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, 2014, 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 #include <lustre_dlm.h>
  44 #include <lustre_fid.h>
  45 #include <obd_class.h>
  46 #include "ldlm_internal.h"
  47
  48 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  49
  50 int ldlm_srv_namespace_nr = 0;
  51 int ldlm_cli_namespace_nr = 0;
  52
  53 struct mutex ldlm_srv_namespace_lock;
  54 struct list_head ldlm_srv_namespace_list;
  55
  56 struct mutex ldlm_cli_namespace_lock;
  57 /* Client Namespaces that have active resources in them.
  58  * Once all resources go away, ldlm_poold moves such namespaces to the
  59  * inactive list */
  60 struct list_head ldlm_cli_active_namespace_list;
  61 /* Client namespaces that don't have any locks in them */
  62 struct list_head ldlm_cli_inactive_namespace_list;
  63
  64 static struct proc_dir_entry *ldlm_type_proc_dir;
  65 static struct proc_dir_entry *ldlm_ns_proc_dir;
  66 struct proc_dir_entry *ldlm_svc_proc_dir;
  67
  68 /* during debug dump certain amount of granted locks for one resource to avoid
  69  * DDOS. */
  70 static unsigned int ldlm_dump_granted_max = 256;
  71
  72 #ifdef CONFIG_PROC_FS
  73 static ssize_t
  74 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
  75                           size_t count, loff_t *off)
  76 {
  77         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  78         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  79         RETURN(count);
  80 }
  81 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
  82
  83 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
  84 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
  85
  86 int ldlm_proc_setup(void)
  87 {
  88         int rc;
  89         struct lprocfs_vars list[] = {
  90                 { .name =       "dump_namespaces",
  91                   .fops =       &ldlm_dump_ns_fops,
  92                   .proc_mode =  0222 },
  93                 { .name =       "dump_granted_max",
  94                   .fops =       &ldlm_rw_uint_fops,
  95                   .data =       &ldlm_dump_granted_max },
  96                 { .name =       "cancel_unused_locks_before_replay",
  97                   .fops =       &ldlm_rw_uint_fops,
  98                   .data =       &ldlm_cancel_unused_locks_before_replay },
  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_ns_resources_seq_show(struct seq_file *m, void *v)
 162 {
 163         struct ldlm_namespace   *ns  = m->private;
 164         __u64                   res = 0;
 165         struct cfs_hash_bd              bd;
 166         int                     i;
 167
 168         /* result is not strictly consistant */
 169         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 170                 res += cfs_hash_bd_count_get(&bd);
 171         return lprocfs_u64_seq_show(m, &res);
 172 }
 173 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
 174
 175 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
 176 {
 177         struct ldlm_namespace   *ns = m->private;
 178         __u64                   locks;
 179
 180         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 181                                         LPROCFS_FIELDS_FLAGS_SUM);
 182         return lprocfs_u64_seq_show(m, &locks);
 183 }
 184 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
 185
 186 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
 187 {
 188         struct ldlm_namespace *ns = m->private;
 189         __u32 *nr = &ns->ns_max_unused;
 190
 191         if (ns_connect_lru_resize(ns))
 192                 nr = &ns->ns_nr_unused;
 193         return lprocfs_uint_seq_show(m, nr);
 194 }
 195
 196 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
 197                                           const char __user *buffer,
 198                                           size_t count, loff_t *off)
 199 {
 200         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 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 LPROC_SEQ_FOPS(lprocfs_lru_size);
 284
 285 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
 286 {
 287         struct ldlm_namespace *ns = m->private;
 288         unsigned int supp = ns_connect_cancelset(ns);
 289
 290         return lprocfs_uint_seq_show(m, &supp);
 291 }
 292
 293 static ssize_t lprocfs_elc_seq_write(struct file *file,
 294                                      const char __user *buffer,
 295                                      size_t count, loff_t *off)
 296 {
 297         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 298         unsigned int supp = -1;
 299         int rc;
 300
 301         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 302         if (rc < 0)
 303                 return rc;
 304
 305         if (supp == 0)
 306                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 307         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 308                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 309         return count;
 310 }
 311 LPROC_SEQ_FOPS(lprocfs_elc);
 312
 313 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 314 {
 315         if (ns->ns_proc_dir_entry == NULL)
 316                 CERROR("dlm namespace %s has no procfs dir?\n",
 317                        ldlm_ns_name(ns));
 318         else
 319                 lprocfs_remove(&ns->ns_proc_dir_entry);
 320
 321         if (ns->ns_stats != NULL)
 322                 lprocfs_free_stats(&ns->ns_stats);
 323 }
 324
 325 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 326 {
 327         struct lprocfs_vars lock_vars[2];
 328         char lock_name[MAX_STRING_SIZE + 1];
 329         struct proc_dir_entry *ns_pde;
 330
 331         LASSERT(ns != NULL);
 332         LASSERT(ns->ns_rs_hash != NULL);
 333
 334         if (ns->ns_proc_dir_entry != NULL) {
 335                 ns_pde = ns->ns_proc_dir_entry;
 336         } else {
 337                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 338                 if (ns_pde == NULL)
