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) 2011, 2012, Whamcloud, Inc.
  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 #ifdef __KERNEL__
  44 # include <lustre_dlm.h>
  45 #else
  46 # include <liblustre.h>
  47 #endif
  48
  49 #include <lustre_fid.h>
  50 #include <obd_class.h>
  51 #include "ldlm_internal.h"
  52
  53 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  54
  55 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
  56 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
  57
  58 cfs_mutex_t ldlm_srv_namespace_lock;
  59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  60
  61 cfs_mutex_t ldlm_cli_namespace_lock;
  62 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  63
  64 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  65 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  66 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  67
  68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  69
  70 /* during debug dump certain amount of granted locks for one resource to avoid
  71  * DDOS. */
  72 unsigned int ldlm_dump_granted_max = 256;
  73
  74 #ifdef LPROCFS
  75 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  76                              unsigned long count, void *data)
  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
  83 int ldlm_proc_setup(void)
  84 {
  85         int rc;
  86         struct lprocfs_vars list[] = {
  87                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  88                 { "dump_granted_max",
  89                   lprocfs_rd_uint, lprocfs_wr_uint,
  90                   &ldlm_dump_granted_max, NULL },
  91                 { "cancel_unused_locks_before_replay",
  92                   lprocfs_rd_uint, lprocfs_wr_uint,
  93                   &ldlm_cancel_unused_locks_before_replay, NULL },
  94                 { NULL }};
  95         ENTRY;
  96         LASSERT(ldlm_ns_proc_dir == NULL);
  97
  98         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  99                                               proc_lustre_root,
 100                                               NULL, NULL);
 101         if (IS_ERR(ldlm_type_proc_dir)) {
 102                 CERROR("LProcFS failed in ldlm-init\n");
 103                 rc = PTR_ERR(ldlm_type_proc_dir);
 104                 GOTO(err, rc);
 105         }
 106
 107         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 108                                             ldlm_type_proc_dir,
 109                                             NULL, NULL);
 110         if (IS_ERR(ldlm_ns_proc_dir)) {
 111                 CERROR("LProcFS failed in ldlm-init\n");
 112                 rc = PTR_ERR(ldlm_ns_proc_dir);
 113                 GOTO(err_type, rc);
 114         }
 115
 116         ldlm_svc_proc_dir = lprocfs_register("services",
 117                                             ldlm_type_proc_dir,
 118                                             NULL, NULL);
 119         if (IS_ERR(ldlm_svc_proc_dir)) {
 120                 CERROR("LProcFS failed in ldlm-init\n");
 121                 rc = PTR_ERR(ldlm_svc_proc_dir);
 122                 GOTO(err_ns, rc);
 123         }
 124
 125         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 126
 127         RETURN(0);
 128
 129 err_ns:
 130         lprocfs_remove(&ldlm_ns_proc_dir);
 131 err_type:
 132         lprocfs_remove(&ldlm_type_proc_dir);
 133 err:
 134         ldlm_svc_proc_dir = NULL;
 135         RETURN(rc);
 136 }
 137
 138 void ldlm_proc_cleanup(void)
 139 {
 140         if (ldlm_svc_proc_dir)
 141                 lprocfs_remove(&ldlm_svc_proc_dir);
 142
 143         if (ldlm_ns_proc_dir)
 144                 lprocfs_remove(&ldlm_ns_proc_dir);
 145
 146         if (ldlm_type_proc_dir)
 147                 lprocfs_remove(&ldlm_type_proc_dir);
 148 }
 149
 150 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 151                                    int count, int *eof, void *data)
 152 {
 153         struct ldlm_namespace *ns  = data;
 154         __u64                  res = 0;
 155         cfs_hash_bd_t          bd;
 156         int                    i;
 157
 158         /* result is not strictly consistant */
 159         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 160                 res += cfs_hash_bd_count_get(&bd);
 161         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 162 }
 163
 164 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 165                                int count, int *eof, void *data)
 166 {
 167         struct ldlm_namespace *ns = data;
 168         __u64                  locks;
 169
 170         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 171                                         LPROCFS_FIELDS_FLAGS_SUM);
 172         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 173 }
 174
 175 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 176                                int count, int *eof, void *data)
 177 {
 178         struct ldlm_namespace *ns = data;
 179         __u32 *nr = &ns->ns_max_unused;
 180
 181         if (ns_connect_lru_resize(ns))
 182                 nr = &ns->ns_nr_unused;
 183         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 184 }
 185
 186 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 187                                unsigned long count, void *data)
 188 {
 189         struct ldlm_namespace *ns = data;
 190         char dummy[MAX_STRING_SIZE + 1], *end;
 191         unsigned long tmp;
 192         int lru_resize;
 193
 194         dummy[MAX_STRING_SIZE] = '\0';
 195         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 196                 return -EFAULT;
 197
 198         if (strncmp(dummy, "clear", 5) == 0) {
 199                 CDEBUG(D_DLMTRACE,
 200                        "dropping all unused locks from namespace %s\n",
 201                        ldlm_ns_name(ns));
 202                 if (ns_connect_lru_resize(ns)) {
 203                         int canceled, unused  = ns->ns_nr_unused;
 204
 205                         /* Try to cancel all @ns_nr_unused locks. */
 206                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 207                                                    LDLM_CANCEL_PASSED);
 208                         if (canceled < unused) {
 209                                 CDEBUG(D_DLMTRACE,
 210                                        "not all requested locks are canceled, "
 211                                        "requested: %d, canceled: %d\n", unused,
 212                                        canceled);
