lustre/ldlm/ldlm_resource.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  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_semaphore_t ldlm_srv_namespace_lock;
  59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  60
  61 cfs_semaphore_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 #ifdef LPROCFS
  71 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  72                              unsigned long count, void *data)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78
  79 int ldlm_proc_setup(void)
  80 {
  81         int rc;
  82         struct lprocfs_vars list[] = {
  83                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  84                 { "cancel_unused_locks_before_replay",
  85                   lprocfs_rd_uint, lprocfs_wr_uint,
  86                   &ldlm_cancel_unused_locks_before_replay, NULL },
  87                 { NULL }};
  88         ENTRY;
  89         LASSERT(ldlm_ns_proc_dir == NULL);
  90
  91         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  92                                               proc_lustre_root,
  93                                               NULL, NULL);
  94         if (IS_ERR(ldlm_type_proc_dir)) {
  95                 CERROR("LProcFS failed in ldlm-init\n");
  96                 rc = PTR_ERR(ldlm_type_proc_dir);
  97                 GOTO(err, rc);
  98         }
  99
 100         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 101                                             ldlm_type_proc_dir,
 102                                             NULL, NULL);
 103         if (IS_ERR(ldlm_ns_proc_dir)) {
 104                 CERROR("LProcFS failed in ldlm-init\n");
 105                 rc = PTR_ERR(ldlm_ns_proc_dir);
 106                 GOTO(err_type, rc);
 107         }
 108
 109         ldlm_svc_proc_dir = lprocfs_register("services",
 110                                             ldlm_type_proc_dir,
 111                                             NULL, NULL);
 112         if (IS_ERR(ldlm_svc_proc_dir)) {
 113                 CERROR("LProcFS failed in ldlm-init\n");
 114                 rc = PTR_ERR(ldlm_svc_proc_dir);
 115                 GOTO(err_ns, rc);
 116         }
 117
 118         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 119
 120         RETURN(0);
 121
 122 err_ns:
 123         lprocfs_remove(&ldlm_ns_proc_dir);
 124 err_type:
 125         lprocfs_remove(&ldlm_type_proc_dir);
 126 err:
 127         ldlm_svc_proc_dir = NULL;
 128         RETURN(rc);
 129 }
 130
 131 void ldlm_proc_cleanup(void)
 132 {
 133         if (ldlm_svc_proc_dir)
 134                 lprocfs_remove(&ldlm_svc_proc_dir);
 135
 136         if (ldlm_ns_proc_dir)
 137                 lprocfs_remove(&ldlm_ns_proc_dir);
 138
 139         if (ldlm_type_proc_dir)
 140                 lprocfs_remove(&ldlm_type_proc_dir);
 141 }
 142
 143 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 144                                    int count, int *eof, void *data)
 145 {
 146         struct ldlm_namespace *ns  = data;
 147         __u64                  res = 0;
 148         cfs_hash_bd_t          bd;
 149         int                    i;
 150
 151         /* result is not strictly consistant */
 152         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 153                 res += cfs_hash_bd_count_get(&bd);
 154         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 155 }
 156
 157 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 158                                int count, int *eof, void *data)
 159 {
 160         struct ldlm_namespace *ns = data;
 161         __u64                  locks;
 162
 163         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 164                                         LPROCFS_FIELDS_FLAGS_SUM);
 165         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 166 }
 167
 168 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 169                                int count, int *eof, void *data)
 170 {
 171         struct ldlm_namespace *ns = data;
 172         __u32 *nr = &ns->ns_max_unused;
 173
 174         if (ns_connect_lru_resize(ns))
 175                 nr = &ns->ns_nr_unused;
 176         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 177 }
 178
 179 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 180                                unsigned long count, void *data)
 181 {
 182         struct ldlm_namespace *ns = data;
 183         char dummy[MAX_STRING_SIZE + 1], *end;
 184         unsigned long tmp;
 185         int lru_resize;
 186
 187         dummy[MAX_STRING_SIZE] = '\0';
 188         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 189                 return -EFAULT;
 190
 191         if (strncmp(dummy, "clear", 5) == 0) {
 192                 CDEBUG(D_DLMTRACE,
 193                        "dropping all unused locks from namespace %s\n",
 194                        ldlm_ns_name(ns));
 195                 if (ns_connect_lru_resize(ns)) {
 196                         int canceled, unused  = ns->ns_nr_unused;
 197
 198                         /* Try to cancel all @ns_nr_unused locks. */
 199                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 200                                                    LDLM_CANCEL_PASSED);
 201                         if (canceled < unused) {
 202                                 CDEBUG(D_DLMTRACE,
 203                                        "not all requested locks are canceled, "
 204                                        "requested: %d, canceled: %d\n", unused,
 205                                        canceled);
 206                                 return -EINVAL;
 207                         }
 208                 } else {
 209                         tmp = ns->ns_max_unused;
 210                         ns->ns_max_unused = 0;
 211                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 212                         ns->ns_max_unused = tmp;
 213                 }
 214                 return count;
 215         }
 216
 217         tmp = simple_strtoul(dummy, &end, 0);
 218         if (dummy == end) {
 219                 CERROR("invalid value written\n");
 220                 return -EINVAL;
 221         }
 222         lru_resize = (tmp == 0);
 223
 224         if (ns_connect_lru_resize(ns)) {
 225                 if (!lru_resize)