 339                         return -ENOMEM;
 340                 ns->ns_proc_dir_entry = ns_pde;
 341         }
 342
 343         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 344         if (ns->ns_stats == NULL)
 345                 return -ENOMEM;
 346
 347         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 348                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 349
 350         lock_name[MAX_STRING_SIZE] = '\0';
 351
 352         memset(lock_vars, 0, sizeof(lock_vars));
 353         lock_vars[0].name = lock_name;
 354
 355         ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
 356                      &lprocfs_ns_resources_fops);
 357         ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
 358                      &lprocfs_ns_locks_fops);
 359
 360         if (ns_is_client(ns)) {
 361                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
 362                              &ns->ns_nr_unused, &ldlm_uint_fops);
 363                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
 364                              &lprocfs_lru_size_fops);
 365                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
 366                              &ns->ns_max_age, &ldlm_rw_uint_fops);
 367                 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
 368                              ns, &lprocfs_elc_fops);
 369         } else {
 370                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 371                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 372                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 373                              &ns->ns_timeouts, &ldlm_uint_fops);
 374                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 375                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 376                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 377                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 378                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 379                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 380                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 381                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 382         }
 383         return 0;
 384 }
 385 #undef MAX_STRING_SIZE
 386 #else /* CONFIG_PROC_FS */
 387
 388 #define ldlm_namespace_proc_unregister(ns)      ({;})
 389 #define ldlm_namespace_proc_register(ns)        ({0;})
 390
 391 #endif /* CONFIG_PROC_FS */
 392
 393 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 394                                   const void *key, unsigned mask)
 395 {
 396         const struct ldlm_res_id     *id  = key;
 397         unsigned                val = 0;
 398         unsigned                i;
 399
 400         for (i = 0; i < RES_NAME_SIZE; i++)
 401                 val += id->name[i];
 402         return val & mask;
 403 }
 404
 405 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 406                                       const void *key, unsigned mask)
 407 {
 408         const struct ldlm_res_id *id = key;
 409         struct lu_fid       fid;
 410         __u32               hash;
 411         __u32               val;
 412
 413         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 414         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 415         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 416
 417         hash = fid_flatten32(&fid);
 418         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 419         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 420                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 421                 hash += (val >> 5) + (val << 11);
 422         } else {
 423                 val = fid_oid(&fid);
 424         }
 425         hash = hash_long(hash, hs->hs_bkt_bits);
 426         /* give me another random factor */
 427         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 428
 429         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 430         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 431
 432         return hash & mask;
 433 }
 434
 435 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 436 {
 437         struct ldlm_resource   *res;
 438
 439         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 440         return &res->lr_name;
 441 }
 442
 443 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 444 {
 445         struct ldlm_resource   *res;
 446
 447         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 448         return ldlm_res_eq((const struct ldlm_res_id *)key,
 449                            (const struct ldlm_res_id *)&res->lr_name);
 450 }
 451
 452 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 453 {
 454         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 455 }
 456
 457 static void
 458 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 459 {
 460         struct ldlm_resource *res;
 461
 462         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 463         ldlm_resource_getref(res);
 464 }
 465
 466 static void
 467 ldlm_res_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 468 {
 469         struct ldlm_resource *res;
 470
 471         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 472         /* cfs_hash_for_each_nolock is the only chance we call it */
 473         ldlm_resource_putref_locked(res);
 474 }
 475
 476 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 477 {
 478         struct ldlm_resource *res;
 479
 480         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 481         ldlm_resource_putref(res);
 482 }
 483
 484 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 485         .hs_hash        = ldlm_res_hop_hash,
 486         .hs_key         = ldlm_res_hop_key,
 487         .hs_keycmp      = ldlm_res_hop_keycmp,
 488         .hs_keycpy      = NULL,
 489         .hs_object      = ldlm_res_hop_object,
