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         return 0;
 370 }
 371 #undef MAX_STRING_SIZE
 372 #else /* LPROCFS */
 373
 374 #define ldlm_namespace_proc_unregister(ns)      ({;})
 375 #define ldlm_namespace_proc_register(ns)        ({0;})
 376
 377 #endif /* LPROCFS */
 378
 379 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs, void *key, unsigned mask)
 380 {
 381         struct ldlm_res_id     *id  = key;
 382         unsigned                val = 0;
 383         unsigned                i;
 384
 385         for (i = 0; i < RES_NAME_SIZE; i++)
 386                 val += id->name[i];
 387         return val & mask;
 388 }
 389
 390 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs, void *key, unsigned mask)
 391 {
 392         struct ldlm_res_id *id = key;
 393         struct lu_fid       fid;
 394         __u64               hash;
 395
 396         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 397         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF];
 398         fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF];
 399
 400         hash = fid_flatten(&fid);
 401         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 402         /* ignore a few low bits */
 403         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0)
 404                 hash += id->name[LUSTRE_RES_ID_HSH_OFF] >> 5;
 405         else
 406                 hash = hash >> 5;
 407         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 408         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 409
 410         return hash & mask;
 411 }
 412
 413 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 414 {
 415         struct ldlm_resource   *res;
 416
 417         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 418         return &res->lr_name;
 419 }
 420
 421 static int ldlm_res_eq(const struct ldlm_res_id *res0,
 422                        const struct ldlm_res_id *res1)
 423 {
 424         return !memcmp(res0, res1, sizeof(*res0));
 425 }
 426
 427 static int ldlm_res_hop_keycmp(void *key, cfs_hlist_node_t *hnode)
 428 {
 429         struct ldlm_resource   *res;
 430
 431         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 432         return ldlm_res_eq((const struct ldlm_res_id *)key,
 433                            (const struct ldlm_res_id *)&res->lr_name);
 434 }
 435
 436 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 437 {
 438         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 439 }
 440
 441 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 442 {
 443         struct ldlm_resource *res;
 444
 445         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 446         ldlm_resource_getref(res);
 447         LDLM_RESOURCE_ADDREF(res);
 448 }
 449
 450 static void ldlm_res_hop_put_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         /* cfs_hash_for_each_nolock is the only chance we call it */
 456         LDLM_RESOURCE_DELREF(res);
 457         ldlm_resource_putref_locked(res);
 458 }
 459
 460 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 461 {
 462         struct ldlm_resource *res;
 463
 464         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 465         LDLM_RESOURCE_DELREF(res);
 466         ldlm_resource_putref(res);
 467 }
 468
 469 cfs_hash_ops_t ldlm_ns_hash_ops = {
 470         .hs_hash        = ldlm_res_hop_hash,
 471         .hs_key         = ldlm_res_hop_key,
 472         .hs_keycmp      = ldlm_res_hop_keycmp,
 473         .hs_keycpy      = NULL,
 474         .hs_object      = ldlm_res_hop_object,
 475         .hs_get         = ldlm_res_hop_get_locked,
 476         .hs_put_locked  = ldlm_res_hop_put_locked,
 477         .hs_put         = ldlm_res_hop_put
 478 };
 479
 480 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 481         .hs_hash        = ldlm_res_hop_fid_hash,
 482         .hs_key         = ldlm_res_hop_key,
 483         .hs_keycmp      = ldlm_res_hop_keycmp,
 484         .hs_keycpy      = NULL,
 485         .hs_object      = ldlm_res_hop_object,
 486         .hs_get         = ldlm_res_hop_get_locked,
 487         .hs_put_locked  = ldlm_res_hop_put_locked,
 488         .hs_put         = ldlm_res_hop_put
 489 };
 490
 491 typedef struct {
 492         ldlm_ns_type_t  nsd_type;
 493         /** hash bucket bits */
 494         unsigned        nsd_bkt_bits;
 495         /** hash bits */
 496         unsigned        nsd_all_bits;
 497         /** hash operations */
 498         cfs_hash_ops_t *nsd_hops;
 499 } ldlm_ns_hash_def_t;
 500
 501 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 502 {
 503         {
 504                 .nsd_type       = LDLM_NS_TYPE_MDC,
 505                 .nsd_bkt_bits   = 11,
 506                 .nsd_all_bits   = 15,
 507                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 508         },
 509         {
 510                 .nsd_type       = LDLM_NS_TYPE_MDT,
 511                 .nsd_bkt_bits   = 14,
 512                 .nsd_all_bits   = 21,
 513                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 514         },
 515         {
 516                 .nsd_type       = LDLM_NS_TYPE_OSC,
 517                 .nsd_bkt_bits   = 8,
 518                 .nsd_all_bits   = 12,
