lustre/ldlm/ldlm_resource.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ldlm/ldlm_resource.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43 #ifdef __KERNEL__
  44 # include <lustre_dlm.h>
  45 #else
  46 # include <liblustre.h>
  47 #endif
  48
  49 #include <lustre_fid.h>
  50 #include <obd_class.h>
  51 #include "ldlm_internal.h"
  52
  53 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  54
  55 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
  56 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
  57
  58 cfs_semaphore_t ldlm_srv_namespace_lock;
  59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  60
  61 cfs_semaphore_t ldlm_cli_namespace_lock;
  62 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  63
  64 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  65 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  66 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  67
  68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  69
  70 #ifdef LPROCFS
  71 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  72                              unsigned long count, void *data)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78
  79 int ldlm_proc_setup(void)
  80 {
  81         int rc;
  82         struct lprocfs_vars list[] = {
  83                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  84                 { "cancel_unused_locks_before_replay",
  85                   lprocfs_rd_uint, lprocfs_wr_uint,
  86                   &ldlm_cancel_unused_locks_before_replay, NULL },
  87                 { NULL }};
  88         ENTRY;
  89         LASSERT(ldlm_ns_proc_dir == NULL);
  90
  91         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  92                                               proc_lustre_root,
  93                                               NULL, NULL);
  94         if (IS_ERR(ldlm_type_proc_dir)) {
  95                 CERROR("LProcFS failed in ldlm-init\n");
  96                 rc = PTR_ERR(ldlm_type_proc_dir);
  97                 GOTO(err, rc);
  98         }
  99
 100         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 101                                             ldlm_type_proc_dir,
 102                                             NULL, NULL);
 103         if (IS_ERR(ldlm_ns_proc_dir)) {
 104                 CERROR("LProcFS failed in ldlm-init\n");
 105                 rc = PTR_ERR(ldlm_ns_proc_dir);
 106                 GOTO(err_type, rc);
 107         }
 108
 109         ldlm_svc_proc_dir = lprocfs_register("services",
 110                                             ldlm_type_proc_dir,
 111                                             NULL, NULL);
 112         if (IS_ERR(ldlm_svc_proc_dir)) {
 113                 CERROR("LProcFS failed in ldlm-init\n");
 114                 rc = PTR_ERR(ldlm_svc_proc_dir);
 115                 GOTO(err_ns, rc);
 116         }
 117
 118         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 119
 120         RETURN(0);
 121
 122 err_ns:
 123         lprocfs_remove(&ldlm_ns_proc_dir);
 124 err_type:
 125         lprocfs_remove(&ldlm_type_proc_dir);
 126 err:
 127         ldlm_svc_proc_dir = NULL;
 128         RETURN(rc);
 129 }
 130
 131 void ldlm_proc_cleanup(void)
 132 {
 133         if (ldlm_svc_proc_dir)
 134                 lprocfs_remove(&ldlm_svc_proc_dir);
 135
 136         if (ldlm_ns_proc_dir)
 137                 lprocfs_remove(&ldlm_ns_proc_dir);
 138
 139         if (ldlm_type_proc_dir)
 140                 lprocfs_remove(&ldlm_type_proc_dir);
 141 }
 142
 143 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 144                                    int count, int *eof, void *data)
 145 {
 146         struct ldlm_namespace *ns  = data;
 147         __u64                  res = 0;
 148         cfs_hash_bd_t          bd;
 149         int                    i;
 150
 151         /* result is not strictly consistant */
 152         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 153                 res += cfs_hash_bd_count_get(&bd);
 154         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 155 }
 156
 157 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 158                                int count, int *eof, void *data)
 159 {
 160         struct ldlm_namespace *ns = data;
 161         __u64                  locks;
 162
 163         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 164                                         LPROCFS_FIELDS_FLAGS_SUM);
 165         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 166 }
 167
 168 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 169                                int count, int *eof, void *data)
 170 {
 171         struct ldlm_namespace *ns = data;
 172         __u32 *nr = &ns->ns_max_unused;
 173
 174         if (ns_connect_lru_resize(ns))
 175                 nr = &ns->ns_nr_unused;
 176         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 177 }
 178
 179 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 180                                unsigned long count, void *data)
 181 {
 182         struct ldlm_namespace *ns = data;
 183         char dummy[MAX_STRING_SIZE + 1], *end;
 184         unsigned long tmp;
 185         int lru_resize;
 186
 187         dummy[MAX_STRING_SIZE] = '\0';
 188         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 189                 return -EFAULT;
 190
 191         if (strncmp(dummy, "clear", 5) == 0) {
 192                 CDEBUG(D_DLMTRACE,
 193                        "dropping all unused locks from namespace %s\n",
 194                        ldlm_ns_name(ns));
 195                 if (ns_connect_lru_resize(ns)) {
 196                         int canceled, unused  = ns->ns_nr_unused;
 197
 198                         /* Try to cancel all @ns_nr_unused locks. */
 199                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 200                                                    LDLM_CANCEL_PASSED);
 201                         if (canceled < unused) {
 202                                 CDEBUG(D_DLMTRACE,
 203                                        "not all requested locks are canceled, "
 204                                        "requested: %d, canceled: %d\n", unused,
 205                                        canceled);
 206                                 return -EINVAL;
 207                         }
 208                 } else {