 490         .hs_get         = ldlm_res_hop_get_locked,
 491         .hs_put_locked  = ldlm_res_hop_put_locked,
 492         .hs_put         = ldlm_res_hop_put
 493 };
 494
 495 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 496         .hs_hash        = ldlm_res_hop_fid_hash,
 497         .hs_key         = ldlm_res_hop_key,
 498         .hs_keycmp      = ldlm_res_hop_keycmp,
 499         .hs_keycpy      = NULL,
 500         .hs_object      = ldlm_res_hop_object,
 501         .hs_get         = ldlm_res_hop_get_locked,
 502         .hs_put_locked  = ldlm_res_hop_put_locked,
 503         .hs_put         = ldlm_res_hop_put
 504 };
 505
 506 typedef struct {
 507         ldlm_ns_type_t  nsd_type;
 508         /** hash bucket bits */
 509         unsigned        nsd_bkt_bits;
 510         /** hash bits */
 511         unsigned        nsd_all_bits;
 512         /** hash operations */
 513         struct cfs_hash_ops *nsd_hops;
 514 } ldlm_ns_hash_def_t;
 515
 516 static ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 517 {
 518         {
 519                 .nsd_type       = LDLM_NS_TYPE_MDC,
 520                 .nsd_bkt_bits   = 11,
 521                 .nsd_all_bits   = 16,
 522                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 523         },
 524         {
 525                 .nsd_type       = LDLM_NS_TYPE_MDT,
 526                 .nsd_bkt_bits   = 14,
 527                 .nsd_all_bits   = 21,
 528                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 529         },
 530         {
 531                 .nsd_type       = LDLM_NS_TYPE_OSC,
 532                 .nsd_bkt_bits   = 8,
 533                 .nsd_all_bits   = 12,
 534                 .nsd_hops       = &ldlm_ns_hash_ops,
 535         },
 536         {
 537                 .nsd_type       = LDLM_NS_TYPE_OST,
 538                 .nsd_bkt_bits   = 11,
 539                 .nsd_all_bits   = 17,
 540                 .nsd_hops       = &ldlm_ns_hash_ops,
 541         },
 542         {
 543                 .nsd_type       = LDLM_NS_TYPE_MGC,
 544                 .nsd_bkt_bits   = 4,
 545                 .nsd_all_bits   = 4,
 546                 .nsd_hops       = &ldlm_ns_hash_ops,
 547         },
 548         {
 549                 .nsd_type       = LDLM_NS_TYPE_MGT,
 550                 .nsd_bkt_bits   = 4,
 551                 .nsd_all_bits   = 4,
 552                 .nsd_hops       = &ldlm_ns_hash_ops,
 553         },
 554         {
 555                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 556         },
 557 };
 558
 559 /**
 560  * Create and initialize new empty namespace.
 561  */
 562 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 563                                           ldlm_side_t client,
 564                                           ldlm_appetite_t apt,
 565                                           ldlm_ns_type_t ns_type)
 566 {
 567         struct ldlm_namespace *ns = NULL;
 568         struct ldlm_ns_bucket *nsb;
 569         ldlm_ns_hash_def_t    *nsd;
 570         struct cfs_hash_bd          bd;
 571         int                    idx;
 572         int                    rc;
 573         ENTRY;
 574
 575         LASSERT(obd != NULL);
 576
 577         rc = ldlm_get_ref();
 578         if (rc) {
 579                 CERROR("ldlm_get_ref failed: %d\n", rc);
 580                 RETURN(NULL);
 581         }
 582
 583         for (idx = 0;;idx++) {
 584                 nsd = &ldlm_ns_hash_defs[idx];
 585                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 586                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 587                         GOTO(out_ref, NULL);
 588                 }
 589
 590                 if (nsd->nsd_type == ns_type)
 591                         break;
 592         }
 593
 594         OBD_ALLOC_PTR(ns);
 595         if (!ns)
 596                 GOTO(out_ref, NULL);
 597
 598         ns->ns_rs_hash = cfs_hash_create(name,
 599                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 600                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 601                                          CFS_HASH_MIN_THETA,
 602                                          CFS_HASH_MAX_THETA,
 603                                          nsd->nsd_hops,
 604                                          CFS_HASH_DEPTH |
 605                                          CFS_HASH_BIGNAME |
 606                                          CFS_HASH_SPIN_BKTLOCK |
 607                                          CFS_HASH_NO_ITEMREF);
 608         if (ns->ns_rs_hash == NULL)
 609                 GOTO(out_ns, NULL);
 610
 611         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 612                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 613                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 614                 nsb->nsb_namespace = ns;
 615         }
 616
 617         ns->ns_obd      = obd;
 618         ns->ns_appetite = apt;
 619         ns->ns_client   = client;
 620
 621         INIT_LIST_HEAD(&ns->ns_list_chain);
 622         INIT_LIST_HEAD(&ns->ns_unused_list);
 623         spin_lock_init(&ns->ns_lock);
 624         atomic_set(&ns->ns_bref, 0);
 625         init_waitqueue_head(&ns->ns_waitq);
 626
 627         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 628         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 629         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 630
 631         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 632         ns->ns_nr_unused          = 0;
 633         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 634         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 635         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 636         ns->ns_timeouts           = 0;
 637         ns->ns_orig_connect_flags = 0;
 638         ns->ns_connect_flags      = 0;
 639         ns->ns_stopping           = 0;
 640         rc = ldlm_namespace_proc_register(ns);
