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, 2013, 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 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  54
  55 int ldlm_srv_namespace_nr = 0;
  56 int ldlm_cli_namespace_nr = 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 /* Client Namespaces that have active resources in them.
  63  * Once all resources go away, ldlm_poold moves such namespaces to the
  64  * inactive list */
  65 CFS_LIST_HEAD(ldlm_cli_active_namespace_list);
  66 /* Client namespaces that don't have any locks in them */
  67 CFS_LIST_HEAD(ldlm_cli_inactive_namespace_list);
  68
  69 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  70 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  71 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  72
  73 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  74
  75 /* during debug dump certain amount of granted locks for one resource to avoid
  76  * DDOS. */
  77 unsigned int ldlm_dump_granted_max = 256;
  78
  79 #ifdef LPROCFS
  80 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  81                              unsigned long count, void *data)
  82 {
  83         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  84         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  85         RETURN(count);
  86 }
  87
  88 int ldlm_proc_setup(void)
  89 {
  90         int rc;
  91         struct lprocfs_vars list[] = {
  92                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  93                 { "dump_granted_max",
  94                   lprocfs_rd_uint, lprocfs_wr_uint,
  95                   &ldlm_dump_granted_max, NULL },
  96                 { "cancel_unused_locks_before_replay",
  97                   lprocfs_rd_uint, lprocfs_wr_uint,
  98                   &ldlm_cancel_unused_locks_before_replay, NULL },
  99                 { NULL }};
 100         ENTRY;
 101         LASSERT(ldlm_ns_proc_dir == NULL);
 102
 103         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 104                                               proc_lustre_root,
 105                                               NULL, NULL);
 106         if (IS_ERR(ldlm_type_proc_dir)) {
 107                 CERROR("LProcFS failed in ldlm-init\n");
 108                 rc = PTR_ERR(ldlm_type_proc_dir);
 109                 GOTO(err, rc);
 110         }
 111
 112         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 113                                             ldlm_type_proc_dir,
 114                                             NULL, NULL);
 115         if (IS_ERR(ldlm_ns_proc_dir)) {
 116                 CERROR("LProcFS failed in ldlm-init\n");
 117                 rc = PTR_ERR(ldlm_ns_proc_dir);
 118                 GOTO(err_type, rc);
 119         }
 120
 121         ldlm_svc_proc_dir = lprocfs_register("services",
 122                                             ldlm_type_proc_dir,
 123                                             NULL, NULL);
 124         if (IS_ERR(ldlm_svc_proc_dir)) {
 125                 CERROR("LProcFS failed in ldlm-init\n");
 126                 rc = PTR_ERR(ldlm_svc_proc_dir);
 127                 GOTO(err_ns, rc);
 128         }
 129
 130         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 131         if (rc != 0) {
 132                 CERROR("LProcFS failed in ldlm-init\n");
 133                 GOTO(err_svc, rc);
 134         }
 135
 136         RETURN(0);
 137
 138 err_svc:
 139         lprocfs_remove(&ldlm_svc_proc_dir);
 140 err_ns:
 141         lprocfs_remove(&ldlm_ns_proc_dir);
 142 err_type:
 143         lprocfs_remove(&ldlm_type_proc_dir);
 144 err:
 145         ldlm_svc_proc_dir = NULL;
 146         RETURN(rc);
 147 }
 148
 149 void ldlm_proc_cleanup(void)
 150 {
 151         if (ldlm_svc_proc_dir)
 152                 lprocfs_remove(&ldlm_svc_proc_dir);
 153
 154         if (ldlm_ns_proc_dir)
 155                 lprocfs_remove(&ldlm_ns_proc_dir);
 156
 157         if (ldlm_type_proc_dir)
 158                 lprocfs_remove(&ldlm_type_proc_dir);
 159 }
 160
 161 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 162                                    int count, int *eof, void *data)
 163 {
 164         struct ldlm_namespace *ns  = data;
 165         __u64                  res = 0;
 166         cfs_hash_bd_t          bd;
 167         int                    i;
 168
 169         /* result is not strictly consistant */
 170         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 171                 res += cfs_hash_bd_count_get(&bd);
 172         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 173 }
 174
 175 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 176                                int count, int *eof, void *data)
 177 {
 178         struct ldlm_namespace *ns = data;
 179         __u64                  locks;
 180
 181         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 182                                         LPROCFS_FIELDS_FLAGS_SUM);
 183         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 184 }
 185
 186 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 187                                int count, int *eof, void *data)
 188 {
 189         struct ldlm_namespace *ns = data;
 190         __u32 *nr = &ns->ns_max_unused;
 191
 192         if (ns_connect_lru_resize(ns))
 193                 nr = &ns->ns_nr_unused;
 194         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 195 }
 196
 197 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 198                                unsigned long count, void *data)
 199 {
 200         struct ldlm_namespace *ns = data;
 201         char dummy[MAX_STRING_SIZE + 1], *end;
 202         unsigned long tmp;
 203         int lru_resize;
 204
 205         dummy[MAX_STRING_SIZE] = '\0';
 206         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 207                 return -EFAULT;
 208
 209         if (strncmp(dummy, "clear", 5) == 0) {
 210                 CDEBUG(D_DLMTRACE,
 211                        "dropping all unused locks from namespace %s\n",
 212                        ldlm_ns_name(ns));
 213                 if (ns_connect_lru_resize(ns)) {
 214                         int canceled, unused  = ns->ns_nr_unused;
 215
 216                         /* Try to cancel all @ns_nr_unused locks. */
 217                         canceled = ldlm_cancel_lru(ns, unused, 0,
 218                                                    LDLM_CANCEL_PASSED);
 219                         if (canceled < unused) {
 220                                 CDEBUG(D_DLMTRACE,
 221                                        "not all requested locks are canceled, "
 222                                        "requested: %d, canceled: %d\n", unused,
 223                                        canceled);
 224                                 return -EINVAL;
 225                         }
 226                 } else {
 227                         tmp = ns->ns_max_unused;
 228                         ns->ns_max_unused = 0;
 229                         ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
 230                         ns->ns_max_unused = tmp;
 231                 }
 232                 return count;
 233         }
 234
 235         tmp = simple_strtoul(dummy, &end, 0);
 236         if (dummy == end) {