 209                         tmp = ns->ns_max_unused;
 210                         ns->ns_max_unused = 0;
 211                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 212                         ns->ns_max_unused = tmp;
 213                 }
 214                 return count;
 215         }
 216
 217         tmp = simple_strtoul(dummy, &end, 0);
 218         if (dummy == end) {
 219                 CERROR("invalid value written\n");
 220                 return -EINVAL;
 221         }
 222         lru_resize = (tmp == 0);
 223
 224         if (ns_connect_lru_resize(ns)) {
 225                 if (!lru_resize)
 226                         ns->ns_max_unused = (unsigned int)tmp;
 227
 228                 if (tmp > ns->ns_nr_unused)
 229                         tmp = ns->ns_nr_unused;
 230                 tmp = ns->ns_nr_unused - tmp;
 231
 232                 CDEBUG(D_DLMTRACE,
 233                        "changing namespace %s unused locks from %u to %u\n",
 234                        ldlm_ns_name(ns), ns->ns_nr_unused,
 235                        (unsigned int)tmp);
 236                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 237
 238                 if (!lru_resize) {
 239                         CDEBUG(D_DLMTRACE,
 240                                "disable lru_resize for namespace %s\n",
 241                                ldlm_ns_name(ns));
 242                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 243                 }
 244         } else {
 245                 CDEBUG(D_DLMTRACE,
 246                        "changing namespace %s max_unused from %u to %u\n",
 247                        ldlm_ns_name(ns), ns->ns_max_unused,
 248                        (unsigned int)tmp);
 249                 ns->ns_max_unused = (unsigned int)tmp;
 250                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 251
 252                 /* Make sure that originally lru resize was supported before
 253                  * turning it on here. */
 254                 if (lru_resize &&
 255                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 256                         CDEBUG(D_DLMTRACE,
 257                                "enable lru_resize for namespace %s\n",
 258                                ldlm_ns_name(ns));
 259                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 260                 }
 261         }
 262
 263         return count;
 264 }
 265
 266 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 267 {
 268         struct proc_dir_entry *dir;
 269
 270         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 271         if (dir == NULL) {
 272                 CERROR("dlm namespace %s has no procfs dir?\n",
 273                        ldlm_ns_name(ns));
 274         } else {
 275                 lprocfs_remove(&dir);
 276         }
 277
 278         if (ns->ns_stats != NULL)
 279                 lprocfs_free_stats(&ns->ns_stats);
 280 }
 281
 282 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 283 {
 284         struct lprocfs_vars lock_vars[2];
 285         char lock_name[MAX_STRING_SIZE + 1];
 286
 287         LASSERT(ns != NULL);
 288         LASSERT(ns->ns_rs_hash != NULL);
 289
 290         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 291         if (ns->ns_stats == NULL)
 292                 return -ENOMEM;
 293
 294         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 295                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 296
 297         lock_name[MAX_STRING_SIZE] = '\0';
 298
 299         memset(lock_vars, 0, sizeof(lock_vars));
 300         lock_vars[0].name = lock_name;
 301
 302         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 303                  ldlm_ns_name(ns));
 304         lock_vars[0].data = ns;
 305         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 306         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 307
 308         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 309                  ldlm_ns_name(ns));
 310         lock_vars[0].data = ns;
 311         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 312         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 313
 314         if (ns_is_client(ns)) {
 315                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 316                          ldlm_ns_name(ns));
 317                 lock_vars[0].data = &ns->ns_nr_unused;
 318                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 319                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 320
 321                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 322                          ldlm_ns_name(ns));
 323                 lock_vars[0].data = ns;
 324                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 325                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 326                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 327
 328                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 329                          ldlm_ns_name(ns));
 330                 lock_vars[0].data = &ns->ns_max_age;
 331                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 332                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 333                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 334         } else {
 335                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 336                          ldlm_ns_name(ns));
 337                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 338                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 339                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 340                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 341
 342                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 343                          ldlm_ns_name(ns));
 344                 lock_vars[0].data = &ns->ns_timeouts;
 345                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 346                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 347
 348                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 349                          ldlm_ns_name(ns));
 350                 lock_vars[0].data = &ns->ns_max_nolock_size;
 351                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 352                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 354