 226                         ns->ns_max_unused = (unsigned int)tmp;
 227
 228                 if (tmp > ns->ns_nr_unused)
 229                         tmp = ns->ns_nr_unused;
 230                 tmp = ns->ns_nr_unused - tmp;
 231
 232                 CDEBUG(D_DLMTRACE,
 233                        "changing namespace %s unused locks from %u to %u\n",
 234                        ldlm_ns_name(ns), ns->ns_nr_unused,
 235                        (unsigned int)tmp);
 236                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 237
 238                 if (!lru_resize) {
 239                         CDEBUG(D_DLMTRACE,
 240                                "disable lru_resize for namespace %s\n",
 241                                ldlm_ns_name(ns));
 242                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 243                 }
 244         } else {
 245                 CDEBUG(D_DLMTRACE,
 246                        "changing namespace %s max_unused from %u to %u\n",
 247                        ldlm_ns_name(ns), ns->ns_max_unused,
 248                        (unsigned int)tmp);
 249                 ns->ns_max_unused = (unsigned int)tmp;
 250                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 251
 252                 /* Make sure that originally lru resize was supported before
 253                  * turning it on here. */
 254                 if (lru_resize &&
 255                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 256                         CDEBUG(D_DLMTRACE,
 257                                "enable lru_resize for namespace %s\n",
 258                                ldlm_ns_name(ns));
 259                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 260                 }
 261         }
 262
 263         return count;
 264 }
 265
 266 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 267 {
 268         struct proc_dir_entry *dir;
 269
 270         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 271         if (dir == NULL) {
 272                 CERROR("dlm namespace %s has no procfs dir?\n",
 273                        ldlm_ns_name(ns));
 274         } else {
 275                 lprocfs_remove(&dir);
 276         }
 277
 278         if (ns->ns_stats != NULL)
 279                 lprocfs_free_stats(&ns->ns_stats);
 280 }
 281
 282 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 283 {
 284         struct lprocfs_vars lock_vars[2];
 285         char lock_name[MAX_STRING_SIZE + 1];
 286
 287         LASSERT(ns != NULL);
 288         LASSERT(ns->ns_rs_hash != NULL);
 289
 290         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 291         if (ns->ns_stats == NULL)
 292                 return -ENOMEM;
 293
 294         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 295                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 296
 297         lock_name[MAX_STRING_SIZE] = '\0';
 298
 299         memset(lock_vars, 0, sizeof(lock_vars));
 300         lock_vars[0].name = lock_name;
 301
 302         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 303                  ldlm_ns_name(ns));
 304         lock_vars[0].data = ns;
 305         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 306         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 307
 308         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 309                  ldlm_ns_name(ns));
 310         lock_vars[0].data = ns;
 311         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 312         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 313
 314         if (ns_is_client(ns)) {
 315                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 316                          ldlm_ns_name(ns));
 317                 lock_vars[0].data = &ns->ns_nr_unused;
 318                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 319                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 320
 321                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 322                          ldlm_ns_name(ns));
 323                 lock_vars[0].data = ns;
 324                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 325                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 326                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 327
 328                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 329                          ldlm_ns_name(ns));
 330                 lock_vars[0].data = &ns->ns_max_age;
 331                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 332                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 333                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 334         } else {
 335                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 336                          ldlm_ns_name(ns));
 337                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 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
 342                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 343                          ldlm_ns_name(ns));
 344                 lock_vars[0].data = &ns->ns_timeouts;
 345                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 346                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 347
 348                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 349                          ldlm_ns_name(ns));
 350                 lock_vars[0].data = &ns->ns_max_nolock_size;
 351                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 352                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 354
 355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 356                          ldlm_ns_name(ns));
 357                 lock_vars[0].data = &ns->ns_contention_time;
 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/contended_locks",
 363                          ldlm_ns_name(ns));
 364                 lock_vars[0].data = &ns->ns_contended_locks;
 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/max_parallel_ast",
 370                          ldlm_ns_name(ns));
 371                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 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         return 0;
 377 }
 378 #undef MAX_STRING_SIZE
 379 #else /* LPROCFS */
 380