 213                                 return -EINVAL;
 214                         }
 215                 } else {
 216                         tmp = ns->ns_max_unused;
 217                         ns->ns_max_unused = 0;
 218                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 219                         ns->ns_max_unused = tmp;
 220                 }
 221                 return count;
 222         }
 223
 224         tmp = simple_strtoul(dummy, &end, 0);
 225         if (dummy == end) {
 226                 CERROR("invalid value written\n");
 227                 return -EINVAL;
 228         }
 229         lru_resize = (tmp == 0);
 230
 231         if (ns_connect_lru_resize(ns)) {
 232                 if (!lru_resize)
 233                         ns->ns_max_unused = (unsigned int)tmp;
 234
 235                 if (tmp > ns->ns_nr_unused)
 236                         tmp = ns->ns_nr_unused;
 237                 tmp = ns->ns_nr_unused - tmp;
 238
 239                 CDEBUG(D_DLMTRACE,
 240                        "changing namespace %s unused locks from %u to %u\n",
 241                        ldlm_ns_name(ns), ns->ns_nr_unused,
 242                        (unsigned int)tmp);
 243                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 244
 245                 if (!lru_resize) {
 246                         CDEBUG(D_DLMTRACE,
 247                                "disable lru_resize for namespace %s\n",
 248                                ldlm_ns_name(ns));
 249                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 250                 }
 251         } else {
 252                 CDEBUG(D_DLMTRACE,
 253                        "changing namespace %s max_unused from %u to %u\n",
 254                        ldlm_ns_name(ns), ns->ns_max_unused,
 255                        (unsigned int)tmp);
 256                 ns->ns_max_unused = (unsigned int)tmp;
 257                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 258
 259                 /* Make sure that originally lru resize was supported before
 260                  * turning it on here. */
 261                 if (lru_resize &&
 262                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 263                         CDEBUG(D_DLMTRACE,
 264                                "enable lru_resize for namespace %s\n",
 265                                ldlm_ns_name(ns));
 266                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 267                 }
 268         }
 269
 270         return count;
 271 }
 272
 273 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 274 {
 275         struct proc_dir_entry *dir;
 276
 277         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 278         if (dir == NULL) {
 279                 CERROR("dlm namespace %s has no procfs dir?\n",
 280                        ldlm_ns_name(ns));
 281         } else {
 282                 lprocfs_remove(&dir);
 283         }
 284
 285         if (ns->ns_stats != NULL)
 286                 lprocfs_free_stats(&ns->ns_stats);
 287 }
 288
 289 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 290 {
 291         struct lprocfs_vars lock_vars[2];
 292         char lock_name[MAX_STRING_SIZE + 1];
 293
 294         LASSERT(ns != NULL);
 295         LASSERT(ns->ns_rs_hash != NULL);
 296
 297         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 298         if (ns->ns_stats == NULL)
 299                 return -ENOMEM;
 300
 301         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 302                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 303
 304         lock_name[MAX_STRING_SIZE] = '\0';
 305
 306         memset(lock_vars, 0, sizeof(lock_vars));
 307         lock_vars[0].name = lock_name;
 308
 309         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 310                  ldlm_ns_name(ns));
 311         lock_vars[0].data = ns;
 312         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 313         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 314
 315         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 316                  ldlm_ns_name(ns));
 317         lock_vars[0].data = ns;
 318         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 319         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 320
 321         if (ns_is_client(ns)) {
 322                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 323                          ldlm_ns_name(ns));
 324                 lock_vars[0].data = &ns->ns_nr_unused;
 325                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 326                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 327
 328                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 329                          ldlm_ns_name(ns));
 330                 lock_vars[0].data = ns;
 331                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 332                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 333                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 334
 335                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 336                          ldlm_ns_name(ns));
 337                 lock_vars[0].data = &ns->ns_max_age;
 338                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 339                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 340                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 341         } else {
 342                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 343                          ldlm_ns_name(ns));
 344                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 345                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 346                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 347                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 348
 349                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 350                          ldlm_ns_name(ns));
 351                 lock_vars[0].data = &ns->ns_timeouts;
 352                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 354
 355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 356                          ldlm_ns_name(ns));
 357                 lock_vars[0].data = &ns->ns_max_nolock_size;
 358                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 359                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 360                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 361
