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  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/ldlm/ldlm_resource.c
  39  *
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  * Author: Peter Braam <braam@clusterfs.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45 #ifdef __KERNEL__
  46 # include <lustre_dlm.h>
  47 #else
  48 # include <liblustre.h>
  49 #endif
  50
  51 #include <lustre_fid.h>
  52 #include <obd_class.h>
  53 #include "ldlm_internal.h"
  54
  55 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  56
  57 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
  58 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
  59
  60 cfs_semaphore_t ldlm_srv_namespace_lock;
  61 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  62
  63 cfs_semaphore_t ldlm_cli_namespace_lock;
  64 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  65
  66 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  67 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  68 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  69
  70 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  71
  72 #ifdef LPROCFS
  73 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  74                              unsigned long count, void *data)
  75 {
  76         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  77         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  78         RETURN(count);
  79 }
  80
  81 int ldlm_proc_setup(void)
  82 {
  83         int rc;
  84         struct lprocfs_vars list[] = {
  85                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  86                 { "cancel_unused_locks_before_replay",
  87                   lprocfs_rd_uint, lprocfs_wr_uint,
  88                   &ldlm_cancel_unused_locks_before_replay, NULL },
  89                 { NULL }};
  90         ENTRY;
  91         LASSERT(ldlm_ns_proc_dir == NULL);
  92
  93         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  94                                               proc_lustre_root,
  95                                               NULL, NULL);
  96         if (IS_ERR(ldlm_type_proc_dir)) {
  97                 CERROR("LProcFS failed in ldlm-init\n");
  98                 rc = PTR_ERR(ldlm_type_proc_dir);
  99                 GOTO(err, rc);
 100         }
 101
 102         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 103                                             ldlm_type_proc_dir,
 104                                             NULL, NULL);
 105         if (IS_ERR(ldlm_ns_proc_dir)) {
 106                 CERROR("LProcFS failed in ldlm-init\n");
 107                 rc = PTR_ERR(ldlm_ns_proc_dir);
 108                 GOTO(err_type, rc);
 109         }
 110
 111         ldlm_svc_proc_dir = lprocfs_register("services",
 112                                             ldlm_type_proc_dir,
 113                                             NULL, NULL);
 114         if (IS_ERR(ldlm_svc_proc_dir)) {
 115                 CERROR("LProcFS failed in ldlm-init\n");
 116                 rc = PTR_ERR(ldlm_svc_proc_dir);
 117                 GOTO(err_ns, rc);
 118         }
 119
 120         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 121
 122         RETURN(0);
 123
 124 err_ns:
 125         lprocfs_remove(&ldlm_ns_proc_dir);
 126 err_type:
 127         lprocfs_remove(&ldlm_type_proc_dir);
 128 err:
 129         ldlm_svc_proc_dir = NULL;
 130         RETURN(rc);
 131 }
 132
 133 void ldlm_proc_cleanup(void)
 134 {
 135         if (ldlm_svc_proc_dir)
 136                 lprocfs_remove(&ldlm_svc_proc_dir);
 137
 138         if (ldlm_ns_proc_dir)
 139                 lprocfs_remove(&ldlm_ns_proc_dir);
 140
 141         if (ldlm_type_proc_dir)
 142                 lprocfs_remove(&ldlm_type_proc_dir);
 143 }
 144
 145 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 146                                    int count, int *eof, void *data)
 147 {
 148         struct ldlm_namespace *ns  = data;
 149         __u64                  res = 0;
 150         cfs_hash_bd_t          bd;
 151         int                    i;
 152
 153         /* result is not strictly consistant */
 154         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 155                 res += cfs_hash_bd_count_get(&bd);
 156         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 157 }
 158
 159 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 160                                int count, int *eof, void *data)
 161 {
 162         struct ldlm_namespace *ns = data;
 163         __u64                  locks;
 164
 165         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 166                                         LPROCFS_FIELDS_FLAGS_SUM);
 167         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 168 }
 169
 170 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 171                                int count, int *eof, void *data)
 172 {
 173         struct ldlm_namespace *ns = data;
 174         __u32 *nr = &ns->ns_max_unused;
 175
 176         if (ns_connect_lru_resize(ns))
 177                 nr = &ns->ns_nr_unused;
 178         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 179 }
 180
 181 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 182                                unsigned long count, void *data)
 183 {
 184         struct ldlm_namespace *ns = data;
 185         char dummy[MAX_STRING_SIZE + 1], *end;
 186         unsigned long tmp;
 187         int lru_resize;
 188
 189         dummy[MAX_STRING_SIZE] = '\0';
 190         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 191                 return -EFAULT;
 192
 193         if (strncmp(dummy, "clear", 5) == 0) {
 194                 CDEBUG(D_DLMTRACE,
 195                        "dropping all unused locks from namespace %s\n",
 196                        ldlm_ns_name(ns));
 197                 if (ns_connect_lru_resize(ns)) {
 198                         int canceled, unused  = ns->ns_nr_unused;
 199
 200                         /* Try to cancel all @ns_nr_unused locks. */
 201                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 202                                                    LDLM_CANCEL_PASSED);
 203                         if (canceled < unused) {
 204                                 CDEBUG(D_DLMTRACE,
 205                                        "not all requested locks are canceled, "
 206                                        "requested: %d, canceled: %d\n", unused,
 207                                        canceled);
 208                                 return -EINVAL;
 209                         }
