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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ldlm/ldlm_resource.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43 #include <lustre_dlm.h>
  44 #include <lustre_fid.h>
  45 #include <obd_class.h>
  46 #include "ldlm_internal.h"
  47
  48 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  49 struct kmem_cache *ldlm_interval_tree_slab;
  50
  51 int ldlm_srv_namespace_nr = 0;
  52 int ldlm_cli_namespace_nr = 0;
  53
  54 struct mutex ldlm_srv_namespace_lock;
  55 struct list_head ldlm_srv_namespace_list;
  56
  57 struct mutex ldlm_cli_namespace_lock;
  58 /* Client Namespaces that have active resources in them.
  59  * Once all resources go away, ldlm_poold moves such namespaces to the
  60  * inactive list */
  61 struct list_head ldlm_cli_active_namespace_list;
  62 /* Client namespaces that don't have any locks in them */
  63 struct list_head ldlm_cli_inactive_namespace_list;
  64
  65 static struct proc_dir_entry *ldlm_type_proc_dir;
  66 static struct proc_dir_entry *ldlm_ns_proc_dir;
  67 struct proc_dir_entry *ldlm_svc_proc_dir;
  68
  69 /* during debug dump certain amount of granted locks for one resource to avoid
  70  * DDOS. */
  71 static unsigned int ldlm_dump_granted_max = 256;
  72
  73 #ifdef CONFIG_PROC_FS
  74 static ssize_t
  75 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
  76                           size_t count, loff_t *off)
  77 {
  78         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  79         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  80         RETURN(count);
  81 }
  82 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
  83
  84 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
  85 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
  86
  87 #ifdef HAVE_SERVER_SUPPORT
  88
  89 static int seq_watermark_show(struct seq_file *m, void *data)
  90 {
  91         seq_printf(m, "%llu\n", *(__u64 *)m->private);
  92         return 0;
  93 }
  94
  95 static ssize_t seq_watermark_write(struct file *file,
  96                                    const char __user *buffer, size_t count,
  97                                    loff_t *off)
  98 {
  99         __s64 value;
 100         __u64 watermark;
 101         __u64 *data = ((struct seq_file *)file->private_data)->private;
 102         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 103         int rc;
 104
 105         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 106         if (rc) {
 107                 CERROR("Failed to set %s, rc = %d.\n",
 108                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 109                        rc);
 110                 return rc;
 111         } else if (value != 0 && value < (1 << 20)) {
 112                 CERROR("%s should be greater than 1MB.\n",
 113                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 114                 return -EINVAL;
 115         }
 116         watermark = value >> 20;
 117
 118         if (wm_low) {
 119                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 120                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 121                                "lock_limit_mb.\n");
 122                         return -EINVAL;
 123                 }
 124
 125                 *data = watermark;
 126                 if (watermark != 0) {
 127                         watermark <<= 20;
 128                         do_div(watermark, sizeof(struct ldlm_lock));
 129                 }
 130                 ldlm_reclaim_threshold = watermark;
 131         } else {
 132                 if (ldlm_reclaim_threshold_mb != 0 &&
 133                     watermark < ldlm_reclaim_threshold_mb) {
 134                         CERROR("lock_limit_mb must be greater than "
 135                                "lock_reclaim_threshold_mb.\n");
 136                         return -EINVAL;
 137                 }
 138
 139                 *data = watermark;
 140                 if (watermark != 0) {
 141                         watermark <<= 20;
 142                         do_div(watermark, sizeof(struct ldlm_lock));
 143                 }
 144                 ldlm_lock_limit = watermark;
 145         }
 146
 147         return count;
 148 }
 149
 150 static int seq_watermark_open(struct inode *inode, struct file *file)
 151 {
 152         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 153 }
 154
 155 static const struct file_operations ldlm_watermark_fops = {
 156         .owner          = THIS_MODULE,
 157         .open           = seq_watermark_open,
 158         .read           = seq_read,
 159         .write          = seq_watermark_write,
 160         .llseek         = seq_lseek,
 161         .release        = lprocfs_single_release,
 162 };
 163
 164 static int seq_granted_show(struct seq_file *m, void *data)
 165 {
 166         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 167                    (struct percpu_counter *)m->private));
 168         return 0;
 169 }
 170
 171 static int seq_granted_open(struct inode *inode, struct file *file)
 172 {
 173         return single_open(file, seq_granted_show, PDE_DATA(inode));
 174 }
 175
 176 static const struct file_operations ldlm_granted_fops = {
 177         .owner  = THIS_MODULE,
 178         .open   = seq_granted_open,
 179         .read   = seq_read,
 180         .llseek = seq_lseek,
 181         .release = seq_release,
 182 };
 183
 184 #endif /* HAVE_SERVER_SUPPORT */
 185
 186 int ldlm_proc_setup(void)
 187 {
 188         int rc;
 189         struct lprocfs_vars list[] = {
 190                 { .name =       "dump_namespaces",
 191                   .fops =       &ldlm_dump_ns_fops,
 192                   .proc_mode =  0222 },
 193                 { .name =       "dump_granted_max",
 194                   .fops =       &ldlm_rw_uint_fops,
 195                   .data =       &ldlm_dump_granted_max },
 196                 { .name =       "cancel_unused_locks_before_replay",
 197                   .fops =       &ldlm_rw_uint_fops,
 198                   .data =       &ldlm_cancel_unused_locks_before_replay },
 199 #ifdef HAVE_SERVER_SUPPORT
 200                 { .name =       "lock_reclaim_threshold_mb",
 201                   .fops =       &ldlm_watermark_fops,
 202                   .data =       &ldlm_reclaim_threshold_mb },
 203                 { .name =       "lock_limit_mb",
 204                   .fops =       &ldlm_watermark_fops,
 205                   .data =       &ldlm_lock_limit_mb },
 206                 { .name =       "lock_granted_count",
 207                   .fops =       &ldlm_granted_fops,
 208                   .data =       &ldlm_granted_total },
 209 #endif
 210                 { NULL }};
 211         ENTRY;
 212         LASSERT(ldlm_ns_proc_dir == NULL);
 213
 214         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 215                                               proc_lustre_root,
 216                                               NULL, NULL);
 217         if (IS_ERR(ldlm_type_proc_dir)) {
 218                 CERROR("LProcFS failed in ldlm-init\n");
 219                 rc = PTR_ERR(ldlm_type_proc_dir);
 220                 GOTO(err, rc);