 237                 CERROR("invalid value written\n");
 238                 return -EINVAL;
 239         }
 240         lru_resize = (tmp == 0);
 241
 242         if (ns_connect_lru_resize(ns)) {
 243                 if (!lru_resize)
 244                         ns->ns_max_unused = (unsigned int)tmp;
 245
 246                 if (tmp > ns->ns_nr_unused)
 247                         tmp = ns->ns_nr_unused;
 248                 tmp = ns->ns_nr_unused - tmp;
 249
 250                 CDEBUG(D_DLMTRACE,
 251                        "changing namespace %s unused locks from %u to %u\n",
 252                        ldlm_ns_name(ns), ns->ns_nr_unused,
 253                        (unsigned int)tmp);
 254                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
 255
 256                 if (!lru_resize) {
 257                         CDEBUG(D_DLMTRACE,
 258                                "disable lru_resize for namespace %s\n",
 259                                ldlm_ns_name(ns));
 260                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 261                 }
 262         } else {
 263                 CDEBUG(D_DLMTRACE,
 264                        "changing namespace %s max_unused from %u to %u\n",
 265                        ldlm_ns_name(ns), ns->ns_max_unused,
 266                        (unsigned int)tmp);
 267                 ns->ns_max_unused = (unsigned int)tmp;
 268                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
 269
 270                 /* Make sure that LRU resize was originally supported before
 271                  * turning it on here. */
 272                 if (lru_resize &&
 273                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 274                         CDEBUG(D_DLMTRACE,
 275                                "enable lru_resize for namespace %s\n",
 276                                ldlm_ns_name(ns));
 277                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 278                 }
 279         }
 280
 281         return count;
 282 }
 283
 284 static int lprocfs_rd_elc(char *page, char **start, off_t off,
 285                           int count, int *eof, void *data)
 286 {
 287         struct ldlm_namespace *ns = data;
 288         unsigned int supp = ns_connect_cancelset(ns);
 289
 290         return lprocfs_rd_uint(page, start, off, count, eof, &supp);
 291 }
 292
 293 static int lprocfs_wr_elc(struct file *file, const char *buffer,
 294                                unsigned long count, void *data)
 295 {
 296         struct ldlm_namespace *ns = data;
 297         unsigned int supp = -1;
 298         int rc;
 299
 300         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 301         if (rc < 0)
 302                 return rc;
 303
 304         if (supp == 0)
 305                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 306         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 307                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 308         return count;
 309 }
 310
 311 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 312 {
 313         struct proc_dir_entry *dir;
 314
 315         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 316         if (dir == NULL) {
 317                 CERROR("dlm namespace %s has no procfs dir?\n",
 318                        ldlm_ns_name(ns));
 319         } else {
 320                 lprocfs_remove(&dir);
 321         }
 322
 323         if (ns->ns_stats != NULL)
 324                 lprocfs_free_stats(&ns->ns_stats);
 325 }
 326
 327 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 328 {
 329         struct lprocfs_vars lock_vars[2];
 330         char lock_name[MAX_STRING_SIZE + 1];
 331
 332         LASSERT(ns != NULL);
 333         LASSERT(ns->ns_rs_hash != NULL);
 334
 335         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 336         if (ns->ns_stats == NULL)
 337                 return -ENOMEM;
 338
 339         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 340                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 341
 342         lock_name[MAX_STRING_SIZE] = '\0';
 343
 344         memset(lock_vars, 0, sizeof(lock_vars));
 345         lock_vars[0].name = lock_name;
 346
 347         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 348                  ldlm_ns_name(ns));
 349         lock_vars[0].data = ns;
 350         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 351         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 352
 353         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 354                  ldlm_ns_name(ns));
 355         lock_vars[0].data = ns;
 356         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 357         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 358
 359         if (ns_is_client(ns)) {
 360                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 361                          ldlm_ns_name(ns));
 362                 lock_vars[0].data = &ns->ns_nr_unused;
 363                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 364                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 365
 366                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 367                          ldlm_ns_name(ns));
 368                 lock_vars[0].data = ns;
 369                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 370                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 371                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 372
 373                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 374                          ldlm_ns_name(ns));
 375                 lock_vars[0].data = &ns->ns_max_age;
 376                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 377                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 378                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 379
 380                 snprintf(lock_name, MAX_STRING_SIZE, "%s/early_lock_cancel",
 381                          ldlm_ns_name(ns));
 382                 lock_vars[0].data = ns;
 383                 lock_vars[0].read_fptr = lprocfs_rd_elc;
 384                 lock_vars[0].write_fptr = lprocfs_wr_elc;
 385                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 386         } else {
 387                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 388                          ldlm_ns_name(ns));
 389                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 390                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 391                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 392                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 393
 394                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 395                          ldlm_ns_name(ns));
 396                 lock_vars[0].data = &ns->ns_timeouts;
 397                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 398                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 399
 400                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 401                          ldlm_ns_name(ns));
 402                 lock_vars[0].data = &ns->ns_max_nolock_size;
