lustre/ldlm/ldlm_resource.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ldlm/ldlm_resource.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43 #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 struct mutex ldlm_srv_namespace_lock;
  59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  60
  61 struct mutex ldlm_cli_namespace_lock;
  62 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  63
  64 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  65 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  66 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  67
  68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  69
  70 /* during debug dump certain amount of granted locks for one resource to avoid
  71  * DDOS. */
  72 unsigned int ldlm_dump_granted_max = 256;
  73
  74 #ifdef LPROCFS
  75 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  76                              unsigned long count, void *data)
  77 {
  78         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  79         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  80         RETURN(count);
  81 }
  82
  83 int ldlm_proc_setup(void)
  84 {
  85         int rc;
  86         struct lprocfs_vars list[] = {
  87                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  88                 { "dump_granted_max",
  89                   lprocfs_rd_uint, lprocfs_wr_uint,
  90                   &ldlm_dump_granted_max, NULL },
  91                 { "cancel_unused_locks_before_replay",
  92                   lprocfs_rd_uint, lprocfs_wr_uint,
  93                   &ldlm_cancel_unused_locks_before_replay, NULL },
  94                 { NULL }};
  95         ENTRY;
  96         LASSERT(ldlm_ns_proc_dir == NULL);
  97
  98         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  99                                               proc_lustre_root,
 100                                               NULL, NULL);
 101         if (IS_ERR(ldlm_type_proc_dir)) {
 102                 CERROR("LProcFS failed in ldlm-init\n");
 103                 rc = PTR_ERR(ldlm_type_proc_dir);
 104                 GOTO(err, rc);
 105         }
 106
 107         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 108                                             ldlm_type_proc_dir,
 109                                             NULL, NULL);
 110         if (IS_ERR(ldlm_ns_proc_dir)) {
 111                 CERROR("LProcFS failed in ldlm-init\n");
 112                 rc = PTR_ERR(ldlm_ns_proc_dir);
 113                 GOTO(err_type, rc);
 114         }
 115
 116         ldlm_svc_proc_dir = lprocfs_register("services",
 117                                             ldlm_type_proc_dir,
 118                                             NULL, NULL);
 119         if (IS_ERR(ldlm_svc_proc_dir)) {
 120                 CERROR("LProcFS failed in ldlm-init\n");
 121                 rc = PTR_ERR(ldlm_svc_proc_dir);
 122                 GOTO(err_ns, rc);
 123         }
 124
 125         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 126
 127         RETURN(0);
 128
 129 err_ns:
 130         lprocfs_remove(&ldlm_ns_proc_dir);
 131 err_type:
 132         lprocfs_remove(&ldlm_type_proc_dir);
 133 err:
 134         ldlm_svc_proc_dir = NULL;
 135         RETURN(rc);
 136 }
 137
 138 void ldlm_proc_cleanup(void)
 139 {
 140         if (ldlm_svc_proc_dir)
 141                 lprocfs_remove(&ldlm_svc_proc_dir);
 142
 143         if (ldlm_ns_proc_dir)
 144                 lprocfs_remove(&ldlm_ns_proc_dir);
 145
 146         if (ldlm_type_proc_dir)
 147                 lprocfs_remove(&ldlm_type_proc_dir);
 148 }
 149
 150 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 151                                    int count, int *eof, void *data)
 152 {
 153         struct ldlm_namespace *ns  = data;
 154         __u64                  res = 0;
 155         cfs_hash_bd_t          bd;
 156         int                    i;
 157
 158         /* result is not strictly consistant */
 159         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 160                 res += cfs_hash_bd_count_get(&bd);
 161         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 162 }
 163
 164 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 165                                int count, int *eof, void *data)
 166 {
 167         struct ldlm_namespace *ns = data;
 168         __u64                  locks;
 169
 170         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 171                                         LPROCFS_FIELDS_FLAGS_SUM);
 172         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 173 }
 174
 175 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 176                                int count, int *eof, void *data)
 177 {
 178         struct ldlm_namespace *ns = data;
 179         __u32 *nr = &ns->ns_max_unused;
 180
 181         if (ns_connect_lru_resize(ns))
 182                 nr = &ns->ns_nr_unused;
 183         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 184 }
 185
 186 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 187                                unsigned long count, void *data)
 188 {
 189         struct ldlm_namespace *ns = data;
 190         char dummy[MAX_STRING_SIZE + 1], *end;
 191         unsigned long tmp;
 192         int lru_resize;
 193
 194         dummy[MAX_STRING_SIZE] = '\0';
 195         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 196                 return -EFAULT;
 197
 198         if (strncmp(dummy, "clear", 5) == 0) {
 199                 CDEBUG(D_DLMTRACE,
 200                        "dropping all unused locks from namespace %s\n",
 201                        ldlm_ns_name(ns));
 202                 if (ns_connect_lru_resize(ns)) {
 203                         int canceled, unused  = ns->ns_nr_unused;
 204
 205                         /* Try to cancel all @ns_nr_unused locks. */
 206                         canceled = ldlm_cancel_lru(ns, unused, 0,
 207                                                    LDLM_CANCEL_PASSED);
 208                         if (canceled < unused) {
 209                                 CDEBUG(D_DLMTRACE,
 210                                        "not all requested locks are canceled, "
 211                                        "requested: %d, canceled: %d\n", unused,
 212                                        canceled);
 213                                 return -EINVAL;
 214                         }
 215                 } else {
 216                         tmp = ns->ns_max_unused;
 217                         ns->ns_max_unused = 0;
 218                         ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
 219                         ns->ns_max_unused = tmp;
 220                 }
 221                 return count;
 222         }
 223
 224         tmp = simple_strtoul(dummy, &end, 0);
 225         if (dummy == end) {
 226                 CERROR("invalid value written\n");
 227                 return -EINVAL;
 228         }
 229         lru_resize = (tmp == 0);
 230
 231         if (ns_connect_lru_resize(ns)) {
 232                 if (!lru_resize)
 233                         ns->ns_max_unused = (unsigned int)tmp;
 234
 235                 if (tmp > ns->ns_nr_unused)