 362                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 363                          ldlm_ns_name(ns));
 364                 lock_vars[0].data = &ns->ns_contention_time;
 365                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 366                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 367                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 368
 369                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 370                          ldlm_ns_name(ns));
 371                 lock_vars[0].data = &ns->ns_contended_locks;
 372                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 373                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 374                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 375
 376                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 377                          ldlm_ns_name(ns));
 378                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 379                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 380                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 381                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 382         }
 383         return 0;
 384 }
 385 #undef MAX_STRING_SIZE
 386 #else /* LPROCFS */
 387
 388 #define ldlm_namespace_proc_unregister(ns)      ({;})
 389 #define ldlm_namespace_proc_register(ns)        ({0;})
 390
 391 #endif /* LPROCFS */
 392
 393 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 394                                   const void *key, unsigned mask)
 395 {
 396         const struct ldlm_res_id     *id  = key;
 397         unsigned                val = 0;
 398         unsigned                i;
 399
 400         for (i = 0; i < RES_NAME_SIZE; i++)
 401                 val += id->name[i];
 402         return val & mask;
 403 }
 404
 405 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 406                                       const void *key, unsigned mask)
 407 {
 408         const struct ldlm_res_id *id = key;
 409         struct lu_fid       fid;
 410         __u32               hash;
 411         __u32               val;
 412
 413         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 414         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 415         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 416
 417         hash = fid_flatten32(&fid);
 418         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 419         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 420                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 421                 hash += (val >> 5) + (val << 11);
 422         } else {
 423                 val = fid_oid(&fid);
 424         }
 425         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 426         /* give me another random factor */
 427         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 428
 429         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 430         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 431
 432         return hash & mask;
 433 }
 434
 435 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 436 {
 437         struct ldlm_resource   *res;
 438
 439         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 440         return &res->lr_name;
 441 }
 442
 443 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 444 {
 445         struct ldlm_resource   *res;
 446
 447         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 448         return ldlm_res_eq((const struct ldlm_res_id *)key,
 449                            (const struct ldlm_res_id *)&res->lr_name);
 450 }
 451
 452 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 453 {
 454         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 455 }
 456
 457 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 458 {
 459         struct ldlm_resource *res;
 460
 461         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 462         ldlm_resource_getref(res);
 463 }
 464
 465 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 466 {
 467         struct ldlm_resource *res;
 468
 469         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 470         /* cfs_hash_for_each_nolock is the only chance we call it */
 471         ldlm_resource_putref_locked(res);
 472 }
 473
 474 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 475 {
 476         struct ldlm_resource *res;
 477
 478         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 479         ldlm_resource_putref(res);
 480 }
 481
 482 cfs_hash_ops_t ldlm_ns_hash_ops = {
 483         .hs_hash        = ldlm_res_hop_hash,
 484         .hs_key         = ldlm_res_hop_key,
 485         .hs_keycmp      = ldlm_res_hop_keycmp,
 486         .hs_keycpy      = NULL,
 487         .hs_object      = ldlm_res_hop_object,
 488         .hs_get         = ldlm_res_hop_get_locked,
 489         .hs_put_locked  = ldlm_res_hop_put_locked,
 490         .hs_put         = ldlm_res_hop_put
 491 };
 492
 493 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 494         .hs_hash        = ldlm_res_hop_fid_hash,
 495         .hs_key         = ldlm_res_hop_key,
 496         .hs_keycmp      = ldlm_res_hop_keycmp,
 497         .hs_keycpy      = NULL,
 498         .hs_object      = ldlm_res_hop_object,
 499         .hs_get         = ldlm_res_hop_get_locked,
 500         .hs_put_locked  = ldlm_res_hop_put_locked,
 501         .hs_put         = ldlm_res_hop_put
 502 };
 503
 504 typedef struct {
 505         ldlm_ns_type_t  nsd_type;
 506         /** hash bucket bits */
 507         unsigned        nsd_bkt_bits;
 508         /** hash bits */
 509         unsigned        nsd_all_bits;
 510         /** hash operations */
 511         cfs_hash_ops_t *nsd_hops;
 512 } ldlm_ns_hash_def_t;
 513
 514 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 515 {
 516         {
 517                 .nsd_type       = LDLM_NS_TYPE_MDC,
 518                 .nsd_bkt_bits   = 11,
 519                 .nsd_all_bits   = 16,
 520                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 521         },
 522         {
 523                 .nsd_type       = LDLM_NS_TYPE_MDT,
 524                 .nsd_bkt_bits   = 14,
 525                 .nsd_all_bits   = 21,
 526                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 527         },
 528         {
 529                 .nsd_type       = LDLM_NS_TYPE_OSC,
 530                 .nsd_bkt_bits   = 8,
 531                 .nsd_all_bits   = 12,
 532                 .nsd_hops       = &ldlm_ns_hash_ops,
 533         },
 534         {