 519                 .nsd_hops       = &ldlm_ns_hash_ops,
 520         },
 521         {
 522                 .nsd_type       = LDLM_NS_TYPE_OST,
 523                 .nsd_bkt_bits   = 11,
 524                 .nsd_all_bits   = 17,
 525                 .nsd_hops       = &ldlm_ns_hash_ops,
 526         },
 527         {
 528                 .nsd_type       = LDLM_NS_TYPE_MGC,
 529                 .nsd_bkt_bits   = 4,
 530                 .nsd_all_bits   = 4,
 531                 .nsd_hops       = &ldlm_ns_hash_ops,
 532         },
 533         {
 534                 .nsd_type       = LDLM_NS_TYPE_MGT,
 535                 .nsd_bkt_bits   = 4,
 536                 .nsd_all_bits   = 4,
 537                 .nsd_hops       = &ldlm_ns_hash_ops,
 538         },
 539         {
 540                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 541         },
 542 };
 543
 544 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 545                                           ldlm_side_t client,
 546                                           ldlm_appetite_t apt,
 547                                           ldlm_ns_type_t ns_type)
 548 {
 549         struct ldlm_namespace *ns = NULL;
 550         struct ldlm_ns_bucket *nsb;
 551         ldlm_ns_hash_def_t    *nsd;
 552         cfs_hash_bd_t          bd;
 553         int                    idx;
 554         int                    rc;
 555         ENTRY;
 556
 557         LASSERT(obd != NULL);
 558
 559         rc = ldlm_get_ref();
 560         if (rc) {
 561                 CERROR("ldlm_get_ref failed: %d\n", rc);
 562                 RETURN(NULL);
 563         }
 564
 565         for (idx = 0;;idx++) {
 566                 nsd = &ldlm_ns_hash_defs[idx];
 567                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 568                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 569                         GOTO(out_ref, NULL);
 570                 }
 571
 572                 if (nsd->nsd_type == ns_type)
 573                         break;
 574         }
 575
 576         OBD_ALLOC_PTR(ns);
 577         if (!ns)
 578                 GOTO(out_ref, NULL);
 579
 580         ns->ns_rs_hash = cfs_hash_create(name,
 581                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 582                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 583                                          CFS_HASH_MIN_THETA,
 584                                          CFS_HASH_MAX_THETA,
 585                                          nsd->nsd_hops,
 586                                          CFS_HASH_DEPTH |
 587                                          CFS_HASH_BIGNAME |
 588                                          CFS_HASH_SPIN_BKTLOCK |
 589                                          CFS_HASH_NO_ITEMREF);
 590         if (ns->ns_rs_hash == NULL)
 591                 GOTO(out_ns, NULL);
 592
 593         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 594                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 595                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 596                 nsb->nsb_namespace = ns;
 597         }
 598
 599         ns->ns_obd      = obd;
 600         ns->ns_appetite = apt;
 601         ns->ns_client   = client;
 602
 603         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 604         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 605         cfs_spin_lock_init(&ns->ns_lock);
 606         cfs_atomic_set(&ns->ns_bref, 0);
 607         cfs_waitq_init(&ns->ns_waitq);
 608
 609         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 610         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 611         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 612
 613         ns->ns_nr_unused          = 0;
 614         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 615         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 616         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 617         ns->ns_timeouts           = 0;
 618         ns->ns_orig_connect_flags = 0;
 619         ns->ns_connect_flags      = 0;
 620         rc = ldlm_namespace_proc_register(ns);
 621         if (rc != 0) {
 622                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 623                 GOTO(out_hash, rc);
 624         }
 625
 626         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 627         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 628         if (rc) {
 629                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 630                 GOTO(out_proc, rc);
 631         }
 632
 633         ldlm_namespace_register(ns, client);
 634         RETURN(ns);
 635 out_proc:
 636         ldlm_namespace_proc_unregister(ns);
 637         ldlm_namespace_cleanup(ns, 0);
 638 out_hash:
 639         cfs_hash_putref(ns->ns_rs_hash);
 640 out_ns:
 641         OBD_FREE_PTR(ns);
 642 out_ref:
 643         ldlm_put_ref();
 644         RETURN(NULL);
 645 }
 646
 647 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 648
 649 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 650  * This is currently only used for recovery, and we make certain assumptions
