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 cfs_mem_cache_t *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
 132         RETURN(0);
 133
 134 err_ns:
 135         lprocfs_remove(&ldlm_ns_proc_dir);
 136 err_type:
 137         lprocfs_remove(&ldlm_type_proc_dir);
 138 err:
 139         ldlm_svc_proc_dir = NULL;
 140         RETURN(rc);
 141 }
 142
 143 void ldlm_proc_cleanup(void)
 144 {
 145         if (ldlm_svc_proc_dir)
 146                 lprocfs_remove(&ldlm_svc_proc_dir);
 147
 148         if (ldlm_ns_proc_dir)
 149                 lprocfs_remove(&ldlm_ns_proc_dir);
 150
 151         if (ldlm_type_proc_dir)
 152                 lprocfs_remove(&ldlm_type_proc_dir);
 153 }
 154
 155 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 156                                    int count, int *eof, void *data)
 157 {
 158         struct ldlm_namespace *ns  = data;
 159         __u64                  res = 0;
 160         cfs_hash_bd_t          bd;
 161         int                    i;
 162
 163         /* result is not strictly consistant */
 164         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 165                 res += cfs_hash_bd_count_get(&bd);
 166         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 167 }
 168
 169 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 170                                int count, int *eof, void *data)
 171 {
 172         struct ldlm_namespace *ns = data;
 173         __u64                  locks;
 174
 175         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 176                                         LPROCFS_FIELDS_FLAGS_SUM);
 177         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 178 }
 179
 180 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 181                                int count, int *eof, void *data)
 182 {
 183         struct ldlm_namespace *ns = data;
 184         __u32 *nr = &ns->ns_max_unused;
 185
 186         if (ns_connect_lru_resize(ns))
 187                 nr = &ns->ns_nr_unused;
 188         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 189 }
 190
 191 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 192                                unsigned long count, void *data)
 193 {
 194         struct ldlm_namespace *ns = data;
 195         char dummy[MAX_STRING_SIZE + 1], *end;
 196         unsigned long tmp;
 197         int lru_resize;
 198
 199         dummy[MAX_STRING_SIZE] = '\0';
 200         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 201                 return -EFAULT;
 202
 203         if (strncmp(dummy, "clear", 5) == 0) {
 204                 CDEBUG(D_DLMTRACE,
 205                        "dropping all unused locks from namespace %s\n",
 206                        ldlm_ns_name(ns));
 207                 if (ns_connect_lru_resize(ns)) {
 208                         int canceled, unused  = ns->ns_nr_unused;
 209
 210                         /* Try to cancel all @ns_nr_unused locks. */
 211                         canceled = ldlm_cancel_lru(ns, unused, 0,
 212                                                    LDLM_CANCEL_PASSED);
 213                         if (canceled < unused) {
 214                                 CDEBUG(D_DLMTRACE,
 215                                        "not all requested locks are canceled, "
 216                                        "requested: %d, canceled: %d\n", unused,
 217                                        canceled);
 218                                 return -EINVAL;
 219                         }
 220                 } else {
 221                         tmp = ns->ns_max_unused;
 222                         ns->ns_max_unused = 0;
 223                         ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
 224                         ns->ns_max_unused = tmp;
 225                 }
 226                 return count;
 227         }
 228
 229         tmp = simple_strtoul(dummy, &end, 0);
 230         if (dummy == end) {
 231                 CERROR("invalid value written\n");
 232                 return -EINVAL;
 233         }
 234         lru_resize = (tmp == 0);
 235
 236         if (ns_connect_lru_resize(ns)) {
 237                 if (!lru_resize)
 238                         ns->ns_max_unused = (unsigned int)tmp;
 239
 240                 if (tmp > ns->ns_nr_unused)
 241                         tmp = ns->ns_nr_unused;
 242                 tmp = ns->ns_nr_unused - tmp;
 243
 244                 CDEBUG(D_DLMTRACE,
 245                        "changing namespace %s unused locks from %u to %u\n",
 246                        ldlm_ns_name(ns), ns->ns_nr_unused,
 247                        (unsigned int)tmp);
 248                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
 249
 250                 if (!lru_resize) {
 251                         CDEBUG(D_DLMTRACE,
 252                                "disable lru_resize for namespace %s\n",
 253                                ldlm_ns_name(ns));
 254                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 255                 }
 256         } else {
 257                 CDEBUG(D_DLMTRACE,
 258                        "changing namespace %s max_unused from %u to %u\n",
 259                        ldlm_ns_name(ns), ns->ns_max_unused,
 260                        (unsigned int)tmp);
 261                 ns->ns_max_unused = (unsigned int)tmp;
 262                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
 263
 264                 /* Make sure that LRU resize was originally supported before
 265                  * turning it on here. */
 266                 if (lru_resize &&
 267                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 268                         CDEBUG(D_DLMTRACE,
 269                                "enable lru_resize for namespace %s\n",
 270                                ldlm_ns_name(ns));
 271                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 272                 }
 273         }
 274
 275         return count;
 276 }
 277
 278 static int lprocfs_rd_elc(char *page, char **start, off_t off,
 279                           int count, int *eof, void *data)
 280 {
 281         struct ldlm_namespace *ns = data;
 282         unsigned int supp = ns_connect_cancelset(ns);
 283
 284         return lprocfs_rd_uint(page, start, off, count, eof, &supp);
 285 }
 286
 287 static int lprocfs_wr_elc(struct file *file, const char *buffer,
 288                                unsigned long count, void *data)
 289 {
 290         struct ldlm_namespace *ns = data;
 291         unsigned int supp = -1;
 292         int rc;
 293
 294         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 295         if (rc < 0)
 296                 return rc;
 297
 298         if (supp == 0)
 299                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 300         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 301                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 302         return count;
 303 }
 304