 210                 } else {
 211                         tmp = ns->ns_max_unused;
 212                         ns->ns_max_unused = 0;
 213                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 214                         ns->ns_max_unused = tmp;
 215                 }
 216                 return count;
 217         }
 218
 219         tmp = simple_strtoul(dummy, &end, 0);
 220         if (dummy == end) {
 221                 CERROR("invalid value written\n");
 222                 return -EINVAL;
 223         }
 224         lru_resize = (tmp == 0);
 225
 226         if (ns_connect_lru_resize(ns)) {
 227                 if (!lru_resize)
 228                         ns->ns_max_unused = (unsigned int)tmp;
 229
 230                 if (tmp > ns->ns_nr_unused)
 231                         tmp = ns->ns_nr_unused;
 232                 tmp = ns->ns_nr_unused - tmp;
 233
 234                 CDEBUG(D_DLMTRACE,
 235                        "changing namespace %s unused locks from %u to %u\n",
 236                        ldlm_ns_name(ns), ns->ns_nr_unused,
 237                        (unsigned int)tmp);
 238                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 239
 240                 if (!lru_resize) {
 241                         CDEBUG(D_DLMTRACE,
 242                                "disable lru_resize for namespace %s\n",
 243                                ldlm_ns_name(ns));
 244                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 245                 }
 246         } else {
 247                 CDEBUG(D_DLMTRACE,
 248                        "changing namespace %s max_unused from %u to %u\n",
 249                        ldlm_ns_name(ns), ns->ns_max_unused,
 250                        (unsigned int)tmp);
 251                 ns->ns_max_unused = (unsigned int)tmp;
 252                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 253
 254                 /* Make sure that originally lru resize was supported before
 255                  * turning it on here. */
 256                 if (lru_resize &&
 257                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 258                         CDEBUG(D_DLMTRACE,
 259                                "enable lru_resize for namespace %s\n",
 260                                ldlm_ns_name(ns));
 261                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 262                 }
 263         }
 264
 265         return count;
 266 }
 267
 268 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 269 {
 270         struct proc_dir_entry *dir;
 271
 272         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 273         if (dir == NULL) {
 274                 CERROR("dlm namespace %s has no procfs dir?\n",
 275                        ldlm_ns_name(ns));
 276         } else {
 277                 lprocfs_remove(&dir);
 278         }
 279
 280         if (ns->ns_stats != NULL)
 281                 lprocfs_free_stats(&ns->ns_stats);
 282 }
 283
 284 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 285 {
 286         struct lprocfs_vars lock_vars[2];
 287         char lock_name[MAX_STRING_SIZE + 1];
 288
 289         LASSERT(ns != NULL);
 290         LASSERT(ns->ns_rs_hash != NULL);
 291
 292         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 293         if (ns->ns_stats == NULL)
 294                 return -ENOMEM;
 295
 296         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 297                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 298
 299         lock_name[MAX_STRING_SIZE] = '\0';
 300
 301         memset(lock_vars, 0, sizeof(lock_vars));
 302         lock_vars[0].name = lock_name;
 303
 304         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 305                  ldlm_ns_name(ns));
 306         lock_vars[0].data = ns;
 307         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 308         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 309
 310         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 311                  ldlm_ns_name(ns));
 312         lock_vars[0].data = ns;
 313         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 314         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 315
 316         if (ns_is_client(ns)) {
 317                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 318                          ldlm_ns_name(ns));
 319                 lock_vars[0].data = &ns->ns_nr_unused;
 320                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 321                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 322
 323                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 324                          ldlm_ns_name(ns));
 325                 lock_vars[0].data = ns;
 326                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 327                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 328                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 329
 330                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 331                          ldlm_ns_name(ns));
 332                 lock_vars[0].data = &ns->ns_max_age;
 333                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 334                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 335                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 336         } else {
 337                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 338                          ldlm_ns_name(ns));
 339                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 340                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 341                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 342                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 343
 344                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 345                          ldlm_ns_name(ns));
 346                 lock_vars[0].data = &ns->ns_timeouts;
 347                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 348                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 349
 350                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 351                          ldlm_ns_name(ns));
 352                 lock_vars[0].data = &ns->ns_max_nolock_size;
 353                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 354                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 355                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 356
 357                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 358                          ldlm_ns_name(ns));