 651  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 652 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 653                              int flags)
 654 {
 655         cfs_list_t *tmp;
 656         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 657         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 658
 659         do {
 660                 struct ldlm_lock *lock = NULL;
 661
 662                 /* first, we look for non-cleaned-yet lock
 663                  * all cleaned locks are marked by CLEANED flag */
 664                 lock_res(res);
 665                 cfs_list_for_each(tmp, q) {
 666                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 667                                               l_res_link);
 668                         if (lock->l_flags & LDLM_FL_CLEANED) {
 669                                 lock = NULL;
 670                                 continue;
 671                         }
 672                         LDLM_LOCK_GET(lock);
 673                         lock->l_flags |= LDLM_FL_CLEANED;
 674                         break;
 675                 }
 676
 677                 if (lock == NULL) {
 678                         unlock_res(res);
 679                         break;
 680                 }
 681
 682                 /* Set CBPENDING so nothing in the cancellation path
 683                  * can match this lock */
 684                 lock->l_flags |= LDLM_FL_CBPENDING;
 685                 lock->l_flags |= LDLM_FL_FAILED;
 686                 lock->l_flags |= flags;
 687
 688                 /* ... without sending a CANCEL message for local_only. */
 689                 if (local_only)
 690                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 691
 692                 if (local_only && (lock->l_readers || lock->l_writers)) {
 693                         /* This is a little bit gross, but much better than the
 694                          * alternative: pretend that we got a blocking AST from
 695                          * the server, so that when the lock is decref'd, it
 696                          * will go away ... */
 697                         unlock_res(res);
 698                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 699                         if (lock->l_completion_ast)
 700                                 lock->l_completion_ast(lock, 0, NULL);
 701                         LDLM_LOCK_RELEASE(lock);
 702                         continue;
 703                 }
 704
 705                 if (client) {
 706                         struct lustre_handle lockh;
 707
 708                         unlock_res(res);
 709                         ldlm_lock2handle(lock, &lockh);
 710                         rc = ldlm_cli_cancel(&lockh);
 711                         if (rc)
 712                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 713                 } else {
 714                         ldlm_resource_unlink_lock(lock);
 715                         unlock_res(res);
 716                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 717                                    "client node");
 718                         ldlm_lock_destroy(lock);
 719                 }
 720                 LDLM_LOCK_RELEASE(lock);
 721         } while (1);
 722 }
 723
 724 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 725                                cfs_hlist_node_t *hnode, void *arg)
 726 {
 727         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 728         int    flags = (int)(unsigned long)arg;
 729
 730         cleanup_resource(res, &res->lr_granted, flags);
 731         cleanup_resource(res, &res->lr_converting, flags);
 732         cleanup_resource(res, &res->lr_waiting, flags);
 733
 734         return 0;
 735 }
 736
 737 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 738                                   cfs_hlist_node_t *hnode, void *arg)
 739 {
 740         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 741
 742         CERROR("Namespace %s resource refcount nonzero "
 743                "(%d) after lock cleanup; forcing "
 744                "cleanup.\n",
 745                ldlm_ns_name(ldlm_res_to_ns(res)),
 746                cfs_atomic_read(&res->lr_refcount) - 1);
 747
 748         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 749                LPU64") (rc: %d)\n", res,
 750                res->lr_name.name[0], res->lr_name.name[1],
 751                res->lr_name.name[2], res->lr_name.name[3],
 752                cfs_atomic_read(&res->lr_refcount) - 1);
 753         return 0;
 754 }
 755
 756 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 757 {
 758         if (ns == NULL) {
 759                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 760                 return ELDLM_OK;
 761         }
 762
 763         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 764                                  (void *)(unsigned long)flags);
 765         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 766         return ELDLM_OK;
 767 }
 768
 769 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 770 {
 771         ENTRY;
 772
 773         /* At shutdown time, don't call the cancellation callback */