 305 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 306 {
 307         struct proc_dir_entry *dir;
 308
 309         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 310         if (dir == NULL) {
 311                 CERROR("dlm namespace %s has no procfs dir?\n",
 312                        ldlm_ns_name(ns));
 313         } else {
 314                 lprocfs_remove(&dir);
 315         }
 316
 317         if (ns->ns_stats != NULL)
 318                 lprocfs_free_stats(&ns->ns_stats);
 319 }
 320
 321 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 322 {
 323         struct lprocfs_vars lock_vars[2];
 324         char lock_name[MAX_STRING_SIZE + 1];
 325
 326         LASSERT(ns != NULL);
 327         LASSERT(ns->ns_rs_hash != NULL);
 328
 329         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 330         if (ns->ns_stats == NULL)
 331                 return -ENOMEM;
 332
 333         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 334                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 335
 336         lock_name[MAX_STRING_SIZE] = '\0';
 337
 338         memset(lock_vars, 0, sizeof(lock_vars));
 339         lock_vars[0].name = lock_name;
 340
 341         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 342                  ldlm_ns_name(ns));
 343         lock_vars[0].data = ns;
 344         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 345         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 346
 347         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 348                  ldlm_ns_name(ns));
 349         lock_vars[0].data = ns;
 350         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 351         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 352
 353         if (ns_is_client(ns)) {
 354                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 355                          ldlm_ns_name(ns));
 356                 lock_vars[0].data = &ns->ns_nr_unused;
 357                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 358                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 359
 360                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 361                          ldlm_ns_name(ns));
 362                 lock_vars[0].data = ns;
 363                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 364                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 365                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 366
 367                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 368                          ldlm_ns_name(ns));
 369                 lock_vars[0].data = &ns->ns_max_age;
 370                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 371                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 372                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 373
 374                 snprintf(lock_name, MAX_STRING_SIZE, "%s/early_lock_cancel",
 375                          ldlm_ns_name(ns));
 376                 lock_vars[0].data = ns;
 377                 lock_vars[0].read_fptr = lprocfs_rd_elc;
 378                 lock_vars[0].write_fptr = lprocfs_wr_elc;
 379                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 380         } else {
 381                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 382                          ldlm_ns_name(ns));
 383                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 384                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 385                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 386                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 387
 388                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 389                          ldlm_ns_name(ns));
 390                 lock_vars[0].data = &ns->ns_timeouts;
 391                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 392                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 393
 394                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 395                          ldlm_ns_name(ns));
 396                 lock_vars[0].data = &ns->ns_max_nolock_size;
 397                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 398                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 399                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 400
 401                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 402                          ldlm_ns_name(ns));
 403                 lock_vars[0].data = &ns->ns_contention_time;
 404                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 405                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 406                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 407
 408                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 409                          ldlm_ns_name(ns));
 410                 lock_vars[0].data = &ns->ns_contended_locks;
 411                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 412                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 413                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 414
 415                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 416                          ldlm_ns_name(ns));
 417                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 418                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 419                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 420                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 421         }
 422         return 0;
 423 }
 424 #undef MAX_STRING_SIZE
 425 #else /* LPROCFS */
 426
 427 #define ldlm_namespace_proc_unregister(ns)      ({;})
 428 #define ldlm_namespace_proc_register(ns)        ({0;})
 429
 430 #endif /* LPROCFS */
 431
 432 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 433                                   const void *key, unsigned mask)
 434 {
 435         const struct ldlm_res_id     *id  = key;
 436         unsigned                val = 0;
 437         unsigned                i;
 438
 439         for (i = 0; i < RES_NAME_SIZE; i++)
 440                 val += id->name[i];
 441         return val & mask;
 442 }
 443
 444 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 445                                       const void *key, unsigned mask)
 446 {
 447         const struct ldlm_res_id *id = key;
 448         struct lu_fid       fid;
 449         __u32               hash;
 450         __u32               val;
 451
 452         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 453         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 454         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 455
 456         hash = fid_flatten32(&fid);
 457         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 458         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 459                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 460                 hash += (val >> 5) + (val << 11);