 236                         tmp = ns->ns_nr_unused;
 237                 tmp = ns->ns_nr_unused - tmp;
 238
 239                 CDEBUG(D_DLMTRACE,
 240                        "changing namespace %s unused locks from %u to %u\n",
 241                        ldlm_ns_name(ns), ns->ns_nr_unused,
 242                        (unsigned int)tmp);
 243                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
 244
 245                 if (!lru_resize) {
 246                         CDEBUG(D_DLMTRACE,
 247                                "disable lru_resize for namespace %s\n",
 248                                ldlm_ns_name(ns));
 249                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 250                 }
 251         } else {
 252                 CDEBUG(D_DLMTRACE,
 253                        "changing namespace %s max_unused from %u to %u\n",
 254                        ldlm_ns_name(ns), ns->ns_max_unused,
 255                        (unsigned int)tmp);
 256                 ns->ns_max_unused = (unsigned int)tmp;
 257                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
 258
 259                 /* Make sure that LRU resize was originally supported before
 260                  * turning it on here. */
 261                 if (lru_resize &&
 262                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 263                         CDEBUG(D_DLMTRACE,
 264                                "enable lru_resize for namespace %s\n",
 265                                ldlm_ns_name(ns));
 266                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 267                 }
 268         }
 269
 270         return count;
 271 }
 272
 273 static int lprocfs_rd_elc(char *page, char **start, off_t off,
 274                           int count, int *eof, void *data)
 275 {
 276         struct ldlm_namespace *ns = data;
 277         unsigned int supp = ns_connect_cancelset(ns);
 278
 279         return lprocfs_rd_uint(page, start, off, count, eof, &supp);
 280 }
 281
 282 static int lprocfs_wr_elc(struct file *file, const char *buffer,
 283                                unsigned long count, void *data)
 284 {
 285         struct ldlm_namespace *ns = data;
 286         unsigned int supp = -1;
 287         int rc;
 288
 289         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 290         if (rc < 0)
 291                 return rc;
 292
 293         if (supp == 0)
 294                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 295         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 296                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 297         return count;
 298 }
 299
 300 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 301 {
 302         struct proc_dir_entry *dir;
 303
 304         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 305         if (dir == NULL) {
 306                 CERROR("dlm namespace %s has no procfs dir?\n",
 307                        ldlm_ns_name(ns));
 308         } else {
 309                 lprocfs_remove(&dir);
 310         }
 311
 312         if (ns->ns_stats != NULL)
 313                 lprocfs_free_stats(&ns->ns_stats);
 314 }
 315
 316 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 317 {
 318         struct lprocfs_vars lock_vars[2];
 319         char lock_name[MAX_STRING_SIZE + 1];
 320
 321         LASSERT(ns != NULL);
 322         LASSERT(ns->ns_rs_hash != NULL);
 323
 324         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 325         if (ns->ns_stats == NULL)
 326                 return -ENOMEM;
 327
 328         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 329                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 330
 331         lock_name[MAX_STRING_SIZE] = '\0';
 332
 333         memset(lock_vars, 0, sizeof(lock_vars));
 334         lock_vars[0].name = lock_name;
 335
 336         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 337                  ldlm_ns_name(ns));
 338         lock_vars[0].data = ns;
 339         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 340         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 341
 342         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 343                  ldlm_ns_name(ns));
 344         lock_vars[0].data = ns;
 345         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 346         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 347
 348         if (ns_is_client(ns)) {
 349                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 350                          ldlm_ns_name(ns));
 351                 lock_vars[0].data = &ns->ns_nr_unused;
 352                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 354
 355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 356                          ldlm_ns_name(ns));
 357                 lock_vars[0].data = ns;
 358                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 359                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 360                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 361
 362                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 363                          ldlm_ns_name(ns));
 364                 lock_vars[0].data = &ns->ns_max_age;
 365                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 366                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 367                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 368
 369                 snprintf(lock_name, MAX_STRING_SIZE, "%s/early_lock_cancel",
 370                          ldlm_ns_name(ns));
 371                 lock_vars[0].data = ns;
 372                 lock_vars[0].read_fptr = lprocfs_rd_elc;
 373                 lock_vars[0].write_fptr = lprocfs_wr_elc;
 374                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 375         } else {
 376                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 377                          ldlm_ns_name(ns));
 378                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 379                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 380                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 381                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 382
 383                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 384                          ldlm_ns_name(ns));
 385                 lock_vars[0].data = &ns->ns_timeouts;
 386                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 387                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 388
 389                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 390                          ldlm_ns_name(ns));
 391                 lock_vars[0].data = &ns->ns_max_nolock_size;
 392                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 393                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 394                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 395
 396                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 397                          ldlm_ns_name(ns));
 398                 lock_vars[0].data = &ns->ns_contention_time;