 774         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 775
 776         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 777                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 778                 int rc;
 779                 CDEBUG(D_DLMTRACE,
 780                        "dlm namespace %s free waiting on refcount %d\n",
 781                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 782 force_wait:
 783                 if (force)
 784                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 785
 786                 rc = l_wait_event(ns->ns_waitq,
 787                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 788
 789                 /* Forced cleanups should be able to reclaim all references,
 790                  * so it's safe to wait forever... we can't leak locks... */
 791                 if (force && rc == -ETIMEDOUT) {
 792                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 793                                        "namespace with %d resources in use, "
 794                                        "(rc=%d)\n", ldlm_ns_name(ns),
 795                                        cfs_atomic_read(&ns->ns_bref), rc);
 796                         GOTO(force_wait, rc);
 797                 }
 798
 799                 if (cfs_atomic_read(&ns->ns_bref)) {
 800                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 801                                        "with %d resources in use, (rc=%d)\n",
 802                                        ldlm_ns_name(ns),
 803                                        cfs_atomic_read(&ns->ns_bref), rc);
 804                         RETURN(ELDLM_NAMESPACE_EXISTS);
 805                 }
 806                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 807                        ldlm_ns_name(ns));
 808         }
 809
 810         RETURN(ELDLM_OK);
 811 }
 812
 813 /**
 814  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 815  * for freeing. Waits for refc == 0.
 816  *
 817  * The following is done:
 818  * (0) Unregister \a ns from its list to make inaccessible for potential users
 819  * like pools thread and others;
 820  * (1) Clear all locks in \a ns.
 821  */
 822 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 823                                struct obd_import *imp,
 824                                int force)
 825 {
 826         int rc;
 827         ENTRY;
 828         if (!ns) {
 829                 EXIT;
 830                 return;
 831         }
 832
 833
 834         /*
 835          * Can fail with -EINTR when force == 0 in which case try harder.
 836          */
 837         rc = __ldlm_namespace_free(ns, force);
 838         if (rc != ELDLM_OK) {
 839                 if (imp) {
 840                         ptlrpc_disconnect_import(imp, 0);
 841                         ptlrpc_invalidate_import(imp);
 842                 }
 843
 844                 /*
 845                  * With all requests dropped and the import inactive
 846                  * we are gaurenteed all reference will be dropped.
 847                  */
 848                 rc = __ldlm_namespace_free(ns, 1);
 849                 LASSERT(rc == 0);
 850         }
 851         EXIT;
 852 }
 853
 854 /**
 855  * Performs freeing memory structures related to \a ns. This is only done when
 856  * ldlm_namespce_free_prior() successfully removed all resources referencing
 857  * \a ns and its refc == 0.
 858  */
 859 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 860 {
 861         ENTRY;
 862         if (!ns) {
 863                 EXIT;
 864                 return;
 865         }
 866
 867
 868         /*
 869          * Make sure that nobody can find this ns in its list.
 870          */
 871         ldlm_namespace_unregister(ns, ns->ns_client);
 872         /*
 873          * Fini pool _before_ parent proc dir is removed. This is important as
 874          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 875          * it after @dir may cause oops.
 876          */
 877         ldlm_pool_fini(&ns->ns_pool);
 878
 879         ldlm_namespace_proc_unregister(ns);
 880         cfs_hash_putref(ns->ns_rs_hash);
 881         /*
 882          * Namespace \a ns should be not on list in this time, otherwise this
 883          * will cause issues realted to using freed \a ns in pools thread.
 884          */
 885         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 886         OBD_FREE_PTR(ns);
 887         ldlm_put_ref();
 888         EXIT;
 889 }
 890
 891
 892 /* Cleanup the resource, and free namespace.
 893  * bug 12864:
 894  * Deadlock issue:
 895  * proc1: destroy import
 896  *        class_disconnect_export(grab cl_sem) ->
 897  *              -> ldlm_namespace_free ->
 898  *              -> lprocfs_remove(grab _lprocfs_lock).
 899  * proc2: read proc info
 900  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 901  *              -> osc_rd_active, etc(grab cl_sem).
 902  *
 903  * So that I have to split the ldlm_namespace_free into two parts - the first
 904  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 905  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 906  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 907  * held.