 359                 lock_vars[0].data = &ns->ns_contention_time;
 360                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 361                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 362                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 363
 364                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 365                          ldlm_ns_name(ns));
 366                 lock_vars[0].data = &ns->ns_contended_locks;
 367                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 368                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 369                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 370
 371                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 372                          ldlm_ns_name(ns));
 373                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 374                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 375                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 376                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 377         }
 378         return 0;
 379 }
 380 #undef MAX_STRING_SIZE
 381 #else /* LPROCFS */
 382
 383 #define ldlm_namespace_proc_unregister(ns)      ({;})
 384 #define ldlm_namespace_proc_register(ns)        ({0;})
 385
 386 #endif /* LPROCFS */
 387
 388 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 389                                   const void *key, unsigned mask)
 390 {
 391         const struct ldlm_res_id     *id  = key;
 392         unsigned                val = 0;
 393         unsigned                i;
 394
 395         for (i = 0; i < RES_NAME_SIZE; i++)
 396                 val += id->name[i];
 397         return val & mask;
 398 }
 399
 400 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 401                                       const void *key, unsigned mask)
 402 {
 403         const struct ldlm_res_id *id = key;
 404         struct lu_fid       fid;
 405         __u32               hash;
 406         __u32               val;
 407
 408         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 409         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF];
 410         fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF];
 411
 412         hash = fid_flatten32(&fid);
 413         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 414         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 415                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 416                 hash += (val >> 5) + (val << 11);
 417         } else {
 418                 val = fid_oid(&fid);
 419         }
 420         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 421         /* give me another random factor */
 422         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 423
 424         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 425         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 426
 427         return hash & mask;
 428 }
 429
 430 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 431 {
 432         struct ldlm_resource   *res;
 433
 434         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 435         return &res->lr_name;
 436 }
 437
 438 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 439 {
 440         struct ldlm_resource   *res;
 441
 442         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 443         return ldlm_res_eq((const struct ldlm_res_id *)key,
 444                            (const struct ldlm_res_id *)&res->lr_name);
 445 }
 446
 447 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 448 {
 449         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 450 }
 451
 452 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 453 {
 454         struct ldlm_resource *res;
 455
 456         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 457         ldlm_resource_getref(res);
 458         LDLM_RESOURCE_ADDREF(res);
 459 }
 460
 461 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 462 {
 463         struct ldlm_resource *res;
 464
 465         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 466         /* cfs_hash_for_each_nolock is the only chance we call it */
 467         LDLM_RESOURCE_DELREF(res);
 468         ldlm_resource_putref_locked(res);
 469 }
 470
 471 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 472 {
 473         struct ldlm_resource *res;
 474
 475         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 476         LDLM_RESOURCE_DELREF(res);
 477         ldlm_resource_putref(res);
 478 }
 479
 480 cfs_hash_ops_t ldlm_ns_hash_ops = {
 481         .hs_hash        = ldlm_res_hop_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 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 492         .hs_hash        = ldlm_res_hop_fid_hash,
 493         .hs_key         = ldlm_res_hop_key,
 494         .hs_keycmp      = ldlm_res_hop_keycmp,
 495         .hs_keycpy      = NULL,
 496         .hs_object      = ldlm_res_hop_object,
 497         .hs_get         = ldlm_res_hop_get_locked,
 498         .hs_put_locked  = ldlm_res_hop_put_locked,
 499         .hs_put         = ldlm_res_hop_put
 500 };
 501
 502 typedef struct {
 503         ldlm_ns_type_t  nsd_type;
 504         /** hash bucket bits */
 505         unsigned        nsd_bkt_bits;
 506         /** hash bits */
 507         unsigned        nsd_all_bits;
 508         /** hash operations */
 509         cfs_hash_ops_t *nsd_hops;
 510 } ldlm_ns_hash_def_t;
 511
 512 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 513 {
 514         {
 515                 .nsd_type       = LDLM_NS_TYPE_MDC,
 516                 .nsd_bkt_bits   = 11,
 517                 .nsd_all_bits   = 16,
 518                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 519         },
 520         {
 521                 .nsd_type       = LDLM_NS_TYPE_MDT,
 522                 .nsd_bkt_bits   = 14,
 523                 .nsd_all_bits   = 21,
 524                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 525         },
 526         {
 527                 .nsd_type       = LDLM_NS_TYPE_OSC,
 528                 .nsd_bkt_bits   = 8,
 529                 .nsd_all_bits   = 12,
 530                 .nsd_hops       = &ldlm_ns_hash_ops,
 531         },
 532         {
 533                 .nsd_type       = LDLM_NS_TYPE_OST,
 534                 .nsd_bkt_bits   = 11,
 535                 .nsd_all_bits   = 17,
 536                 .nsd_hops       = &ldlm_ns_hash_ops,
 537         },
 538         {