 399                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 400                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 401                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 402
 403                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 404                          ldlm_ns_name(ns));
 405                 lock_vars[0].data = &ns->ns_contended_locks;
 406                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 407                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 408                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 409
 410                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 411                          ldlm_ns_name(ns));
 412                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 413                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 414                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 415                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 416         }
 417         return 0;
 418 }
 419 #undef MAX_STRING_SIZE
 420 #else /* LPROCFS */
 421
 422 #define ldlm_namespace_proc_unregister(ns)      ({;})
 423 #define ldlm_namespace_proc_register(ns)        ({0;})
 424
 425 #endif /* LPROCFS */
 426
 427 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 428                                   const void *key, unsigned mask)
 429 {
 430         const struct ldlm_res_id     *id  = key;
 431         unsigned                val = 0;
 432         unsigned                i;
 433
 434         for (i = 0; i < RES_NAME_SIZE; i++)
 435                 val += id->name[i];
 436         return val & mask;
 437 }
 438
 439 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 440                                       const void *key, unsigned mask)
 441 {
 442         const struct ldlm_res_id *id = key;
 443         struct lu_fid       fid;
 444         __u32               hash;
 445         __u32               val;
 446
 447         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 448         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 449         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 450
 451         hash = fid_flatten32(&fid);
 452         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 453         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 454                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 455                 hash += (val >> 5) + (val << 11);
 456         } else {
 457                 val = fid_oid(&fid);
 458         }
 459         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 460         /* give me another random factor */
 461         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 462
 463         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 464         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 465
 466         return hash & mask;
 467 }
 468
 469 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 470 {
 471         struct ldlm_resource   *res;
 472
 473         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 474         return &res->lr_name;
 475 }
 476
 477 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 478 {
 479         struct ldlm_resource   *res;
 480
 481         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 482         return ldlm_res_eq((const struct ldlm_res_id *)key,
 483                            (const struct ldlm_res_id *)&res->lr_name);
 484 }
 485
 486 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 487 {
 488         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 489 }
 490
 491 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 492 {
 493         struct ldlm_resource *res;
 494
 495         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 496         ldlm_resource_getref(res);
 497 }
 498
 499 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 500 {
 501         struct ldlm_resource *res;
 502
 503         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 504         /* cfs_hash_for_each_nolock is the only chance we call it */
 505         ldlm_resource_putref_locked(res);
 506 }
 507
 508 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 509 {
 510         struct ldlm_resource *res;
 511
 512         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 513         ldlm_resource_putref(res);
 514 }
 515
 516 cfs_hash_ops_t ldlm_ns_hash_ops = {
 517         .hs_hash        = ldlm_res_hop_hash,
 518         .hs_key         = ldlm_res_hop_key,
 519         .hs_keycmp      = ldlm_res_hop_keycmp,
 520         .hs_keycpy      = NULL,
 521         .hs_object      = ldlm_res_hop_object,
 522         .hs_get         = ldlm_res_hop_get_locked,
 523         .hs_put_locked  = ldlm_res_hop_put_locked,
 524         .hs_put         = ldlm_res_hop_put
 525 };
 526
 527 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 528         .hs_hash        = ldlm_res_hop_fid_hash,
 529         .hs_key         = ldlm_res_hop_key,
 530         .hs_keycmp      = ldlm_res_hop_keycmp,
 531         .hs_keycpy      = NULL,
 532         .hs_object      = ldlm_res_hop_object,
 533         .hs_get         = ldlm_res_hop_get_locked,
 534         .hs_put_locked  = ldlm_res_hop_put_locked,
 535         .hs_put         = ldlm_res_hop_put
 536 };
 537
 538 typedef struct {
 539         ldlm_ns_type_t  nsd_type;
 540         /** hash bucket bits */
 541         unsigned        nsd_bkt_bits;
 542         /** hash bits */
 543         unsigned        nsd_all_bits;
 544         /** hash operations */
 545         cfs_hash_ops_t *nsd_hops;
 546 } ldlm_ns_hash_def_t;
 547
 548 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 549 {
 550         {
 551                 .nsd_type       = LDLM_NS_TYPE_MDC,
 552                 .nsd_bkt_bits   = 11,
 553                 .nsd_all_bits   = 16,
 554                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 555         },
 556         {
 557                 .nsd_type       = LDLM_NS_TYPE_MDT,
 558                 .nsd_bkt_bits   = 14,
 559                 .nsd_all_bits   = 21,
 560                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 561         },
 562         {
 563                 .nsd_type       = LDLM_NS_TYPE_OSC,
 564                 .nsd_bkt_bits   = 8,
 565                 .nsd_all_bits   = 12,
 566                 .nsd_hops       = &ldlm_ns_hash_ops,
 567         },
 568         {
 569                 .nsd_type       = LDLM_NS_TYPE_OST,
 570                 .nsd_bkt_bits   = 11,
 571                 .nsd_all_bits   = 17,
 572                 .nsd_hops       = &ldlm_ns_hash_ops,
 573         },
 574         {
 575                 .nsd_type       = LDLM_NS_TYPE_MGC,
 576                 .nsd_bkt_bits   = 4,
 577                 .nsd_all_bits   = 4,
 578                 .nsd_hops       = &ldlm_ns_hash_ops,
 579         },
 580         {
 581                 .nsd_type       = LDLM_NS_TYPE_MGT,
 582                 .nsd_bkt_bits   = 4,
 583                 .nsd_all_bits   = 4,
 584                 .nsd_hops       = &ldlm_ns_hash_ops,
 585         },
 586         {
 587                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 588         },
 589 };
 590
 591 /**
 592  * Create and initialize new empty namespace.