 355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 356                          ldlm_ns_name(ns));
 357                 lock_vars[0].data = &ns->ns_contention_time;
 358                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 359                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 360                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 361
 362                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 363                          ldlm_ns_name(ns));
 364                 lock_vars[0].data = &ns->ns_contended_locks;
 365                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 366                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 367                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 368         }
 369         return 0;
 370 }
 371 #undef MAX_STRING_SIZE
 372 #else /* LPROCFS */
 373
 374 #define ldlm_namespace_proc_unregister(ns)      ({;})
 375 #define ldlm_namespace_proc_register(ns)        ({0;})
 376
 377 #endif /* LPROCFS */
 378
 379 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 380                                   const void *key, unsigned mask)
 381 {
 382         const struct ldlm_res_id     *id  = key;
 383         unsigned                val = 0;
 384         unsigned                i;
 385
 386         for (i = 0; i < RES_NAME_SIZE; i++)
 387                 val += id->name[i];
 388         return val & mask;
 389 }
 390
 391 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 392                                       const void *key, unsigned mask)
 393 {
 394         const struct ldlm_res_id *id = key;
 395         struct lu_fid       fid;
 396         __u64               hash;
 397
 398         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 399         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF];
 400         fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF];
 401
 402         hash = fid_flatten(&fid);
 403         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 404         /* ignore a few low bits */
 405         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0)
 406                 hash += id->name[LUSTRE_RES_ID_HSH_OFF] >> 5;
 407         else
 408                 hash = hash >> 5;
 409         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 410         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 411
 412         return hash & mask;
 413 }
 414
 415 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 416 {
 417         struct ldlm_resource   *res;
 418
 419         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 420         return &res->lr_name;
 421 }
 422
 423 static int ldlm_res_eq(const struct ldlm_res_id *res0,
 424                        const struct ldlm_res_id *res1)
 425 {
 426         return !memcmp(res0, res1, sizeof(*res0));
 427 }
 428
 429 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 430 {
 431         struct ldlm_resource   *res;
 432
 433         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 434         return ldlm_res_eq((const struct ldlm_res_id *)key,
 435                            (const struct ldlm_res_id *)&res->lr_name);
 436 }
 437
 438 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 439 {
 440         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 441 }
 442
 443 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 444 {
 445         struct ldlm_resource *res;
 446
 447         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 448         ldlm_resource_getref(res);
 449         LDLM_RESOURCE_ADDREF(res);
 450 }
 451
 452 static void ldlm_res_hop_put_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         /* cfs_hash_for_each_nolock is the only chance we call it */
 458         LDLM_RESOURCE_DELREF(res);
 459         ldlm_resource_putref_locked(res);
 460 }
 461
 462 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 463 {
 464         struct ldlm_resource *res;
 465
 466         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 467         LDLM_RESOURCE_DELREF(res);
 468         ldlm_resource_putref(res);
 469 }
 470
 471 cfs_hash_ops_t ldlm_ns_hash_ops = {
 472         .hs_hash        = ldlm_res_hop_hash,
 473         .hs_key         = ldlm_res_hop_key,
 474         .hs_keycmp      = ldlm_res_hop_keycmp,
 475         .hs_keycpy      = NULL,
 476         .hs_object      = ldlm_res_hop_object,
 477         .hs_get         = ldlm_res_hop_get_locked,
 478         .hs_put_locked  = ldlm_res_hop_put_locked,
 479         .hs_put         = ldlm_res_hop_put
 480 };
 481
 482 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 483         .hs_hash        = ldlm_res_hop_fid_hash,
 484         .hs_key         = ldlm_res_hop_key,
 485         .hs_keycmp      = ldlm_res_hop_keycmp,
 486         .hs_keycpy      = NULL,
 487         .hs_object      = ldlm_res_hop_object,
 488         .hs_get         = ldlm_res_hop_get_locked,
 489         .hs_put_locked  = ldlm_res_hop_put_locked,
 490         .hs_put         = ldlm_res_hop_put
 491 };
 492
 493 typedef struct {
 494         ldlm_ns_type_t  nsd_type;
 495         /** hash bucket bits */
 496         unsigned        nsd_bkt_bits;
 497         /** hash bits */
 498         unsigned        nsd_all_bits;
 499         /** hash operations */
 500         cfs_hash_ops_t *nsd_hops;
 501 } ldlm_ns_hash_def_t;
 502
 503 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 504 {
 505         {
 506                 .nsd_type       = LDLM_NS_TYPE_MDC,
 507                 .nsd_bkt_bits   = 11,
 508                 .nsd_all_bits   = 15,
 509                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 510         },
 511         {
 512                 .nsd_type       = LDLM_NS_TYPE_MDT,
 513                 .nsd_bkt_bits   = 14,
 514                 .nsd_all_bits   = 21,
 515                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 516         },
 517         {
 518                 .nsd_type       = LDLM_NS_TYPE_OSC,
 519                 .nsd_bkt_bits   = 8,
 520                 .nsd_all_bits   = 12,
 521                 .nsd_hops       = &ldlm_ns_hash_ops,
 522         },
 523         {
 524                 .nsd_type       = LDLM_NS_TYPE_OST,
 525                 .nsd_bkt_bits   = 11,
 526                 .nsd_all_bits   = 17,