 461         } else {
 462                 val = fid_oid(&fid);
 463         }
 464         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 465         /* give me another random factor */
 466         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
 467
 468         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 469         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 470
 471         return hash & mask;
 472 }
 473
 474 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 475 {
 476         struct ldlm_resource   *res;
 477
 478         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 479         return &res->lr_name;
 480 }
 481
 482 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 483 {
 484         struct ldlm_resource   *res;
 485
 486         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 487         return ldlm_res_eq((const struct ldlm_res_id *)key,
 488                            (const struct ldlm_res_id *)&res->lr_name);
 489 }
 490
 491 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 492 {
 493         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 494 }
 495
 496 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 497 {
 498         struct ldlm_resource *res;
 499
 500         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 501         ldlm_resource_getref(res);
 502 }
 503
 504 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 505 {
 506         struct ldlm_resource *res;
 507
 508         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 509         /* cfs_hash_for_each_nolock is the only chance we call it */
 510         ldlm_resource_putref_locked(res);
 511 }
 512
 513 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 514 {
 515         struct ldlm_resource *res;
 516
 517         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 518         ldlm_resource_putref(res);
 519 }
 520
 521 cfs_hash_ops_t ldlm_ns_hash_ops = {
 522         .hs_hash        = ldlm_res_hop_hash,
 523         .hs_key         = ldlm_res_hop_key,
 524         .hs_keycmp      = ldlm_res_hop_keycmp,
 525         .hs_keycpy      = NULL,
 526         .hs_object      = ldlm_res_hop_object,
 527         .hs_get         = ldlm_res_hop_get_locked,
 528         .hs_put_locked  = ldlm_res_hop_put_locked,
 529         .hs_put         = ldlm_res_hop_put
 530 };
 531
 532 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 533         .hs_hash        = ldlm_res_hop_fid_hash,
 534         .hs_key         = ldlm_res_hop_key,
 535         .hs_keycmp      = ldlm_res_hop_keycmp,
 536         .hs_keycpy      = NULL,
 537         .hs_object      = ldlm_res_hop_object,
 538         .hs_get         = ldlm_res_hop_get_locked,
 539         .hs_put_locked  = ldlm_res_hop_put_locked,
 540         .hs_put         = ldlm_res_hop_put
 541 };
 542
 543 typedef struct {
 544         ldlm_ns_type_t  nsd_type;
 545         /** hash bucket bits */
 546         unsigned        nsd_bkt_bits;
 547         /** hash bits */
 548         unsigned        nsd_all_bits;
 549         /** hash operations */
 550         cfs_hash_ops_t *nsd_hops;
 551 } ldlm_ns_hash_def_t;
 552
 553 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 554 {
 555         {
 556                 .nsd_type       = LDLM_NS_TYPE_MDC,
 557                 .nsd_bkt_bits   = 11,
 558                 .nsd_all_bits   = 16,
 559                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 560         },
 561         {
 562                 .nsd_type       = LDLM_NS_TYPE_MDT,
 563                 .nsd_bkt_bits   = 14,
 564                 .nsd_all_bits   = 21,
 565                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 566         },
 567         {
 568                 .nsd_type       = LDLM_NS_TYPE_OSC,
 569                 .nsd_bkt_bits   = 8,
 570                 .nsd_all_bits   = 12,
 571                 .nsd_hops       = &ldlm_ns_hash_ops,
 572         },
 573         {
 574                 .nsd_type       = LDLM_NS_TYPE_OST,
 575                 .nsd_bkt_bits   = 11,
 576                 .nsd_all_bits   = 17,
 577                 .nsd_hops       = &ldlm_ns_hash_ops,
 578         },
 579         {
 580                 .nsd_type       = LDLM_NS_TYPE_MGC,
 581                 .nsd_bkt_bits   = 4,
 582                 .nsd_all_bits   = 4,
 583                 .nsd_hops       = &ldlm_ns_hash_ops,
 584         },
 585         {
 586                 .nsd_type       = LDLM_NS_TYPE_MGT,
 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_UNKNOWN,
 593         },
 594 };
 595
 596 /**
 597  * Create and initialize new empty namespace.
 598  */
 599 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 600                                           ldlm_side_t client,
 601                                           ldlm_appetite_t apt,
 602                                           ldlm_ns_type_t ns_type)
 603 {
 604         struct ldlm_namespace *ns = NULL;
 605         struct ldlm_ns_bucket *nsb;
 606         ldlm_ns_hash_def_t    *nsd;
 607         cfs_hash_bd_t          bd;
 608         int                    idx;
 609         int                    rc;
 610         ENTRY;
 611
 612         LASSERT(obd != NULL);
 613
 614         rc = ldlm_get_ref();
 615         if (rc) {
 616                 CERROR("ldlm_get_ref failed: %d\n", rc);
 617                 RETURN(NULL);
 618         }
 619
 620         for (idx = 0;;idx++) {
 621                 nsd = &ldlm_ns_hash_defs[idx];
 622                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 623                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 624                         GOTO(out_ref, NULL);
 625                 }
 626
 627                 if (nsd->nsd_type == ns_type)
 628                         break;
 629         }
 630
 631         OBD_ALLOC_PTR(ns);
 632         if (!ns)
 633                 GOTO(out_ref, NULL);
 634
 635         ns->ns_rs_hash = cfs_hash_create(name,
 636                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 637                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 638                                          CFS_HASH_MIN_THETA,
 639                                          CFS_HASH_MAX_THETA,
 640                                          nsd->nsd_hops,
 641                                          CFS_HASH_DEPTH |
 642                                          CFS_HASH_BIGNAME |
 643                                          CFS_HASH_SPIN_BKTLOCK |
 644                                          CFS_HASH_NO_ITEMREF);
 645         if (ns->ns_rs_hash == NULL)
 646                 GOTO(out_ns, NULL);
 647
 648         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 649                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 650                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 651                 nsb->nsb_namespace = ns;
 652         }
 653
 654         ns->ns_obd      = obd;