 535                 .nsd_type       = LDLM_NS_TYPE_OST,
 536                 .nsd_bkt_bits   = 11,
 537                 .nsd_all_bits   = 17,
 538                 .nsd_hops       = &ldlm_ns_hash_ops,
 539         },
 540         {
 541                 .nsd_type       = LDLM_NS_TYPE_MGC,
 542                 .nsd_bkt_bits   = 4,
 543                 .nsd_all_bits   = 4,
 544                 .nsd_hops       = &ldlm_ns_hash_ops,
 545         },
 546         {
 547                 .nsd_type       = LDLM_NS_TYPE_MGT,
 548                 .nsd_bkt_bits   = 4,
 549                 .nsd_all_bits   = 4,
 550                 .nsd_hops       = &ldlm_ns_hash_ops,
 551         },
 552         {
 553                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 554         },
 555 };
 556
 557 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 558                                           ldlm_side_t client,
 559                                           ldlm_appetite_t apt,
 560                                           ldlm_ns_type_t ns_type)
 561 {
 562         struct ldlm_namespace *ns = NULL;
 563         struct ldlm_ns_bucket *nsb;
 564         ldlm_ns_hash_def_t    *nsd;
 565         cfs_hash_bd_t          bd;
 566         int                    idx;
 567         int                    rc;
 568         ENTRY;
 569
 570         LASSERT(obd != NULL);
 571
 572         rc = ldlm_get_ref();
 573         if (rc) {
 574                 CERROR("ldlm_get_ref failed: %d\n", rc);
 575                 RETURN(NULL);
 576         }
 577
 578         for (idx = 0;;idx++) {
 579                 nsd = &ldlm_ns_hash_defs[idx];
 580                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 581                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 582                         GOTO(out_ref, NULL);
 583                 }
 584
 585                 if (nsd->nsd_type == ns_type)
 586                         break;
 587         }
 588
 589         OBD_ALLOC_PTR(ns);
 590         if (!ns)
 591                 GOTO(out_ref, NULL);
 592
 593         ns->ns_rs_hash = cfs_hash_create(name,
 594                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 595                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 596                                          CFS_HASH_MIN_THETA,
 597                                          CFS_HASH_MAX_THETA,
 598                                          nsd->nsd_hops,
 599                                          CFS_HASH_DEPTH |
 600                                          CFS_HASH_BIGNAME |
 601                                          CFS_HASH_SPIN_BKTLOCK |
 602                                          CFS_HASH_NO_ITEMREF);
 603         if (ns->ns_rs_hash == NULL)
 604                 GOTO(out_ns, NULL);
 605
 606         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 607                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 608                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 609                 nsb->nsb_namespace = ns;
 610         }
 611
 612         ns->ns_obd      = obd;
 613         ns->ns_appetite = apt;
 614         ns->ns_client   = client;
 615
 616         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 617         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 618         cfs_spin_lock_init(&ns->ns_lock);
 619         cfs_atomic_set(&ns->ns_bref, 0);
 620         cfs_waitq_init(&ns->ns_waitq);
 621
 622         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 623         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 624         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 625
 626         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 627         ns->ns_nr_unused          = 0;
 628         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 629         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 630         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 631         ns->ns_timeouts           = 0;
 632         ns->ns_orig_connect_flags = 0;
 633         ns->ns_connect_flags      = 0;
 634         ns->ns_stopping           = 0;
 635         rc = ldlm_namespace_proc_register(ns);
 636         if (rc != 0) {
 637                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 638                 GOTO(out_hash, rc);
 639         }
 640
 641         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 642         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 643         if (rc) {
 644                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 645                 GOTO(out_proc, rc);
 646         }
 647
 648         ldlm_namespace_register(ns, client);
 649         RETURN(ns);
 650 out_proc:
 651         ldlm_namespace_proc_unregister(ns);
 652         ldlm_namespace_cleanup(ns, 0);
 653 out_hash:
 654         cfs_hash_putref(ns->ns_rs_hash);
 655 out_ns:
 656         OBD_FREE_PTR(ns);
 657 out_ref:
 658         ldlm_put_ref();
 659         RETURN(NULL);
 660 }
 661
 662 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 663
 664 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 665  * This is currently only used for recovery, and we make certain assumptions
 666  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 667 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 668                              int flags)
 669 {
 670         cfs_list_t *tmp;
 671         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 672         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 673
 674         do {
 675                 struct ldlm_lock *lock = NULL;
 676
 677                 /* first, we look for non-cleaned-yet lock
 678                  * all cleaned locks are marked by CLEANED flag */
 679                 lock_res(res);
 680                 cfs_list_for_each(tmp, q) {
 681                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 682                                               l_res_link);
 683                         if (lock->l_flags & LDLM_FL_CLEANED) {
 684                                 lock = NULL;
 685                                 continue;
 686                         }
 687                         LDLM_LOCK_GET(lock);
 688                         lock->l_flags |= LDLM_FL_CLEANED;
 689                         break;
 690                 }
 691
 692                 if (lock == NULL) {
 693                         unlock_res(res);
 694                         break;
 695                 }
 696
 697                 /* Set CBPENDING so nothing in the cancellation path
 698                  * can match this lock */