 403                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 404                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 405                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 406
 407                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 408                          ldlm_ns_name(ns));
 409                 lock_vars[0].data = &ns->ns_contention_time;
 410                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 411                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 412                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 413
 414                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 415                          ldlm_ns_name(ns));
 416                 lock_vars[0].data = &ns->ns_contended_locks;
 417                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 418                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 419                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 420
 421                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 422                          ldlm_ns_name(ns));
 423                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 424                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 425                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 426                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 427         }
 428         return 0;
 429 }
 430 #undef MAX_STRING_SIZE
 431 #else /* LPROCFS */
 432
 433 #define ldlm_namespace_proc_unregister(ns)      ({;})
 434 #define ldlm_namespace_proc_register(ns)        ({0;})
 435
 436 #endif /* LPROCFS */
 437
 438 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 439                                   const void *key, unsigned mask)
 440 {
 441         const struct ldlm_res_id     *id  = key;
 442         unsigned                val = 0;
 443         unsigned                i;
 444
 445         for (i = 0; i < RES_NAME_SIZE; i++)
 446                 val += id->name[i];
 447         return val & mask;
 448 }
 449
 450 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 451                                       const void *key, unsigned mask)
 452 {
 453         const struct ldlm_res_id *id = key;
 454         struct lu_fid       fid;
 455         __u32               hash;
 456         __u32               val;
 457
 458         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 459         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 460         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 461
 462         hash = fid_flatten32(&fid);
 463         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 464         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 465                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 466                 hash += (val >> 5) + (val << 11);
 467         } else {
 468                 val = fid_oid(&fid);
 469         }
 470         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 471         /* give me another random factor */
 472         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 473
 474         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 475         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 476
 477         return hash & mask;
 478 }
 479
 480 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 481 {
 482         struct ldlm_resource   *res;
 483
 484         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 485         return &res->lr_name;
 486 }
 487
 488 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 489 {
 490         struct ldlm_resource   *res;
 491
 492         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 493         return ldlm_res_eq((const struct ldlm_res_id *)key,
 494                            (const struct ldlm_res_id *)&res->lr_name);
 495 }
 496
 497 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 498 {
 499         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 500 }
 501
 502 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 503 {
 504         struct ldlm_resource *res;
 505
 506         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 507         ldlm_resource_getref(res);
 508 }
 509
 510 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 511 {
 512         struct ldlm_resource *res;
 513
 514         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 515         /* cfs_hash_for_each_nolock is the only chance we call it */
 516         ldlm_resource_putref_locked(res);
 517 }
 518
 519 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 520 {
 521         struct ldlm_resource *res;
 522
 523         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 524         ldlm_resource_putref(res);
 525 }
 526
 527 cfs_hash_ops_t ldlm_ns_hash_ops = {
 528         .hs_hash        = ldlm_res_hop_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 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 539         .hs_hash        = ldlm_res_hop_fid_hash,
 540         .hs_key         = ldlm_res_hop_key,
 541         .hs_keycmp      = ldlm_res_hop_keycmp,
 542         .hs_keycpy      = NULL,
 543         .hs_object      = ldlm_res_hop_object,
 544         .hs_get         = ldlm_res_hop_get_locked,
 545         .hs_put_locked  = ldlm_res_hop_put_locked,
 546         .hs_put         = ldlm_res_hop_put
 547 };
 548
 549 typedef struct {
 550         ldlm_ns_type_t  nsd_type;
 551         /** hash bucket bits */
 552         unsigned        nsd_bkt_bits;
 553         /** hash bits */
 554         unsigned        nsd_all_bits;
 555         /** hash operations */
 556         cfs_hash_ops_t *nsd_hops;
 557 } ldlm_ns_hash_def_t;
 558
 559 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 560 {
 561         {
 562                 .nsd_type       = LDLM_NS_TYPE_MDC,
 563                 .nsd_bkt_bits   = 11,
 564                 .nsd_all_bits   = 16,
 565                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 566         },
 567         {
 568                 .nsd_type       = LDLM_NS_TYPE_MDT,
 569                 .nsd_bkt_bits   = 14,
 570                 .nsd_all_bits   = 21,
 571                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 572         },
 573         {
 574                 .nsd_type       = LDLM_NS_TYPE_OSC,
 575                 .nsd_bkt_bits   = 8,
 576                 .nsd_all_bits   = 12,
 577                 .nsd_hops       = &ldlm_ns_hash_ops,
 578         },
 579         {
 580                 .nsd_type       = LDLM_NS_TYPE_OST,
 581                 .nsd_bkt_bits   = 11,
 582                 .nsd_all_bits   = 17,
 583                 .nsd_hops       = &ldlm_ns_hash_ops,
 584         },
 585         {
 586                 .nsd_type       = LDLM_NS_TYPE_MGC,
 587                 .nsd_bkt_bits   = 4,
 588                 .nsd_all_bits   = 4,
 589                 .nsd_hops       = &ldlm_ns_hash_ops,
 590         },
 591         {
 592                 .nsd_type       = LDLM_NS_TYPE_MGT,
 593                 .nsd_bkt_bits   = 4,
 594                 .nsd_all_bits   = 4,
 595                 .nsd_hops       = &ldlm_ns_hash_ops,
 596         },
 597         {
 598                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 599         },
 600 };
 601
 602 /**
 603  * Create and initialize new empty namespace.