 221         }
 222
 223         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 224                                             ldlm_type_proc_dir,
 225                                             NULL, NULL);
 226         if (IS_ERR(ldlm_ns_proc_dir)) {
 227                 CERROR("LProcFS failed in ldlm-init\n");
 228                 rc = PTR_ERR(ldlm_ns_proc_dir);
 229                 GOTO(err_type, rc);
 230         }
 231
 232         ldlm_svc_proc_dir = lprocfs_register("services",
 233                                              ldlm_type_proc_dir,
 234                                              NULL, NULL);
 235         if (IS_ERR(ldlm_svc_proc_dir)) {
 236                 CERROR("LProcFS failed in ldlm-init\n");
 237                 rc = PTR_ERR(ldlm_svc_proc_dir);
 238                 GOTO(err_ns, rc);
 239         }
 240
 241         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 242         if (rc != 0) {
 243                 CERROR("LProcFS failed in ldlm-init\n");
 244                 GOTO(err_svc, rc);
 245         }
 246
 247         RETURN(0);
 248
 249 err_svc:
 250         lprocfs_remove(&ldlm_svc_proc_dir);
 251 err_ns:
 252         lprocfs_remove(&ldlm_ns_proc_dir);
 253 err_type:
 254         lprocfs_remove(&ldlm_type_proc_dir);
 255 err:
 256         ldlm_svc_proc_dir = NULL;
 257         RETURN(rc);
 258 }
 259
 260 void ldlm_proc_cleanup(void)
 261 {
 262         if (ldlm_svc_proc_dir)
 263                 lprocfs_remove(&ldlm_svc_proc_dir);
 264
 265         if (ldlm_ns_proc_dir)
 266                 lprocfs_remove(&ldlm_ns_proc_dir);
 267
 268         if (ldlm_type_proc_dir)
 269                 lprocfs_remove(&ldlm_type_proc_dir);
 270 }
 271
 272 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
 273 {
 274         struct ldlm_namespace   *ns  = m->private;
 275         __u64                   res = 0;
 276         struct cfs_hash_bd              bd;
 277         int                     i;
 278
 279         /* result is not strictly consistant */
 280         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 281                 res += cfs_hash_bd_count_get(&bd);
 282         return lprocfs_u64_seq_show(m, &res);
 283 }
 284 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
 285
 286 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
 287 {
 288         struct ldlm_namespace   *ns = m->private;
 289         __u64                   locks;
 290
 291         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 292                                         LPROCFS_FIELDS_FLAGS_SUM);
 293         return lprocfs_u64_seq_show(m, &locks);
 294 }
 295 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
 296
 297 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
 298 {
 299         struct ldlm_namespace *ns = m->private;
 300         __u32 *nr = &ns->ns_max_unused;
 301
 302         if (ns_connect_lru_resize(ns))
 303                 nr = &ns->ns_nr_unused;
 304         return lprocfs_uint_seq_show(m, nr);
 305 }
 306
 307 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
 308                                           const char __user *buffer,
 309                                           size_t count, loff_t *off)
 310 {
 311         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 312         char dummy[MAX_STRING_SIZE + 1], *end;
 313         unsigned long tmp;
 314         int lru_resize;
 315
 316         dummy[MAX_STRING_SIZE] = '\0';
 317         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 318                 return -EFAULT;
 319
 320         if (strncmp(dummy, "clear", 5) == 0) {
 321                 CDEBUG(D_DLMTRACE,
 322                        "dropping all unused locks from namespace %s\n",
 323                        ldlm_ns_name(ns));
 324                 if (ns_connect_lru_resize(ns)) {
 325                         int canceled, unused  = ns->ns_nr_unused;
 326
 327                         /* Try to cancel all @ns_nr_unused locks. */
 328                         canceled = ldlm_cancel_lru(ns, unused, 0,
 329                                                    LDLM_LRU_FLAG_PASSED);
 330                         if (canceled < unused) {
 331                                 CDEBUG(D_DLMTRACE,
 332                                        "not all requested locks are canceled, "
 333                                        "requested: %d, canceled: %d\n", unused,
 334                                        canceled);
 335                                 return -EINVAL;
 336                         }
 337                 } else {
 338                         tmp = ns->ns_max_unused;
 339                         ns->ns_max_unused = 0;
 340                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED);
 341                         ns->ns_max_unused = tmp;
 342                 }
 343                 return count;
 344         }
 345
 346         tmp = simple_strtoul(dummy, &end, 0);
 347         if (dummy == end) {
 348                 CERROR("invalid value written\n");
 349                 return -EINVAL;
 350         }
 351         lru_resize = (tmp == 0);
 352
 353         if (ns_connect_lru_resize(ns)) {
 354                 if (!lru_resize)
 355                         ns->ns_max_unused = tmp;
 356
 357                 if (tmp > ns->ns_nr_unused)
 358                         tmp = ns->ns_nr_unused;
 359                 tmp = ns->ns_nr_unused - tmp;
 360
 361                 CDEBUG(D_DLMTRACE,
 362                        "changing namespace %s unused locks from %u to %u\n",
 363                        ldlm_ns_name(ns), ns->ns_nr_unused,
 364                        (unsigned int)tmp);
 365                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 366
 367                 if (!lru_resize) {
 368                         CDEBUG(D_DLMTRACE,
 369                                "disable lru_resize for namespace %s\n",
 370                                ldlm_ns_name(ns));
 371                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 372                 }
 373         } else {
 374                 CDEBUG(D_DLMTRACE,
 375                        "changing namespace %s max_unused from %u to %u\n",
 376                        ldlm_ns_name(ns), ns->ns_max_unused,
 377                        (unsigned int)tmp);
 378                 ns->ns_max_unused = (unsigned int)tmp;
 379                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 380
 381                 /* Make sure that LRU resize was originally supported before
 382                  * turning it on here. */
 383                 if (lru_resize &&
 384                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 385                         CDEBUG(D_DLMTRACE,
 386                                "enable lru_resize for namespace %s\n",
 387                                ldlm_ns_name(ns));
 388                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 389                 }
 390         }
 391
 392         return count;
 393 }
 394 LPROC_SEQ_FOPS(lprocfs_lru_size);
 395
 396 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
 397 {
 398         struct ldlm_namespace *ns = m->private;
 399         unsigned int supp = ns_connect_cancelset(ns);