 593  */
 594 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 595                                           ldlm_side_t client,
 596                                           ldlm_appetite_t apt,
 597                                           ldlm_ns_type_t ns_type)
 598 {
 599         struct ldlm_namespace *ns = NULL;
 600         struct ldlm_ns_bucket *nsb;
 601         ldlm_ns_hash_def_t    *nsd;
 602         cfs_hash_bd_t          bd;
 603         int                    idx;
 604         int                    rc;
 605         ENTRY;
 606
 607         LASSERT(obd != NULL);
 608
 609         rc = ldlm_get_ref();
 610         if (rc) {
 611                 CERROR("ldlm_get_ref failed: %d\n", rc);
 612                 RETURN(NULL);
 613         }
 614
 615         for (idx = 0;;idx++) {
 616                 nsd = &ldlm_ns_hash_defs[idx];
 617                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 618                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 619                         GOTO(out_ref, NULL);
 620                 }
 621
 622                 if (nsd->nsd_type == ns_type)
 623                         break;
 624         }
 625
 626         OBD_ALLOC_PTR(ns);
 627         if (!ns)
 628                 GOTO(out_ref, NULL);
 629
 630         ns->ns_rs_hash = cfs_hash_create(name,
 631                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 632                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 633                                          CFS_HASH_MIN_THETA,
 634                                          CFS_HASH_MAX_THETA,
 635                                          nsd->nsd_hops,
 636                                          CFS_HASH_DEPTH |
 637                                          CFS_HASH_BIGNAME |
 638                                          CFS_HASH_SPIN_BKTLOCK |
 639                                          CFS_HASH_NO_ITEMREF);
 640         if (ns->ns_rs_hash == NULL)
 641                 GOTO(out_ns, NULL);
 642
 643         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 644                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 645                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 646                 nsb->nsb_namespace = ns;
 647         }
 648
 649         ns->ns_obd      = obd;
 650         ns->ns_appetite = apt;
 651         ns->ns_client   = client;
 652
 653         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 654         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 655         spin_lock_init(&ns->ns_lock);
 656         cfs_atomic_set(&ns->ns_bref, 0);
 657         cfs_waitq_init(&ns->ns_waitq);
 658
 659         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 660         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 661         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 662
 663         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 664         ns->ns_nr_unused          = 0;
 665         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 666         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 667         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 668         ns->ns_timeouts           = 0;
 669         ns->ns_orig_connect_flags = 0;
 670         ns->ns_connect_flags      = 0;
 671         ns->ns_stopping           = 0;
 672         rc = ldlm_namespace_proc_register(ns);
 673         if (rc != 0) {
 674                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 675                 GOTO(out_hash, rc);
 676         }
 677
 678         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 679         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 680         if (rc) {
 681                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 682                 GOTO(out_proc, rc);
 683         }
 684
 685         ldlm_namespace_register(ns, client);
 686         RETURN(ns);
 687 out_proc:
 688         ldlm_namespace_proc_unregister(ns);
 689         ldlm_namespace_cleanup(ns, 0);
 690 out_hash:
 691         cfs_hash_putref(ns->ns_rs_hash);
 692 out_ns:
 693         OBD_FREE_PTR(ns);
 694 out_ref:
 695         ldlm_put_ref();
 696         RETURN(NULL);
 697 }
 698 EXPORT_SYMBOL(ldlm_namespace_new);
 699
 700 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 701
 702 /**
 703  * Cancel and destroy all locks on a resource.
 704  *
 705  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 706  * clean up.  This is currently only used for recovery, and we make
 707  * certain assumptions as a result--notably, that we shouldn't cancel
 708  * locks with refs.