 604  */
 605 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 606                                           ldlm_side_t client,
 607                                           ldlm_appetite_t apt,
 608                                           ldlm_ns_type_t ns_type)
 609 {
 610         struct ldlm_namespace *ns = NULL;
 611         struct ldlm_ns_bucket *nsb;
 612         ldlm_ns_hash_def_t    *nsd;
 613         cfs_hash_bd_t          bd;
 614         int                    idx;
 615         int                    rc;
 616         ENTRY;
 617
 618         LASSERT(obd != NULL);
 619
 620         rc = ldlm_get_ref();
 621         if (rc) {
 622                 CERROR("ldlm_get_ref failed: %d\n", rc);
 623                 RETURN(NULL);
 624         }
 625
 626         for (idx = 0;;idx++) {
 627                 nsd = &ldlm_ns_hash_defs[idx];
 628                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 629                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 630                         GOTO(out_ref, NULL);
 631                 }
 632
 633                 if (nsd->nsd_type == ns_type)
 634                         break;
 635         }
 636
 637         OBD_ALLOC_PTR(ns);
 638         if (!ns)
 639                 GOTO(out_ref, NULL);
 640
 641         ns->ns_rs_hash = cfs_hash_create(name,
 642                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 643                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 644                                          CFS_HASH_MIN_THETA,
 645                                          CFS_HASH_MAX_THETA,
 646                                          nsd->nsd_hops,
 647                                          CFS_HASH_DEPTH |
 648                                          CFS_HASH_BIGNAME |
 649                                          CFS_HASH_SPIN_BKTLOCK |
 650                                          CFS_HASH_NO_ITEMREF);
 651         if (ns->ns_rs_hash == NULL)
 652                 GOTO(out_ns, NULL);
 653
 654         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 655                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 656                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 657                 nsb->nsb_namespace = ns;
 658         }
 659
 660         ns->ns_obd      = obd;
 661         ns->ns_appetite = apt;
 662         ns->ns_client   = client;
 663
 664         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 665         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 666         spin_lock_init(&ns->ns_lock);
 667         cfs_atomic_set(&ns->ns_bref, 0);
 668         init_waitqueue_head(&ns->ns_waitq);
 669
 670         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 671         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 672         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 673
 674         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 675         ns->ns_nr_unused          = 0;
 676         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 677         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 678         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 679         ns->ns_timeouts           = 0;
 680         ns->ns_orig_connect_flags = 0;
 681         ns->ns_connect_flags      = 0;
 682         ns->ns_stopping           = 0;
 683         rc = ldlm_namespace_proc_register(ns);
 684         if (rc != 0) {
 685                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 686                 GOTO(out_hash, rc);
 687         }
 688
 689         idx = ldlm_namespace_nr_read(client);
 690         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 691         if (rc) {
 692                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 693                 GOTO(out_proc, rc);
 694         }
 695
 696         ldlm_namespace_register(ns, client);
 697         RETURN(ns);
 698 out_proc:
 699         ldlm_namespace_proc_unregister(ns);
 700         ldlm_namespace_cleanup(ns, 0);
 701 out_hash:
 702         cfs_hash_putref(ns->ns_rs_hash);
 703 out_ns:
 704         OBD_FREE_PTR(ns);
 705 out_ref:
 706         ldlm_put_ref();
 707         RETURN(NULL);
 708 }
 709 EXPORT_SYMBOL(ldlm_namespace_new);
 710
 711 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 712
 713 /**
 714  * Cancel and destroy all locks on a resource.
 715  *
 716  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 717  * clean up.  This is currently only used for recovery, and we make
 718  * certain assumptions as a result--notably, that we shouldn't cancel
 719  * locks with refs.
 720  */
 721 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 722                              __u64 flags)
 723 {
 724         cfs_list_t *tmp;
 725         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 726         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 727
 728         do {
 729                 struct ldlm_lock *lock = NULL;
 730
 731                 /* First, we look for non-cleaned-yet lock
 732                  * all cleaned locks are marked by CLEANED flag. */
 733                 lock_res(res);
 734                 cfs_list_for_each(tmp, q) {
 735                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 736                                               l_res_link);
 737                         if (ldlm_is_cleaned(lock)) {
 738                                 lock = NULL;
 739                                 continue;
 740                         }
 741                         LDLM_LOCK_GET(lock);
 742                         ldlm_set_cleaned(lock);
 743                         break;
 744                 }
 745
 746                 if (lock == NULL) {
 747                         unlock_res(res);
 748                         break;
 749                 }
 750
 751                 /* Set CBPENDING so nothing in the cancellation path
 752                  * can match this lock. */
 753                 ldlm_set_cbpending(lock);
 754                 ldlm_set_failed(lock);
 755                 lock->l_flags |= flags;
 756
 757                 /* ... without sending a CANCEL message for local_only. */
 758                 if (local_only)
 759                         ldlm_set_local_only(lock);
 760
 761                 if (local_only && (lock->l_readers || lock->l_writers)) {
 762                         /* This is a little bit gross, but much better than the
 763                          * alternative: pretend that we got a blocking AST from
 764                          * the server, so that when the lock is decref'd, it
 765                          * will go away ... */
 766                         unlock_res(res);
 767                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 768                         if (lock->l_completion_ast)
 769                                 lock->l_completion_ast(lock, 0, NULL);
 770                         LDLM_LOCK_RELEASE(lock);
 771                         continue;
 772                 }
 773
 774                 if (client) {
 775                         struct lustre_handle lockh;
 776
 777                         unlock_res(res);
 778                         ldlm_lock2handle(lock, &lockh);
 779                         rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 780                         if (rc)
 781                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 782                 } else {
 783                         ldlm_resource_unlink_lock(lock);
 784                         unlock_res(res);
 785                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 786                                    "client node");
 787                         ldlm_lock_destroy(lock);
 788                 }
 789                 LDLM_LOCK_RELEASE(lock);
 790         } while (1);
 791 }
 792
 793 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 794                                cfs_hlist_node_t *hnode, void *arg)
 795 {
 796         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 797         __u64 flags = *(__u64 *)arg;
 798
 799         cleanup_resource(res, &res->lr_granted, flags);
 800         cleanup_resource(res, &res->lr_converting, flags);
 801         cleanup_resource(res, &res->lr_waiting, flags);
 802
 803         return 0;
 804 }
 805
 806 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 807                                   cfs_hlist_node_t *hnode, void *arg)
 808 {
 809         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 810
 811         lock_res(res);
 812         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 813                "(%d) after lock cleanup; forcing cleanup.\n",
 814                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 815                cfs_atomic_read(&res->lr_refcount) - 1);
 816
 817         ldlm_resource_dump(D_ERROR, res);
 818         unlock_res(res);
 819         return 0;
 820 }
 821
 822 /**
 823  * Cancel and destroy all locks in the namespace.