 527                 .nsd_hops       = &ldlm_ns_hash_ops,
 528         },
 529         {
 530                 .nsd_type       = LDLM_NS_TYPE_MGC,
 531                 .nsd_bkt_bits   = 4,
 532                 .nsd_all_bits   = 4,
 533                 .nsd_hops       = &ldlm_ns_hash_ops,
 534         },
 535         {
 536                 .nsd_type       = LDLM_NS_TYPE_MGT,
 537                 .nsd_bkt_bits   = 4,
 538                 .nsd_all_bits   = 4,
 539                 .nsd_hops       = &ldlm_ns_hash_ops,
 540         },
 541         {
 542                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 543         },
 544 };
 545
 546 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 547                                           ldlm_side_t client,
 548                                           ldlm_appetite_t apt,
 549                                           ldlm_ns_type_t ns_type)
 550 {
 551         struct ldlm_namespace *ns = NULL;
 552         struct ldlm_ns_bucket *nsb;
 553         ldlm_ns_hash_def_t    *nsd;
 554         cfs_hash_bd_t          bd;
 555         int                    idx;
 556         int                    rc;
 557         ENTRY;
 558
 559         LASSERT(obd != NULL);
 560
 561         rc = ldlm_get_ref();
 562         if (rc) {
 563                 CERROR("ldlm_get_ref failed: %d\n", rc);
 564                 RETURN(NULL);
 565         }
 566
 567         for (idx = 0;;idx++) {
 568                 nsd = &ldlm_ns_hash_defs[idx];
 569                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 570                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 571                         GOTO(out_ref, NULL);
 572                 }
 573
 574                 if (nsd->nsd_type == ns_type)
 575                         break;
 576         }
 577
 578         OBD_ALLOC_PTR(ns);
 579         if (!ns)
 580                 GOTO(out_ref, NULL);
 581
 582         ns->ns_rs_hash = cfs_hash_create(name,
 583                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 584                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 585                                          CFS_HASH_MIN_THETA,
 586                                          CFS_HASH_MAX_THETA,
 587                                          nsd->nsd_hops,
 588                                          CFS_HASH_DEPTH |
 589                                          CFS_HASH_BIGNAME |
 590                                          CFS_HASH_SPIN_BKTLOCK |
 591                                          CFS_HASH_NO_ITEMREF);
 592         if (ns->ns_rs_hash == NULL)
 593                 GOTO(out_ns, NULL);
 594
 595         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 596                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 597                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 598                 nsb->nsb_namespace = ns;
 599         }
 600
 601         ns->ns_obd      = obd;
 602         ns->ns_appetite = apt;
 603         ns->ns_client   = client;
 604
 605         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 606         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 607         cfs_spin_lock_init(&ns->ns_lock);
 608         cfs_atomic_set(&ns->ns_bref, 0);
 609         cfs_waitq_init(&ns->ns_waitq);
 610
 611         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 612         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 613         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 614
 615         ns->ns_nr_unused          = 0;
 616         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 617         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 618         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 619         ns->ns_timeouts           = 0;
 620         ns->ns_orig_connect_flags = 0;
 621         ns->ns_connect_flags      = 0;
 622         rc = ldlm_namespace_proc_register(ns);
 623         if (rc != 0) {
 624                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 625                 GOTO(out_hash, rc);
 626         }
 627
 628         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 629         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 630         if (rc) {
 631                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 632                 GOTO(out_proc, rc);
 633         }
 634
 635         ldlm_namespace_register(ns, client);
 636         RETURN(ns);
 637 out_proc:
 638         ldlm_namespace_proc_unregister(ns);
 639         ldlm_namespace_cleanup(ns, 0);
 640 out_hash:
 641         cfs_hash_putref(ns->ns_rs_hash);
 642 out_ns:
 643         OBD_FREE_PTR(ns);
 644 out_ref:
 645         ldlm_put_ref();
 646         RETURN(NULL);
 647 }
 648
 649 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 650
 651 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 652  * This is currently only used for recovery, and we make certain assumptions
 653  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 654 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 655                              int flags)
 656 {
 657         cfs_list_t *tmp;
 658         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 659         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 660
 661         do {
 662                 struct ldlm_lock *lock = NULL;
 663
 664                 /* first, we look for non-cleaned-yet lock
 665                  * all cleaned locks are marked by CLEANED flag */
 666                 lock_res(res);
 667                 cfs_list_for_each(tmp, q) {
 668                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 669                                               l_res_link);
 670                         if (lock->l_flags & LDLM_FL_CLEANED) {
 671                                 lock = NULL;
 672                                 continue;
 673                         }
 674                         LDLM_LOCK_GET(lock);
 675                         lock->l_flags |= LDLM_FL_CLEANED;
 676                         break;
 677                 }
 678
 679                 if (lock == NULL) {
 680                         unlock_res(res);
 681                         break;
 682                 }
 683
 684                 /* Set CBPENDING so nothing in the cancellation path
 685                  * can match this lock */
 686                 lock->l_flags |= LDLM_FL_CBPENDING;