 655         ns->ns_appetite = apt;
 656         ns->ns_client   = client;
 657
 658         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 659         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 660         spin_lock_init(&ns->ns_lock);
 661         cfs_atomic_set(&ns->ns_bref, 0);
 662         cfs_waitq_init(&ns->ns_waitq);
 663
 664         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 665         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 666         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 667
 668         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 669         ns->ns_nr_unused          = 0;
 670         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 671         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 672         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 673         ns->ns_timeouts           = 0;
 674         ns->ns_orig_connect_flags = 0;
 675         ns->ns_connect_flags      = 0;
 676         ns->ns_stopping           = 0;
 677         rc = ldlm_namespace_proc_register(ns);
 678         if (rc != 0) {
 679                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 680                 GOTO(out_hash, rc);
 681         }
 682
 683         idx = ldlm_namespace_nr_read(client);
 684         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 685         if (rc) {
 686                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 687                 GOTO(out_proc, rc);
 688         }
 689
 690         ldlm_namespace_register(ns, client);
 691         RETURN(ns);
 692 out_proc:
 693         ldlm_namespace_proc_unregister(ns);
 694         ldlm_namespace_cleanup(ns, 0);
 695 out_hash:
 696         cfs_hash_putref(ns->ns_rs_hash);
 697 out_ns:
 698         OBD_FREE_PTR(ns);
 699 out_ref:
 700         ldlm_put_ref();
 701         RETURN(NULL);
 702 }
 703 EXPORT_SYMBOL(ldlm_namespace_new);
 704
 705 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 706
 707 /**
 708  * Cancel and destroy all locks on a resource.
 709  *
 710  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 711  * clean up.  This is currently only used for recovery, and we make
 712  * certain assumptions as a result--notably, that we shouldn't cancel
 713  * locks with refs.
 714  */
 715 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 716                              __u64 flags)
 717 {
 718         cfs_list_t *tmp;
 719         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 720         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 721
 722         do {
 723                 struct ldlm_lock *lock = NULL;
 724
 725                 /* First, we look for non-cleaned-yet lock
 726                  * all cleaned locks are marked by CLEANED flag. */
 727                 lock_res(res);
 728                 cfs_list_for_each(tmp, q) {
 729                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 730                                               l_res_link);
 731                         if (lock->l_flags & LDLM_FL_CLEANED) {
 732                                 lock = NULL;
 733                                 continue;
 734                         }
 735                         LDLM_LOCK_GET(lock);
 736                         lock->l_flags |= LDLM_FL_CLEANED;
 737                         break;
 738                 }
 739
 740                 if (lock == NULL) {
 741                         unlock_res(res);
 742                         break;
 743                 }
 744
 745                 /* Set CBPENDING so nothing in the cancellation path
 746                  * can match this lock. */
 747                 lock->l_flags |= LDLM_FL_CBPENDING;
 748                 lock->l_flags |= LDLM_FL_FAILED;
 749                 lock->l_flags |= flags;
 750
 751                 /* ... without sending a CANCEL message for local_only. */
 752                 if (local_only)
 753                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 754
 755                 if (local_only && (lock->l_readers || lock->l_writers)) {
 756                         /* This is a little bit gross, but much better than the
 757                          * alternative: pretend that we got a blocking AST from
 758                          * the server, so that when the lock is decref'd, it
 759                          * will go away ... */
 760                         unlock_res(res);
 761                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 762                         if (lock->l_completion_ast)
 763                                 lock->l_completion_ast(lock, 0, NULL);
 764                         LDLM_LOCK_RELEASE(lock);
 765                         continue;
 766                 }
 767
 768                 if (client) {
 769                         struct lustre_handle lockh;
 770
 771                         unlock_res(res);
 772                         ldlm_lock2handle(lock, &lockh);
 773                         rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 774                         if (rc)
 775                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 776                 } else {
 777                         ldlm_resource_unlink_lock(lock);
 778                         unlock_res(res);
 779                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 780                                    "client node");
 781                         ldlm_lock_destroy(lock);
 782                 }
 783                 LDLM_LOCK_RELEASE(lock);
 784         } while (1);
 785 }
 786
 787 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 788                                cfs_hlist_node_t *hnode, void *arg)
 789 {
 790         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 791         __u64 flags = *(__u64 *)arg;
 792
 793         cleanup_resource(res, &res->lr_granted, flags);
 794         cleanup_resource(res, &res->lr_converting, flags);
 795         cleanup_resource(res, &res->lr_waiting, flags);
 796
 797         return 0;
 798 }
 799
 800 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 801                                   cfs_hlist_node_t *hnode, void *arg)
 802 {
 803         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 804
 805         lock_res(res);
 806         CERROR("Namespace %s resource refcount nonzero "
 807                "(%d) after lock cleanup; forcing "
 808                "cleanup.\n",
 809                ldlm_ns_name(ldlm_res_to_ns(res)),
 810                cfs_atomic_read(&res->lr_refcount) - 1);
 811
 812         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 813                LPU64") (rc: %d)\n", res,
 814                res->lr_name.name[0], res->lr_name.name[1],
 815                res->lr_name.name[2], res->lr_name.name[3],
 816                cfs_atomic_read(&res->lr_refcount) - 1);
 817
 818         ldlm_resource_dump(D_ERROR, res);
 819         unlock_res(res);
 820         return 0;
 821 }
 822
 823 /**
 824  * Cancel and destroy all locks in the namespace.