 381 #define ldlm_namespace_proc_unregister(ns)      ({;})
 382 #define ldlm_namespace_proc_register(ns)        ({0;})
 383
 384 #endif /* LPROCFS */
 385
 386 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 387                                   const void *key, unsigned mask)
 388 {
 389         const struct ldlm_res_id     *id  = key;
 390         unsigned                val = 0;
 391         unsigned                i;
 392
 393         for (i = 0; i < RES_NAME_SIZE; i++)
 394                 val += id->name[i];
 395         return val & mask;
 396 }
 397
 398 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 399                                       const void *key, unsigned mask)
 400 {
 401         const struct ldlm_res_id *id = key;
 402         struct lu_fid       fid;
 403         __u32               hash;
 404         __u32               val;
 405
 406         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 407         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF];
 408         fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF];
 409
 410         hash = fid_flatten32(&fid);
 411         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 412         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 413                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 414                 hash += (val >> 5) + (val << 11);
 415         } else {
 416                 val = fid_oid(&fid);
 417         }
 418         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 419         /* give me another random factor */
 420         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 421
 422         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 423         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 424
 425         return hash & mask;
 426 }
 427
 428 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 429 {
 430         struct ldlm_resource   *res;
 431
 432         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 433         return &res->lr_name;
 434 }
 435
 436 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 437 {
 438         struct ldlm_resource   *res;
 439
 440         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 441         return ldlm_res_eq((const struct ldlm_res_id *)key,
 442                            (const struct ldlm_res_id *)&res->lr_name);
 443 }
 444
 445 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 446 {
 447         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 448 }
 449
 450 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 451 {
 452         struct ldlm_resource *res;
 453
 454         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 455         ldlm_resource_getref(res);
 456         LDLM_RESOURCE_ADDREF(res);
 457 }
 458
 459 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 460 {
 461         struct ldlm_resource *res;
 462
 463         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 464         /* cfs_hash_for_each_nolock is the only chance we call it */
 465         LDLM_RESOURCE_DELREF(res);
 466         ldlm_resource_putref_locked(res);
 467 }
 468
 469 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 470 {
 471         struct ldlm_resource *res;
 472
 473         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 474         LDLM_RESOURCE_DELREF(res);
 475         ldlm_resource_putref(res);
 476 }
 477
 478 cfs_hash_ops_t ldlm_ns_hash_ops = {
 479         .hs_hash        = ldlm_res_hop_hash,
 480         .hs_key         = ldlm_res_hop_key,
 481         .hs_keycmp      = ldlm_res_hop_keycmp,
 482         .hs_keycpy      = NULL,
 483         .hs_object      = ldlm_res_hop_object,
 484         .hs_get         = ldlm_res_hop_get_locked,
 485         .hs_put_locked  = ldlm_res_hop_put_locked,
 486         .hs_put         = ldlm_res_hop_put
 487 };
 488
 489 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 490         .hs_hash        = ldlm_res_hop_fid_hash,
 491         .hs_key         = ldlm_res_hop_key,
 492         .hs_keycmp      = ldlm_res_hop_keycmp,
 493         .hs_keycpy      = NULL,
 494         .hs_object      = ldlm_res_hop_object,
 495         .hs_get         = ldlm_res_hop_get_locked,
 496         .hs_put_locked  = ldlm_res_hop_put_locked,
 497         .hs_put         = ldlm_res_hop_put
 498 };
 499
 500 typedef struct {
 501         ldlm_ns_type_t  nsd_type;
 502         /** hash bucket bits */
 503         unsigned        nsd_bkt_bits;
 504         /** hash bits */
 505         unsigned        nsd_all_bits;
 506         /** hash operations */
 507         cfs_hash_ops_t *nsd_hops;
 508 } ldlm_ns_hash_def_t;
 509
 510 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 511 {
 512         {
 513                 .nsd_type       = LDLM_NS_TYPE_MDC,
 514                 .nsd_bkt_bits   = 11,
 515                 .nsd_all_bits   = 16,
 516                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 517         },
 518         {
 519                 .nsd_type       = LDLM_NS_TYPE_MDT,
 520                 .nsd_bkt_bits   = 14,
 521                 .nsd_all_bits   = 21,
 522                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 523         },
 524         {
 525                 .nsd_type       = LDLM_NS_TYPE_OSC,
 526                 .nsd_bkt_bits   = 8,
 527                 .nsd_all_bits   = 12,
 528                 .nsd_hops       = &ldlm_ns_hash_ops,
 529         },
 530         {
 531                 .nsd_type       = LDLM_NS_TYPE_OST,
 532                 .nsd_bkt_bits   = 11,
 533                 .nsd_all_bits   = 17,
 534                 .nsd_hops       = &ldlm_ns_hash_ops,
 535         },
 536         {
 537                 .nsd_type       = LDLM_NS_TYPE_MGC,
 538                 .nsd_bkt_bits   = 4,
 539                 .nsd_all_bits   = 4,
 540                 .nsd_hops       = &ldlm_ns_hash_ops,
 541         },
 542         {
 543                 .nsd_type       = LDLM_NS_TYPE_MGT,
 544                 .nsd_bkt_bits   = 4,
 545                 .nsd_all_bits   = 4,
 546                 .nsd_hops       = &ldlm_ns_hash_ops,
 547         },
 548         {
 549                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 550         },
 551 };
 552
 553 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 554                                           ldlm_side_t client,