 400
 401         return lprocfs_uint_seq_show(m, &supp);
 402 }
 403
 404 static ssize_t lprocfs_elc_seq_write(struct file *file,
 405                                      const char __user *buffer,
 406                                      size_t count, loff_t *off)
 407 {
 408         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 409         unsigned int supp = -1;
 410         int rc;
 411
 412         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 413         if (rc < 0)
 414                 return rc;
 415
 416         if (supp == 0)
 417                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 418         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 419                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 420         return count;
 421 }
 422 LPROC_SEQ_FOPS(lprocfs_elc);
 423
 424 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 425 {
 426         if (ns->ns_proc_dir_entry == NULL)
 427                 CERROR("dlm namespace %s has no procfs dir?\n",
 428                        ldlm_ns_name(ns));
 429         else
 430                 lprocfs_remove(&ns->ns_proc_dir_entry);
 431
 432         if (ns->ns_stats != NULL)
 433                 lprocfs_free_stats(&ns->ns_stats);
 434 }
 435
 436 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 437 {
 438         struct lprocfs_vars lock_vars[2];
 439         char lock_name[MAX_STRING_SIZE + 1];
 440         struct proc_dir_entry *ns_pde;
 441
 442         LASSERT(ns != NULL);
 443         LASSERT(ns->ns_rs_hash != NULL);
 444
 445         if (ns->ns_proc_dir_entry != NULL) {
 446                 ns_pde = ns->ns_proc_dir_entry;
 447         } else {
 448                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 449                 if (ns_pde == NULL)
 450                         return -ENOMEM;
 451                 ns->ns_proc_dir_entry = ns_pde;
 452         }
 453
 454         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 455         if (ns->ns_stats == NULL)
 456                 return -ENOMEM;
 457
 458         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 459                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 460
 461         lock_name[MAX_STRING_SIZE] = '\0';
 462
 463         memset(lock_vars, 0, sizeof(lock_vars));
 464         lock_vars[0].name = lock_name;
 465
 466         ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
 467                      &lprocfs_ns_resources_fops);
 468         ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
 469                      &lprocfs_ns_locks_fops);
 470
 471         if (ns_is_client(ns)) {
 472                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
 473                              &ns->ns_nr_unused, &ldlm_uint_fops);
 474                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
 475                              &lprocfs_lru_size_fops);
 476                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
 477                              &ns->ns_max_age, &ldlm_rw_uint_fops);
 478                 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
 479                              ns, &lprocfs_elc_fops);
 480         } else {
 481                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 482                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 483                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 484                              &ns->ns_timeouts, &ldlm_uint_fops);
 485                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 486                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 487                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 488                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 489                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 490                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 491                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 492                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 493         }
 494         return 0;
 495 }
 496 #undef MAX_STRING_SIZE
 497 #else /* CONFIG_PROC_FS */
 498
 499 #define ldlm_namespace_proc_unregister(ns)      ({;})
 500 #define ldlm_namespace_proc_register(ns)        ({0;})
 501
 502 #endif /* CONFIG_PROC_FS */
 503
 504 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 505                                   const void *key, unsigned mask)
 506 {
 507         const struct ldlm_res_id     *id  = key;
 508         unsigned                val = 0;
 509         unsigned                i;
 510
 511         for (i = 0; i < RES_NAME_SIZE; i++)
 512                 val += id->name[i];
 513         return val & mask;
 514 }
 515
 516 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 517                                       const void *key, unsigned mask)
 518 {
 519         const struct ldlm_res_id *id = key;
 520         struct lu_fid       fid;
 521         __u32               hash;
 522         __u32               val;
 523
 524         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 525         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 526         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 527
 528         hash = fid_flatten32(&fid);
 529         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 530         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 531                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 532                 hash += (val >> 5) + (val << 11);
 533         } else {
 534                 val = fid_oid(&fid);
 535         }
 536         hash = hash_long(hash, hs->hs_bkt_bits);
 537         /* give me another random factor */
 538         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 539
 540         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 541         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 542
 543         return hash & mask;
 544 }
 545
 546 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 547 {
 548         struct ldlm_resource   *res;
 549
 550         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 551         return &res->lr_name;
 552 }
 553
 554 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 555 {
 556         struct ldlm_resource   *res;
 557
 558         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 559         return ldlm_res_eq((const struct ldlm_res_id *)key,
 560                            (const struct ldlm_res_id *)&res->lr_name);
 561 }
 562
 563 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 564 {
 565         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 566 }
 567
 568 static void
 569 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 570 {
 571         struct ldlm_resource *res;
 572
 573         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 574         ldlm_resource_getref(res);
 575 }
 576
 577 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 578 {
 579         struct ldlm_resource *res;
 580
 581         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 582         ldlm_resource_putref(res);
 583 }
 584