 825  *
 826  * Typically used during evictions when server notified client that it was
 827  * evicted and all of its state needs to be destroyed.
 828  * Also used during shutdown.
 829  */
 830 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 831 {
 832         if (ns == NULL) {
 833                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 834                 return ELDLM_OK;
 835         }
 836
 837         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
 838         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 839         return ELDLM_OK;
 840 }
 841 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 842
 843 /**
 844  * Attempts to free namespace.
 845  *
 846  * Only used when namespace goes away, like during an unmount.
 847  */
 848 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 849 {
 850         ENTRY;
 851
 852         /* At shutdown time, don't call the cancellation callback */
 853         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 854
 855         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 856                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 857                 int rc;
 858                 CDEBUG(D_DLMTRACE,
 859                        "dlm namespace %s free waiting on refcount %d\n",
 860                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 861 force_wait:
 862                 if (force)
 863                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 864
 865                 rc = l_wait_event(ns->ns_waitq,
 866                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 867
 868                 /* Forced cleanups should be able to reclaim all references,
 869                  * so it's safe to wait forever... we can't leak locks... */
 870                 if (force && rc == -ETIMEDOUT) {
 871                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 872                                        "namespace with %d resources in use, "
 873                                        "(rc=%d)\n", ldlm_ns_name(ns),
 874                                        cfs_atomic_read(&ns->ns_bref), rc);
 875                         GOTO(force_wait, rc);
 876                 }
 877
 878                 if (cfs_atomic_read(&ns->ns_bref)) {
 879                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 880                                        "with %d resources in use, (rc=%d)\n",
 881                                        ldlm_ns_name(ns),
 882                                        cfs_atomic_read(&ns->ns_bref), rc);
 883                         RETURN(ELDLM_NAMESPACE_EXISTS);
 884                 }
 885                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 886                        ldlm_ns_name(ns));
 887         }
 888
 889         RETURN(ELDLM_OK);
 890 }
 891
 892 /**
 893  * Performs various cleanups for passed \a ns to make it drop refc and be
 894  * ready for freeing. Waits for refc == 0.
 895  *
 896  * The following is done:
 897  * (0) Unregister \a ns from its list to make inaccessible for potential
 898  * users like pools thread and others;
 899  * (1) Clear all locks in \a ns.
 900  */
 901 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 902                                struct obd_import *imp,
 903                                int force)
 904 {
 905         int rc;
 906         ENTRY;
 907         if (!ns) {
 908                 EXIT;
 909                 return;
 910         }
 911
 912         spin_lock(&ns->ns_lock);
 913         ns->ns_stopping = 1;
 914         spin_unlock(&ns->ns_lock);
 915
 916         /*
 917          * Can fail with -EINTR when force == 0 in which case try harder.
 918          */
 919         rc = __ldlm_namespace_free(ns, force);
 920         if (rc != ELDLM_OK) {
 921                 if (imp) {
 922                         ptlrpc_disconnect_import(imp, 0);
 923                         ptlrpc_invalidate_import(imp);
 924                 }
 925
 926                 /*
 927                  * With all requests dropped and the import inactive
 928                  * we are gaurenteed all reference will be dropped.
 929                  */
 930                 rc = __ldlm_namespace_free(ns, 1);
 931                 LASSERT(rc == 0);
 932         }
 933         EXIT;
 934 }
 935
 936 /**
 937  * Performs freeing memory structures related to \a ns. This is only done
 938  * when ldlm_namespce_free_prior() successfully removed all resources
 939  * referencing \a ns and its refc == 0.
 940  */
 941 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 942 {
 943         ENTRY;
 944         if (!ns) {
 945                 EXIT;
 946                 return;
 947         }
 948
 949         /* Make sure that nobody can find this ns in its list. */
 950         ldlm_namespace_unregister(ns, ns->ns_client);
 951         /* Fini pool _before_ parent proc dir is removed. This is important as
 952          * ldlm_pool_fini() removes own proc dir which is child to @dir.
 953          * Removing it after @dir may cause oops. */
 954         ldlm_pool_fini(&ns->ns_pool);
 955
 956         ldlm_namespace_proc_unregister(ns);
 957         cfs_hash_putref(ns->ns_rs_hash);
 958         /* Namespace \a ns should be not on list at this time, otherwise
 959          * this will cause issues related to using freed \a ns in poold
 960          * thread. */
 961         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 962         OBD_FREE_PTR(ns);
 963         ldlm_put_ref();
 964         EXIT;
 965 }
 966
 967 /**
 968  * Cleanup the resource, and free namespace.
 969  * bug 12864:
 970  * Deadlock issue:
 971  * proc1: destroy import
 972  *        class_disconnect_export(grab cl_sem) ->
 973  *              -> ldlm_namespace_free ->
 974  *              -> lprocfs_remove(grab _lprocfs_lock).
 975  * proc2: read proc info
 976  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 977  *              -> osc_rd_active, etc(grab cl_sem).