 908  */
 909 void ldlm_namespace_free(struct ldlm_namespace *ns,
 910                          struct obd_import *imp,
 911                          int force)
 912 {
 913         ldlm_namespace_free_prior(ns, imp, force);
 914         ldlm_namespace_free_post(ns);
 915 }
 916
 917 void ldlm_namespace_get(struct ldlm_namespace *ns)
 918 {
 919         cfs_atomic_inc(&ns->ns_bref);
 920 }
 921
 922 void ldlm_namespace_put(struct ldlm_namespace *ns)
 923 {
 924         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 925                 cfs_waitq_signal(&ns->ns_waitq);
 926                 cfs_spin_unlock(&ns->ns_lock);
 927         }
 928 }
 929
 930 /* Register @ns in the list of namespaces */
 931 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 932 {
 933         cfs_mutex_down(ldlm_namespace_lock(client));
 934         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 935         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 936         cfs_atomic_inc(ldlm_namespace_nr(client));
 937         cfs_mutex_up(ldlm_namespace_lock(client));
 938 }
 939
 940 /* Unregister @ns from the list of namespaces */
 941 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 942 {
 943         cfs_mutex_down(ldlm_namespace_lock(client));
 944         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 945         /*
 946          * Some asserts and possibly other parts of code still using
 947          * list_empty(&ns->ns_list_chain). This is why it is important
 948          * to use list_del_init() here.
 949          */
 950         cfs_list_del_init(&ns->ns_list_chain);
 951         cfs_atomic_dec(ldlm_namespace_nr(client));
 952         cfs_mutex_up(ldlm_namespace_lock(client));
 953 }
 954
 955 /* Should be called under ldlm_namespace_lock(client) taken */
 956 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 957 {
 958         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 959         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 960         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 961 }
 962
 963 /* Should be called under ldlm_namespace_lock(client) taken */
 964 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 965 {
 966         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 967         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 968         return container_of(ldlm_namespace_list(client)->next,
 969                 struct ldlm_namespace, ns_list_chain);
 970 }
 971
 972 static struct ldlm_resource *ldlm_resource_new(void)
 973 {
 974         struct ldlm_resource *res;
 975         int idx;
 976
 977         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 978         if (res == NULL)
 979                 return NULL;
 980
 981         CFS_INIT_LIST_HEAD(&res->lr_granted);
 982         CFS_INIT_LIST_HEAD(&res->lr_converting);
 983         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 984
 985         /* initialize interval trees for each lock mode*/
 986         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 987                 res->lr_itree[idx].lit_size = 0;
 988                 res->lr_itree[idx].lit_mode = 1 << idx;
 989                 res->lr_itree[idx].lit_root = NULL;
 990         }
 991
 992         cfs_atomic_set(&res->lr_refcount, 1);
 993         cfs_spin_lock_init(&res->lr_lock);
 994         lu_ref_init(&res->lr_reference);
 995
 996         /* one who creates the resource must unlock
 997          * the semaphore after lvb initialization */
 998         cfs_init_mutex_locked(&res->lr_lvb_sem);
 999
1000         return res;
1001 }
1002
1003 /* Args: unlocked namespace
1004  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1005  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1006 struct ldlm_resource *
1007 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1008                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1009 {
1010         cfs_hlist_node_t     *hnode;
1011         struct ldlm_resource *res;
1012         cfs_hash_bd_t         bd;
1013         __u64                 version;
1014
1015         LASSERT(ns != NULL);
1016         LASSERT(parent == NULL);
1017         LASSERT(ns->ns_rs_hash != NULL);
1018         LASSERT(name->name[0] != 0);
1019
1020         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1021         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1022         if (hnode != NULL) {
1023                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1024                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1025                 /* synchronize WRT resource creation */
1026                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1027                         cfs_down(&res->lr_lvb_sem);
1028                         cfs_up(&res->lr_lvb_sem);
1029                 }
1030                 return res;
1031         }
1032
1033         version = cfs_hash_bd_version_get(&bd);
1034         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1035
1036         if (create == 0)
1037                 return NULL;
1038
1039         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1040                  "type: %d\n", type);
1041         res = ldlm_resource_new();
1042         if (!res)
1043                 return NULL;
1044