 539                 .nsd_type       = LDLM_NS_TYPE_MGC,
 540                 .nsd_bkt_bits   = 4,
 541                 .nsd_all_bits   = 4,
 542                 .nsd_hops       = &ldlm_ns_hash_ops,
 543         },
 544         {
 545                 .nsd_type       = LDLM_NS_TYPE_MGT,
 546                 .nsd_bkt_bits   = 4,
 547                 .nsd_all_bits   = 4,
 548                 .nsd_hops       = &ldlm_ns_hash_ops,
 549         },
 550         {
 551                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 552         },
 553 };
 554
 555 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 556                                           ldlm_side_t client,
 557                                           ldlm_appetite_t apt,
 558                                           ldlm_ns_type_t ns_type)
 559 {
 560         struct ldlm_namespace *ns = NULL;
 561         struct ldlm_ns_bucket *nsb;
 562         ldlm_ns_hash_def_t    *nsd;
 563         cfs_hash_bd_t          bd;
 564         int                    idx;
 565         int                    rc;
 566         ENTRY;
 567
 568         LASSERT(obd != NULL);
 569
 570         rc = ldlm_get_ref();
 571         if (rc) {
 572                 CERROR("ldlm_get_ref failed: %d\n", rc);
 573                 RETURN(NULL);
 574         }
 575
 576         for (idx = 0;;idx++) {
 577                 nsd = &ldlm_ns_hash_defs[idx];
 578                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 579                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 580                         GOTO(out_ref, NULL);
 581                 }
 582
 583                 if (nsd->nsd_type == ns_type)
 584                         break;
 585         }
 586
 587         OBD_ALLOC_PTR(ns);
 588         if (!ns)
 589                 GOTO(out_ref, NULL);
 590
 591         ns->ns_rs_hash = cfs_hash_create(name,
 592                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 593                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 594                                          CFS_HASH_MIN_THETA,
 595                                          CFS_HASH_MAX_THETA,
 596                                          nsd->nsd_hops,
 597                                          CFS_HASH_DEPTH |
 598                                          CFS_HASH_BIGNAME |
 599                                          CFS_HASH_SPIN_BKTLOCK |
 600                                          CFS_HASH_NO_ITEMREF);
 601         if (ns->ns_rs_hash == NULL)
 602                 GOTO(out_ns, NULL);
 603
 604         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 605                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 606                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 607                 nsb->nsb_namespace = ns;
 608         }
 609
 610         ns->ns_obd      = obd;
 611         ns->ns_appetite = apt;
 612         ns->ns_client   = client;
 613
 614         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 615         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 616         cfs_spin_lock_init(&ns->ns_lock);
 617         cfs_atomic_set(&ns->ns_bref, 0);
 618         cfs_waitq_init(&ns->ns_waitq);
 619
 620         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 621         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 622         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 623
 624         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 625         ns->ns_nr_unused          = 0;
 626         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 627         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 628         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 629         ns->ns_timeouts           = 0;
 630         ns->ns_orig_connect_flags = 0;
 631         ns->ns_connect_flags      = 0;
 632         ns->ns_stopping           = 0;
 633         rc = ldlm_namespace_proc_register(ns);
 634         if (rc != 0) {
 635                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 636                 GOTO(out_hash, rc);
 637         }
 638
 639         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 640         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 641         if (rc) {
 642                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 643                 GOTO(out_proc, rc);
 644         }
 645
 646         ldlm_namespace_register(ns, client);
 647         RETURN(ns);
 648 out_proc:
 649         ldlm_namespace_proc_unregister(ns);
 650         ldlm_namespace_cleanup(ns, 0);
 651 out_hash:
 652         cfs_hash_putref(ns->ns_rs_hash);
 653 out_ns:
 654         OBD_FREE_PTR(ns);
 655 out_ref:
 656         ldlm_put_ref();
 657         RETURN(NULL);
 658 }
 659
 660 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 661
 662 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 663  * This is currently only used for recovery, and we make certain assumptions
 664  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 665 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 666                              int flags)
 667 {
 668         cfs_list_t *tmp;
 669         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 670         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 671
 672         do {
 673                 struct ldlm_lock *lock = NULL;
 674
 675                 /* first, we look for non-cleaned-yet lock
 676                  * all cleaned locks are marked by CLEANED flag */
 677                 lock_res(res);
 678                 cfs_list_for_each(tmp, q) {
 679                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 680                                               l_res_link);
 681                         if (lock->l_flags & LDLM_FL_CLEANED) {
 682                                 lock = NULL;
 683                                 continue;
 684                         }
 685                         LDLM_LOCK_GET(lock);
 686                         lock->l_flags |= LDLM_FL_CLEANED;
 687                         break;
 688                 }
 689
 690                 if (lock == NULL) {
 691                         unlock_res(res);
 692                         break;
 693                 }
 694
 695                 /* Set CBPENDING so nothing in the cancellation path
 696                  * can match this lock */
 697                 lock->l_flags |= LDLM_FL_CBPENDING;
 698                 lock->l_flags |= LDLM_FL_FAILED;
 699                 lock->l_flags |= flags;
 700
 701                 /* ... without sending a CANCEL message for local_only. */
 702                 if (local_only)