 978  *
 979  * So that I have to split the ldlm_namespace_free into two parts - the first
 980  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 981  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 982  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 983  * held.
 984  */
 985 void ldlm_namespace_free(struct ldlm_namespace *ns,
 986                          struct obd_import *imp,
 987                          int force)
 988 {
 989         ldlm_namespace_free_prior(ns, imp, force);
 990         ldlm_namespace_free_post(ns);
 991 }
 992 EXPORT_SYMBOL(ldlm_namespace_free);
 993
 994 void ldlm_namespace_get(struct ldlm_namespace *ns)
 995 {
 996         cfs_atomic_inc(&ns->ns_bref);
 997 }
 998 EXPORT_SYMBOL(ldlm_namespace_get);
 999
1000 /* This is only for callers that care about refcount */
1001 int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1002 {
1003         return cfs_atomic_inc_return(&ns->ns_bref);
1004 }
1005
1006 void ldlm_namespace_put(struct ldlm_namespace *ns)
1007 {
1008         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1009                 cfs_waitq_signal(&ns->ns_waitq);
1010                 spin_unlock(&ns->ns_lock);
1011         }
1012 }
1013 EXPORT_SYMBOL(ldlm_namespace_put);
1014
1015 /** Register \a ns in the list of namespaces */
1016 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
1017 {
1018         mutex_lock(ldlm_namespace_lock(client));
1019         LASSERT(cfs_list_empty(&ns->ns_list_chain));
1020         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1021         ldlm_namespace_nr_inc(client);
1022         mutex_unlock(ldlm_namespace_lock(client));
1023 }
1024
1025 /** Unregister \a ns from the list of namespaces. */
1026 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1027 {
1028         mutex_lock(ldlm_namespace_lock(client));
1029         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1030         /* Some asserts and possibly other parts of the code are still
1031          * using list_empty(&ns->ns_list_chain). This is why it is
1032          * important to use list_del_init() here. */
1033         cfs_list_del_init(&ns->ns_list_chain);
1034         ldlm_namespace_nr_dec(client);
1035         mutex_unlock(ldlm_namespace_lock(client));
1036 }
1037
1038 /** Should be called with ldlm_namespace_lock(client) taken. */
1039 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1040                                        ldlm_side_t client)
1041 {
1042         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1043         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1044         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1045 }
1046
1047 /** Should be called with ldlm_namespace_lock(client) taken. */
1048 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1049                                          ldlm_side_t client)
1050 {
1051         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1052         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1053         cfs_list_move_tail(&ns->ns_list_chain,
1054                            ldlm_namespace_inactive_list(client));
1055 }
1056
1057 /** Should be called with ldlm_namespace_lock(client) taken. */
1058 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1059 {
1060         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1061         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
1062         return container_of(ldlm_namespace_list(client)->next,
1063                 struct ldlm_namespace, ns_list_chain);
1064 }
1065
1066 /** Create and initialize new resource. */
1067 static struct ldlm_resource *ldlm_resource_new(void)
1068 {
1069         struct ldlm_resource *res;
1070         int idx;
1071
1072         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
1073         if (res == NULL)
1074                 return NULL;
1075
1076         CFS_INIT_LIST_HEAD(&res->lr_granted);
1077         CFS_INIT_LIST_HEAD(&res->lr_converting);
1078         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1079
1080         /* Initialize interval trees for each lock mode. */
1081         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1082                 res->lr_itree[idx].lit_size = 0;
1083                 res->lr_itree[idx].lit_mode = 1 << idx;
1084                 res->lr_itree[idx].lit_root = NULL;
1085         }
1086
1087         cfs_atomic_set(&res->lr_refcount, 1);
1088         spin_lock_init(&res->lr_lock);
1089         lu_ref_init(&res->lr_reference);
1090
1091         /* The creator of the resource must unlock the mutex after LVB
1092          * initialization. */
1093         mutex_init(&res->lr_lvb_mutex);
1094         mutex_lock(&res->lr_lvb_mutex);
1095
1096         return res;
1097 }
1098
1099 /**
1100  * Return a reference to resource with given name, creating it if necessary.