 699                 lock->l_flags |= LDLM_FL_CBPENDING;
 700                 lock->l_flags |= LDLM_FL_FAILED;
 701                 lock->l_flags |= flags;
 702
 703                 /* ... without sending a CANCEL message for local_only. */
 704                 if (local_only)
 705                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 706
 707                 if (local_only && (lock->l_readers || lock->l_writers)) {
 708                         /* This is a little bit gross, but much better than the
 709                          * alternative: pretend that we got a blocking AST from
 710                          * the server, so that when the lock is decref'd, it
 711                          * will go away ... */
 712                         unlock_res(res);
 713                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 714                         if (lock->l_completion_ast)
 715                                 lock->l_completion_ast(lock, 0, NULL);
 716                         LDLM_LOCK_RELEASE(lock);
 717                         continue;
 718                 }
 719
 720                 if (client) {
 721                         struct lustre_handle lockh;
 722
 723                         unlock_res(res);
 724                         ldlm_lock2handle(lock, &lockh);
 725                         rc = ldlm_cli_cancel(&lockh);
 726                         if (rc)
 727                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 728                 } else {
 729                         ldlm_resource_unlink_lock(lock);
 730                         unlock_res(res);
 731                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 732                                    "client node");
 733                         ldlm_lock_destroy(lock);
 734                 }
 735                 LDLM_LOCK_RELEASE(lock);
 736         } while (1);
 737 }
 738
 739 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 740                                cfs_hlist_node_t *hnode, void *arg)
 741 {
 742         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 743         int    flags = (int)(unsigned long)arg;
 744
 745         cleanup_resource(res, &res->lr_granted, flags);
 746         cleanup_resource(res, &res->lr_converting, flags);
 747         cleanup_resource(res, &res->lr_waiting, flags);
 748
 749         return 0;
 750 }
 751
 752 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 753                                   cfs_hlist_node_t *hnode, void *arg)
 754 {
 755         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 756
 757         CERROR("Namespace %s resource refcount nonzero "
 758                "(%d) after lock cleanup; forcing "
 759                "cleanup.\n",
 760                ldlm_ns_name(ldlm_res_to_ns(res)),
 761                cfs_atomic_read(&res->lr_refcount) - 1);
 762
 763         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 764                LPU64") (rc: %d)\n", res,
 765                res->lr_name.name[0], res->lr_name.name[1],
 766                res->lr_name.name[2], res->lr_name.name[3],
 767                cfs_atomic_read(&res->lr_refcount) - 1);
 768         return 0;
 769 }
 770
 771 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 772 {
 773         if (ns == NULL) {
 774                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 775                 return ELDLM_OK;
 776         }
 777
 778         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 779                                  (void *)(unsigned long)flags);
 780         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 781         return ELDLM_OK;
 782 }
 783
 784 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 785 {
 786         ENTRY;
 787
 788         /* At shutdown time, don't call the cancellation callback */
 789         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 790
 791         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 792                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 793                 int rc;
 794                 CDEBUG(D_DLMTRACE,
 795                        "dlm namespace %s free waiting on refcount %d\n",
 796                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 797 force_wait:
 798                 if (force)
 799                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 800
 801                 rc = l_wait_event(ns->ns_waitq,
 802                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 803
 804                 /* Forced cleanups should be able to reclaim all references,
 805                  * so it's safe to wait forever... we can't leak locks... */
 806                 if (force && rc == -ETIMEDOUT) {
 807                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 808                                        "namespace with %d resources in use, "
 809                                        "(rc=%d)\n", ldlm_ns_name(ns),
 810                                        cfs_atomic_read(&ns->ns_bref), rc);
 811                         GOTO(force_wait, rc);
 812                 }
 813
 814                 if (cfs_atomic_read(&ns->ns_bref)) {
 815                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 816                                        "with %d resources in use, (rc=%d)\n",
 817                                        ldlm_ns_name(ns),
 818                                        cfs_atomic_read(&ns->ns_bref), rc);
 819                         RETURN(ELDLM_NAMESPACE_EXISTS);
 820                 }
 821                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 822                        ldlm_ns_name(ns));
 823         }
 824
 825         RETURN(ELDLM_OK);
 826 }
 827
 828 /**
 829  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 830  * for freeing. Waits for refc == 0.
 831  *
 832  * The following is done:
 833  * (0) Unregister \a ns from its list to make inaccessible for potential users
 834  * like pools thread and others;
 835  * (1) Clear all locks in \a ns.
 836  */
 837 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 838                                struct obd_import *imp,
 839                                int force)
 840 {
 841         int rc;
 842         ENTRY;
 843         if (!ns) {
 844                 EXIT;
 845                 return;
 846         }
 847
 848         cfs_spin_lock(&ns->ns_lock);
 849         ns->ns_stopping = 1;
 850         cfs_spin_unlock(&ns->ns_lock);
 851
 852         /*
 853          * Can fail with -EINTR when force == 0 in which case try harder.