1045         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1046         res->lr_name       = *name;
1047         res->lr_type       = type;
1048         res->lr_most_restr = LCK_NL;
1049
1050         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1051         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1052                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1053
1054         if (hnode != NULL) {
1055                 /* someone won the race and added the resource before */
1056                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1057                 /* clean lu_ref for failed resource */
1058                 lu_ref_fini(&res->lr_reference);
1059                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1060
1061                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1062                 /* synchronize WRT resource creation */
1063                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1064                         cfs_down(&res->lr_lvb_sem);
1065                         cfs_up(&res->lr_lvb_sem);
1066                 }
1067                 return res;
1068         }
1069         /* we won! let's add the resource */
1070         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1071         if (cfs_hash_bd_count_get(&bd) == 1)
1072                 ldlm_namespace_get(ns);
1073
1074         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1075         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1076                 int rc;
1077
1078                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1079                 rc = ns->ns_lvbo->lvbo_init(res);
1080                 if (rc)
1081                         CERROR("lvbo_init failed for resource "
1082                                LPU64": rc %d\n", name->name[0], rc);
1083                 /* we create resource with locked lr_lvb_sem */
1084                 cfs_up(&res->lr_lvb_sem);
1085         }
1086
1087         return res;
1088 }
1089
1090 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1091 {
1092         LASSERT(res != NULL);
1093         LASSERT(res != LP_POISON);
1094         cfs_atomic_inc(&res->lr_refcount);
1095         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1096                cfs_atomic_read(&res->lr_refcount));
1097         return res;
1098 }
1099
1100 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1101                                          struct ldlm_resource *res)
1102 {
1103         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1104
1105         if (!cfs_list_empty(&res->lr_granted)) {
1106                 ldlm_resource_dump(D_ERROR, res);
1107                 LBUG();
1108         }
1109
1110         if (!cfs_list_empty(&res->lr_converting)) {
1111                 ldlm_resource_dump(D_ERROR, res);
1112                 LBUG();
1113         }
1114
1115         if (!cfs_list_empty(&res->lr_waiting)) {
1116                 ldlm_resource_dump(D_ERROR, res);
1117                 LBUG();
1118         }
1119
1120         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1121                                bd, &res->lr_hash);
1122         lu_ref_fini(&res->lr_reference);
1123         if (cfs_hash_bd_count_get(bd) == 0)
1124                 ldlm_namespace_put(nsb->nsb_namespace);
1125 }
1126
1127 /* Returns 1 if the resource was freed, 0 if it remains. */
1128 int ldlm_resource_putref(struct ldlm_resource *res)
1129 {
1130         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1131         int ref = cfs_atomic_read(&res->lr_refcount);
1132         cfs_hash_bd_t   bd;
1133
1134         CDEBUG(D_INFO, "putref res: %p count: %d\n", res, ref - 1);
1135         LASSERTF(ref > 0 && ref < LI_POISON, "%d", ref);
1136         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1137         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1138                 __ldlm_resource_putref_final(&bd, res);
1139                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1140                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1141                         ns->ns_lvbo->lvbo_free(res);
1142                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1143                 return 1;
1144         }
1145         return 0;
1146 }
1147
1148 /* Returns 1 if the resource was freed, 0 if it remains. */
1149 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1150 {
1151         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1152         int ref = cfs_atomic_read(&res->lr_refcount);
1153
1154         CDEBUG(D_INFO, "putref res: %p count: %d\n", res, ref - 1);
1155         LASSERTF(ref > 0 && ref < LI_POISON, "%d", ref);
1156         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1157                 cfs_hash_bd_t bd;
1158
1159                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1160                                 &res->lr_name, &bd);
1161                 __ldlm_resource_putref_final(&bd, res);
1162                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1163                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1164                  * so we should never be here while calling cfs_hash_del,
1165                  * cfs_hash_for_each_nolock is the only case we can get
1166                  * here, which is safe to release cfs_hash_bd_lock.