 585 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 586         .hs_hash        = ldlm_res_hop_hash,
 587         .hs_key         = ldlm_res_hop_key,
 588         .hs_keycmp      = ldlm_res_hop_keycmp,
 589         .hs_keycpy      = NULL,
 590         .hs_object      = ldlm_res_hop_object,
 591         .hs_get         = ldlm_res_hop_get_locked,
 592         .hs_put         = ldlm_res_hop_put
 593 };
 594
 595 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 596         .hs_hash        = ldlm_res_hop_fid_hash,
 597         .hs_key         = ldlm_res_hop_key,
 598         .hs_keycmp      = ldlm_res_hop_keycmp,
 599         .hs_keycpy      = NULL,
 600         .hs_object      = ldlm_res_hop_object,
 601         .hs_get         = ldlm_res_hop_get_locked,
 602         .hs_put         = ldlm_res_hop_put
 603 };
 604
 605 typedef struct ldlm_ns_hash_def {
 606         enum ldlm_ns_type       nsd_type;
 607         /** hash bucket bits */
 608         unsigned                nsd_bkt_bits;
 609         /** hash bits */
 610         unsigned                nsd_all_bits;
 611         /** hash operations */
 612         struct cfs_hash_ops *nsd_hops;
 613 } ldlm_ns_hash_def_t;
 614
 615 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 616 {
 617         {
 618                 .nsd_type       = LDLM_NS_TYPE_MDC,
 619                 .nsd_bkt_bits   = 11,
 620                 .nsd_all_bits   = 16,
 621                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 622         },
 623         {
 624                 .nsd_type       = LDLM_NS_TYPE_MDT,
 625                 .nsd_bkt_bits   = 14,
 626                 .nsd_all_bits   = 21,
 627                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 628         },
 629         {
 630                 .nsd_type       = LDLM_NS_TYPE_OSC,
 631                 .nsd_bkt_bits   = 8,
 632                 .nsd_all_bits   = 12,
 633                 .nsd_hops       = &ldlm_ns_hash_ops,
 634         },
 635         {
 636                 .nsd_type       = LDLM_NS_TYPE_OST,
 637                 .nsd_bkt_bits   = 11,
 638                 .nsd_all_bits   = 17,
 639                 .nsd_hops       = &ldlm_ns_hash_ops,
 640         },
 641         {
 642                 .nsd_type       = LDLM_NS_TYPE_MGC,
 643                 .nsd_bkt_bits   = 4,
 644                 .nsd_all_bits   = 4,
 645                 .nsd_hops       = &ldlm_ns_hash_ops,
 646         },
 647         {
 648                 .nsd_type       = LDLM_NS_TYPE_MGT,
 649                 .nsd_bkt_bits   = 4,
 650                 .nsd_all_bits   = 4,
 651                 .nsd_hops       = &ldlm_ns_hash_ops,
 652         },
 653         {
 654                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 655         },
 656 };
 657
 658 /**
 659  * Create and initialize new empty namespace.
 660  */
 661 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 662                                           enum ldlm_side client,
 663                                           enum ldlm_appetite apt,
 664                                           enum ldlm_ns_type ns_type)
 665 {
 666         struct ldlm_namespace *ns = NULL;
 667         struct ldlm_ns_bucket *nsb;
 668         struct ldlm_ns_hash_def *nsd;
 669         struct cfs_hash_bd bd;
 670         int idx;
 671         int rc;
 672         ENTRY;
 673
 674         LASSERT(obd != NULL);
 675
 676         rc = ldlm_get_ref();
 677         if (rc) {
 678                 CERROR("ldlm_get_ref failed: %d\n", rc);
 679                 RETURN(NULL);
 680         }
 681
 682         for (idx = 0;;idx++) {
 683                 nsd = &ldlm_ns_hash_defs[idx];
 684                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 685                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 686                         GOTO(out_ref, NULL);
 687                 }
 688
 689                 if (nsd->nsd_type == ns_type)
 690                         break;
 691         }
 692
 693         OBD_ALLOC_PTR(ns);
 694         if (!ns)
 695                 GOTO(out_ref, NULL);
 696
 697         ns->ns_rs_hash = cfs_hash_create(name,
 698                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 699                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 700                                          CFS_HASH_MIN_THETA,
 701                                          CFS_HASH_MAX_THETA,
 702                                          nsd->nsd_hops,
 703                                          CFS_HASH_DEPTH |
 704                                          CFS_HASH_BIGNAME |
 705                                          CFS_HASH_SPIN_BKTLOCK |
 706                                          CFS_HASH_NO_ITEMREF);
 707         if (ns->ns_rs_hash == NULL)
 708                 GOTO(out_ns, NULL);
 709
 710         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 711                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 712                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 713                 nsb->nsb_namespace = ns;
 714                 nsb->nsb_reclaim_start = 0;
 715         }
 716
 717         ns->ns_obd      = obd;
 718         ns->ns_appetite = apt;
 719         ns->ns_client   = client;
 720
 721         INIT_LIST_HEAD(&ns->ns_list_chain);
 722         INIT_LIST_HEAD(&ns->ns_unused_list);
 723         spin_lock_init(&ns->ns_lock);
 724         atomic_set(&ns->ns_bref, 0);
 725         init_waitqueue_head(&ns->ns_waitq);
 726
 727         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 728         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 729         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 730
 731         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 732         ns->ns_nr_unused          = 0;
 733         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 734         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 735         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 736         ns->ns_timeouts           = 0;
 737         ns->ns_orig_connect_flags = 0;
 738         ns->ns_connect_flags      = 0;
 739         ns->ns_stopping           = 0;
 740         ns->ns_reclaim_start      = 0;
 741         rc = ldlm_namespace_proc_register(ns);
 742         if (rc != 0) {
 743                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 744                 GOTO(out_hash, rc);
 745         }
 746
 747         idx = ldlm_namespace_nr_read(client);
 748         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 749         if (rc) {
 750                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 751                 GOTO(out_proc, rc);
 752         }
 753
 754         ldlm_namespace_register(ns, client);
 755         RETURN(ns);
 756 out_proc:
 757         ldlm_namespace_proc_unregister(ns);
 758         ldlm_namespace_cleanup(ns, 0);
 759 out_hash:
 760         cfs_hash_putref(ns->ns_rs_hash);
 761 out_ns:
 762         OBD_FREE_PTR(ns);
 763 out_ref:
 764         ldlm_put_ref();
 765         RETURN(NULL);
 766 }
 767 EXPORT_SYMBOL(ldlm_namespace_new);
 768
 769 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 770
 771 /**
 772  * Cancel and destroy all locks on a resource.