 703                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 704
 705                 if (local_only && (lock->l_readers || lock->l_writers)) {
 706                         /* This is a little bit gross, but much better than the
 707                          * alternative: pretend that we got a blocking AST from
 708                          * the server, so that when the lock is decref'd, it
 709                          * will go away ... */
 710                         unlock_res(res);
 711                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 712                         if (lock->l_completion_ast)
 713                                 lock->l_completion_ast(lock, 0, NULL);
 714                         LDLM_LOCK_RELEASE(lock);
 715                         continue;
 716                 }
 717
 718                 if (client) {
 719                         struct lustre_handle lockh;
 720
 721                         unlock_res(res);
 722                         ldlm_lock2handle(lock, &lockh);
 723                         rc = ldlm_cli_cancel(&lockh);
 724                         if (rc)
 725                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 726                 } else {
 727                         ldlm_resource_unlink_lock(lock);
 728                         unlock_res(res);
 729                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 730                                    "client node");
 731                         ldlm_lock_destroy(lock);
 732                 }
 733                 LDLM_LOCK_RELEASE(lock);
 734         } while (1);
 735 }
 736
 737 static int ldlm_resource_clean(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         int    flags = (int)(unsigned long)arg;
 742
 743         cleanup_resource(res, &res->lr_granted, flags);
 744         cleanup_resource(res, &res->lr_converting, flags);
 745         cleanup_resource(res, &res->lr_waiting, flags);
 746
 747         return 0;
 748 }
 749
 750 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 751                                   cfs_hlist_node_t *hnode, void *arg)
 752 {
 753         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 754
 755         CERROR("Namespace %s resource refcount nonzero "
 756                "(%d) after lock cleanup; forcing "
 757                "cleanup.\n",
 758                ldlm_ns_name(ldlm_res_to_ns(res)),
 759                cfs_atomic_read(&res->lr_refcount) - 1);
 760
 761         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 762                LPU64") (rc: %d)\n", res,
 763                res->lr_name.name[0], res->lr_name.name[1],
 764                res->lr_name.name[2], res->lr_name.name[3],
 765                cfs_atomic_read(&res->lr_refcount) - 1);
 766         return 0;
 767 }
 768
 769 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 770 {
 771         if (ns == NULL) {
 772                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 773                 return ELDLM_OK;
 774         }
 775
 776         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 777                                  (void *)(unsigned long)flags);
 778         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 779         return ELDLM_OK;
 780 }
 781
 782 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 783 {
 784         ENTRY;
 785
 786         /* At shutdown time, don't call the cancellation callback */
 787         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 788
 789         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 790                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 791                 int rc;
 792                 CDEBUG(D_DLMTRACE,
 793                        "dlm namespace %s free waiting on refcount %d\n",
 794                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 795 force_wait:
 796                 if (force)
 797                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 798
 799                 rc = l_wait_event(ns->ns_waitq,
 800                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 801
 802                 /* Forced cleanups should be able to reclaim all references,
 803                  * so it's safe to wait forever... we can't leak locks... */
 804                 if (force && rc == -ETIMEDOUT) {
 805                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 806                                        "namespace with %d resources in use, "
 807                                        "(rc=%d)\n", ldlm_ns_name(ns),
 808                                        cfs_atomic_read(&ns->ns_bref), rc);
 809                         GOTO(force_wait, rc);
 810                 }
 811
 812                 if (cfs_atomic_read(&ns->ns_bref)) {
 813                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 814                                        "with %d resources in use, (rc=%d)\n",
 815                                        ldlm_ns_name(ns),
 816                                        cfs_atomic_read(&ns->ns_bref), rc);
 817                         RETURN(ELDLM_NAMESPACE_EXISTS);
 818                 }
 819                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 820                        ldlm_ns_name(ns));
 821         }
 822
 823         RETURN(ELDLM_OK);
 824 }
 825
 826 /**
 827  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 828  * for freeing. Waits for refc == 0.
 829  *
 830  * The following is done:
 831  * (0) Unregister \a ns from its list to make inaccessible for potential users
 832  * like pools thread and others;
 833  * (1) Clear all locks in \a ns.
 834  */
 835 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 836                                struct obd_import *imp,
 837                                int force)
 838 {
 839         int rc;
 840         ENTRY;
 841         if (!ns) {
 842                 EXIT;
 843                 return;
 844         }
 845
 846         cfs_spin_lock(&ns->ns_lock);
 847         ns->ns_stopping = 1;
 848         cfs_spin_unlock(&ns->ns_lock);
 849
 850         /*
 851          * Can fail with -EINTR when force == 0 in which case try harder.
 852          */
 853         rc = __ldlm_namespace_free(ns, force);
 854         if (rc != ELDLM_OK) {
 855                 if (imp) {
 856                         ptlrpc_disconnect_import(imp, 0);
 857                         ptlrpc_invalidate_import(imp);
 858                 }
 859
 860                 /*
 861                  * With all requests dropped and the import inactive
 862                  * we are gaurenteed all reference will be dropped.