 709  */
 710 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 711                              __u64 flags)
 712 {
 713         cfs_list_t *tmp;
 714         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 715         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 716
 717         do {
 718                 struct ldlm_lock *lock = NULL;
 719
 720                 /* First, we look for non-cleaned-yet lock
 721                  * all cleaned locks are marked by CLEANED flag. */
 722                 lock_res(res);
 723                 cfs_list_for_each(tmp, q) {
 724                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 725                                               l_res_link);
 726                         if (lock->l_flags & LDLM_FL_CLEANED) {
 727                                 lock = NULL;
 728                                 continue;
 729                         }
 730                         LDLM_LOCK_GET(lock);
 731                         lock->l_flags |= LDLM_FL_CLEANED;
 732                         break;
 733                 }
 734
 735                 if (lock == NULL) {
 736                         unlock_res(res);
 737                         break;
 738                 }
 739
 740                 /* Set CBPENDING so nothing in the cancellation path
 741                  * can match this lock. */
 742                 lock->l_flags |= LDLM_FL_CBPENDING;
 743                 lock->l_flags |= LDLM_FL_FAILED;
 744                 lock->l_flags |= flags;
 745
 746                 /* ... without sending a CANCEL message for local_only. */
 747                 if (local_only)
 748                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 749
 750                 if (local_only && (lock->l_readers || lock->l_writers)) {
 751                         /* This is a little bit gross, but much better than the
 752                          * alternative: pretend that we got a blocking AST from
 753                          * the server, so that when the lock is decref'd, it
 754                          * will go away ... */
 755                         unlock_res(res);
 756                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 757                         if (lock->l_completion_ast)
 758                                 lock->l_completion_ast(lock, 0, NULL);
 759                         LDLM_LOCK_RELEASE(lock);
 760                         continue;
 761                 }
 762
 763                 if (client) {
 764                         struct lustre_handle lockh;
 765
 766                         unlock_res(res);
 767                         ldlm_lock2handle(lock, &lockh);
 768                         rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 769                         if (rc)
 770                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 771                 } else {
 772                         ldlm_resource_unlink_lock(lock);
 773                         unlock_res(res);
 774                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 775                                    "client node");
 776                         ldlm_lock_destroy(lock);
 777                 }
 778                 LDLM_LOCK_RELEASE(lock);
 779         } while (1);
 780 }
 781
 782 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 783                                cfs_hlist_node_t *hnode, void *arg)
 784 {
 785         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 786         __u64 flags = *(__u64 *)arg;
 787
 788         cleanup_resource(res, &res->lr_granted, flags);
 789         cleanup_resource(res, &res->lr_converting, flags);
 790         cleanup_resource(res, &res->lr_waiting, flags);
 791
 792         return 0;
 793 }
 794
 795 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 796                                   cfs_hlist_node_t *hnode, void *arg)
 797 {
 798         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 799
 800         lock_res(res);
 801         CERROR("Namespace %s resource refcount nonzero "
 802                "(%d) after lock cleanup; forcing "
 803                "cleanup.\n",
 804                ldlm_ns_name(ldlm_res_to_ns(res)),
 805                cfs_atomic_read(&res->lr_refcount) - 1);
 806
 807         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 808                LPU64") (rc: %d)\n", res,
 809                res->lr_name.name[0], res->lr_name.name[1],
 810                res->lr_name.name[2], res->lr_name.name[3],
 811                cfs_atomic_read(&res->lr_refcount) - 1);
 812
 813         ldlm_resource_dump(D_ERROR, res);
 814         unlock_res(res);
 815         return 0;
 816 }
 817
 818 /**
 819  * Cancel and destroy all locks in the namespace.
 820  *
 821  * Typically used during evictions when server notified client that it was
 822  * evicted and all of its state needs to be destroyed.
 823  * Also used during shutdown.
 824  */
 825 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 826 {
 827         if (ns == NULL) {
 828                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 829                 return ELDLM_OK;
 830         }
 831
 832         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
 833         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 834         return ELDLM_OK;
 835 }
 836 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 837
 838 /**
 839  * Attempts to free namespace.
 840  *
 841  * Only used when namespace goes away, like during an unmount.
 842  */
 843 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 844 {
 845         ENTRY;
 846
 847         /* At shutdown time, don't call the cancellation callback */
 848         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 849
 850         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 851                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 852                 int rc;
 853                 CDEBUG(D_DLMTRACE,
 854                        "dlm namespace %s free waiting on refcount %d\n",
 855                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 856 force_wait:
 857                 if (force)
 858                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 859
 860                 rc = l_wait_event(ns->ns_waitq,
 861                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 862
 863                 /* Forced cleanups should be able to reclaim all references,
 864                  * so it's safe to wait forever... we can't leak locks... */
 865                 if (force && rc == -ETIMEDOUT) {
 866                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 867                                        "namespace with %d resources in use, "
 868                                        "(rc=%d)\n", ldlm_ns_name(ns),
 869                                        cfs_atomic_read(&ns->ns_bref), rc);
 870                         GOTO(force_wait, rc);
 871                 }
 872
 873                 if (cfs_atomic_read(&ns->ns_bref)) {
 874                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 875                                        "with %d resources in use, (rc=%d)\n",
 876                                        ldlm_ns_name(ns),
 877                                        cfs_atomic_read(&ns->ns_bref), rc);
 878                         RETURN(ELDLM_NAMESPACE_EXISTS);
 879                 }
 880                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 881                        ldlm_ns_name(ns));
 882         }
 883
 884         RETURN(ELDLM_OK);
 885 }
 886
 887 /**
 888  * Performs various cleanups for passed \a ns to make it drop refc and be
 889  * ready for freeing. Waits for refc == 0.
 890  *
 891  * The following is done:
 892  * (0) Unregister \a ns from its list to make inaccessible for potential
 893  * users like pools thread and others;
 894  * (1) Clear all locks in \a ns.
 895  */
 896 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 897                                struct obd_import *imp,
 898                                int force)
 899 {
 900         int rc;
 901         ENTRY;
 902         if (!ns) {
 903                 EXIT;
 904                 return;
 905         }
 906
 907         spin_lock(&ns->ns_lock);
 908         ns->ns_stopping = 1;
 909         spin_unlock(&ns->ns_lock);
 910
 911         /*
 912          * Can fail with -EINTR when force == 0 in which case try harder.
 913          */
 914         rc = __ldlm_namespace_free(ns, force);
 915         if (rc != ELDLM_OK) {
 916                 if (imp) {
 917                         ptlrpc_disconnect_import(imp, 0);
 918                         ptlrpc_invalidate_import(imp);
 919                 }
 920
 921                 /*
 922                  * With all requests dropped and the import inactive
 923                  * we are gaurenteed all reference will be dropped.