 824  *
 825  * Typically used during evictions when server notified client that it was
 826  * evicted and all of its state needs to be destroyed.
 827  * Also used during shutdown.
 828  */
 829 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 830 {
 831         if (ns == NULL) {
 832                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 833                 return ELDLM_OK;
 834         }
 835
 836         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
 837         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 838         return ELDLM_OK;
 839 }
 840 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 841
 842 /**
 843  * Attempts to free namespace.
 844  *
 845  * Only used when namespace goes away, like during an unmount.
 846  */
 847 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 848 {
 849         ENTRY;
 850
 851         /* At shutdown time, don't call the cancellation callback */
 852         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 853
 854         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 855                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 856                 int rc;
 857                 CDEBUG(D_DLMTRACE,
 858                        "dlm namespace %s free waiting on refcount %d\n",
 859                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 860 force_wait:
 861                 if (force)
 862                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 863
 864                 rc = l_wait_event(ns->ns_waitq,
 865                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 866
 867                 /* Forced cleanups should be able to reclaim all references,
 868                  * so it's safe to wait forever... we can't leak locks... */
 869                 if (force && rc == -ETIMEDOUT) {
 870                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 871                                        "namespace with %d resources in use, "
 872                                        "(rc=%d)\n", ldlm_ns_name(ns),
 873                                        cfs_atomic_read(&ns->ns_bref), rc);
 874                         GOTO(force_wait, rc);
 875                 }
 876
 877                 if (cfs_atomic_read(&ns->ns_bref)) {
 878                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 879                                        "with %d resources in use, (rc=%d)\n",
 880                                        ldlm_ns_name(ns),
 881                                        cfs_atomic_read(&ns->ns_bref), rc);
 882                         RETURN(ELDLM_NAMESPACE_EXISTS);
 883                 }
 884                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 885                        ldlm_ns_name(ns));
 886         }
 887
 888         RETURN(ELDLM_OK);
 889 }
 890
 891 /**
 892  * Performs various cleanups for passed \a ns to make it drop refc and be
 893  * ready for freeing. Waits for refc == 0.
 894  *
 895  * The following is done:
 896  * (0) Unregister \a ns from its list to make inaccessible for potential
 897  * users like pools thread and others;
 898  * (1) Clear all locks in \a ns.
 899  */
 900 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 901                                struct obd_import *imp,
 902                                int force)
 903 {
 904         int rc;
 905         ENTRY;
 906         if (!ns) {
 907                 EXIT;
 908                 return;
 909         }
 910
 911         spin_lock(&ns->ns_lock);
 912         ns->ns_stopping = 1;
 913         spin_unlock(&ns->ns_lock);
 914
 915         /*
 916          * Can fail with -EINTR when force == 0 in which case try harder.
 917          */
 918         rc = __ldlm_namespace_free(ns, force);
 919         if (rc != ELDLM_OK) {
 920                 if (imp) {
 921                         ptlrpc_disconnect_import(imp, 0);
 922                         ptlrpc_invalidate_import(imp);
 923                 }
 924
 925                 /*
 926                  * With all requests dropped and the import inactive
 927                  * we are gaurenteed all reference will be dropped.
 928                  */
 929                 rc = __ldlm_namespace_free(ns, 1);
 930                 LASSERT(rc == 0);
 931         }
 932         EXIT;
 933 }
 934
 935 /**
 936  * Performs freeing memory structures related to \a ns. This is only done
 937  * when ldlm_namespce_free_prior() successfully removed all resources
 938  * referencing \a ns and its refc == 0.
 939  */
 940 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 941 {
 942         ENTRY;
 943         if (!ns) {
 944                 EXIT;
 945                 return;
 946         }
 947
 948         /* Make sure that nobody can find this ns in its list. */
 949         ldlm_namespace_unregister(ns, ns->ns_client);
 950         /* Fini pool _before_ parent proc dir is removed. This is important as
 951          * ldlm_pool_fini() removes own proc dir which is child to @dir.