 687                 lock->l_flags |= LDLM_FL_FAILED;
 688                 lock->l_flags |= flags;
 689
 690                 /* ... without sending a CANCEL message for local_only. */
 691                 if (local_only)
 692                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 693
 694                 if (local_only && (lock->l_readers || lock->l_writers)) {
 695                         /* This is a little bit gross, but much better than the
 696                          * alternative: pretend that we got a blocking AST from
 697                          * the server, so that when the lock is decref'd, it
 698                          * will go away ... */
 699                         unlock_res(res);
 700                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 701                         if (lock->l_completion_ast)
 702                                 lock->l_completion_ast(lock, 0, NULL);
 703                         LDLM_LOCK_RELEASE(lock);
 704                         continue;
 705                 }
 706
 707                 if (client) {
 708                         struct lustre_handle lockh;
 709
 710                         unlock_res(res);
 711                         ldlm_lock2handle(lock, &lockh);
 712                         rc = ldlm_cli_cancel(&lockh);
 713                         if (rc)
 714                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 715                 } else {
 716                         ldlm_resource_unlink_lock(lock);
 717                         unlock_res(res);
 718                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 719                                    "client node");
 720                         ldlm_lock_destroy(lock);
 721                 }
 722                 LDLM_LOCK_RELEASE(lock);
 723         } while (1);
 724 }
 725
 726 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 727                                cfs_hlist_node_t *hnode, void *arg)
 728 {
 729         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 730         int    flags = (int)(unsigned long)arg;
 731
 732         cleanup_resource(res, &res->lr_granted, flags);
 733         cleanup_resource(res, &res->lr_converting, flags);
 734         cleanup_resource(res, &res->lr_waiting, flags);
 735
 736         return 0;
 737 }
 738
 739 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 740                                   cfs_hlist_node_t *hnode, void *arg)
 741 {
 742         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 743
 744         CERROR("Namespace %s resource refcount nonzero "
 745                "(%d) after lock cleanup; forcing "
 746                "cleanup.\n",
 747                ldlm_ns_name(ldlm_res_to_ns(res)),
 748                cfs_atomic_read(&res->lr_refcount) - 1);
 749
 750         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 751                LPU64") (rc: %d)\n", res,
 752                res->lr_name.name[0], res->lr_name.name[1],
 753                res->lr_name.name[2], res->lr_name.name[3],
 754                cfs_atomic_read(&res->lr_refcount) - 1);
 755         return 0;
 756 }
 757
 758 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 759 {
 760         if (ns == NULL) {
 761                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 762                 return ELDLM_OK;
 763         }
 764
 765         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 766                                  (void *)(unsigned long)flags);
 767         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 768         return ELDLM_OK;
 769 }
 770
 771 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 772 {
 773         ENTRY;
 774
 775         /* At shutdown time, don't call the cancellation callback */
 776         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 777
 778         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 779                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 780                 int rc;
 781                 CDEBUG(D_DLMTRACE,
 782                        "dlm namespace %s free waiting on refcount %d\n",
 783                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 784 force_wait:
 785                 if (force)
 786                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 787
 788                 rc = l_wait_event(ns->ns_waitq,
 789                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 790
 791                 /* Forced cleanups should be able to reclaim all references,
 792                  * so it's safe to wait forever... we can't leak locks... */
 793                 if (force && rc == -ETIMEDOUT) {
 794                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 795                                        "namespace with %d resources in use, "
 796                                        "(rc=%d)\n", ldlm_ns_name(ns),
 797                                        cfs_atomic_read(&ns->ns_bref), rc);
 798                         GOTO(force_wait, rc);
 799                 }
 800
 801                 if (cfs_atomic_read(&ns->ns_bref)) {
 802                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 803                                        "with %d resources in use, (rc=%d)\n",
 804                                        ldlm_ns_name(ns),
 805                                        cfs_atomic_read(&ns->ns_bref), rc);
 806                         RETURN(ELDLM_NAMESPACE_EXISTS);
 807                 }
 808                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 809                        ldlm_ns_name(ns));
 810         }
 811
 812         RETURN(ELDLM_OK);
 813 }
 814
 815 /**
 816  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 817  * for freeing. Waits for refc == 0.
 818  *
 819  * The following is done:
 820  * (0) Unregister \a ns from its list to make inaccessible for potential users
 821  * like pools thread and others;
 822  * (1) Clear all locks in \a ns.
 823  */
 824 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 825                                struct obd_import *imp,
 826                                int force)
 827 {
 828         int rc;
 829         ENTRY;
 830         if (!ns) {
 831                 EXIT;
 832                 return;
 833         }
 834
 835
 836         /*
 837          * Can fail with -EINTR when force == 0 in which case try harder.