 924                  */
 925                 rc = __ldlm_namespace_free(ns, 1);
 926                 LASSERT(rc == 0);
 927         }
 928         EXIT;
 929 }
 930
 931 /**
 932  * Performs freeing memory structures related to \a ns. This is only done
 933  * when ldlm_namespce_free_prior() successfully removed all resources
 934  * referencing \a ns and its refc == 0.
 935  */
 936 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 937 {
 938         ENTRY;
 939         if (!ns) {
 940                 EXIT;
 941                 return;
 942         }
 943
 944         /* Make sure that nobody can find this ns in its list. */
 945         ldlm_namespace_unregister(ns, ns->ns_client);
 946         /* Fini pool _before_ parent proc dir is removed. This is important as
 947          * ldlm_pool_fini() removes own proc dir which is child to @dir.
 948          * Removing it after @dir may cause oops. */
 949         ldlm_pool_fini(&ns->ns_pool);
 950
 951         ldlm_namespace_proc_unregister(ns);
 952         cfs_hash_putref(ns->ns_rs_hash);
 953         /* Namespace \a ns should be not on list at this time, otherwise
 954          * this will cause issues related to using freed \a ns in poold
 955          * thread. */
 956         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 957         OBD_FREE_PTR(ns);
 958         ldlm_put_ref();
 959         EXIT;
 960 }
 961
 962 /**
 963  * Cleanup the resource, and free namespace.
 964  * bug 12864:
 965  * Deadlock issue:
 966  * proc1: destroy import
 967  *        class_disconnect_export(grab cl_sem) ->
 968  *              -> ldlm_namespace_free ->
 969  *              -> lprocfs_remove(grab _lprocfs_lock).
 970  * proc2: read proc info
 971  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 972  *              -> osc_rd_active, etc(grab cl_sem).
 973  *
 974  * So that I have to split the ldlm_namespace_free into two parts - the first
 975  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 976  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 977  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 978  * held.
 979  */
 980 void ldlm_namespace_free(struct ldlm_namespace *ns,
 981                          struct obd_import *imp,
 982                          int force)
 983 {
 984         ldlm_namespace_free_prior(ns, imp, force);
 985         ldlm_namespace_free_post(ns);
 986 }
 987 EXPORT_SYMBOL(ldlm_namespace_free);
 988
 989 void ldlm_namespace_get(struct ldlm_namespace *ns)
 990 {
 991         cfs_atomic_inc(&ns->ns_bref);
 992 }
 993 EXPORT_SYMBOL(ldlm_namespace_get);
 994
 995 void ldlm_namespace_put(struct ldlm_namespace *ns)
 996 {
 997         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 998                 cfs_waitq_signal(&ns->ns_waitq);
 999                 spin_unlock(&ns->ns_lock);
1000         }
1001 }
1002 EXPORT_SYMBOL(ldlm_namespace_put);
1003
1004 /** Register \a ns in the list of namespaces */
1005 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
1006 {
1007         mutex_lock(ldlm_namespace_lock(client));
1008         LASSERT(cfs_list_empty(&ns->ns_list_chain));
1009         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
1010         cfs_atomic_inc(ldlm_namespace_nr(client));
1011         mutex_unlock(ldlm_namespace_lock(client));
1012 }
1013
1014 /** Unregister \a ns from the list of namespaces. */
1015 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1016 {
1017         mutex_lock(ldlm_namespace_lock(client));
1018         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1019         /* Some asserts and possibly other parts of the code are still
1020          * using list_empty(&ns->ns_list_chain). This is why it is
1021          * important to use list_del_init() here. */
1022         cfs_list_del_init(&ns->ns_list_chain);
1023         cfs_atomic_dec(ldlm_namespace_nr(client));
1024         mutex_unlock(ldlm_namespace_lock(client));
1025 }
1026
1027 /** Should be called with ldlm_namespace_lock(client) taken. */
1028 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
1029 {
1030         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1031         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1032         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1033 }
1034
1035 /** Should be called with ldlm_namespace_lock(client) taken. */
1036 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1037 {
1038         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1039         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
1040         return container_of(ldlm_namespace_list(client)->next,
1041                 struct ldlm_namespace, ns_list_chain);
1042 }
1043
1044 /** Create and initialize new resource. */
1045 static struct ldlm_resource *ldlm_resource_new(void)
1046 {
1047         struct ldlm_resource *res;
1048         int idx;
1049
1050         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
1051         if (res == NULL)
1052                 return NULL;
1053
1054         CFS_INIT_LIST_HEAD(&res->lr_granted);
1055         CFS_INIT_LIST_HEAD(&res->lr_converting);
1056         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1057
1058         /* Initialize interval trees for each lock mode. */
1059         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1060                 res->lr_itree[idx].lit_size = 0;
1061                 res->lr_itree[idx].lit_mode = 1 << idx;
1062                 res->lr_itree[idx].lit_root = NULL;
1063         }
1064
1065         cfs_atomic_set(&res->lr_refcount, 1);
1066         spin_lock_init(&res->lr_lock);
1067         lu_ref_init(&res->lr_reference);
1068
1069         /* The creator of the resource must unlock the mutex after LVB
1070          * initialization. */
1071         mutex_init(&res->lr_lvb_mutex);
1072         mutex_lock(&res->lr_lvb_mutex);
1073
1074         return res;
1075 }
1076
1077 /**
1078  * Return a reference to resource with given name, creating it if necessary.