 952          * Removing it after @dir may cause oops. */
 953         ldlm_pool_fini(&ns->ns_pool);
 954
 955         ldlm_namespace_proc_unregister(ns);
 956         cfs_hash_putref(ns->ns_rs_hash);
 957         /* Namespace \a ns should be not on list at this time, otherwise
 958          * this will cause issues related to using freed \a ns in poold
 959          * thread. */
 960         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 961         OBD_FREE_PTR(ns);
 962         ldlm_put_ref();
 963         EXIT;
 964 }
 965
 966 /**
 967  * Cleanup the resource, and free namespace.
 968  * bug 12864:
 969  * Deadlock issue:
 970  * proc1: destroy import
 971  *        class_disconnect_export(grab cl_sem) ->
 972  *              -> ldlm_namespace_free ->
 973  *              -> lprocfs_remove(grab _lprocfs_lock).
 974  * proc2: read proc info
 975  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 976  *              -> osc_rd_active, etc(grab cl_sem).
 977  *
 978  * So that I have to split the ldlm_namespace_free into two parts - the first
 979  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 980  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 981  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 982  * held.
 983  */
 984 void ldlm_namespace_free(struct ldlm_namespace *ns,
 985                          struct obd_import *imp,
 986                          int force)
 987 {
 988         ldlm_namespace_free_prior(ns, imp, force);
 989         ldlm_namespace_free_post(ns);
 990 }
 991 EXPORT_SYMBOL(ldlm_namespace_free);
 992
 993 void ldlm_namespace_get(struct ldlm_namespace *ns)
 994 {
 995         cfs_atomic_inc(&ns->ns_bref);
 996 }
 997 EXPORT_SYMBOL(ldlm_namespace_get);
 998
 999 /* This is only for callers that care about refcount */
1000 int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1001 {
1002         return cfs_atomic_inc_return(&ns->ns_bref);
1003 }
1004
1005 void ldlm_namespace_put(struct ldlm_namespace *ns)
1006 {
1007         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1008                 wake_up(&ns->ns_waitq);
1009                 spin_unlock(&ns->ns_lock);
1010         }
1011 }
1012 EXPORT_SYMBOL(ldlm_namespace_put);
1013
1014 /** Register \a ns in the list of namespaces */
1015 void ldlm_namespace_register(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         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1020         ldlm_namespace_nr_inc(client);
1021         mutex_unlock(ldlm_namespace_lock(client));
1022 }
1023
1024 /** Unregister \a ns from the list of namespaces. */
1025 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1026 {
1027         mutex_lock(ldlm_namespace_lock(client));
1028         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1029         /* Some asserts and possibly other parts of the code are still
1030          * using list_empty(&ns->ns_list_chain). This is why it is
1031          * important to use list_del_init() here. */
1032         cfs_list_del_init(&ns->ns_list_chain);
1033         ldlm_namespace_nr_dec(client);
1034         mutex_unlock(ldlm_namespace_lock(client));
1035 }
1036
1037 /** Should be called with ldlm_namespace_lock(client) taken. */
1038 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1039                                        ldlm_side_t client)
1040 {
1041         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1042         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1043         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1044 }
1045
1046 /** Should be called with ldlm_namespace_lock(client) taken. */
1047 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1048                                          ldlm_side_t client)
1049 {
1050         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1051         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1052         cfs_list_move_tail(&ns->ns_list_chain,
1053                            ldlm_namespace_inactive_list(client));
1054 }
1055
1056 /** Should be called with ldlm_namespace_lock(client) taken. */
1057 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1058 {
1059         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1060         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
1061         return container_of(ldlm_namespace_list(client)->next,
1062                             struct ldlm_namespace, ns_list_chain);
1063 }
1064
1065 /** Create and initialize new resource. */
1066 static struct ldlm_resource *ldlm_resource_new(void)
1067 {
1068         struct ldlm_resource *res;
1069         int idx;
1070
1071         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO);
1072         if (res == NULL)
1073                 return NULL;
1074
1075         CFS_INIT_LIST_HEAD(&res->lr_granted);
1076         CFS_INIT_LIST_HEAD(&res->lr_converting);
1077         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1078
1079         /* Initialize interval trees for each lock mode. */
1080         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1081                 res->lr_itree[idx].lit_size = 0;
1082                 res->lr_itree[idx].lit_mode = 1 << idx;
1083                 res->lr_itree[idx].lit_root = NULL;
1084         }
1085
1086         cfs_atomic_set(&res->lr_refcount, 1);
1087         spin_lock_init(&res->lr_lock);
1088         lu_ref_init(&res->lr_reference);
1089
1090         /* The creator of the resource must unlock the mutex after LVB
1091          * initialization. */
1092         mutex_init(&res->lr_lvb_mutex);
1093         mutex_lock(&res->lr_lvb_mutex);
1094
1095         return res;
1096 }
1097
1098 /**
1099  * Return a reference to resource with given name, creating it if necessary.