 838          */
 839         rc = __ldlm_namespace_free(ns, force);
 840         if (rc != ELDLM_OK) {
 841                 if (imp) {
 842                         ptlrpc_disconnect_import(imp, 0);
 843                         ptlrpc_invalidate_import(imp);
 844                 }
 845
 846                 /*
 847                  * With all requests dropped and the import inactive
 848                  * we are gaurenteed all reference will be dropped.
 849                  */
 850                 rc = __ldlm_namespace_free(ns, 1);
 851                 LASSERT(rc == 0);
 852         }
 853         EXIT;
 854 }
 855
 856 /**
 857  * Performs freeing memory structures related to \a ns. This is only done when
 858  * ldlm_namespce_free_prior() successfully removed all resources referencing
 859  * \a ns and its refc == 0.
 860  */
 861 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 862 {
 863         ENTRY;
 864         if (!ns) {
 865                 EXIT;
 866                 return;
 867         }
 868
 869
 870         /*
 871          * Make sure that nobody can find this ns in its list.
 872          */
 873         ldlm_namespace_unregister(ns, ns->ns_client);
 874         /*
 875          * Fini pool _before_ parent proc dir is removed. This is important as
 876          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 877          * it after @dir may cause oops.
 878          */
 879         ldlm_pool_fini(&ns->ns_pool);
 880
 881         ldlm_namespace_proc_unregister(ns);
 882         cfs_hash_putref(ns->ns_rs_hash);
 883         /*
 884          * Namespace \a ns should be not on list in this time, otherwise this
 885          * will cause issues realted to using freed \a ns in pools thread.
 886          */
 887         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 888         OBD_FREE_PTR(ns);
 889         ldlm_put_ref();
 890         EXIT;
 891 }
 892
 893
 894 /* Cleanup the resource, and free namespace.
 895  * bug 12864:
 896  * Deadlock issue:
 897  * proc1: destroy import
 898  *        class_disconnect_export(grab cl_sem) ->
 899  *              -> ldlm_namespace_free ->
 900  *              -> lprocfs_remove(grab _lprocfs_lock).
 901  * proc2: read proc info
 902  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 903  *              -> osc_rd_active, etc(grab cl_sem).
 904  *
 905  * So that I have to split the ldlm_namespace_free into two parts - the first
 906  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 907  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 908  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 909  * held.
 910  */
 911 void ldlm_namespace_free(struct ldlm_namespace *ns,
 912                          struct obd_import *imp,
 913                          int force)
 914 {
 915         ldlm_namespace_free_prior(ns, imp, force);
 916         ldlm_namespace_free_post(ns);
 917 }
 918
 919 void ldlm_namespace_get(struct ldlm_namespace *ns)
 920 {
 921         cfs_atomic_inc(&ns->ns_bref);
 922 }
 923
 924 void ldlm_namespace_put(struct ldlm_namespace *ns)
 925 {
 926         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 927                 cfs_waitq_signal(&ns->ns_waitq);
 928                 cfs_spin_unlock(&ns->ns_lock);
 929         }
 930 }
 931
 932 /* Register @ns in the list of namespaces */
 933 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 934 {
 935         cfs_mutex_down(ldlm_namespace_lock(client));
 936         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 937         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 938         cfs_atomic_inc(ldlm_namespace_nr(client));
 939         cfs_mutex_up(ldlm_namespace_lock(client));
 940 }
 941
 942 /* Unregister @ns from the list of namespaces */
 943 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 944 {
 945         cfs_mutex_down(ldlm_namespace_lock(client));
 946         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 947         /*
 948          * Some asserts and possibly other parts of code still using
 949          * list_empty(&ns->ns_list_chain). This is why it is important
 950          * to use list_del_init() here.
