.
1252  * Args: namespace with ns_lock unlocked
1253  * Locks: takes and releases NS hash-lock and res->lr_lock
1254  * Returns: referenced, unlocked ldlm_resource or NULL
1255  */
1256 struct ldlm_resource *
1257 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1258                   const struct ldlm_res_id *name, enum ldlm_type type,
1259                   int create)
1260 {
1261         struct hlist_node       *hnode;
1262         struct ldlm_resource    *res = NULL;
1263         struct cfs_hash_bd              bd;
1264         __u64                   version;
1265         int                     ns_refcount = 0;
1266
1267         LASSERT(ns != NULL);
1268         LASSERT(parent == NULL);
1269         LASSERT(ns->ns_rs_hash != NULL);
1270         LASSERT(name->name[0] != 0);
1271
1272         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1273         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1274         if (hnode != NULL) {
1275                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1276                 GOTO(found, res);
1277         }
1278
1279         version = cfs_hash_bd_version_get(&bd);
1280         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1281
1282         if (create == 0)
1283                 return ERR_PTR(-ENOENT);
1284
1285         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1286                  "type: %d\n", type);
1287         res = ldlm_resource_new(type);
1288         if (res == NULL)
1289                 return ERR_PTR(-ENOMEM);
1290
1291         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1292         res->lr_name = *name;
1293         res->lr_type = type;
1294
1295         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1296         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1297                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1298
1299         if (hnode != NULL) {
1300                 /* Someone won the race and already added the resource. */
1301                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1302                 /* Clean lu_ref for failed resource. */
1303                 lu_ref_fini(&res->lr_reference);
1304                 if (res->lr_itree != NULL)
1305                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1306                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1307                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1308 found:
1309                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1310                 return res;
1311         }
1312         /* We won! Let's add the resource. */
1313         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1314         if (cfs_hash_bd_count_get(&bd) == 1)
1315                 ns_refcount = ldlm_namespace_get_return(ns);
1316
1317         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1318
1319         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1320
1321         /* Let's see if we happened to be the very first resource in this
1322          * namespace. If so, and this is a client namespace, we need to move
1323          * the namespace into the active namespaces list to be patrolled by
1324          * the ldlm_poold. */
1325         if (ns_is_client(ns) && ns_refcount == 1) {
1326                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1327                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1328                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1329         }
1330
1331         return res;
1332 }
1333 EXPORT_SYMBOL(ldlm_resource_get);
1334
1335 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1336 {
1337         LASSERT(res != NULL);
1338         LASSERT(res != LP_POISON);
1339         atomic_inc(&res->lr_refcount);
1340         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1341                atomic_read(&res->lr_refcount));
1342         return res;
1343 }
1344
1345 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1346                                          struct ldlm_resource *res)
1347 {
1348         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1349
1350         if (!list_empty(&res->lr_granted)) {
1351                 ldlm_resource_dump(D_ERROR, res);
1352                 LBUG();
1353         }
1354
1355         if (!list_empty(&res->lr_converting)) {
1356                 ldlm_resource_dump(D_ERROR, res);
1357                 LBUG();
1358         }
1359
1360         if (!list_empty(&res->lr_waiting)) {
1361                 ldlm_resource_dump(D_ERROR, res);
1362                 LBUG();
1363         }
1364
1365         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1366                                bd, &res->lr_hash);
1367         lu_ref_fini(&res->lr_reference);
1368         if (cfs_hash_bd_count_get(bd) == 0)
1369                 ldlm_namespace_put(nsb->nsb_namespace);
1370 }
1371
1372 /* Returns 1 if the resource was freed, 0 if it remains. */
1373 int ldlm_resource_putref(struct ldlm_resource *res)
1374 {
1375         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1376         struct cfs_hash_bd   bd;
1377
1378         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1379         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1380                res, atomic_read(&res->lr_refcount) - 1);
1381
1382         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1383         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1384                 __ldlm_resource_putref_final(&bd, res);
1385                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1386                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1387                         ns->ns_lvbo->lvbo_free(res);
1388                 if (res->lr_itree != NULL)
1389                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1390                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1391                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1392                 return 1;
1393         }
1394         return 0;
1395 }
1396 EXPORT_SYMBOL(ldlm_resource_putref);
1397
1398 /**
1399  * Add a lock into a given resource into specified lock list.