 641         if (rc != 0) {
 642                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 643                 GOTO(out_hash, rc);
 644         }
 645
 646         idx = ldlm_namespace_nr_read(client);
 647         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 648         if (rc) {
 649                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 650                 GOTO(out_proc, rc);
 651         }
 652
 653         ldlm_namespace_register(ns, client);
 654         RETURN(ns);
 655 out_proc:
 656         ldlm_namespace_proc_unregister(ns);
 657         ldlm_namespace_cleanup(ns, 0);
 658 out_hash:
 659         cfs_hash_putref(ns->ns_rs_hash);
 660 out_ns:
 661         OBD_FREE_PTR(ns);
 662 out_ref:
 663         ldlm_put_ref();
 664         RETURN(NULL);
 665 }
 666 EXPORT_SYMBOL(ldlm_namespace_new);
 667
 668 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 669
 670 /**
 671  * Cancel and destroy all locks on a resource.
 672  *
 673  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 674  * clean up.  This is currently only used for recovery, and we make
 675  * certain assumptions as a result--notably, that we shouldn't cancel
 676  * locks with refs.
 677  */
 678 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 679                              __u64 flags)
 680 {
 681         struct list_head *tmp;
 682         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 683         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 684
 685         do {
 686                 struct ldlm_lock *lock = NULL;
 687
 688                 /* First, we look for non-cleaned-yet lock
 689                  * all cleaned locks are marked by CLEANED flag. */
 690                 lock_res(res);
 691                 list_for_each(tmp, q) {
 692                         lock = list_entry(tmp, struct ldlm_lock,
 693                                           l_res_link);
 694                         if (ldlm_is_cleaned(lock)) {
 695                                 lock = NULL;
 696                                 continue;
 697                         }
 698                         LDLM_LOCK_GET(lock);
 699                         ldlm_set_cleaned(lock);
 700                         break;
 701                 }
 702
 703                 if (lock == NULL) {
 704                         unlock_res(res);
 705                         break;
 706                 }
 707
 708                 /* Set CBPENDING so nothing in the cancellation path
 709                  * can match this lock. */
 710                 ldlm_set_cbpending(lock);
 711                 ldlm_set_failed(lock);
 712                 lock->l_flags |= flags;
 713
 714                 /* ... without sending a CANCEL message for local_only. */
 715                 if (local_only)
 716                         ldlm_set_local_only(lock);
 717
 718                 if (local_only && (lock->l_readers || lock->l_writers)) {
 719                         /* This is a little bit gross, but much better than the
 720                          * alternative: pretend that we got a blocking AST from
 721                          * the server, so that when the lock is decref'd, it
 722                          * will go away ... */
 723                         unlock_res(res);
 724                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 725                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 726                                 set_current_state(TASK_UNINTERRUPTIBLE);
 727                                 schedule_timeout(cfs_time_seconds(4));
 728                                 set_current_state(TASK_RUNNING);
 729                         }
 730                         if (lock->l_completion_ast)
 731                                 lock->l_completion_ast(lock,
 732                                                        LDLM_FL_FAILED, NULL);
 733                         LDLM_LOCK_RELEASE(lock);
 734                         continue;
 735                 }
 736
 737                 if (client) {
 738                         struct lustre_handle lockh;
 739
 740                         unlock_res(res);
 741                         ldlm_lock2handle(lock, &lockh);
 742                         rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 743                         if (rc)
 744                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 745                 } else {
 746                         ldlm_resource_unlink_lock(lock);
 747                         unlock_res(res);
 748                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 749                                    "client node");
 750                         ldlm_lock_destroy(lock);
 751                 }
 752                 LDLM_LOCK_RELEASE(lock);
 753         } while (1);
 754 }
 755
 756 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 757                                struct hlist_node *hnode, void *arg)
 758 {
 759         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 760         __u64 flags = *(__u64 *)arg;
 761
 762         cleanup_resource(res, &res->lr_granted, flags);
 763         cleanup_resource(res, &res->lr_converting, flags);
 764         cleanup_resource(res, &res->lr_waiting, flags);
 765
 766         return 0;
 767 }
 768
 769 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 770                                   struct hlist_node *hnode, void *arg)
 771 {
 772         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 773
 774         lock_res(res);
 775         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 776                "(%d) after lock cleanup; forcing cleanup.\n",
 777                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 778                atomic_read(&res->lr_refcount) - 1);
 779
 780         ldlm_resource_dump(D_ERROR, res);
 781         unlock_res(res);
 782         return 0;
 783 }
 784
 785 /**
 786  * Cancel and destroy all locks in the namespace.