 854          */
 855         rc = __ldlm_namespace_free(ns, force);
 856         if (rc != ELDLM_OK) {
 857                 if (imp) {
 858                         ptlrpc_disconnect_import(imp, 0);
 859                         ptlrpc_invalidate_import(imp);
 860                 }
 861
 862                 /*
 863                  * With all requests dropped and the import inactive
 864                  * we are gaurenteed all reference will be dropped.
 865                  */
 866                 rc = __ldlm_namespace_free(ns, 1);
 867                 LASSERT(rc == 0);
 868         }
 869         EXIT;
 870 }
 871
 872 /**
 873  * Performs freeing memory structures related to \a ns. This is only done when
 874  * ldlm_namespce_free_prior() successfully removed all resources referencing
 875  * \a ns and its refc == 0.
 876  */
 877 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 878 {
 879         ENTRY;
 880         if (!ns) {
 881                 EXIT;
 882                 return;
 883         }
 884
 885
 886         /*
 887          * Make sure that nobody can find this ns in its list.
 888          */
 889         ldlm_namespace_unregister(ns, ns->ns_client);
 890         /*
 891          * Fini pool _before_ parent proc dir is removed. This is important as
 892          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 893          * it after @dir may cause oops.
 894          */
 895         ldlm_pool_fini(&ns->ns_pool);
 896
 897         ldlm_namespace_proc_unregister(ns);
 898         cfs_hash_putref(ns->ns_rs_hash);
 899         /*
 900          * Namespace \a ns should be not on list in this time, otherwise this
 901          * will cause issues realted to using freed \a ns in pools thread.
 902          */
 903         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 904         OBD_FREE_PTR(ns);
 905         ldlm_put_ref();
 906         EXIT;
 907 }
 908
 909
 910 /* Cleanup the resource, and free namespace.
 911  * bug 12864:
 912  * Deadlock issue:
 913  * proc1: destroy import
 914  *        class_disconnect_export(grab cl_sem) ->
 915  *              -> ldlm_namespace_free ->
 916  *              -> lprocfs_remove(grab _lprocfs_lock).
 917  * proc2: read proc info
 918  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 919  *              -> osc_rd_active, etc(grab cl_sem).
 920  *
 921  * So that I have to split the ldlm_namespace_free into two parts - the first
 922  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 923  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 924  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 925  * held.
 926  */
 927 void ldlm_namespace_free(struct ldlm_namespace *ns,
 928                          struct obd_import *imp,
 929                          int force)
 930 {
 931         ldlm_namespace_free_prior(ns, imp, force);
 932         ldlm_namespace_free_post(ns);
 933 }
 934
 935 void ldlm_namespace_get(struct ldlm_namespace *ns)
 936 {
 937         cfs_atomic_inc(&ns->ns_bref);
 938 }
 939
 940 void ldlm_namespace_put(struct ldlm_namespace *ns)
 941 {
 942         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 943                 cfs_waitq_signal(&ns->ns_waitq);
 944                 cfs_spin_unlock(&ns->ns_lock);
 945         }
 946 }
 947
 948 /* Register @ns in the list of namespaces */
 949 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 950 {
 951         cfs_mutex_lock(ldlm_namespace_lock(client));
 952         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 953         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 954         cfs_atomic_inc(ldlm_namespace_nr(client));
 955         cfs_mutex_unlock(ldlm_namespace_lock(client));
 956 }
 957
 958 /* Unregister @ns from the list of namespaces */
 959 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 960 {
 961         cfs_mutex_lock(ldlm_namespace_lock(client));
 962         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 963         /*
 964          * Some asserts and possibly other parts of code still using
 965          * list_empty(&ns->ns_list_chain). This is why it is important
 966          * to use list_del_init() here.