 555                                           ldlm_appetite_t apt,
 556                                           ldlm_ns_type_t ns_type)
 557 {
 558         struct ldlm_namespace *ns = NULL;
 559         struct ldlm_ns_bucket *nsb;
 560         ldlm_ns_hash_def_t    *nsd;
 561         cfs_hash_bd_t          bd;
 562         int                    idx;
 563         int                    rc;
 564         ENTRY;
 565
 566         LASSERT(obd != NULL);
 567
 568         rc = ldlm_get_ref();
 569         if (rc) {
 570                 CERROR("ldlm_get_ref failed: %d\n", rc);
 571                 RETURN(NULL);
 572         }
 573
 574         for (idx = 0;;idx++) {
 575                 nsd = &ldlm_ns_hash_defs[idx];
 576                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 577                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 578                         GOTO(out_ref, NULL);
 579                 }
 580
 581                 if (nsd->nsd_type == ns_type)
 582                         break;
 583         }
 584
 585         OBD_ALLOC_PTR(ns);
 586         if (!ns)
 587                 GOTO(out_ref, NULL);
 588
 589         ns->ns_rs_hash = cfs_hash_create(name,
 590                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 591                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 592                                          CFS_HASH_MIN_THETA,
 593                                          CFS_HASH_MAX_THETA,
 594                                          nsd->nsd_hops,
 595                                          CFS_HASH_DEPTH |
 596                                          CFS_HASH_BIGNAME |
 597                                          CFS_HASH_SPIN_BKTLOCK |
 598                                          CFS_HASH_NO_ITEMREF);
 599         if (ns->ns_rs_hash == NULL)
 600                 GOTO(out_ns, NULL);
 601
 602         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 603                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 604                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 605                 nsb->nsb_namespace = ns;
 606         }
 607
 608         ns->ns_obd      = obd;
 609         ns->ns_appetite = apt;
 610         ns->ns_client   = client;
 611
 612         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 613         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 614         cfs_spin_lock_init(&ns->ns_lock);
 615         cfs_atomic_set(&ns->ns_bref, 0);
 616         cfs_waitq_init(&ns->ns_waitq);
 617
 618         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 619         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 620         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 621
 622         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 623         ns->ns_nr_unused          = 0;
 624         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 625         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 626         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 627         ns->ns_timeouts           = 0;
 628         ns->ns_orig_connect_flags = 0;
 629         ns->ns_connect_flags      = 0;
 630         ns->ns_stopping           = 0;
 631         rc = ldlm_namespace_proc_register(ns);
 632         if (rc != 0) {
 633                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 634                 GOTO(out_hash, rc);
 635         }
 636
 637         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 638         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 639         if (rc) {
 640                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 641                 GOTO(out_proc, rc);
 642         }
 643
 644         ldlm_namespace_register(ns, client);
 645         RETURN(ns);
 646 out_proc:
 647         ldlm_namespace_proc_unregister(ns);
 648         ldlm_namespace_cleanup(ns, 0);
 649 out_hash:
 650         cfs_hash_putref(ns->ns_rs_hash);
 651 out_ns:
 652         OBD_FREE_PTR(ns);
 653 out_ref:
 654         ldlm_put_ref();
 655         RETURN(NULL);
 656 }
 657
 658 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 659
 660 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 661  * This is currently only used for recovery, and we make certain assumptions
 662  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 663 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 664                              int flags)
 665 {
 666         cfs_list_t *tmp;
 667         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 668         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 669
 670         do {
 671                 struct ldlm_lock *lock = NULL;
 672
 673                 /* first, we look for non-cleaned-yet lock
 674                  * all cleaned locks are marked by CLEANED flag */
 675                 lock_res(res);
 676                 cfs_list_for_each(tmp, q) {
 677                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 678                                               l_res_link);
 679                         if (lock->l_flags & LDLM_FL_CLEANED) {
 680                                 lock = NULL;
 681                                 continue;
 682                         }
 683                         LDLM_LOCK_GET(lock);
 684                         lock->l_flags |= LDLM_FL_CLEANED;
 685                         break;
 686                 }
 687
 688                 if (lock == NULL) {
 689                         unlock_res(res);
 690                         break;
 691                 }
 692
 693                 /* Set CBPENDING so nothing in the cancellation path
 694                  * can match this lock */
 695                 lock->l_flags |= LDLM_FL_CBPENDING;
 696                 lock->l_flags |= LDLM_FL_FAILED;
 697                 lock->l_flags |= flags;
 698
 699                 /* ... without sending a CANCEL message for local_only. */
 700                 if (local_only)
 701                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 702
 703                 if (local_only && (lock->l_readers || lock->l_writers)) {
 704                         /* This is a little bit gross, but much better than the
 705                          * alternative: pretend that we got a blocking AST from
 706                          * the server, so that when the lock is decref'd, it