 787  *
 788  * Typically used during evictions when server notified client that it was
 789  * evicted and all of its state needs to be destroyed.
 790  * Also used during shutdown.
 791  */
 792 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 793 {
 794         if (ns == NULL) {
 795                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 796                 return ELDLM_OK;
 797         }
 798
 799         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
 800         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 801         return ELDLM_OK;
 802 }
 803 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 804
 805 /**
 806  * Attempts to free namespace.
 807  *
 808  * Only used when namespace goes away, like during an unmount.
 809  */
 810 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 811 {
 812         ENTRY;
 813
 814         /* At shutdown time, don't call the cancellation callback */
 815         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 816
 817         if (atomic_read(&ns->ns_bref) > 0) {
 818                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 819                 int rc;
 820                 CDEBUG(D_DLMTRACE,
 821                        "dlm namespace %s free waiting on refcount %d\n",
 822                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
 823 force_wait:
 824                 if (force)
 825                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
 826                                           MSEC_PER_SEC) / 4, NULL, NULL);
 827
 828                 rc = l_wait_event(ns->ns_waitq,
 829                                   atomic_read(&ns->ns_bref) == 0, &lwi);
 830
 831                 /* Forced cleanups should be able to reclaim all references,
 832                  * so it's safe to wait forever... we can't leak locks... */
 833                 if (force && rc == -ETIMEDOUT) {
 834                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 835                                        "namespace with %d resources in use, "
 836                                        "(rc=%d)\n", ldlm_ns_name(ns),
 837                                        atomic_read(&ns->ns_bref), rc);
 838                         GOTO(force_wait, rc);
 839                 }
 840
 841                 if (atomic_read(&ns->ns_bref)) {
 842                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 843                                        "with %d resources in use, (rc=%d)\n",
 844                                        ldlm_ns_name(ns),
 845                                        atomic_read(&ns->ns_bref), rc);
 846                         RETURN(ELDLM_NAMESPACE_EXISTS);
 847                 }
 848                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 849                        ldlm_ns_name(ns));
 850         }
 851
 852         RETURN(ELDLM_OK);
 853 }
 854
 855 /**
 856  * Performs various cleanups for passed \a ns to make it drop refc and be
 857  * ready for freeing. Waits for refc == 0.
 858  *
 859  * The following is done:
 860  * (0) Unregister \a ns from its list to make inaccessible for potential
 861  * users like pools thread and others;
 862  * (1) Clear all locks in \a ns.
 863  */
 864 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 865                                struct obd_import *imp,
 866                                int force)
 867 {
 868         int rc;
 869         ENTRY;
 870         if (!ns) {
 871                 EXIT;
 872                 return;
 873         }
 874
 875         spin_lock(&ns->ns_lock);
 876         ns->ns_stopping = 1;
 877         spin_unlock(&ns->ns_lock);
 878
 879         /*
 880          * Can fail with -EINTR when force == 0 in which case try harder.
 881          */
 882         rc = __ldlm_namespace_free(ns, force);
 883         if (rc != ELDLM_OK) {
 884                 if (imp) {
 885                         ptlrpc_disconnect_import(imp, 0);
 886                         ptlrpc_invalidate_import(imp);
 887                 }
 888
 889                 /*
 890                  * With all requests dropped and the import inactive
 891                  * we are gaurenteed all reference will be dropped.