 863                  */
 864                 rc = __ldlm_namespace_free(ns, 1);
 865                 LASSERT(rc == 0);
 866         }
 867         EXIT;
 868 }
 869
 870 /**
 871  * Performs freeing memory structures related to \a ns. This is only done when
 872  * ldlm_namespce_free_prior() successfully removed all resources referencing
 873  * \a ns and its refc == 0.
 874  */
 875 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 876 {
 877         ENTRY;
 878         if (!ns) {
 879                 EXIT;
 880                 return;
 881         }
 882
 883
 884         /*
 885          * Make sure that nobody can find this ns in its list.
 886          */
 887         ldlm_namespace_unregister(ns, ns->ns_client);
 888         /*
 889          * Fini pool _before_ parent proc dir is removed. This is important as
 890          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 891          * it after @dir may cause oops.
 892          */
 893         ldlm_pool_fini(&ns->ns_pool);
 894
 895         ldlm_namespace_proc_unregister(ns);
 896         cfs_hash_putref(ns->ns_rs_hash);
 897         /*
 898          * Namespace \a ns should be not on list in this time, otherwise this
 899          * will cause issues realted to using freed \a ns in pools thread.
 900          */
 901         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 902         OBD_FREE_PTR(ns);
 903         ldlm_put_ref();
 904         EXIT;
 905 }
 906
 907
 908 /* Cleanup the resource, and free namespace.
 909  * bug 12864:
 910  * Deadlock issue:
 911  * proc1: destroy import
 912  *        class_disconnect_export(grab cl_sem) ->
 913  *              -> ldlm_namespace_free ->
 914  *              -> lprocfs_remove(grab _lprocfs_lock).
 915  * proc2: read proc info
 916  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 917  *              -> osc_rd_active, etc(grab cl_sem).
 918  *
 919  * So that I have to split the ldlm_namespace_free into two parts - the first
 920  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 921  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 922  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 923  * held.
 924  */
 925 void ldlm_namespace_free(struct ldlm_namespace *ns,
 926                          struct obd_import *imp,
 927                          int force)
 928 {
 929         ldlm_namespace_free_prior(ns, imp, force);
 930         ldlm_namespace_free_post(ns);
 931 }
 932
 933 void ldlm_namespace_get(struct ldlm_namespace *ns)
 934 {
 935         cfs_atomic_inc(&ns->ns_bref);
 936 }
 937
 938 void ldlm_namespace_put(struct ldlm_namespace *ns)
 939 {
 940         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 941                 cfs_waitq_signal(&ns->ns_waitq);
 942                 cfs_spin_unlock(&ns->ns_lock);
 943         }
 944 }
 945
 946 /* Register @ns in the list of namespaces */
 947 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 948 {
 949         cfs_mutex_down(ldlm_namespace_lock(client));
 950         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 951         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 952         cfs_atomic_inc(ldlm_namespace_nr(client));
 953         cfs_mutex_up(ldlm_namespace_lock(client));
 954 }
 955
 956 /* Unregister @ns from the list of namespaces */
 957 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 958 {
 959         cfs_mutex_down(ldlm_namespace_lock(client));
 960         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 961         /*
 962          * Some asserts and possibly other parts of code still using
 963          * list_empty(&ns->ns_list_chain). This is why it is important
 964          * to use list_del_init() here.