 951          */
 952         cfs_list_del_init(&ns->ns_list_chain);
 953         cfs_atomic_dec(ldlm_namespace_nr(client));
 954         cfs_mutex_up(ldlm_namespace_lock(client));
 955 }
 956
 957 /* Should be called under ldlm_namespace_lock(client) taken */
 958 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 959 {
 960         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 961         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 962         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 963 }
 964
 965 /* Should be called under ldlm_namespace_lock(client) taken */
 966 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 967 {
 968         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 969         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 970         return container_of(ldlm_namespace_list(client)->next,
 971                 struct ldlm_namespace, ns_list_chain);
 972 }
 973
 974 static struct ldlm_resource *ldlm_resource_new(void)
 975 {
 976         struct ldlm_resource *res;
 977         int idx;
 978
 979         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 980         if (res == NULL)
 981                 return NULL;
 982
 983         CFS_INIT_LIST_HEAD(&res->lr_granted);
 984         CFS_INIT_LIST_HEAD(&res->lr_converting);
 985         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 986
 987         /* initialize interval trees for each lock mode*/
 988         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 989                 res->lr_itree[idx].lit_size = 0;
 990                 res->lr_itree[idx].lit_mode = 1 << idx;
 991                 res->lr_itree[idx].lit_root = NULL;
 992         }
 993
 994         cfs_atomic_set(&res->lr_refcount, 1);
 995         cfs_spin_lock_init(&res->lr_lock);
 996         lu_ref_init(&res->lr_reference);
 997
 998         /* one who creates the resource must unlock
 999          * the semaphore after lvb initialization */
1000         cfs_init_mutex_locked(&res->lr_lvb_sem);
1001
1002         return res;
1003 }
1004
1005 /* Args: unlocked namespace
1006  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1007  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1008 struct ldlm_resource *
1009 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1010                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1011 {
1012         cfs_hlist_node_t     *hnode;
1013         struct ldlm_resource *res;
1014         cfs_hash_bd_t         bd;
1015         __u64                 version;
1016
1017         LASSERT(ns != NULL);
1018         LASSERT(parent == NULL);
1019         LASSERT(ns->ns_rs_hash != NULL);
1020         LASSERT(name->name[0] != 0);
1021
1022         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1023         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1024         if (hnode != NULL) {
1025                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1026                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1027                 /* synchronize WRT resource creation */
1028                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1029                         cfs_down(&res->lr_lvb_sem);
1030                         cfs_up(&res->lr_lvb_sem);
1031                 }
1032                 return res;
1033         }
1034
1035         version = cfs_hash_bd_version_get(&bd);
1036         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1037
1038         if (create == 0)
1039                 return NULL;
1040
1041         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1042                  "type: %d\n", type);
1043         res = ldlm_resource_new();
1044         if (!res)
1045                 return NULL;
1046
1047         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1048         res->lr_name       = *name;
1049         res->lr_type       = type;
1050         res->lr_most_restr = LCK_NL;
1051
1052         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1053         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1054                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1055
1056         if (hnode != NULL) {
1057                 /* someone won the race and added the resource before */
1058                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1059                 /* clean lu_ref for failed resource */
1060                 lu_ref_fini(&res->lr_reference);
1061                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1062
1063                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1064                 /* synchronize WRT resource creation */
1065                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1066                         cfs_down(&res->lr_lvb_sem);
1067                         cfs_up(&res->lr_lvb_sem);
1068                 }
1069                 return res;
1070         }
1071         /* we won! let's add the resource */
1072         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1073         if (cfs_hash_bd_count_get(&bd) == 1)
1074                 ldlm_namespace_get(ns);
1075
1076         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1077         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1078                 int rc;
1079
1080                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1081                 rc = ns->ns_lvbo->lvbo_init(res);
1082                 if (rc)
1083                         CERROR("lvbo_init failed for resource "
1084                                LPU64": rc %d\n", name->name[0], rc);
1085                 /* we create resource with locked lr_lvb_sem */
1086                 cfs_up(&res->lr_lvb_sem);
1087         }
1088
1089         return res;
1090 }
1091
1092 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1093 {
1094         LASSERT(res != NULL);
1095         LASSERT(res != LP_POISON);
1096         cfs_atomic_inc(&res->lr_refcount);
1097         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1098                cfs_atomic_read(&res->lr_refcount));
1099         return res;
1100 }
1101
1102 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1103                                          struct ldlm_resource *res)
1104 {
1105         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1106
1107         if (!cfs_list_empty(&res->lr_granted)) {
1108                 ldlm_resource_dump(D_ERROR, res);
1109                 LBUG();
1110         }
1111
1112         if (!cfs_list_empty(&res->lr_converting)) {
1113                 ldlm_resource_dump(D_ERROR, res);
1114                 LBUG();
1115         }
1116
1117         if (!cfs_list_empty(&res->lr_waiting)) {
1118                 ldlm_resource_dump(D_ERROR, res);
1119                 LBUG();
1120         }
1121
1122         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1123                                bd, &res->lr_hash);
1124         lu_ref_fini(&res->lr_reference);
1125         if (cfs_hash_bd_count_get(bd) == 0)
1126                 ldlm_namespace_put(nsb->nsb_namespace);
1127 }
1128
1129 /* Returns 1 if the resource was freed, 0 if it remains. */
1130 int ldlm_resource_putref(struct ldlm_resource *res)
1131 {
1132         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1133         cfs_hash_bd_t   bd;
1134
1135         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1136         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1137                res, cfs_atomic_read(&res->lr_refcount) - 1);
1138
1139         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1140         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1141                 __ldlm_resource_putref_final(&bd, res);
1142                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1143                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1144                         ns->ns_lvbo->lvbo_free(res);
1145                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1146                 return 1;
1147         }
1148         return 0;
1149 }
1150
1151 /* Returns 1 if the resource was freed, 0 if it remains. */
1152 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1153 {
1154         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1155
1156         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1157         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1158                res, cfs_atomic_read(&res->lr_refcount) - 1);
1159
1160         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1161                 cfs_hash_bd_t bd;
1162
1163                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1164                                 &res->lr_name, &bd);
1165                 __ldlm_resource_putref_final(&bd, res);
1166                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1167                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1168                  * so we should never be here while calling cfs_hash_del,
1169                  * cfs_hash_for_each_nolock is the only case we can get
1170                  * here, which is safe to release cfs_hash_bd_lock.