 892                  */
 893                 rc = __ldlm_namespace_free(ns, 1);
 894                 LASSERT(rc == 0);
 895         }
 896         EXIT;
 897 }
 898
 899 /**
 900  * Performs freeing memory structures related to \a ns. This is only done
 901  * when ldlm_namespce_free_prior() successfully removed all resources
 902  * referencing \a ns and its refc == 0.
 903  */
 904 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 905 {
 906         ENTRY;
 907         if (!ns) {
 908                 EXIT;
 909                 return;
 910         }
 911
 912         /* Make sure that nobody can find this ns in its list. */
 913         ldlm_namespace_unregister(ns, ns->ns_client);
 914         /* Fini pool _before_ parent proc dir is removed. This is important as
 915          * ldlm_pool_fini() removes own proc dir which is child to @dir.
 916          * Removing it after @dir may cause oops. */
 917         ldlm_pool_fini(&ns->ns_pool);
 918
 919         ldlm_namespace_proc_unregister(ns);
 920         cfs_hash_putref(ns->ns_rs_hash);
 921         /* Namespace \a ns should be not on list at this time, otherwise
 922          * this will cause issues related to using freed \a ns in poold
 923          * thread. */
 924         LASSERT(list_empty(&ns->ns_list_chain));
 925         OBD_FREE_PTR(ns);
 926         ldlm_put_ref();
 927         EXIT;
 928 }
 929
 930 /**
 931  * Cleanup the resource, and free namespace.
 932  * bug 12864:
 933  * Deadlock issue:
 934  * proc1: destroy import
 935  *        class_disconnect_export(grab cl_sem) ->
 936  *              -> ldlm_namespace_free ->
 937  *              -> lprocfs_remove(grab _lprocfs_lock).
 938  * proc2: read proc info
 939  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 940  *              -> osc_rd_active, etc(grab cl_sem).
 941  *
 942  * So that I have to split the ldlm_namespace_free into two parts - the first
 943  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 944  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 945  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 946  * held.
 947  */
 948 void ldlm_namespace_free(struct ldlm_namespace *ns,
 949                          struct obd_import *imp,
 950                          int force)
 951 {
 952         ldlm_namespace_free_prior(ns, imp, force);
 953         ldlm_namespace_free_post(ns);
 954 }
 955 EXPORT_SYMBOL(ldlm_namespace_free);
 956
 957 void ldlm_namespace_get(struct ldlm_namespace *ns)
 958 {
 959         atomic_inc(&ns->ns_bref);
 960 }
 961
 962 /* This is only for callers that care about refcount */
 963 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
 964 {
 965         return atomic_inc_return(&ns->ns_bref);
 966 }
 967
 968 void ldlm_namespace_put(struct ldlm_namespace *ns)
 969 {
 970         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 971                 wake_up(&ns->ns_waitq);
 972                 spin_unlock(&ns->ns_lock);
 973         }
 974 }
 975
 976 /** Register \a ns in the list of namespaces */
 977 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 978 {
 979         mutex_lock(ldlm_namespace_lock(client));
 980         LASSERT(list_empty(&ns->ns_list_chain));
 981         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
 982         ldlm_namespace_nr_inc(client);
 983         mutex_unlock(ldlm_namespace_lock(client));
 984 }
 985
 986 /** Unregister \a ns from the list of namespaces. */
 987 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 988 {
 989         mutex_lock(ldlm_namespace_lock(client));
 990         LASSERT(!list_empty(&ns->ns_list_chain));
 991         /* Some asserts and possibly other parts of the code are still
 992          * using list_empty(&ns->ns_list_chain). This is why it is
 993          * important to use list_del_init() here. */
 994         list_del_init(&ns->ns_list_chain);
 995         ldlm_namespace_nr_dec(client);
 996         mutex_unlock(ldlm_namespace_lock(client));
 997 }
 998
 999 /** Should be called with ldlm_namespace_lock(client) taken. */
1000 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1001                                        ldlm_side_t client)
1002 {
1003         LASSERT(!list_empty(&ns->ns_list_chain));
1004         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1005         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1006 }
1007
1008 /** Should be called with ldlm_namespace_lock(client) taken. */
1009 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1010                                          ldlm_side_t client)
1011 {
1012         LASSERT(!list_empty(&ns->ns_list_chain));
1013         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1014         list_move_tail(&ns->ns_list_chain,
1015                        ldlm_namespace_inactive_list(client));
1016 }
1017
1018 /** Should be called with ldlm_namespace_lock(client) taken. */
1019 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1020 {
1021         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1022         LASSERT(!list_empty(ldlm_namespace_list(client)));
1023         return container_of(ldlm_namespace_list(client)->next,
1024                             struct ldlm_namespace, ns_list_chain);
1025 }
1026
1027 /** Create and initialize new resource. */
1028 static struct ldlm_resource *ldlm_resource_new(void)
1029 {
1030         struct ldlm_resource *res;
1031         int idx;