1101  * Args: namespace with ns_lock unlocked
1102  * Locks: takes and releases NS hash-lock and res->lr_lock
1103  * Returns: referenced, unlocked ldlm_resource or NULL
1104  */
1105 struct ldlm_resource *
1106 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1107                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1108 {
1109         cfs_hlist_node_t     *hnode;
1110         struct ldlm_resource *res;
1111         cfs_hash_bd_t         bd;
1112         __u64                 version;
1113         int                   ns_refcount = 0;
1114
1115         LASSERT(ns != NULL);
1116         LASSERT(parent == NULL);
1117         LASSERT(ns->ns_rs_hash != NULL);
1118         LASSERT(name->name[0] != 0);
1119
1120         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1121         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1122         if (hnode != NULL) {
1123                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1124                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1125                 /* Synchronize with regard to resource creation. */
1126                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1127                         mutex_lock(&res->lr_lvb_mutex);
1128                         mutex_unlock(&res->lr_lvb_mutex);
1129                 }
1130
1131                 if (unlikely(res->lr_lvb_len < 0)) {
1132                         ldlm_resource_putref(res);
1133                         res = NULL;
1134                 }
1135                 return res;
1136         }
1137
1138         version = cfs_hash_bd_version_get(&bd);
1139         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1140
1141         if (create == 0)
1142                 return NULL;
1143
1144         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1145                  "type: %d\n", type);
1146         res = ldlm_resource_new();
1147         if (!res)
1148                 return NULL;
1149
1150         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1151         res->lr_name       = *name;
1152         res->lr_type       = type;
1153         res->lr_most_restr = LCK_NL;
1154
1155         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1156         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1157                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1158
1159         if (hnode != NULL) {
1160                 /* Someone won the race and already added the resource. */
1161                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1162                 /* Clean lu_ref for failed resource. */
1163                 lu_ref_fini(&res->lr_reference);
1164                 /* We have taken lr_lvb_mutex. Drop it. */
1165                 mutex_unlock(&res->lr_lvb_mutex);
1166                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1167
1168                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1169                 /* Synchronize with regard to resource creation. */
1170                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1171                         mutex_lock(&res->lr_lvb_mutex);
1172                         mutex_unlock(&res->lr_lvb_mutex);
1173                 }
1174
1175                 if (unlikely(res->lr_lvb_len < 0)) {
1176                         ldlm_resource_putref(res);
1177                         res = NULL;
1178                 }
1179                 return res;
1180         }
1181         /* We won! Let's add the resource. */
1182         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1183         if (cfs_hash_bd_count_get(&bd) == 1)
1184                 ns_refcount = ldlm_namespace_get_return(ns);
1185
1186         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1187         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1188                 int rc;
1189
1190                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1191                 rc = ns->ns_lvbo->lvbo_init(res);
1192                 if (rc < 0) {
1193                         CERROR("%s: lvbo_init failed for resource "LPX64":"
1194                                LPX64": rc = %d\n", ns->ns_obd->obd_name,
1195                                name->name[0], name->name[1], rc);
1196                         if (res->lr_lvb_data) {
1197                                 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1198                                 res->lr_lvb_data = NULL;
1199                         }
1200                         res->lr_lvb_len = rc;
1201                         mutex_unlock(&res->lr_lvb_mutex);
1202                         ldlm_resource_putref(res);
1203                         return NULL;
1204                 }
1205         }
1206
1207         /* We create resource with locked lr_lvb_mutex. */
1208         mutex_unlock(&res->lr_lvb_mutex);
1209
1210         /* Let's see if we happened to be the very first resource in this
1211          * namespace. If so, and this is a client namespace, we need to move
1212          * the namespace into the active namespaces list to be patrolled by
1213          * the ldlm_poold.
1214          * A notable exception, for quota namespaces qsd_lib.c already took a
1215          * namespace reference, so it won't be participating in all of this,
1216          * but I guess that's ok since we have no business cancelling quota
1217          * locks anyway */
1218         if (ns_is_client(ns) && ns_refcount == 1) {
1219                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1220                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1221                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1222         }
1223
1224         return res;
1225 }
1226 EXPORT_SYMBOL(ldlm_resource_get);
1227
1228 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1229 {
1230         LASSERT(res != NULL);
1231         LASSERT(res != LP_POISON);
1232         cfs_atomic_inc(&res->lr_refcount);
1233         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1234                cfs_atomic_read(&res->lr_refcount));
1235         return res;
1236 }
1237
1238 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1239                                          struct ldlm_resource *res)
1240 {
1241         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1242
1243         if (!cfs_list_empty(&res->lr_granted)) {
1244                 ldlm_resource_dump(D_ERROR, res);
1245                 LBUG();
1246         }
1247
1248         if (!cfs_list_empty(&res->lr_converting)) {
1249                 ldlm_resource_dump(D_ERROR, res);
1250                 LBUG();
1251         }
1252
1253         if (!cfs_list_empty(&res->lr_waiting)) {
1254                 ldlm_resource_dump(D_ERROR, res);
1255                 LBUG();
1256         }
1257
1258         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1259                                bd, &res->lr_hash);
1260         lu_ref_fini(&res->lr_reference);
1261         if (cfs_hash_bd_count_get(bd) == 0)
1262                 ldlm_namespace_put(nsb->nsb_namespace);
1263 }
1264
1265 /* Returns 1 if the resource was freed, 0 if it remains. */
1266 int ldlm_resource_putref(struct ldlm_resource *res)
1267 {
1268         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1269         cfs_hash_bd_t   bd;
1270
1271         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1272         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1273                res, cfs_atomic_read(&res->lr_refcount) - 1);
1274
1275         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1276         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1277                 __ldlm_resource_putref_final(&bd, res);
1278                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1279                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1280                         ns->ns_lvbo->lvbo_free(res);
1281                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1282                 return 1;
1283         }
1284         return 0;
1285 }
1286 EXPORT_SYMBOL(ldlm_resource_putref);
1287
1288 /* Returns 1 if the resource was freed, 0 if it remains. */
1289 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1290 {
1291         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1292
1293         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1294         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1295                res, cfs_atomic_read(&res->lr_refcount) - 1);
1296
1297         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1298                 cfs_hash_bd_t bd;
1299
1300                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1301                                 &res->lr_name, &bd);
1302                 __ldlm_resource_putref_final(&bd, res);
1303                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1304                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1305                  * so we should never be here while calling cfs_hash_del,
1306                  * cfs_hash_for_each_nolock is the only case we can get
1307                  * here, which is safe to release cfs_hash_bd_lock.