1100  * Args: namespace with ns_lock unlocked
1101  * Locks: takes and releases NS hash-lock and res->lr_lock
1102  * Returns: referenced, unlocked ldlm_resource or NULL
1103  */
1104 struct ldlm_resource *
1105 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1106                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1107 {
1108         cfs_hlist_node_t     *hnode;
1109         struct ldlm_resource *res;
1110         cfs_hash_bd_t         bd;
1111         __u64                 version;
1112         int                   ns_refcount = 0;
1113
1114         LASSERT(ns != NULL);
1115         LASSERT(parent == NULL);
1116         LASSERT(ns->ns_rs_hash != NULL);
1117         LASSERT(name->name[0] != 0);
1118
1119         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1120         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1121         if (hnode != NULL) {
1122                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1123                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1124                 /* Synchronize with regard to resource creation. */
1125                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1126                         mutex_lock(&res->lr_lvb_mutex);
1127                         mutex_unlock(&res->lr_lvb_mutex);
1128                 }
1129
1130                 if (unlikely(res->lr_lvb_len < 0)) {
1131                         ldlm_resource_putref(res);
1132                         res = NULL;
1133                 }
1134                 return res;
1135         }
1136
1137         version = cfs_hash_bd_version_get(&bd);
1138         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1139
1140         if (create == 0)
1141                 return NULL;
1142
1143         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1144                  "type: %d\n", type);
1145         res = ldlm_resource_new();
1146         if (!res)
1147                 return NULL;
1148
1149         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1150         res->lr_name       = *name;
1151         res->lr_type       = type;
1152         res->lr_most_restr = LCK_NL;
1153
1154         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1155         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1156                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1157
1158         if (hnode != NULL) {
1159                 /* Someone won the race and already added the resource. */
1160                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1161                 /* Clean lu_ref for failed resource. */
1162                 lu_ref_fini(&res->lr_reference);
1163                 /* We have taken lr_lvb_mutex. Drop it. */
1164                 mutex_unlock(&res->lr_lvb_mutex);
1165                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1166
1167                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1168                 /* Synchronize with regard to resource creation. */
1169                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1170                         mutex_lock(&res->lr_lvb_mutex);
1171                         mutex_unlock(&res->lr_lvb_mutex);
1172                 }
1173
1174                 if (unlikely(res->lr_lvb_len < 0)) {
1175                         ldlm_resource_putref(res);
1176                         res = NULL;
1177                 }
1178                 return res;
1179         }
1180         /* We won! Let's add the resource. */
1181         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1182         if (cfs_hash_bd_count_get(&bd) == 1)
1183                 ns_refcount = ldlm_namespace_get_return(ns);
1184
1185         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1186         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1187                 int rc;
1188
1189                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1190                 rc = ns->ns_lvbo->lvbo_init(res);
1191                 if (rc < 0) {
1192                         CERROR("%s: lvbo_init failed for resource "LPX64":"
1193                                LPX64": rc = %d\n", ns->ns_obd->obd_name,
1194                                name->name[0], name->name[1], rc);
1195                         if (res->lr_lvb_data) {
1196                                 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1197                                 res->lr_lvb_data = NULL;
1198                         }
1199                         res->lr_lvb_len = rc;
1200                         mutex_unlock(&res->lr_lvb_mutex);
1201                         ldlm_resource_putref(res);
1202                         return NULL;
1203                 }
1204         }
1205
1206         /* We create resource with locked lr_lvb_mutex. */
1207         mutex_unlock(&res->lr_lvb_mutex);
1208
1209         /* Let's see if we happened to be the very first resource in this
1210          * namespace. If so, and this is a client namespace, we need to move
1211          * the namespace into the active namespaces list to be patrolled by
1212          * the ldlm_poold. */
1213         if (ns_is_client(ns) && ns_refcount == 1) {
1214                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1215                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1216                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1217         }
1218
1219         return res;
1220 }
1221 EXPORT_SYMBOL(ldlm_resource_get);
1222
1223 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1224 {
1225         LASSERT(res != NULL);
1226         LASSERT(res != LP_POISON);
1227         cfs_atomic_inc(&res->lr_refcount);
1228         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1229                cfs_atomic_read(&res->lr_refcount));
1230         return res;
1231 }
1232
1233 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1234                                          struct ldlm_resource *res)
1235 {
1236         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1237
1238         if (!cfs_list_empty(&res->lr_granted)) {
1239                 ldlm_resource_dump(D_ERROR, res);
1240                 LBUG();
1241         }
1242
1243         if (!cfs_list_empty(&res->lr_converting)) {
1244                 ldlm_resource_dump(D_ERROR, res);
1245                 LBUG();
1246         }
1247
1248         if (!cfs_list_empty(&res->lr_waiting)) {
1249                 ldlm_resource_dump(D_ERROR, res);
1250                 LBUG();
1251         }
1252
1253         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1254                                bd, &res->lr_hash);
1255         lu_ref_fini(&res->lr_reference);
1256         if (cfs_hash_bd_count_get(bd) == 0)
1257                 ldlm_namespace_put(nsb->nsb_namespace);
1258 }
1259
1260 /* Returns 1 if the resource was freed, 0 if it remains. */
1261 int ldlm_resource_putref(struct ldlm_resource *res)
1262 {
1263         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1264         cfs_hash_bd_t   bd;
1265
1266         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1267         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1268                res, cfs_atomic_read(&res->lr_refcount) - 1);
1269
1270         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1271         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1272                 __ldlm_resource_putref_final(&bd, res);
1273                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1274                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1275                         ns->ns_lvbo->lvbo_free(res);
1276                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1277                 return 1;
1278         }
1279         return 0;
1280 }
1281 EXPORT_SYMBOL(ldlm_resource_putref);
1282
1283 /* Returns 1 if the resource was freed, 0 if it remains. */
1284 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1285 {
1286         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1287
1288         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1289         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1290                res, cfs_atomic_read(&res->lr_refcount) - 1);
1291
1292         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1293                 cfs_hash_bd_t bd;
1294
1295                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1296                                 &res->lr_name, &bd);
1297                 __ldlm_resource_putref_final(&bd, res);
1298                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1299                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1300                  * so we should never be here while calling cfs_hash_del,
1301                  * cfs_hash_for_each_nolock is the only case we can get
1302                  * here, which is safe to release cfs_hash_bd_lock.