 965          */
 966         cfs_list_del_init(&ns->ns_list_chain);
 967         cfs_atomic_dec(ldlm_namespace_nr(client));
 968         cfs_mutex_up(ldlm_namespace_lock(client));
 969 }
 970
 971 /* Should be called under ldlm_namespace_lock(client) taken */
 972 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 973 {
 974         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 975         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 976         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 977 }
 978
 979 /* Should be called under ldlm_namespace_lock(client) taken */
 980 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 981 {
 982         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 983         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 984         return container_of(ldlm_namespace_list(client)->next,
 985                 struct ldlm_namespace, ns_list_chain);
 986 }
 987
 988 static struct ldlm_resource *ldlm_resource_new(void)
 989 {
 990         struct ldlm_resource *res;
 991         int idx;
 992
 993         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 994         if (res == NULL)
 995                 return NULL;
 996
 997         CFS_INIT_LIST_HEAD(&res->lr_granted);
 998         CFS_INIT_LIST_HEAD(&res->lr_converting);
 999         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1000
1001         /* initialize interval trees for each lock mode*/
1002         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1003                 res->lr_itree[idx].lit_size = 0;
1004                 res->lr_itree[idx].lit_mode = 1 << idx;
1005                 res->lr_itree[idx].lit_root = NULL;
1006         }
1007
1008         cfs_atomic_set(&res->lr_refcount, 1);
1009         cfs_spin_lock_init(&res->lr_lock);
1010         lu_ref_init(&res->lr_reference);
1011
1012         /* one who creates the resource must unlock
1013          * the semaphore after lvb initialization */
1014         cfs_init_mutex_locked(&res->lr_lvb_sem);
1015
1016         return res;
1017 }
1018
1019 /* Args: unlocked namespace
1020  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1021  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1022 struct ldlm_resource *
1023 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1024                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1025 {
1026         cfs_hlist_node_t     *hnode;
1027         struct ldlm_resource *res;
1028         cfs_hash_bd_t         bd;
1029         __u64                 version;
1030
1031         LASSERT(ns != NULL);
1032         LASSERT(parent == NULL);
1033         LASSERT(ns->ns_rs_hash != NULL);
1034         LASSERT(name->name[0] != 0);
1035
1036         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1037         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1038         if (hnode != NULL) {
1039                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1040                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1041                 /* synchronize WRT resource creation */
1042                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1043                         cfs_down(&res->lr_lvb_sem);
1044                         cfs_up(&res->lr_lvb_sem);
1045                 }
1046                 return res;
1047         }
1048
1049         version = cfs_hash_bd_version_get(&bd);
1050         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1051
1052         if (create == 0)
1053                 return NULL;
1054
1055         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1056                  "type: %d\n", type);
1057         res = ldlm_resource_new();
1058         if (!res)
1059                 return NULL;
1060
1061         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1062         res->lr_name       = *name;
1063         res->lr_type       = type;
1064         res->lr_most_restr = LCK_NL;
1065
1066         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1067         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1068                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1069
1070         if (hnode != NULL) {
1071                 /* someone won the race and added the resource before */
1072                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1073                 /* clean lu_ref for failed resource */
1074                 lu_ref_fini(&res->lr_reference);
1075                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1076
1077                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1078                 /* synchronize WRT resource creation */
1079                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1080                         cfs_down(&res->lr_lvb_sem);
1081                         cfs_up(&res->lr_lvb_sem);
1082                 }
1083                 return res;
1084         }
1085         /* we won! let's add the resource */
1086         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1087         if (cfs_hash_bd_count_get(&bd) == 1)
1088                 ldlm_namespace_get(ns);
1089
1090         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1091         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1092                 int rc;
1093
1094                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1095                 rc = ns->ns_lvbo->lvbo_init(res);
1096                 if (rc)
1097                         CERROR("lvbo_init failed for resource "
1098                                LPU64": rc %d\n", name->name[0], rc);
1099                 /* we create resource with locked lr_lvb_sem */
1100                 cfs_up(&res->lr_lvb_sem);
1101         }
1102
1103         return res;
1104 }
1105
1106 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1107 {
1108         LASSERT(res != NULL);
1109         LASSERT(res != LP_POISON);
1110         cfs_atomic_inc(&res->lr_refcount);
1111         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1112                cfs_atomic_read(&res->lr_refcount));
1113         return res;
1114 }
1115
1116 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1117                                          struct ldlm_resource *res)
1118 {
1119         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1120
1121         if (!cfs_list_empty(&res->lr_granted)) {
1122                 ldlm_resource_dump(D_ERROR, res);
1123                 LBUG();
1124         }
1125
1126         if (!cfs_list_empty(&res->lr_converting)) {
1127                 ldlm_resource_dump(D_ERROR, res);
1128                 LBUG();
1129         }
1130
1131         if (!cfs_list_empty(&res->lr_waiting)) {
1132                 ldlm_resource_dump(D_ERROR, res);
1133                 LBUG();
1134         }
1135
1136         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1137                                bd, &res->lr_hash);
1138         lu_ref_fini(&res->lr_reference);
1139         if (cfs_hash_bd_count_get(bd) == 0)
1140                 ldlm_namespace_put(nsb->nsb_namespace);
1141 }
1142
1143 /* Returns 1 if the resource was freed, 0 if it remains. */
1144 int ldlm_resource_putref(struct ldlm_resource *res)
1145 {
1146         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1147         cfs_hash_bd_t   bd;
1148
1149         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1150         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1151                res, cfs_atomic_read(&res->lr_refcount) - 1);
1152
1153         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1154         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1155                 __ldlm_resource_putref_final(&bd, res);
1156                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1157                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1158                         ns->ns_lvbo->lvbo_free(res);
1159                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1160                 return 1;
1161         }
1162         return 0;
1163 }
1164
1165 /* Returns 1 if the resource was freed, 0 if it remains. */
1166 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1167 {
1168         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1169
1170         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1171         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1172                res, cfs_atomic_read(&res->lr_refcount) - 1);
1173
1174         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1175                 cfs_hash_bd_t bd;
1176
1177                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1178                                 &res->lr_name, &bd);
1179                 __ldlm_resource_putref_final(&bd, res);
1180                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1181                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1182                  * so we should never be here while calling cfs_hash_del,
1183                  * cfs_hash_for_each_nolock is the only case we can get
1184                  * here, which is safe to release cfs_hash_bd_lock.