1400  */
1401 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1402                             struct ldlm_lock *lock)
1403 {
1404         check_res_locked(res);
1405
1406         LDLM_DEBUG(lock, "About to add this lock");
1407
1408         if (ldlm_is_destroyed(lock)) {
1409                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1410                 return;
1411         }
1412
1413         LASSERT(list_empty(&lock->l_res_link));
1414
1415         list_add_tail(&lock->l_res_link, head);
1416 }
1417
1418 /**
1419  * Insert a lock into resource after specified lock.
1420  *
1421  * Obtain resource description from the lock we are inserting after.
1422  */
1423 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1424                                      struct ldlm_lock *new)
1425 {
1426         struct ldlm_resource *res = original->l_resource;
1427
1428         check_res_locked(res);
1429
1430         ldlm_resource_dump(D_INFO, res);
1431         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1432
1433         if (ldlm_is_destroyed(new)) {
1434                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1435                 goto out;
1436         }
1437
1438         LASSERT(list_empty(&new->l_res_link));
1439
1440         list_add(&new->l_res_link, &original->l_res_link);
1441  out:;
1442 }
1443
1444 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1445 {
1446         int type = lock->l_resource->lr_type;
1447
1448         check_res_locked(lock->l_resource);
1449         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1450                 ldlm_unlink_lock_skiplist(lock);
1451         else if (type == LDLM_EXTENT)
1452                 ldlm_extent_unlink_lock(lock);
1453         list_del_init(&lock->l_res_link);
1454 }
1455 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1456
1457 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1458 {
1459         desc->lr_type = res->lr_type;
1460         desc->lr_name = res->lr_name;
1461 }
1462
1463 /**
1464  * Print information about all locks in all namespaces on this node to debug
1465  * log.
1466  */
1467 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1468 {
1469         struct list_head *tmp;
1470
1471         if (!((libcfs_debug | D_ERROR) & level))
1472                 return;
1473
1474         mutex_lock(ldlm_namespace_lock(client));
1475
1476         list_for_each(tmp, ldlm_namespace_list(client)) {
1477                 struct ldlm_namespace *ns;
1478
1479                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1480                 ldlm_namespace_dump(level, ns);
1481         }
1482
1483         mutex_unlock(ldlm_namespace_lock(client));
1484 }
1485
1486 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1487                               struct hlist_node *hnode, void *arg)
1488 {
1489         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1490         int    level = (int)(unsigned long)arg;
1491
1492         lock_res(res);
1493         ldlm_resource_dump(level, res);
1494         unlock_res(res);
1495
1496         return 0;
1497 }
1498
1499 /**
1500  * Print information about all locks in this namespace on this node to debug
1501  * log.
1502  */
1503 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1504 {
1505         if (!((libcfs_debug | D_ERROR) & level))
1506                 return;
1507
1508         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1509                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1510                ns_is_client(ns) ? "client" : "server");
1511
1512         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1513                 return;
1514
1515         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1516                                  ldlm_res_hash_dump,
1517                                  (void *)(unsigned long)level, 0);
1518         spin_lock(&ns->ns_lock);
1519         ns->ns_next_dump = cfs_time_shift(10);
1520         spin_unlock(&ns->ns_lock);
1521 }
1522
1523 /**
1524  * Print information about all locks in this resource to debug log.
1525  */
1526 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1527 {
1528         struct ldlm_lock *lock;
1529         unsigned int granted = 0;
1530
1531         CLASSERT(RES_NAME_SIZE == 4);
1532
1533         if (!((libcfs_debug | D_ERROR) & level))
1534                 return;
1535
1536         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1537                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1538
1539         if (!list_empty(&res->lr_granted)) {
1540                 CDEBUG(level, "Granted locks (in reverse order):\n");
1541                 list_for_each_entry_reverse(lock, &res->lr_granted,
1542                                                 l_res_link) {
1543                         LDLM_DEBUG_LIMIT(level, lock, "###");
1544                         if (!(level & D_CANTMASK) &&
1545                             ++granted > ldlm_dump_granted_max) {
1546                                 CDEBUG(level, "only dump %d granted locks to "
1547                                        "avoid DDOS.\n", granted);
1548                                 break;
1549                         }
1550                 }
1551         }
1552         if (!list_empty(&res->lr_converting)) {
1553                 CDEBUG(level, "Converting locks:\n");
1554                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1555                         LDLM_DEBUG_LIMIT(level, lock, "###");
1556         }
1557         if (!list_empty(&res->lr_waiting)) {
1558                 CDEBUG(level, "Waiting locks:\n");
1559                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1560                         LDLM_DEBUG_LIMIT(level, lock, "###");
1561         }
1562 }
1563 EXPORT_SYMBOL(ldlm_resource_dump);