1167                  */
1168                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1169                         ns->ns_lvbo->lvbo_free(res);
1170                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1171
1172                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1173                 return 1;
1174         }
1175         return 0;
1176 }
1177
1178 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1179                             struct ldlm_lock *lock)
1180 {
1181         check_res_locked(res);
1182
1183         ldlm_resource_dump(D_INFO, res);
1184         CDEBUG(D_OTHER, "About to add this lock:\n");
1185         ldlm_lock_dump(D_OTHER, lock, 0);
1186
1187         if (lock->l_destroyed) {
1188                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1189                 return;
1190         }
1191
1192         LASSERT(cfs_list_empty(&lock->l_res_link));
1193
1194         cfs_list_add_tail(&lock->l_res_link, head);
1195 }
1196
1197 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1198                                      struct ldlm_lock *new)
1199 {
1200         struct ldlm_resource *res = original->l_resource;
1201
1202         check_res_locked(res);
1203
1204         ldlm_resource_dump(D_INFO, res);
1205         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1206         ldlm_lock_dump(D_OTHER, new, 0);
1207
1208         if (new->l_destroyed) {
1209                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1210                 goto out;
1211         }
1212
1213         LASSERT(cfs_list_empty(&new->l_res_link));
1214
1215         cfs_list_add(&new->l_res_link, &original->l_res_link);
1216  out:;
1217 }
1218
1219 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1220 {
1221         int type = lock->l_resource->lr_type;
1222
1223         check_res_locked(lock->l_resource);
1224         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1225                 ldlm_unlink_lock_skiplist(lock);
1226         else if (type == LDLM_EXTENT)
1227                 ldlm_extent_unlink_lock(lock);
1228         cfs_list_del_init(&lock->l_res_link);
1229 }
1230
1231 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1232 {
1233         desc->lr_type = res->lr_type;
1234         desc->lr_name = res->lr_name;
1235 }
1236
1237 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1238 {
1239         cfs_list_t *tmp;
1240
1241         if (!((libcfs_debug | D_ERROR) & level))
1242                 return;
1243
1244         cfs_mutex_down(ldlm_namespace_lock(client));
1245
1246         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1247                 struct ldlm_namespace *ns;
1248                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1249                 ldlm_namespace_dump(level, ns);
1250         }
1251
1252         cfs_mutex_up(ldlm_namespace_lock(client));
1253 }
1254
1255 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1256                               cfs_hlist_node_t *hnode, void *arg)
1257 {
1258         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1259         int    level = (int)(unsigned long)arg;
1260
1261         lock_res(res);
1262         ldlm_resource_dump(level, res);
1263         unlock_res(res);
1264
1265         return 0;
1266 }
1267
1268 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1269 {
1270         if (!((libcfs_debug | D_ERROR) & level))
1271                 return;
1272
1273         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1274                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1275                ns_is_client(ns) ? "client" : "server");
1276
1277         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1278                 return;
1279
1280         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1281                                  ldlm_res_hash_dump,
1282                                  (void *)(unsigned long)level);
1283         cfs_spin_lock(&ns->ns_lock);
1284         ns->ns_next_dump = cfs_time_shift(10);
1285         cfs_spin_unlock(&ns->ns_lock);
1286 }
1287
1288 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1289 {
1290         cfs_list_t *tmp;
1291         int pos;
1292
1293         CLASSERT(RES_NAME_SIZE == 4);
1294
1295         if (!((libcfs_debug | D_ERROR) & level))
1296                 return;
1297
1298         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1299                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1300                res->lr_name.name[2], res->lr_name.name[3],
1301                cfs_atomic_read(&res->lr_refcount));
1302
1303         if (!cfs_list_empty(&res->lr_granted)) {
1304                 pos = 0;
1305                 CDEBUG(level, "Granted locks:\n");
1306                 cfs_list_for_each(tmp, &res->lr_granted) {
1307                         struct ldlm_lock *lock;
1308                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1309                                               l_res_link);
1310                         ldlm_lock_dump(level, lock, ++pos);
1311                 }
1312         }
1313         if (!cfs_list_empty(&res->lr_converting)) {
1314                 pos = 0;
1315                 CDEBUG(level, "Converting locks:\n");
1316                 cfs_list_for_each(tmp, &res->lr_converting) {
1317                         struct ldlm_lock *lock;
1318                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1319                                               l_res_link);
1320                         ldlm_lock_dump(level, lock, ++pos);
1321                 }
1322         }
1323         if (!cfs_list_empty(&res->lr_waiting)) {
1324                 pos = 0;
1325                 CDEBUG(level, "Waiting locks:\n");
1326                 cfs_list_for_each(tmp, &res->lr_waiting) {
1327                         struct ldlm_lock *lock;
1328                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1329                                               l_res_link);
1330                         ldlm_lock_dump(level, lock, ++pos);
1331                 }
1332         }
1333 }