1308                  */
1309                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1310                         ns->ns_lvbo->lvbo_free(res);
1311                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1312
1313                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1314                 return 1;
1315         }
1316         return 0;
1317 }
1318
1319 /**
1320  * Add a lock into a given resource into specified lock list.
1321  */
1322 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1323                             struct ldlm_lock *lock)
1324 {
1325         check_res_locked(res);
1326
1327         LDLM_DEBUG(lock, "About to add this lock:\n");
1328
1329         if (lock->l_destroyed) {
1330                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1331                 return;
1332         }
1333
1334         LASSERT(cfs_list_empty(&lock->l_res_link));
1335
1336         cfs_list_add_tail(&lock->l_res_link, head);
1337 }
1338
1339 /**
1340  * Insert a lock into resource after specified lock.
1341  *
1342  * Obtain resource description from the lock we are inserting after.
1343  */
1344 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1345                                      struct ldlm_lock *new)
1346 {
1347         struct ldlm_resource *res = original->l_resource;
1348
1349         check_res_locked(res);
1350
1351         ldlm_resource_dump(D_INFO, res);
1352         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1353
1354         if (new->l_destroyed) {
1355                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1356                 goto out;
1357         }
1358
1359         LASSERT(cfs_list_empty(&new->l_res_link));
1360
1361         cfs_list_add(&new->l_res_link, &original->l_res_link);
1362  out:;
1363 }
1364
1365 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1366 {
1367         int type = lock->l_resource->lr_type;
1368
1369         check_res_locked(lock->l_resource);
1370         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1371                 ldlm_unlink_lock_skiplist(lock);
1372         else if (type == LDLM_EXTENT)
1373                 ldlm_extent_unlink_lock(lock);
1374         cfs_list_del_init(&lock->l_res_link);
1375 }
1376 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1377
1378 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1379 {
1380         desc->lr_type = res->lr_type;
1381         desc->lr_name = res->lr_name;
1382 }
1383
1384 /**
1385  * Print information about all locks in all namespaces on this node to debug
1386  * log.
1387  */
1388 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1389 {
1390         cfs_list_t *tmp;
1391
1392         if (!((libcfs_debug | D_ERROR) & level))
1393                 return;
1394
1395         mutex_lock(ldlm_namespace_lock(client));
1396
1397         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1398                 struct ldlm_namespace *ns;
1399                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1400                 ldlm_namespace_dump(level, ns);
1401         }
1402
1403         mutex_unlock(ldlm_namespace_lock(client));
1404 }
1405 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1406
1407 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1408                               cfs_hlist_node_t *hnode, void *arg)
1409 {
1410         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1411         int    level = (int)(unsigned long)arg;
1412
1413         lock_res(res);
1414         ldlm_resource_dump(level, res);
1415         unlock_res(res);
1416
1417         return 0;
1418 }
1419
1420 /**
1421  * Print information about all locks in this namespace on this node to debug
1422  * log.
1423  */
1424 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1425 {
1426         if (!((libcfs_debug | D_ERROR) & level))
1427                 return;
1428
1429         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1430                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1431                ns_is_client(ns) ? "client" : "server");
1432
1433         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1434                 return;
1435
1436         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1437                                  ldlm_res_hash_dump,
1438                                  (void *)(unsigned long)level);
1439         spin_lock(&ns->ns_lock);
1440         ns->ns_next_dump = cfs_time_shift(10);
1441         spin_unlock(&ns->ns_lock);
1442 }
1443 EXPORT_SYMBOL(ldlm_namespace_dump);
1444
1445 /**
1446  * Print information about all locks in this resource to debug log.
1447  */
1448 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1449 {
1450         struct ldlm_lock *lock;
1451         unsigned int granted = 0;
1452
1453         CLASSERT(RES_NAME_SIZE == 4);
1454
1455         if (!((libcfs_debug | D_ERROR) & level))
1456                 return;
1457
1458         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1459                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1460                res->lr_name.name[2], res->lr_name.name[3],
1461                cfs_atomic_read(&res->lr_refcount));
1462
1463         if (!cfs_list_empty(&res->lr_granted)) {
1464                 CDEBUG(level, "Granted locks (in reverse order):\n");
1465                 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1466                                                 l_res_link) {
1467                         LDLM_DEBUG_LIMIT(level, lock, "###");
1468                         if (!(level & D_CANTMASK) &&
1469                             ++granted > ldlm_dump_granted_max) {
1470                                 CDEBUG(level, "only dump %d granted locks to "
1471                                        "avoid DDOS.\n", granted);
1472                                 break;
1473                         }
1474                 }
1475         }
1476         if (!cfs_list_empty(&res->lr_converting)) {
1477                 CDEBUG(level, "Converting locks:\n");
1478                 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1479                         LDLM_DEBUG_LIMIT(level, lock, "###");
1480         }
1481         if (!cfs_list_empty(&res->lr_waiting)) {
1482                 CDEBUG(level, "Waiting locks:\n");
1483                 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1484                         LDLM_DEBUG_LIMIT(level, lock, "###");
1485         }
1486 }