1303                  */
1304                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1305                         ns->ns_lvbo->lvbo_free(res);
1306                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1307
1308                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1309                 return 1;
1310         }
1311         return 0;
1312 }
1313
1314 /**
1315  * Add a lock into a given resource into specified lock list.
1316  */
1317 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1318                             struct ldlm_lock *lock)
1319 {
1320         check_res_locked(res);
1321
1322         LDLM_DEBUG(lock, "About to add this lock:\n");
1323
1324         if (ldlm_is_destroyed(lock)) {
1325                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1326                 return;
1327         }
1328
1329         LASSERT(cfs_list_empty(&lock->l_res_link));
1330
1331         cfs_list_add_tail(&lock->l_res_link, head);
1332 }
1333
1334 /**
1335  * Insert a lock into resource after specified lock.
1336  *
1337  * Obtain resource description from the lock we are inserting after.
1338  */
1339 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1340                                      struct ldlm_lock *new)
1341 {
1342         struct ldlm_resource *res = original->l_resource;
1343
1344         check_res_locked(res);
1345
1346         ldlm_resource_dump(D_INFO, res);
1347         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1348
1349         if (ldlm_is_destroyed(new)) {
1350                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1351                 goto out;
1352         }
1353
1354         LASSERT(cfs_list_empty(&new->l_res_link));
1355
1356         cfs_list_add(&new->l_res_link, &original->l_res_link);
1357  out:;
1358 }
1359
1360 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1361 {
1362         int type = lock->l_resource->lr_type;
1363
1364         check_res_locked(lock->l_resource);
1365         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1366                 ldlm_unlink_lock_skiplist(lock);
1367         else if (type == LDLM_EXTENT)
1368                 ldlm_extent_unlink_lock(lock);
1369         cfs_list_del_init(&lock->l_res_link);
1370 }
1371 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1372
1373 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1374 {
1375         desc->lr_type = res->lr_type;
1376         desc->lr_name = res->lr_name;
1377 }
1378
1379 /**
1380  * Print information about all locks in all namespaces on this node to debug
1381  * log.
1382  */
1383 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1384 {
1385         cfs_list_t *tmp;
1386
1387         if (!((libcfs_debug | D_ERROR) & level))
1388                 return;
1389
1390         mutex_lock(ldlm_namespace_lock(client));
1391
1392         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1393                 struct ldlm_namespace *ns;
1394                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1395                 ldlm_namespace_dump(level, ns);
1396         }
1397
1398         mutex_unlock(ldlm_namespace_lock(client));
1399 }
1400 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1401
1402 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1403                               cfs_hlist_node_t *hnode, void *arg)
1404 {
1405         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1406         int    level = (int)(unsigned long)arg;
1407
1408         lock_res(res);
1409         ldlm_resource_dump(level, res);
1410         unlock_res(res);
1411
1412         return 0;
1413 }
1414
1415 /**
1416  * Print information about all locks in this namespace on this node to debug
1417  * log.
1418  */
1419 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1420 {
1421         if (!((libcfs_debug | D_ERROR) & level))
1422                 return;
1423
1424         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1425                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1426                ns_is_client(ns) ? "client" : "server");
1427
1428         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1429                 return;
1430
1431         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1432                                  ldlm_res_hash_dump,
1433                                  (void *)(unsigned long)level);
1434         spin_lock(&ns->ns_lock);
1435         ns->ns_next_dump = cfs_time_shift(10);
1436         spin_unlock(&ns->ns_lock);
1437 }
1438 EXPORT_SYMBOL(ldlm_namespace_dump);
1439
1440 /**
1441  * Print information about all locks in this resource to debug log.
1442  */
1443 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1444 {
1445         struct ldlm_lock *lock;
1446         unsigned int granted = 0;
1447
1448         CLASSERT(RES_NAME_SIZE == 4);
1449
1450         if (!((libcfs_debug | D_ERROR) & level))
1451                 return;
1452
1453         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1454                PLDLMRES(res), res, cfs_atomic_read(&res->lr_refcount));
1455
1456         if (!cfs_list_empty(&res->lr_granted)) {
1457                 CDEBUG(level, "Granted locks (in reverse order):\n");
1458                 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1459                                                 l_res_link) {
1460                         LDLM_DEBUG_LIMIT(level, lock, "###");
1461                         if (!(level & D_CANTMASK) &&
1462                             ++granted > ldlm_dump_granted_max) {
1463                                 CDEBUG(level, "only dump %d granted locks to "
1464                                        "avoid DDOS.\n", granted);
1465                                 break;
1466                         }
1467                 }
1468         }
1469         if (!cfs_list_empty(&res->lr_converting)) {
1470                 CDEBUG(level, "Converting locks:\n");
1471                 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1472                         LDLM_DEBUG_LIMIT(level, lock, "###");
1473         }
1474         if (!cfs_list_empty(&res->lr_waiting)) {
1475                 CDEBUG(level, "Waiting locks:\n");
1476                 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1477                         LDLM_DEBUG_LIMIT(level, lock, "###");
1478         }
1479 }