 773  *
 774  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 775  * clean up.  This is currently only used for recovery, and we make
 776  * certain assumptions as a result--notably, that we shouldn't cancel
 777  * locks with refs.
 778  */
 779 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 780                              __u64 flags)
 781 {
 782         struct list_head *tmp;
 783         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 784         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 785
 786         do {
 787                 struct ldlm_lock *lock = NULL;
 788
 789                 /* First, we look for non-cleaned-yet lock
 790                  * all cleaned locks are marked by CLEANED flag. */
 791                 lock_res(res);
 792                 list_for_each(tmp, q) {
 793                         lock = list_entry(tmp, struct ldlm_lock,
 794                                           l_res_link);
 795                         if (ldlm_is_cleaned(lock)) {
 796                                 lock = NULL;
 797                                 continue;
 798                         }
 799                         LDLM_LOCK_GET(lock);
 800                         ldlm_set_cleaned(lock);
 801                         break;
 802                 }
 803
 804                 if (lock == NULL) {
 805                         unlock_res(res);
 806                         break;
 807                 }
 808
 809                 /* Set CBPENDING so nothing in the cancellation path
 810                  * can match this lock. */
 811                 ldlm_set_cbpending(lock);
 812                 ldlm_set_failed(lock);
 813                 lock->l_flags |= flags;
 814
 815                 /* ... without sending a CANCEL message for local_only. */
 816                 if (local_only)
 817                         ldlm_set_local_only(lock);
 818
 819                 if (local_only && (lock->l_readers || lock->l_writers)) {
 820                         /* This is a little bit gross, but much better than the
 821                          * alternative: pretend that we got a blocking AST from
 822                          * the server, so that when the lock is decref'd, it
 823                          * will go away ... */
 824                         unlock_res(res);
 825                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 826                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 827                                 set_current_state(TASK_UNINTERRUPTIBLE);
 828                                 schedule_timeout(cfs_time_seconds(4));
 829                                 set_current_state(TASK_RUNNING);
 830                         }
 831                         if (lock->l_completion_ast)
 832                                 lock->l_completion_ast(lock,
 833                                                        LDLM_FL_FAILED, NULL);
 834                         LDLM_LOCK_RELEASE(lock);
 835                         continue;
 836                 }
 837
 838                 if (client) {
 839                         struct lustre_handle lockh;
 840
 841                         unlock_res(res);
 842                         ldlm_lock2handle(lock, &lockh);
 843                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
 844                         if (rc)
 845                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 846                 } else {
 847                         unlock_res(res);
 848                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 849                                    "client node");
 850                         ldlm_lock_cancel(lock);
 851                 }
 852                 LDLM_LOCK_RELEASE(lock);
 853         } while (1);
 854 }
 855
 856 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 857                                struct hlist_node *hnode, void *arg)
 858 {
 859         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 860         __u64 flags = *(__u64 *)arg;
 861
 862         cleanup_resource(res, &res->lr_granted, flags);
 863         cleanup_resource(res, &res->lr_converting, flags);
 864         cleanup_resource(res, &res->lr_waiting, flags);
 865
 866         return 0;
 867 }
 868
 869 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 870                                   struct hlist_node *hnode, void *arg)
 871 {
 872         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 873
 874         lock_res(res);
 875         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 876                "(%d) after lock cleanup; forcing cleanup.\n",
 877                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 878                atomic_read(&res->lr_refcount) - 1);
 879
 880         ldlm_resource_dump(D_ERROR, res);
 881         unlock_res(res);
 882         return 0;
 883 }
 884
 885 /**
 886  * Cancel and destroy all locks in the namespace.
 887  *
 888  * Typically used during evictions when server notified client that it was
 889  * evicted and all of its state needs to be destroyed.
 890  * Also used during shutdown.
 891  */
 892 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 893 {
 894         if (ns == NULL) {
 895                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 896                 return ELDLM_OK;
 897         }
 898
 899         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 900                                  &flags, 0);
 901         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
 902                                  NULL, 0);
 903         return ELDLM_OK;
 904 }
 905 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 906
 907 /**
 908  * Attempts to free namespace.
 909  *
 910  * Only used when namespace goes away, like during an unmount.