 707                          * will go away ... */
 708                         unlock_res(res);
 709                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 710                         if (lock->l_completion_ast)
 711                                 lock->l_completion_ast(lock, 0, NULL);
 712                         LDLM_LOCK_RELEASE(lock);
 713                         continue;
 714                 }
 715
 716                 if (client) {
 717                         struct lustre_handle lockh;
 718
 719                         unlock_res(res);
 720                         ldlm_lock2handle(lock, &lockh);
 721                         rc = ldlm_cli_cancel(&lockh);
 722                         if (rc)
 723                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 724                 } else {
 725                         ldlm_resource_unlink_lock(lock);
 726                         unlock_res(res);
 727                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 728                                    "client node");
 729                         ldlm_lock_destroy(lock);
 730                 }
 731                 LDLM_LOCK_RELEASE(lock);
 732         } while (1);
 733 }
 734
 735 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 736                                cfs_hlist_node_t *hnode, void *arg)
 737 {
 738         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 739         int    flags = (int)(unsigned long)arg;
 740
 741         cleanup_resource(res, &res->lr_granted, flags);
 742         cleanup_resource(res, &res->lr_converting, flags);
 743         cleanup_resource(res, &res->lr_waiting, flags);
 744
 745         return 0;
 746 }
 747
 748 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 749                                   cfs_hlist_node_t *hnode, void *arg)
 750 {
 751         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 752
 753         CERROR("Namespace %s resource refcount nonzero "
 754                "(%d) after lock cleanup; forcing "
 755                "cleanup.\n",
 756                ldlm_ns_name(ldlm_res_to_ns(res)),
 757                cfs_atomic_read(&res->lr_refcount) - 1);
 758
 759         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 760                LPU64") (rc: %d)\n", res,
 761                res->lr_name.name[0], res->lr_name.name[1],
 762                res->lr_name.name[2], res->lr_name.name[3],
 763                cfs_atomic_read(&res->lr_refcount) - 1);
 764         return 0;
 765 }
 766
 767 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 768 {
 769         if (ns == NULL) {
 770                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 771                 return ELDLM_OK;
 772         }
 773
 774         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 775                                  (void *)(unsigned long)flags);
 776         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 777         return ELDLM_OK;
 778 }
 779
 780 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 781 {
 782         ENTRY;
 783
 784         /* At shutdown time, don't call the cancellation callback */
 785         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 786
 787         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 788                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 789                 int rc;
 790                 CDEBUG(D_DLMTRACE,
 791                        "dlm namespace %s free waiting on refcount %d\n",
 792                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 793 force_wait:
 794                 if (force)
 795                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 796
 797                 rc = l_wait_event(ns->ns_waitq,
 798                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 799
 800                 /* Forced cleanups should be able to reclaim all references,
 801                  * so it's safe to wait forever... we can't leak locks... */
 802                 if (force && rc == -ETIMEDOUT) {
 803                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 804                                        "namespace with %d resources in use, "
 805                                        "(rc=%d)\n", ldlm_ns_name(ns),
 806                                        cfs_atomic_read(&ns->ns_bref), rc);
 807                         GOTO(force_wait, rc);
 808                 }
 809
 810                 if (cfs_atomic_read(&ns->ns_bref)) {
 811                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 812                                        "with %d resources in use, (rc=%d)\n",
 813                                        ldlm_ns_name(ns),
 814                                        cfs_atomic_read(&ns->ns_bref), rc);
 815                         RETURN(ELDLM_NAMESPACE_EXISTS);
 816                 }
 817                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 818                        ldlm_ns_name(ns));
 819         }
 820
 821         RETURN(ELDLM_OK);
 822 }
 823
 824 /**
 825  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 826  * for freeing. Waits for refc == 0.
 827  *
 828  * The following is done:
 829  * (0) Unregister \a ns from its list to make inaccessible for potential users
 830  * like pools thread and others;
 831  * (1) Clear all locks in \a ns.
 832  */
 833 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 834                                struct obd_import *imp,
 835                                int force)
 836 {
 837         int rc;
 838         ENTRY;
 839         if (!ns) {
 840                 EXIT;
 841                 return;
 842         }
 843
 844         cfs_spin_lock(&ns->ns_lock);
 845         ns->ns_stopping = 1;
 846         cfs_spin_unlock(&ns->ns_lock);
 847
 848         /*
 849          * Can fail with -EINTR when force == 0 in which case try harder.
 850          */
 851         rc = __ldlm_namespace_free(ns, force);
 852         if (rc != ELDLM_OK) {
 853                 if (imp) {
 854                         ptlrpc_disconnect_import(imp, 0);
 855                         ptlrpc_invalidate_import(imp);
 856                 }
 857
 858                 /*
 859                  * With all requests dropped and the import inactive
 860                  * we are gaurenteed all reference will be dropped.