1032
1033         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1034         if (res == NULL)
1035                 return NULL;
1036
1037         INIT_LIST_HEAD(&res->lr_granted);
1038         INIT_LIST_HEAD(&res->lr_converting);
1039         INIT_LIST_HEAD(&res->lr_waiting);
1040
1041         /* Initialize interval trees for each lock mode. */
1042         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1043                 res->lr_itree[idx].lit_size = 0;
1044                 res->lr_itree[idx].lit_mode = 1 << idx;
1045                 res->lr_itree[idx].lit_root = NULL;
1046         }
1047
1048         atomic_set(&res->lr_refcount, 1);
1049         spin_lock_init(&res->lr_lock);
1050         lu_ref_init(&res->lr_reference);
1051
1052         /* Since LVB init can be delayed now, there is no longer need to
1053          * immediatelly acquire mutex here. */
1054         mutex_init(&res->lr_lvb_mutex);
1055         res->lr_lvb_initialized = false;
1056
1057         return res;
1058 }
1059
1060 /**
1061  * Return a reference to resource with given name, creating it if necessary.
1062  * Args: namespace with ns_lock unlocked
1063  * Locks: takes and releases NS hash-lock and res->lr_lock
1064  * Returns: referenced, unlocked ldlm_resource or NULL
1065  */
1066 struct ldlm_resource *
1067 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1068                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1069 {
1070         struct hlist_node       *hnode;
1071         struct ldlm_resource    *res = NULL;
1072         struct cfs_hash_bd              bd;
1073         __u64                   version;
1074         int                     ns_refcount = 0;
1075
1076         LASSERT(ns != NULL);
1077         LASSERT(parent == NULL);
1078         LASSERT(ns->ns_rs_hash != NULL);
1079         LASSERT(name->name[0] != 0);
1080
1081         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1082         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1083         if (hnode != NULL) {
1084                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1085                 GOTO(found, res);
1086         }
1087
1088         version = cfs_hash_bd_version_get(&bd);
1089         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1090
1091         if (create == 0)
1092                 return ERR_PTR(-ENOENT);
1093
1094         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1095                  "type: %d\n", type);
1096         res = ldlm_resource_new();
1097         if (res == NULL)
1098                 return ERR_PTR(-ENOMEM);
1099
1100         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1101         res->lr_name       = *name;
1102         res->lr_type       = type;
1103         res->lr_most_restr = LCK_NL;
1104
1105         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1106         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1107                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1108
1109         if (hnode != NULL) {
1110                 /* Someone won the race and already added the resource. */
1111                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1112                 /* Clean lu_ref for failed resource. */
1113                 lu_ref_fini(&res->lr_reference);
1114                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1115 found:
1116                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1117                 return res;
1118         }
1119         /* We won! Let's add the resource. */
1120         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1121         if (cfs_hash_bd_count_get(&bd) == 1)
1122                 ns_refcount = ldlm_namespace_get_return(ns);
1123
1124         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1125
1126         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1127
1128         /* Let's see if we happened to be the very first resource in this
1129          * namespace. If so, and this is a client namespace, we need to move
1130          * the namespace into the active namespaces list to be patrolled by
1131          * the ldlm_poold. */
1132         if (ns_is_client(ns) && ns_refcount == 1) {
1133                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1134                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1135                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1136         }
1137
1138         return res;
1139 }
1140 EXPORT_SYMBOL(ldlm_resource_get);
1141
1142 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1143 {
1144         LASSERT(res != NULL);
1145         LASSERT(res != LP_POISON);
1146         atomic_inc(&res->lr_refcount);
1147         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1148                atomic_read(&res->lr_refcount));
1149         return res;
1150 }
1151
1152 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1153                                          struct ldlm_resource *res)
1154 {
1155         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1156
1157         if (!list_empty(&res->lr_granted)) {
1158                 ldlm_resource_dump(D_ERROR, res);
1159                 LBUG();
1160         }
1161
1162         if (!list_empty(&res->lr_converting)) {
1163                 ldlm_resource_dump(D_ERROR, res);
1164                 LBUG();
1165         }
1166