 967          */
 968         cfs_list_del_init(&ns->ns_list_chain);
 969         cfs_atomic_dec(ldlm_namespace_nr(client));
 970         cfs_mutex_unlock(ldlm_namespace_lock(client));
 971 }
 972
 973 /* Should be called under ldlm_namespace_lock(client) taken */
 974 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 975 {
 976         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 977         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
 978         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 979 }
 980
 981 /* Should be called under ldlm_namespace_lock(client) taken */
 982 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 983 {
 984         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
 985         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 986         return container_of(ldlm_namespace_list(client)->next,
 987                 struct ldlm_namespace, ns_list_chain);
 988 }
 989
 990 static struct ldlm_resource *ldlm_resource_new(void)
 991 {
 992         struct ldlm_resource *res;
 993         int idx;
 994
 995         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 996         if (res == NULL)
 997                 return NULL;
 998
 999         CFS_INIT_LIST_HEAD(&res->lr_granted);
1000         CFS_INIT_LIST_HEAD(&res->lr_converting);
1001         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1002
1003         /* initialize interval trees for each lock mode*/
1004         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1005                 res->lr_itree[idx].lit_size = 0;
1006                 res->lr_itree[idx].lit_mode = 1 << idx;
1007                 res->lr_itree[idx].lit_root = NULL;
1008         }
1009
1010         cfs_atomic_set(&res->lr_refcount, 1);
1011         cfs_spin_lock_init(&res->lr_lock);
1012         lu_ref_init(&res->lr_reference);
1013
1014         /* one who creates the resource must unlock
1015          * the mutex after lvb initialization */
1016         cfs_mutex_init(&res->lr_lvb_mutex);
1017         cfs_mutex_lock(&res->lr_lvb_mutex);
1018
1019         return res;
1020 }
1021
1022 /* Args: unlocked namespace
1023  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1024  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1025 struct ldlm_resource *
1026 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1027                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1028 {
1029         cfs_hlist_node_t     *hnode;
1030         struct ldlm_resource *res;
1031         cfs_hash_bd_t         bd;
1032         __u64                 version;
1033
1034         LASSERT(ns != NULL);
1035         LASSERT(parent == NULL);
1036         LASSERT(ns->ns_rs_hash != NULL);
1037         LASSERT(name->name[0] != 0);
1038
1039         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1040         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1041         if (hnode != NULL) {
1042                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1043                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1044                 /* synchronize WRT resource creation */
1045                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1046                         cfs_mutex_lock(&res->lr_lvb_mutex);
1047                         cfs_mutex_unlock(&res->lr_lvb_mutex);
1048                 }
1049                 return res;
1050         }
1051
1052         version = cfs_hash_bd_version_get(&bd);
1053         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1054
1055         if (create == 0)
1056                 return NULL;
1057
1058         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1059                  "type: %d\n", type);
1060         res = ldlm_resource_new();
1061         if (!res)
1062                 return NULL;
1063
1064         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1065         res->lr_name       = *name;
1066         res->lr_type       = type;
1067         res->lr_most_restr = LCK_NL;
1068
1069         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1070         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1071                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1072
1073         if (hnode != NULL) {
1074                 /* someone won the race and added the resource before */
1075                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1076                 /* clean lu_ref for failed resource */
1077                 lu_ref_fini(&res->lr_reference);
1078                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1079
1080                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1081                 /* synchronize WRT resource creation */
1082                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1083                         cfs_mutex_lock(&res->lr_lvb_mutex);
1084                         cfs_mutex_unlock(&res->lr_lvb_mutex);
1085                 }
1086                 return res;
1087         }
1088         /* we won! let's add the resource */
1089         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1090         if (cfs_hash_bd_count_get(&bd) == 1)
1091                 ldlm_namespace_get(ns);
1092
1093         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1094         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1095                 int rc;
1096
1097                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1098                 rc = ns->ns_lvbo->lvbo_init(res);
1099                 if (rc)
1100                         CERROR("lvbo_init failed for resource "
1101                                LPU64": rc %d\n", name->name[0], rc);
1102                 /* we create resource with locked lr_lvb_mutex */
1103                 cfs_mutex_unlock(&res->lr_lvb_mutex);
1104         }
1105
1106         return res;
1107 }
1108
1109 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1110 {
1111         LASSERT(res != NULL);
1112         LASSERT(res != LP_POISON);
1113         cfs_atomic_inc(&res->lr_refcount);
1114         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1115                cfs_atomic_read(&res->lr_refcount));
1116         return res;
1117 }
1118
1119 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1120                                          struct ldlm_resource *res)
1121 {
1122         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1123
1124         if (!cfs_list_empty(&res->lr_granted)) {
1125                 ldlm_resource_dump(D_ERROR, res);
1126                 LBUG();
1127         }
1128
1129         if (!cfs_list_empty(&res->lr_converting)) {
1130                 ldlm_resource_dump(D_ERROR, res);
1131                 LBUG();
1132         }
1133
1134         if (!cfs_list_empty(&res->lr_waiting)) {
1135                 ldlm_resource_dump(D_ERROR, res);
1136                 LBUG();
1137         }
1138
1139         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1140                                bd, &res->lr_hash);
1141         lu_ref_fini(&res->lr_reference);
1142         if (cfs_hash_bd_count_get(bd) == 0)
1143                 ldlm_namespace_put(nsb->nsb_namespace);
1144 }
1145
1146 /* Returns 1 if the resource was freed, 0 if it remains. */
1147 int ldlm_resource_putref(struct ldlm_resource *res)
1148 {
1149         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1150         cfs_hash_bd_t   bd;
1151
1152         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1153         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1154                res, cfs_atomic_read(&res->lr_refcount) - 1);
1155
1156         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1157         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1158                 __ldlm_resource_putref_final(&bd, res);
1159                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1160                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1161                         ns->ns_lvbo->lvbo_free(res);
1162                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1163                 return 1;
1164         }
1165         return 0;
1166 }
1167
1168 /* Returns 1 if the resource was freed, 0 if it remains. */
1169 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1170 {
1171         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1172
1173         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1174         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1175                res, cfs_atomic_read(&res->lr_refcount) - 1);
1176
1177         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1178                 cfs_hash_bd_t bd;
1179
1180                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1181                                 &res->lr_name, &bd);
1182                 __ldlm_resource_putref_final(&bd, res);
1183                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1184                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1185                  * so we should never be here while calling cfs_hash_del,
1186                  * cfs_hash_for_each_nolock is the only case we can get
1187                  * here, which is safe to release cfs_hash_bd_lock.