 861                  */
 862                 rc = __ldlm_namespace_free(ns, 1);
 863                 LASSERT(rc == 0);
 864         }
 865         EXIT;
 866 }
 867
 868 /**
 869  * Performs freeing memory structures related to \a ns. This is only done when
 870  * ldlm_namespce_free_prior() successfully removed all resources referencing
 871  * \a ns and its refc == 0.
 872  */
 873 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 874 {
 875         ENTRY;
 876         if (!ns) {
 877                 EXIT;
 878                 return;
 879         }
 880
 881
 882         /*
 883          * Make sure that nobody can find this ns in its list.
 884          */
 885         ldlm_namespace_unregister(ns, ns->ns_client);
 886         /*
 887          * Fini pool _before_ parent proc dir is removed. This is important as
 888          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 889          * it after @dir may cause oops.
 890          */
 891         ldlm_pool_fini(&ns->ns_pool);
 892
 893         ldlm_namespace_proc_unregister(ns);
 894         cfs_hash_putref(ns->ns_rs_hash);
 895         /*
 896          * Namespace \a ns should be not on list in this time, otherwise this
 897          * will cause issues realted to using freed \a ns in pools thread.
 898          */
 899         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 900         OBD_FREE_PTR(ns);
 901         ldlm_put_ref();
 902         EXIT;
 903 }
 904
 905
 906 /* Cleanup the resource, and free namespace.
 907  * bug 12864:
 908  * Deadlock issue:
 909  * proc1: destroy import
 910  *        class_disconnect_export(grab cl_sem) ->
 911  *              -> ldlm_namespace_free ->
 912  *              -> lprocfs_remove(grab _lprocfs_lock).
 913  * proc2: read proc info
 914  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 915  *              -> osc_rd_active, etc(grab cl_sem).
 916  *
 917  * So that I have to split the ldlm_namespace_free into two parts - the first
 918  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 919  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 920  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 921  * held.
 922  */
 923 void ldlm_namespace_free(struct ldlm_namespace *ns,
 924                          struct obd_import *imp,
 925                          int force)
 926 {
 927         ldlm_namespace_free_prior(ns, imp, force);
 928         ldlm_namespace_free_post(ns);
 929 }
 930
 931 void ldlm_namespace_get(struct ldlm_namespace *ns)
 932 {
 933         cfs_atomic_inc(&ns->ns_bref);
 934 }
 935
 936 void ldlm_namespace_put(struct ldlm_namespace *ns)
 937 {
 938         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 939                 cfs_waitq_signal(&ns->ns_waitq);
 940                 cfs_spin_unlock(&ns->ns_lock);
 941         }
 942 }
 943
 944 /* Register @ns in the list of namespaces */
 945 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 946 {
 947         cfs_mutex_down(ldlm_namespace_lock(client));
 948         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 949         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 950         cfs_atomic_inc(ldlm_namespace_nr(client));
 951         cfs_mutex_up(ldlm_namespace_lock(client));
 952 }
 953
 954 /* Unregister @ns from the list of namespaces */
 955 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 956 {
 957         cfs_mutex_down(ldlm_namespace_lock(client));
 958         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 959         /*
 960          * Some asserts and possibly other parts of code still using
 961          * list_empty(&ns->ns_list_chain). This is why it is important
 962          * to use list_del_init() here.
 963          */
 964         cfs_list_del_init(&ns->ns_list_chain);
 965         cfs_atomic_dec(ldlm_namespace_nr(client));
 966         cfs_mutex_up(ldlm_namespace_lock(client));
 967 }
 968
 969 /* Should be called under ldlm_namespace_lock(client) taken */
 970 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 971 {
 972         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 973         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 974         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 975 }
 976
 977 /* Should be called under ldlm_namespace_lock(client) taken */
 978 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 979 {
 980         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 981         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 982         return container_of(ldlm_namespace_list(client)->next,
 983                 struct ldlm_namespace, ns_list_chain);
 984 }
 985
 986 static struct ldlm_resource *ldlm_resource_new(void)
 987 {
 988         struct ldlm_resource *res;
 989         int idx;
 990
 991         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 992         if (res == NULL)
 993                 return NULL;
 994
 995         CFS_INIT_LIST_HEAD(&res->lr_granted);
 996         CFS_INIT_LIST_HEAD(&res->lr_converting);
 997         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 998
 999         /* initialize interval trees for each lock mode*/
1000         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1001                 res->lr_itree[idx].lit_size = 0;
1002                 res->lr_itree[idx].lit_mode = 1 << idx;
1003                 res->lr_itree[idx].lit_root = NULL;
1004         }
1005
1006         cfs_atomic_set(&res->lr_refcount, 1);
1007         cfs_spin_lock_init(&res->lr_lock);
1008         lu_ref_init(&res->lr_reference);
1009
1010         /* one who creates the resource must unlock
1011          * the semaphore after lvb initialization */
1012         cfs_init_mutex_locked(&res->lr_lvb_sem);
1013
1014         return res;
1015 }
1016
1017 /* Args: unlocked namespace
1018  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1019  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1020 struct ldlm_resource *
1021 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1022                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1023 {
1024         cfs_hlist_node_t     *hnode;
1025         struct ldlm_resource *res;
1026         cfs_hash_bd_t         bd;
1027         __u64                 version;
1028
1029         LASSERT(ns != NULL);
1030         LASSERT(parent == NULL);
1031         LASSERT(ns->ns_rs_hash != NULL);
1032         LASSERT(name->name[0] != 0);
1033