 911  */
 912 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 913 {
 914         ENTRY;
 915
 916         /* At shutdown time, don't call the cancellation callback */
 917         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 918
 919         if (atomic_read(&ns->ns_bref) > 0) {
 920                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 921                 int rc;
 922                 CDEBUG(D_DLMTRACE,
 923                        "dlm namespace %s free waiting on refcount %d\n",
 924                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
 925 force_wait:
 926                 if (force)
 927                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
 928                                           MSEC_PER_SEC) / 4, NULL, NULL);
 929
 930                 rc = l_wait_event(ns->ns_waitq,
 931                                   atomic_read(&ns->ns_bref) == 0, &lwi);
 932
 933                 /* Forced cleanups should be able to reclaim all references,
 934                  * so it's safe to wait forever... we can't leak locks... */
 935                 if (force && rc == -ETIMEDOUT) {
 936                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 937                                        "namespace with %d resources in use, "
 938                                        "(rc=%d)\n", ldlm_ns_name(ns),
 939                                        atomic_read(&ns->ns_bref), rc);
 940                         GOTO(force_wait, rc);
 941                 }
 942
 943                 if (atomic_read(&ns->ns_bref)) {
 944                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 945                                        "with %d resources in use, (rc=%d)\n",
 946                                        ldlm_ns_name(ns),
 947                                        atomic_read(&ns->ns_bref), rc);
 948                         RETURN(ELDLM_NAMESPACE_EXISTS);
 949                 }
 950                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 951                        ldlm_ns_name(ns));
 952         }
 953
 954         RETURN(ELDLM_OK);
 955 }
 956
 957 /**
 958  * Performs various cleanups for passed \a ns to make it drop refc and be
 959  * ready for freeing. Waits for refc == 0.
 960  *
 961  * The following is done:
 962  * (0) Unregister \a ns from its list to make inaccessible for potential
 963  * users like pools thread and others;
 964  * (1) Clear all locks in \a ns.
 965  */
 966 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 967                                struct obd_import *imp,
 968                                int force)
 969 {
 970         int rc;
 971         ENTRY;
 972         if (!ns) {
 973                 EXIT;
 974                 return;
 975         }
 976
 977         spin_lock(&ns->ns_lock);
 978         ns->ns_stopping = 1;
 979         spin_unlock(&ns->ns_lock);
 980
 981         /*
 982          * Can fail with -EINTR when force == 0 in which case try harder.
 983          */
 984         rc = __ldlm_namespace_free(ns, force);
 985         if (rc != ELDLM_OK) {
 986                 if (imp) {
 987                         ptlrpc_disconnect_import(imp, 0);
 988                         ptlrpc_invalidate_import(imp);
 989                 }
 990
 991                 /*
 992                  * With all requests dropped and the import inactive
 993                  * we are gaurenteed all reference will be dropped.
 994                  */
 995                 rc = __ldlm_namespace_free(ns, 1);
 996                 LASSERT(rc == 0);
 997         }
 998         EXIT;
 999 }
1000 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1001
1002 /**
1003  * Performs freeing memory structures related to \a ns. This is only done
1004  * when ldlm_namespce_free_prior() successfully removed all resources
1005  * referencing \a ns and its refc == 0.
1006  */
1007 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1008 {
1009         ENTRY;
1010         if (!ns) {
1011                 EXIT;
1012                 return;
1013         }
1014
1015         /* Make sure that nobody can find this ns in its list. */
1016         ldlm_namespace_unregister(ns, ns->ns_client);
1017         /* Fini pool _before_ parent proc dir is removed. This is important as
1018          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1019          * Removing it after @dir may cause oops. */
1020         ldlm_pool_fini(&ns->ns_pool);
1021
1022         ldlm_namespace_proc_unregister(ns);
1023         cfs_hash_putref(ns->ns_rs_hash);
1024         /* Namespace \a ns should be not on list at this time, otherwise
1025          * this will cause issues related to using freed \a ns in poold
1026          * thread. */
1027         LASSERT(list_empty(&ns->ns_list_chain));
1028         OBD_FREE_PTR(ns);
1029         ldlm_put_ref();
1030         EXIT;
1031 }
1032 EXPORT_SYMBOL(ldlm_namespace_free_post);
1033
1034 /**
1035  * Cleanup the resource, and free namespace.
1036  * bug 12864:
1037  * Deadlock issue:
1038  * proc1: destroy import
1039  *        class_disconnect_export(grab cl_sem) ->
1040  *              -> ldlm_namespace_free ->
1041  *              -> lprocfs_remove(grab _lprocfs_lock).
1042  * proc2: read proc info
1043  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1044  *              -> osc_rd_active, etc(grab cl_sem).
1045  *
1046  * So that I have to split the ldlm_namespace_free into two parts - the first
1047  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1048  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1049  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1050  * held.