1171                  */
1172                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1173                         ns->ns_lvbo->lvbo_free(res);
1174                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1175
1176                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1177                 return 1;
1178         }
1179         return 0;
1180 }
1181
1182 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1183                             struct ldlm_lock *lock)
1184 {
1185         check_res_locked(res);
1186
1187         ldlm_resource_dump(D_INFO, res);
1188         CDEBUG(D_OTHER, "About to add this lock:\n");
1189         ldlm_lock_dump(D_OTHER, lock, 0);
1190
1191         if (lock->l_destroyed) {
1192                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1193                 return;
1194         }
1195
1196         LASSERT(cfs_list_empty(&lock->l_res_link));
1197
1198         cfs_list_add_tail(&lock->l_res_link, head);
1199 }
1200
1201 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1202                                      struct ldlm_lock *new)
1203 {
1204         struct ldlm_resource *res = original->l_resource;
1205
1206         check_res_locked(res);
1207
1208         ldlm_resource_dump(D_INFO, res);
1209         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1210         ldlm_lock_dump(D_OTHER, new, 0);
1211
1212         if (new->l_destroyed) {
1213                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1214                 goto out;
1215         }
1216
1217         LASSERT(cfs_list_empty(&new->l_res_link));
1218
1219         cfs_list_add(&new->l_res_link, &original->l_res_link);
1220  out:;
1221 }
1222
1223 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1224 {
1225         int type = lock->l_resource->lr_type;
1226
1227         check_res_locked(lock->l_resource);
1228         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1229                 ldlm_unlink_lock_skiplist(lock);
1230         else if (type == LDLM_EXTENT)
1231                 ldlm_extent_unlink_lock(lock);
1232         cfs_list_del_init(&lock->l_res_link);
1233 }
1234
1235 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1236 {
1237         desc->lr_type = res->lr_type;
1238         desc->lr_name = res->lr_name;
1239 }
1240
1241 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1242 {
1243         cfs_list_t *tmp;
1244
1245         if (!((libcfs_debug | D_ERROR) & level))
1246                 return;
1247
1248         cfs_mutex_down(ldlm_namespace_lock(client));
1249
1250         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1251                 struct ldlm_namespace *ns;
1252                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1253                 ldlm_namespace_dump(level, ns);
1254         }
1255
1256         cfs_mutex_up(ldlm_namespace_lock(client));
1257 }
1258
1259 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1260                               cfs_hlist_node_t *hnode, void *arg)
1261 {
1262         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1263         int    level = (int)(unsigned long)arg;
1264
1265         lock_res(res);
1266         ldlm_resource_dump(level, res);
1267         unlock_res(res);
1268
1269         return 0;
1270 }
1271
1272 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1273 {
1274         if (!((libcfs_debug | D_ERROR) & level))
1275                 return;
1276
1277         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1278                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1279                ns_is_client(ns) ? "client" : "server");
1280
1281         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1282                 return;
1283
1284         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1285                                  ldlm_res_hash_dump,
1286                                  (void *)(unsigned long)level);
1287         cfs_spin_lock(&ns->ns_lock);
1288         ns->ns_next_dump = cfs_time_shift(10);
1289         cfs_spin_unlock(&ns->ns_lock);
1290 }
1291
1292 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1293 {
1294         cfs_list_t *tmp;
1295         int pos;
1296
1297         CLASSERT(RES_NAME_SIZE == 4);
1298
1299         if (!((libcfs_debug | D_ERROR) & level))
1300                 return;
1301
1302         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1303                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1304                res->lr_name.name[2], res->lr_name.name[3],
1305                cfs_atomic_read(&res->lr_refcount));
1306
1307         if (!cfs_list_empty(&res->lr_granted)) {
1308                 pos = 0;
1309                 CDEBUG(level, "Granted locks:\n");
1310                 cfs_list_for_each(tmp, &res->lr_granted) {
1311                         struct ldlm_lock *lock;
1312                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1313                                               l_res_link);
1314                         ldlm_lock_dump(level, lock, ++pos);
1315                 }
1316         }
1317         if (!cfs_list_empty(&res->lr_converting)) {
1318                 pos = 0;
1319                 CDEBUG(level, "Converting locks:\n");
1320                 cfs_list_for_each(tmp, &res->lr_converting) {
1321                         struct ldlm_lock *lock;
1322                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1323                                               l_res_link);
1324                         ldlm_lock_dump(level, lock, ++pos);
1325                 }
1326         }
1327         if (!cfs_list_empty(&res->lr_waiting)) {
1328                 pos = 0;
1329                 CDEBUG(level, "Waiting locks:\n");
1330                 cfs_list_for_each(tmp, &res->lr_waiting) {
1331                         struct ldlm_lock *lock;
1332                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1333                                               l_res_link);
1334                         ldlm_lock_dump(level, lock, ++pos);
1335                 }
1336         }
1337 }