1034         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1035         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1036         if (hnode != NULL) {
1037                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1038                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1039                 /* synchronize WRT resource creation */
1040                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1041                         cfs_down(&res->lr_lvb_sem);
1042                         cfs_up(&res->lr_lvb_sem);
1043                 }
1044                 return res;
1045         }
1046
1047         version = cfs_hash_bd_version_get(&bd);
1048         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1049
1050         if (create == 0)
1051                 return NULL;
1052
1053         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1054                  "type: %d\n", type);
1055         res = ldlm_resource_new();
1056         if (!res)
1057                 return NULL;
1058
1059         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1060         res->lr_name       = *name;
1061         res->lr_type       = type;
1062         res->lr_most_restr = LCK_NL;
1063
1064         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1065         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1066                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1067
1068         if (hnode != NULL) {
1069                 /* someone won the race and added the resource before */
1070                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1071                 /* clean lu_ref for failed resource */
1072                 lu_ref_fini(&res->lr_reference);
1073                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1074
1075                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1076                 /* synchronize WRT resource creation */
1077                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1078                         cfs_down(&res->lr_lvb_sem);
1079                         cfs_up(&res->lr_lvb_sem);
1080                 }
1081                 return res;
1082         }
1083         /* we won! let's add the resource */
1084         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1085         if (cfs_hash_bd_count_get(&bd) == 1)
1086                 ldlm_namespace_get(ns);
1087
1088         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1089         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1090                 int rc;
1091
1092                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1093                 rc = ns->ns_lvbo->lvbo_init(res);
1094                 if (rc)
1095                         CERROR("lvbo_init failed for resource "
1096                                LPU64": rc %d\n", name->name[0], rc);
1097                 /* we create resource with locked lr_lvb_sem */
1098                 cfs_up(&res->lr_lvb_sem);
1099         }
1100
1101         return res;
1102 }
1103
1104 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1105 {
1106         LASSERT(res != NULL);
1107         LASSERT(res != LP_POISON);
1108         cfs_atomic_inc(&res->lr_refcount);
1109         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1110                cfs_atomic_read(&res->lr_refcount));
1111         return res;
1112 }
1113
1114 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1115                                          struct ldlm_resource *res)
1116 {
1117         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1118
1119         if (!cfs_list_empty(&res->lr_granted)) {
1120                 ldlm_resource_dump(D_ERROR, res);
1121                 LBUG();
1122         }
1123
1124         if (!cfs_list_empty(&res->lr_converting)) {
1125                 ldlm_resource_dump(D_ERROR, res);
1126                 LBUG();
1127         }
1128
1129         if (!cfs_list_empty(&res->lr_waiting)) {
1130                 ldlm_resource_dump(D_ERROR, res);
1131                 LBUG();
1132         }
1133
1134         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1135                                bd, &res->lr_hash);
1136         lu_ref_fini(&res->lr_reference);
1137         if (cfs_hash_bd_count_get(bd) == 0)
1138                 ldlm_namespace_put(nsb->nsb_namespace);
1139 }
1140
1141 /* Returns 1 if the resource was freed, 0 if it remains. */
1142 int ldlm_resource_putref(struct ldlm_resource *res)
1143 {
1144         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1145         cfs_hash_bd_t   bd;
1146
1147         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1148         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1149                res, cfs_atomic_read(&res->lr_refcount) - 1);
1150
1151         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1152         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1153                 __ldlm_resource_putref_final(&bd, res);
1154                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1155                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1156                         ns->ns_lvbo->lvbo_free(res);
1157                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1158                 return 1;
1159         }
1160         return 0;
1161 }
1162
1163 /* Returns 1 if the resource was freed, 0 if it remains. */
1164 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1165 {
1166         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1167
1168         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1169         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1170                res, cfs_atomic_read(&res->lr_refcount) - 1);
1171
1172         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1173                 cfs_hash_bd_t bd;
1174
1175                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1176                                 &res->lr_name, &bd);
1177                 __ldlm_resource_putref_final(&bd, res);
1178                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1179                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1180                  * so we should never be here while calling cfs_hash_del,
1181                  * cfs_hash_for_each_nolock is the only case we can get
1182                  * here, which is safe to release cfs_hash_bd_lock.