1051  */
1052 void ldlm_namespace_free(struct ldlm_namespace *ns,
1053                          struct obd_import *imp,
1054                          int force)
1055 {
1056         ldlm_namespace_free_prior(ns, imp, force);
1057         ldlm_namespace_free_post(ns);
1058 }
1059 EXPORT_SYMBOL(ldlm_namespace_free);
1060
1061 void ldlm_namespace_get(struct ldlm_namespace *ns)
1062 {
1063         atomic_inc(&ns->ns_bref);
1064 }
1065
1066 /* This is only for callers that care about refcount */
1067 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1068 {
1069         return atomic_inc_return(&ns->ns_bref);
1070 }
1071
1072 void ldlm_namespace_put(struct ldlm_namespace *ns)
1073 {
1074         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1075                 wake_up(&ns->ns_waitq);
1076                 spin_unlock(&ns->ns_lock);
1077         }
1078 }
1079
1080 /** Register \a ns in the list of namespaces */
1081 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1082 {
1083         mutex_lock(ldlm_namespace_lock(client));
1084         LASSERT(list_empty(&ns->ns_list_chain));
1085         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1086         ldlm_namespace_nr_inc(client);
1087         mutex_unlock(ldlm_namespace_lock(client));
1088 }
1089
1090 /** Unregister \a ns from the list of namespaces. */
1091 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1092 {
1093         mutex_lock(ldlm_namespace_lock(client));
1094         LASSERT(!list_empty(&ns->ns_list_chain));
1095         /* Some asserts and possibly other parts of the code are still
1096          * using list_empty(&ns->ns_list_chain). This is why it is
1097          * important to use list_del_init() here. */
1098         list_del_init(&ns->ns_list_chain);
1099         ldlm_namespace_nr_dec(client);
1100         mutex_unlock(ldlm_namespace_lock(client));
1101 }
1102
1103 /** Should be called with ldlm_namespace_lock(client) taken. */
1104 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1105                                           enum ldlm_side client)
1106 {
1107         LASSERT(!list_empty(&ns->ns_list_chain));
1108         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1109         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1110 }
1111
1112 /** Should be called with ldlm_namespace_lock(client) taken. */
1113 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1114                                             enum ldlm_side client)
1115 {
1116         LASSERT(!list_empty(&ns->ns_list_chain));
1117         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1118         list_move_tail(&ns->ns_list_chain,
1119                        ldlm_namespace_inactive_list(client));
1120 }
1121
1122 /** Should be called with ldlm_namespace_lock(client) taken. */
1123 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1124 {
1125         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1126         LASSERT(!list_empty(ldlm_namespace_list(client)));
1127         return container_of(ldlm_namespace_list(client)->next,
1128                             struct ldlm_namespace, ns_list_chain);
1129 }
1130
1131 /** Create and initialize new resource. */
1132 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1133 {
1134         struct ldlm_resource *res;
1135         int idx;
1136
1137         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1138         if (res == NULL)
1139                 return NULL;
1140
1141         if (ldlm_type == LDLM_EXTENT) {
1142                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1143                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1144                 if (res->lr_itree == NULL) {
1145                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1146                         return NULL;
1147                 }
1148                 /* Initialize interval trees for each lock mode. */
1149                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1150                         res->lr_itree[idx].lit_size = 0;
1151                         res->lr_itree[idx].lit_mode = 1 << idx;
1152                         res->lr_itree[idx].lit_root = NULL;
1153                 }
1154         }
1155
1156         INIT_LIST_HEAD(&res->lr_granted);
1157         INIT_LIST_HEAD(&res->lr_converting);
1158         INIT_LIST_HEAD(&res->lr_waiting);
1159
1160         atomic_set(&res->lr_refcount, 1);
1161         spin_lock_init(&res->lr_lock);
1162         lu_ref_init(&res->lr_reference);
1163
1164         /* Since LVB init can be delayed now, there is no longer need to
1165          * immediatelly acquire mutex here. */
1166         mutex_init(&res->lr_lvb_mutex);
1167         res->lr_lvb_initialized = false;
1168
1169         return res;
1170 }
1171
1172 /**
1173  * Return a reference to resource with given name, creating it if necessary.
1174  * Args: namespace with ns_lock unlocked
1175  * Locks: takes and releases NS hash-lock and res->lr_lock
1176  * Returns: referenced, unlocked ldlm_resource or NULL
1177  */
1178 struct ldlm_resource *
1179 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1180                   const struct ldlm_res_id *name, enum ldlm_type type,
1181                   int create)
1182 {
1183         struct hlist_node       *hnode;
1184         struct ldlm_resource    *res = NULL;
1185         struct cfs_hash_bd              bd;
1186         __u64                   version;
1187         int                     ns_refcount = 0;
1188
1189         LASSERT(ns != NULL);
1190         LASSERT(parent == NULL);
1191         LASSERT(ns->ns_rs_hash != NULL);
1192         LASSERT(name->name[0] != 0);
1193
1194         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1195         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1196         if (hnode != NULL) {
1197                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1198                 GOTO(found, res);
1199         }
1200
1201         version = cfs_hash_bd_version_get(&bd);
1202         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1203
1204         if (create == 0)
1205                 return ERR_PTR(-ENOENT);
1206
1207         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1208                  "type: %d\n", type);
1209         res = ldlm_resource_new(type);
1210         if (res == NULL)
1211                 return ERR_PTR(-ENOMEM);
1212
1213         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1214         res->lr_name = *name;
1215         res->lr_type = type;
1216
1217         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1218         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1219                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1220
1221         if (hnode != NULL) {
1222                 /* Someone won the race and already added the resource. */
1223                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1224                 /* Clean lu_ref for failed resource. */
1225                 lu_ref_fini(&res->lr_reference);
1226                 if (res->lr_itree != NULL)
1227                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1228                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1229                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1230 found:
1231                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1232                 return res;
1233         }
1234         /* We won! Let's add the resource. */
1235         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1236         if (cfs_hash_bd_count_get(&bd) == 1)
1237                 ns_refcount = ldlm_namespace_get_return(ns);
1238
1239         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1240
1241         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1242
1243         /* Let's see if we happened to be the very first resource in this
1244          * namespace. If so, and this is a client namespace, we need to move
1245          * the namespace into the active namespaces list to be patrolled by
1246          * the ldlm_poold. */
1247         if (ns_is_client(ns) && ns_refcount == 1) {
1248                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1249                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1250                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1251         }
1252
1253         return res;
1254 }
1255 EXPORT_SYMBOL(ldlm_resource_get);
1256
1257 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1258 {
1259         LASSERT(res != NULL);
1260         LASSERT(res != LP_POISON);
1261         atomic_inc(&res->lr_refcount);
1262         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1263                atomic_read(&res->lr_refcount));
1264         return res;
1265 }
1266
1267 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1268                                          struct ldlm_resource *res)
1269 {
1270         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1271
1272         if (!list_empty(&res->lr_granted)) {
1273                 ldlm_resource_dump(D_ERROR, res);
1274                 LBUG();
1275         }
1276
1277         if (!list_empty(&res->lr_converting)) {
1278                 ldlm_resource_dump(D_ERROR, res);
1279                 LBUG();
1280         }
1281
1282         if (!list_empty(&res->lr_waiting)) {
1283                 ldlm_resource_dump(D_ERROR, res);
1284                 LBUG();
1285         }
1286
1287         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1288                                bd, &res->lr_hash);
1289         lu_ref_fini(&res->lr_reference);
1290         if (cfs_hash_bd_count_get(bd) == 0)
1291                 ldlm_namespace_put(nsb->nsb_namespace);
1292 }
1293
1294 /* Returns 1 if the resource was freed, 0 if it remains. */
1295 int ldlm_resource_putref(struct ldlm_resource *res)
1296 {
1297         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1298         struct cfs_hash_bd   bd;
1299
1300         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1301         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1302                res, atomic_read(&res->lr_refcount) - 1);
1303
1304         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1305         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1306                 __ldlm_resource_putref_final(&bd, res);
1307                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1308                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1309                         ns->ns_lvbo->lvbo_free(res);
1310                 if (res->lr_itree != NULL)
1311                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1312                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1313                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1314                 return 1;
1315         }
1316         return 0;
1317 }
1318 EXPORT_SYMBOL(ldlm_resource_putref);
1319
1320 /**
1321  * Add a lock into a given resource into specified lock list.