1188                  */
1189                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1190                         ns->ns_lvbo->lvbo_free(res);
1191                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1192
1193                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1194                 return 1;
1195         }
1196         return 0;
1197 }
1198
1199 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1200                             struct ldlm_lock *lock)
1201 {
1202         check_res_locked(res);
1203
1204         LDLM_DEBUG(lock, "About to add this lock:\n");
1205
1206         if (lock->l_destroyed) {
1207                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1208                 return;
1209         }
1210
1211         LASSERT(cfs_list_empty(&lock->l_res_link));
1212
1213         cfs_list_add_tail(&lock->l_res_link, head);
1214 }
1215
1216 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1217                                      struct ldlm_lock *new)
1218 {
1219         struct ldlm_resource *res = original->l_resource;
1220
1221         check_res_locked(res);
1222
1223         ldlm_resource_dump(D_INFO, res);
1224         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1225
1226         if (new->l_destroyed) {
1227                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1228                 goto out;
1229         }
1230
1231         LASSERT(cfs_list_empty(&new->l_res_link));
1232
1233         cfs_list_add(&new->l_res_link, &original->l_res_link);
1234  out:;
1235 }
1236
1237 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1238 {
1239         int type = lock->l_resource->lr_type;
1240
1241         check_res_locked(lock->l_resource);
1242         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1243                 ldlm_unlink_lock_skiplist(lock);
1244         else if (type == LDLM_EXTENT)
1245                 ldlm_extent_unlink_lock(lock);
1246         cfs_list_del_init(&lock->l_res_link);
1247 }
1248
1249 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1250 {
1251         desc->lr_type = res->lr_type;
1252         desc->lr_name = res->lr_name;
1253 }
1254
1255 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1256 {
1257         cfs_list_t *tmp;
1258
1259         if (!((libcfs_debug | D_ERROR) & level))
1260                 return;
1261
1262         cfs_mutex_lock(ldlm_namespace_lock(client));
1263
1264         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1265                 struct ldlm_namespace *ns;
1266                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1267                 ldlm_namespace_dump(level, ns);
1268         }
1269
1270         cfs_mutex_unlock(ldlm_namespace_lock(client));
1271 }
1272
1273 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1274                               cfs_hlist_node_t *hnode, void *arg)
1275 {
1276         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1277         int    level = (int)(unsigned long)arg;
1278
1279         lock_res(res);
1280         ldlm_resource_dump(level, res);
1281         unlock_res(res);
1282
1283         return 0;
1284 }
1285
1286 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1287 {
1288         if (!((libcfs_debug | D_ERROR) & level))
1289                 return;
1290
1291         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1292                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1293                ns_is_client(ns) ? "client" : "server");
1294
1295         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1296                 return;
1297
1298         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1299                                  ldlm_res_hash_dump,
1300                                  (void *)(unsigned long)level);
1301         cfs_spin_lock(&ns->ns_lock);
1302         ns->ns_next_dump = cfs_time_shift(10);
1303         cfs_spin_unlock(&ns->ns_lock);
1304 }
1305
1306 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1307 {
1308         struct ldlm_lock *lock;
1309         unsigned int granted = 0;
1310
1311         CLASSERT(RES_NAME_SIZE == 4);
1312
1313         if (!((libcfs_debug | D_ERROR) & level))
1314                 return;
1315
1316         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1317                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1318                res->lr_name.name[2], res->lr_name.name[3],
1319                cfs_atomic_read(&res->lr_refcount));
1320
1321         if (!cfs_list_empty(&res->lr_granted)) {
1322                 CDEBUG(level, "Granted locks (in reverse order):\n");
1323                 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1324                                                 l_res_link) {
1325                         LDLM_DEBUG_LIMIT(level, lock, "###");
1326                         if (!(level & D_CANTMASK) &&
1327                             ++granted > ldlm_dump_granted_max) {
1328                                 CDEBUG(level, "only dump %d granted locks to "
1329                                        "avoid DDOS.\n", granted);
1330                                 break;
1331                         }
1332                 }
1333         }
1334         if (!cfs_list_empty(&res->lr_converting)) {
1335                 CDEBUG(level, "Converting locks:\n");
1336                 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1337                         LDLM_DEBUG_LIMIT(level, lock, "###");
1338         }
1339         if (!cfs_list_empty(&res->lr_waiting)) {
1340                 CDEBUG(level, "Waiting locks:\n");
1341                 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1342                         LDLM_DEBUG_LIMIT(level, lock, "###");
1343         }
1344 }