1183                  */
1184                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1185                         ns->ns_lvbo->lvbo_free(res);
1186                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1187
1188                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1189                 return 1;
1190         }
1191         return 0;
1192 }
1193
1194 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1195                             struct ldlm_lock *lock)
1196 {
1197         check_res_locked(res);
1198
1199         CDEBUG(D_OTHER, "About to add this lock:\n");
1200         ldlm_lock_dump(D_OTHER, lock, 0);
1201
1202         if (lock->l_destroyed) {
1203                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1204                 return;
1205         }
1206
1207         LASSERT(cfs_list_empty(&lock->l_res_link));
1208
1209         cfs_list_add_tail(&lock->l_res_link, head);
1210 }
1211
1212 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1213                                      struct ldlm_lock *new)
1214 {
1215         struct ldlm_resource *res = original->l_resource;
1216
1217         check_res_locked(res);
1218
1219         ldlm_resource_dump(D_INFO, res);
1220         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1221         ldlm_lock_dump(D_OTHER, new, 0);
1222
1223         if (new->l_destroyed) {
1224                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1225                 goto out;
1226         }
1227
1228         LASSERT(cfs_list_empty(&new->l_res_link));
1229
1230         cfs_list_add(&new->l_res_link, &original->l_res_link);
1231  out:;
1232 }
1233
1234 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1235 {
1236         int type = lock->l_resource->lr_type;
1237
1238         check_res_locked(lock->l_resource);
1239         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1240                 ldlm_unlink_lock_skiplist(lock);
1241         else if (type == LDLM_EXTENT)
1242                 ldlm_extent_unlink_lock(lock);
1243         cfs_list_del_init(&lock->l_res_link);
1244 }
1245
1246 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1247 {
1248         desc->lr_type = res->lr_type;
1249         desc->lr_name = res->lr_name;
1250 }
1251
1252 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1253 {
1254         cfs_list_t *tmp;
1255
1256         if (!((libcfs_debug | D_ERROR) & level))
1257                 return;
1258
1259         cfs_mutex_down(ldlm_namespace_lock(client));
1260
1261         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1262                 struct ldlm_namespace *ns;
1263                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1264                 ldlm_namespace_dump(level, ns);
1265         }
1266
1267         cfs_mutex_up(ldlm_namespace_lock(client));
1268 }
1269
1270 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1271                               cfs_hlist_node_t *hnode, void *arg)
1272 {
1273         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1274         int    level = (int)(unsigned long)arg;
1275
1276         lock_res(res);
1277         ldlm_resource_dump(level, res);
1278         unlock_res(res);
1279
1280         return 0;
1281 }
1282
1283 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1284 {
1285         if (!((libcfs_debug | D_ERROR) & level))
1286                 return;
1287
1288         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1289                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1290                ns_is_client(ns) ? "client" : "server");
1291
1292         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1293                 return;
1294
1295         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1296                                  ldlm_res_hash_dump,
1297                                  (void *)(unsigned long)level);
1298         cfs_spin_lock(&ns->ns_lock);
1299         ns->ns_next_dump = cfs_time_shift(10);
1300         cfs_spin_unlock(&ns->ns_lock);
1301 }
1302
1303 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1304 {
1305         cfs_list_t *tmp;
1306         int pos;
1307
1308         CLASSERT(RES_NAME_SIZE == 4);
1309
1310         if (!((libcfs_debug | D_ERROR) & level))
1311                 return;
1312
1313         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1314                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1315                res->lr_name.name[2], res->lr_name.name[3],
1316                cfs_atomic_read(&res->lr_refcount));
1317
1318         if (!cfs_list_empty(&res->lr_granted)) {
1319                 pos = 0;
1320                 CDEBUG(level, "Granted locks:\n");
1321                 cfs_list_for_each(tmp, &res->lr_granted) {
1322                         struct ldlm_lock *lock;
1323                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1324                                               l_res_link);
1325                         ldlm_lock_dump(level, lock, ++pos);
1326                 }
1327         }
1328         if (!cfs_list_empty(&res->lr_converting)) {
1329                 pos = 0;
1330                 CDEBUG(level, "Converting locks:\n");
1331                 cfs_list_for_each(tmp, &res->lr_converting) {
1332                         struct ldlm_lock *lock;
1333                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1334                                               l_res_link);
1335                         ldlm_lock_dump(level, lock, ++pos);
1336                 }
1337         }
1338         if (!cfs_list_empty(&res->lr_waiting)) {
1339                 pos = 0;
1340                 CDEBUG(level, "Waiting locks:\n");
1341                 cfs_list_for_each(tmp, &res->lr_waiting) {
1342                         struct ldlm_lock *lock;
1343                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1344                                               l_res_link);
1345                         ldlm_lock_dump(level, lock, ++pos);
1346                 }
1347         }
1348 }