1322  */
1323 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1324                             struct ldlm_lock *lock)
1325 {
1326         check_res_locked(res);
1327
1328         LDLM_DEBUG(lock, "About to add this lock");
1329
1330         if (ldlm_is_destroyed(lock)) {
1331                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1332                 return;
1333         }
1334
1335         LASSERT(list_empty(&lock->l_res_link));
1336
1337         list_add_tail(&lock->l_res_link, head);
1338 }
1339
1340 /**
1341  * Insert a lock into resource after specified lock.
1342  *
1343  * Obtain resource description from the lock we are inserting after.
1344  */
1345 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1346                                      struct ldlm_lock *new)
1347 {
1348         struct ldlm_resource *res = original->l_resource;
1349
1350         check_res_locked(res);
1351
1352         ldlm_resource_dump(D_INFO, res);
1353         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1354
1355         if (ldlm_is_destroyed(new)) {
1356                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1357                 goto out;
1358         }
1359
1360         LASSERT(list_empty(&new->l_res_link));
1361
1362         list_add(&new->l_res_link, &original->l_res_link);
1363  out:;
1364 }
1365
1366 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1367 {
1368         int type = lock->l_resource->lr_type;
1369
1370         check_res_locked(lock->l_resource);
1371         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1372                 ldlm_unlink_lock_skiplist(lock);
1373         else if (type == LDLM_EXTENT)
1374                 ldlm_extent_unlink_lock(lock);
1375         list_del_init(&lock->l_res_link);
1376 }
1377 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1378
1379 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1380 {
1381         desc->lr_type = res->lr_type;
1382         desc->lr_name = res->lr_name;
1383 }
1384
1385 /**
1386  * Print information about all locks in all namespaces on this node to debug
1387  * log.
1388  */
1389 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1390 {
1391         struct list_head *tmp;
1392
1393         if (!((libcfs_debug | D_ERROR) & level))
1394                 return;
1395
1396         mutex_lock(ldlm_namespace_lock(client));
1397
1398         list_for_each(tmp, ldlm_namespace_list(client)) {
1399                 struct ldlm_namespace *ns;
1400
1401                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1402                 ldlm_namespace_dump(level, ns);
1403         }
1404
1405         mutex_unlock(ldlm_namespace_lock(client));
1406 }
1407
1408 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1409                               struct hlist_node *hnode, void *arg)
1410 {
1411         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1412         int    level = (int)(unsigned long)arg;
1413
1414         lock_res(res);
1415         ldlm_resource_dump(level, res);
1416         unlock_res(res);
1417
1418         return 0;
1419 }
1420
1421 /**
1422  * Print information about all locks in this namespace on this node to debug
1423  * log.
1424  */
1425 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1426 {
1427         if (!((libcfs_debug | D_ERROR) & level))
1428                 return;
1429
1430         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1431                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1432                ns_is_client(ns) ? "client" : "server");
1433
1434         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1435                 return;
1436
1437         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1438                                  ldlm_res_hash_dump,
1439                                  (void *)(unsigned long)level, 0);
1440         spin_lock(&ns->ns_lock);
1441         ns->ns_next_dump = cfs_time_shift(10);
1442         spin_unlock(&ns->ns_lock);
1443 }
1444
1445 /**
1446  * Print information about all locks in this resource to debug log.
1447  */
1448 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1449 {
1450         struct ldlm_lock *lock;
1451         unsigned int granted = 0;
1452
1453         CLASSERT(RES_NAME_SIZE == 4);
1454
1455         if (!((libcfs_debug | D_ERROR) & level))
1456                 return;
1457
1458         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1459                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1460
1461         if (!list_empty(&res->lr_granted)) {
1462                 CDEBUG(level, "Granted locks (in reverse order):\n");
1463                 list_for_each_entry_reverse(lock, &res->lr_granted,
1464                                                 l_res_link) {
1465                         LDLM_DEBUG_LIMIT(level, lock, "###");
1466                         if (!(level & D_CANTMASK) &&
1467                             ++granted > ldlm_dump_granted_max) {
1468                                 CDEBUG(level, "only dump %d granted locks to "
1469                                        "avoid DDOS.\n", granted);
1470                                 break;
1471                         }
1472                 }
1473         }
1474         if (!list_empty(&res->lr_converting)) {
1475                 CDEBUG(level, "Converting locks:\n");
1476                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1477                         LDLM_DEBUG_LIMIT(level, lock, "###");
1478         }
1479         if (!list_empty(&res->lr_waiting)) {
1480                 CDEBUG(level, "Waiting locks:\n");
1481                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1482                         LDLM_DEBUG_LIMIT(level, lock, "###");
1483         }
1484 }
1485 EXPORT_SYMBOL(ldlm_resource_dump);