1079  * Args: namespace with ns_lock unlocked
1080  * Locks: takes and releases NS hash-lock and res->lr_lock
1081  * Returns: referenced, unlocked ldlm_resource or NULL
1082  */
1083 struct ldlm_resource *
1084 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1085                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1086 {
1087         cfs_hlist_node_t     *hnode;
1088         struct ldlm_resource *res;
1089         cfs_hash_bd_t         bd;
1090         __u64                 version;
1091
1092         LASSERT(ns != NULL);
1093         LASSERT(parent == NULL);
1094         LASSERT(ns->ns_rs_hash != NULL);
1095         LASSERT(name->name[0] != 0);
1096
1097         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1098         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1099         if (hnode != NULL) {
1100                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1101                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1102                 /* Synchronize with regard to resource creation. */
1103                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1104                         mutex_lock(&res->lr_lvb_mutex);
1105                         mutex_unlock(&res->lr_lvb_mutex);
1106                 }
1107
1108                 if (unlikely(res->lr_lvb_len < 0)) {
1109                         ldlm_resource_putref(res);
1110                         res = NULL;
1111                 }
1112                 return res;
1113         }
1114
1115         version = cfs_hash_bd_version_get(&bd);
1116         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1117
1118         if (create == 0)
1119                 return NULL;
1120
1121         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1122                  "type: %d\n", type);
1123         res = ldlm_resource_new();
1124         if (!res)
1125                 return NULL;
1126
1127         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1128         res->lr_name       = *name;
1129         res->lr_type       = type;
1130         res->lr_most_restr = LCK_NL;
1131
1132         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1133         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1134                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1135
1136         if (hnode != NULL) {
1137                 /* Someone won the race and already added the resource. */
1138                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1139                 /* Clean lu_ref for failed resource. */
1140                 lu_ref_fini(&res->lr_reference);
1141                 /* We have taken lr_lvb_mutex. Drop it. */
1142                 mutex_unlock(&res->lr_lvb_mutex);
1143                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1144
1145                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1146                 /* Synchronize with regard to resource creation. */
1147                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1148                         mutex_lock(&res->lr_lvb_mutex);
1149                         mutex_unlock(&res->lr_lvb_mutex);
1150                 }
1151
1152                 if (unlikely(res->lr_lvb_len < 0)) {
1153                         ldlm_resource_putref(res);
1154                         res = NULL;
1155                 }
1156                 return res;
1157         }
1158         /* We won! Let's add the resource. */
1159         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1160         if (cfs_hash_bd_count_get(&bd) == 1)
1161                 ldlm_namespace_get(ns);
1162
1163         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1164         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1165                 int rc;
1166
1167                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1168                 rc = ns->ns_lvbo->lvbo_init(res);
1169                 if (rc < 0) {
1170                         CERROR("lvbo_init failed for resource "
1171                                LPU64": rc %d\n", name->name[0], rc);
1172                         if (res->lr_lvb_data) {
1173                                 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1174                                 res->lr_lvb_data = NULL;
1175                         }
1176                         res->lr_lvb_len = rc;
1177                         mutex_unlock(&res->lr_lvb_mutex);
1178                         ldlm_resource_putref(res);
1179                         return NULL;
1180                 }
1181         }
1182
1183         /* We create resource with locked lr_lvb_mutex. */
1184         mutex_unlock(&res->lr_lvb_mutex);
1185
1186         return res;
1187 }
1188 EXPORT_SYMBOL(ldlm_resource_get);
1189
1190 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1191 {
1192         LASSERT(res != NULL);
1193         LASSERT(res != LP_POISON);
1194         cfs_atomic_inc(&res->lr_refcount);
1195         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1196                cfs_atomic_read(&res->lr_refcount));
1197         return res;
1198 }
1199
1200 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1201                                          struct ldlm_resource *res)
1202 {
1203         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1204
1205         if (!cfs_list_empty(&res->lr_granted)) {
1206                 ldlm_resource_dump(D_ERROR, res);
1207                 LBUG();
1208         }
1209
1210         if (!cfs_list_empty(&res->lr_converting)) {
1211                 ldlm_resource_dump(D_ERROR, res);
1212                 LBUG();
1213         }
1214
1215         if (!cfs_list_empty(&res->lr_waiting)) {
1216                 ldlm_resource_dump(D_ERROR, res);
1217                 LBUG();
1218         }
1219
1220         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1221                                bd, &res->lr_hash);
1222         lu_ref_fini(&res->lr_reference);
1223         if (cfs_hash_bd_count_get(bd) == 0)
1224                 ldlm_namespace_put(nsb->nsb_namespace);
1225 }
1226
1227 /* Returns 1 if the resource was freed, 0 if it remains. */
1228 int ldlm_resource_putref(struct ldlm_resource *res)
1229 {
1230         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1231         cfs_hash_bd_t   bd;
1232
1233         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1234         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1235                res, cfs_atomic_read(&res->lr_refcount) - 1);
1236
1237         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1238         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1239                 __ldlm_resource_putref_final(&bd, res);
1240                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1241                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1242                         ns->ns_lvbo->lvbo_free(res);
1243                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1244                 return 1;
1245         }
1246         return 0;
1247 }
1248 EXPORT_SYMBOL(ldlm_resource_putref);
1249
1250 /* Returns 1 if the resource was freed, 0 if it remains. */
1251 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1252 {
1253         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1254
1255         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1256         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1257                res, cfs_atomic_read(&res->lr_refcount) - 1);
1258
1259         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1260                 cfs_hash_bd_t bd;
1261
1262                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1263                                 &res->lr_name, &bd);
1264                 __ldlm_resource_putref_final(&bd, res);
1265                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1266                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1267                  * so we should never be here while calling cfs_hash_del,
1268                  * cfs_hash_for_each_nolock is the only case we can get
1269                  * here, which is safe to release cfs_hash_bd_lock.