1185                  */
1186                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1187                         ns->ns_lvbo->lvbo_free(res);
1188                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1189
1190                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1191                 return 1;
1192         }
1193         return 0;
1194 }
1195
1196 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1197                             struct ldlm_lock *lock)
1198 {
1199         check_res_locked(res);
1200
1201         CDEBUG(D_OTHER, "About to add this lock:\n");
1202         ldlm_lock_dump(D_OTHER, lock, 0);
1203
1204         if (lock->l_destroyed) {
1205                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1206                 return;
1207         }
1208
1209         LASSERT(cfs_list_empty(&lock->l_res_link));
1210
1211         cfs_list_add_tail(&lock->l_res_link, head);
1212 }
1213
1214 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1215                                      struct ldlm_lock *new)
1216 {
1217         struct ldlm_resource *res = original->l_resource;
1218
1219         check_res_locked(res);
1220
1221         ldlm_resource_dump(D_INFO, res);
1222         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1223         ldlm_lock_dump(D_OTHER, new, 0);
1224
1225         if (new->l_destroyed) {
1226                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1227                 goto out;
1228         }
1229
1230         LASSERT(cfs_list_empty(&new->l_res_link));
1231
1232         cfs_list_add(&new->l_res_link, &original->l_res_link);
1233  out:;
1234 }
1235
1236 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1237 {
1238         int type = lock->l_resource->lr_type;
1239
1240         check_res_locked(lock->l_resource);
1241         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1242                 ldlm_unlink_lock_skiplist(lock);
1243         else if (type == LDLM_EXTENT)
1244                 ldlm_extent_unlink_lock(lock);
1245         cfs_list_del_init(&lock->l_res_link);
1246 }
1247
1248 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1249 {
1250         desc->lr_type = res->lr_type;
1251         desc->lr_name = res->lr_name;
1252 }
1253
1254 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1255 {
1256         cfs_list_t *tmp;
1257
1258         if (!((libcfs_debug | D_ERROR) & level))
1259                 return;
1260
1261         cfs_mutex_down(ldlm_namespace_lock(client));
1262
1263         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1264                 struct ldlm_namespace *ns;
1265                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1266                 ldlm_namespace_dump(level, ns);
1267         }
1268
1269         cfs_mutex_up(ldlm_namespace_lock(client));
1270 }
1271
1272 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1273                               cfs_hlist_node_t *hnode, void *arg)
1274 {
1275         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1276         int    level = (int)(unsigned long)arg;
1277
1278         lock_res(res);
1279         ldlm_resource_dump(level, res);
1280         unlock_res(res);
1281
1282         return 0;
1283 }
1284
1285 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1286 {
1287         if (!((libcfs_debug | D_ERROR) & level))
1288                 return;
1289
1290         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1291                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1292                ns_is_client(ns) ? "client" : "server");
1293
1294         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1295                 return;
1296
1297         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1298                                  ldlm_res_hash_dump,
1299                                  (void *)(unsigned long)level);
1300         cfs_spin_lock(&ns->ns_lock);
1301         ns->ns_next_dump = cfs_time_shift(10);
1302         cfs_spin_unlock(&ns->ns_lock);
1303 }
1304
1305 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1306 {
1307         cfs_list_t *tmp;
1308         int pos;
1309
1310         CLASSERT(RES_NAME_SIZE == 4);
1311
1312         if (!((libcfs_debug | D_ERROR) & level))
1313                 return;
1314
1315         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1316                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1317                res->lr_name.name[2], res->lr_name.name[3],
1318                cfs_atomic_read(&res->lr_refcount));
1319
1320         if (!cfs_list_empty(&res->lr_granted)) {
1321                 pos = 0;
1322                 CDEBUG(level, "Granted locks:\n");
1323                 cfs_list_for_each(tmp, &res->lr_granted) {
1324                         struct ldlm_lock *lock;
1325                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1326                                               l_res_link);
1327                         ldlm_lock_dump(level, lock, ++pos);
1328                 }
1329         }
1330         if (!cfs_list_empty(&res->lr_converting)) {
1331                 pos = 0;
1332                 CDEBUG(level, "Converting locks:\n");
1333                 cfs_list_for_each(tmp, &res->lr_converting) {
1334                         struct ldlm_lock *lock;
1335                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1336                                               l_res_link);
1337                         ldlm_lock_dump(level, lock, ++pos);
1338                 }
1339         }
1340         if (!cfs_list_empty(&res->lr_waiting)) {
1341                 pos = 0;
1342                 CDEBUG(level, "Waiting locks:\n");
1343                 cfs_list_for_each(tmp, &res->lr_waiting) {
1344                         struct ldlm_lock *lock;
1345                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1346                                               l_res_link);
1347                         ldlm_lock_dump(level, lock, ++pos);
1348                 }
1349         }
1350 }