1270                  */
1271                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1272                         ns->ns_lvbo->lvbo_free(res);
1273                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1274
1275                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1276                 return 1;
1277         }
1278         return 0;
1279 }
1280
1281 /**
1282  * Add a lock into a given resource into specified lock list.
1283  */
1284 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1285                             struct ldlm_lock *lock)
1286 {
1287         check_res_locked(res);
1288
1289         LDLM_DEBUG(lock, "About to add this lock:\n");
1290
1291         if (lock->l_destroyed) {
1292                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1293                 return;
1294         }
1295
1296         LASSERT(cfs_list_empty(&lock->l_res_link));
1297
1298         cfs_list_add_tail(&lock->l_res_link, head);
1299 }
1300
1301 /**
1302  * Insert a lock into resource after specified lock.
1303  *
1304  * Obtain resource description from the lock we are inserting after.
1305  */
1306 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1307                                      struct ldlm_lock *new)
1308 {
1309         struct ldlm_resource *res = original->l_resource;
1310
1311         check_res_locked(res);
1312
1313         ldlm_resource_dump(D_INFO, res);
1314         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1315
1316         if (new->l_destroyed) {
1317                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1318                 goto out;
1319         }
1320
1321         LASSERT(cfs_list_empty(&new->l_res_link));
1322
1323         cfs_list_add(&new->l_res_link, &original->l_res_link);
1324  out:;
1325 }
1326
1327 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1328 {
1329         int type = lock->l_resource->lr_type;
1330
1331         check_res_locked(lock->l_resource);
1332         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1333                 ldlm_unlink_lock_skiplist(lock);
1334         else if (type == LDLM_EXTENT)
1335                 ldlm_extent_unlink_lock(lock);
1336         cfs_list_del_init(&lock->l_res_link);
1337 }
1338 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1339
1340 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1341 {
1342         desc->lr_type = res->lr_type;
1343         desc->lr_name = res->lr_name;
1344 }
1345
1346 /**
1347  * Print information about all locks in all namespaces on this node to debug
1348  * log.
1349  */
1350 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1351 {
1352         cfs_list_t *tmp;
1353
1354         if (!((libcfs_debug | D_ERROR) & level))
1355                 return;
1356
1357         mutex_lock(ldlm_namespace_lock(client));
1358
1359         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1360                 struct ldlm_namespace *ns;
1361                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1362                 ldlm_namespace_dump(level, ns);
1363         }
1364
1365         mutex_unlock(ldlm_namespace_lock(client));
1366 }
1367 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1368
1369 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1370                               cfs_hlist_node_t *hnode, void *arg)
1371 {
1372         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1373         int    level = (int)(unsigned long)arg;
1374
1375         lock_res(res);
1376         ldlm_resource_dump(level, res);
1377         unlock_res(res);
1378
1379         return 0;
1380 }
1381
1382 /**
1383  * Print information about all locks in this namespace on this node to debug
1384  * log.
1385  */
1386 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1387 {
1388         if (!((libcfs_debug | D_ERROR) & level))
1389                 return;
1390
1391         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1392                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1393                ns_is_client(ns) ? "client" : "server");
1394
1395         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1396                 return;
1397
1398         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1399                                  ldlm_res_hash_dump,
1400                                  (void *)(unsigned long)level);
1401         spin_lock(&ns->ns_lock);
1402         ns->ns_next_dump = cfs_time_shift(10);
1403         spin_unlock(&ns->ns_lock);
1404 }
1405 EXPORT_SYMBOL(ldlm_namespace_dump);
1406
1407 /**
1408  * Print information about all locks in this resource to debug log.
1409  */
1410 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1411 {
1412         struct ldlm_lock *lock;
1413         unsigned int granted = 0;
1414
1415         CLASSERT(RES_NAME_SIZE == 4);
1416
1417         if (!((libcfs_debug | D_ERROR) & level))
1418                 return;
1419
1420         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1421                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1422                res->lr_name.name[2], res->lr_name.name[3],
1423                cfs_atomic_read(&res->lr_refcount));
1424
1425         if (!cfs_list_empty(&res->lr_granted)) {
1426                 CDEBUG(level, "Granted locks (in reverse order):\n");
1427                 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1428                                                 l_res_link) {
1429                         LDLM_DEBUG_LIMIT(level, lock, "###");
1430                         if (!(level & D_CANTMASK) &&
1431                             ++granted > ldlm_dump_granted_max) {
1432                                 CDEBUG(level, "only dump %d granted locks to "
1433                                        "avoid DDOS.\n", granted);
1434                                 break;
1435                         }
1436                 }
1437         }
1438         if (!cfs_list_empty(&res->lr_converting)) {
1439                 CDEBUG(level, "Converting locks:\n");
1440                 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1441                         LDLM_DEBUG_LIMIT(level, lock, "###");
1442         }
1443         if (!cfs_list_empty(&res->lr_waiting)) {
1444                 CDEBUG(level, "Waiting locks:\n");
1445                 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1446                         LDLM_DEBUG_LIMIT(level, lock, "###");
1447         }
1448 }