1167         if (!list_empty(&res->lr_waiting)) {
1168                 ldlm_resource_dump(D_ERROR, res);
1169                 LBUG();
1170         }
1171
1172         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1173                                bd, &res->lr_hash);
1174         lu_ref_fini(&res->lr_reference);
1175         if (cfs_hash_bd_count_get(bd) == 0)
1176                 ldlm_namespace_put(nsb->nsb_namespace);
1177 }
1178
1179 /* Returns 1 if the resource was freed, 0 if it remains. */
1180 int ldlm_resource_putref(struct ldlm_resource *res)
1181 {
1182         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1183         struct cfs_hash_bd   bd;
1184
1185         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1186         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1187                res, atomic_read(&res->lr_refcount) - 1);
1188
1189         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1190         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1191                 __ldlm_resource_putref_final(&bd, res);
1192                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1193                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1194                         ns->ns_lvbo->lvbo_free(res);
1195                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1196                 return 1;
1197         }
1198         return 0;
1199 }
1200 EXPORT_SYMBOL(ldlm_resource_putref);
1201
1202 /* Returns 1 if the resource was freed, 0 if it remains. */
1203 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1204 {
1205         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1206
1207         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1208         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1209                res, atomic_read(&res->lr_refcount) - 1);
1210
1211         if (atomic_dec_and_test(&res->lr_refcount)) {
1212                 struct cfs_hash_bd bd;
1213
1214                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1215                                 &res->lr_name, &bd);
1216                 __ldlm_resource_putref_final(&bd, res);
1217                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1218                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1219                  * so we should never be here while calling cfs_hash_del,
1220                  * cfs_hash_for_each_nolock is the only case we can get
1221                  * here, which is safe to release cfs_hash_bd_lock.
1222                  */
1223                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1224                         ns->ns_lvbo->lvbo_free(res);
1225                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1226
1227                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1228                 return 1;
1229         }
1230         return 0;
1231 }
1232
1233 /**
1234  * Add a lock into a given resource into specified lock list.
1235  */
1236 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1237                             struct ldlm_lock *lock)
1238 {
1239         check_res_locked(res);
1240
1241         LDLM_DEBUG(lock, "About to add this lock:\n");
1242
1243         if (ldlm_is_destroyed(lock)) {
1244                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1245                 return;
1246         }
1247
1248         LASSERT(list_empty(&lock->l_res_link));
1249
1250         list_add_tail(&lock->l_res_link, head);
1251 }
1252
1253 /**
1254  * Insert a lock into resource after specified lock.
1255  *
1256  * Obtain resource description from the lock we are inserting after.
1257  */
1258 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1259                                      struct ldlm_lock *new)
1260 {
1261         struct ldlm_resource *res = original->l_resource;
1262
1263         check_res_locked(res);
1264
1265         ldlm_resource_dump(D_INFO, res);
1266         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1267
1268         if (ldlm_is_destroyed(new)) {
1269                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1270                 goto out;
1271         }
1272
1273         LASSERT(list_empty(&new->l_res_link));
1274
1275         list_add(&new->l_res_link, &original->l_res_link);
1276  out:;
1277 }
1278
1279 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1280 {
1281         int type = lock->l_resource->lr_type;
1282
1283         check_res_locked(lock->l_resource);
1284         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1285                 ldlm_unlink_lock_skiplist(lock);
1286         else if (type == LDLM_EXTENT)
1287                 ldlm_extent_unlink_lock(lock);
1288         list_del_init(&lock->l_res_link);
1289 }
1290 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1291
1292 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1293 {
1294         desc->lr_type = res->lr_type;
1295         desc->lr_name = res->lr_name;
1296 }
1297
1298 /**
1299  * Print information about all locks in all namespaces on this node to debug
1300  * log.
1301  */
1302 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1303 {
1304         struct list_head *tmp;
1305
1306         if (!((libcfs_debug | D_ERROR) & level))
1307                 return;
1308
1309         mutex_lock(ldlm_namespace_lock(client));
1310
1311         list_for_each(tmp, ldlm_namespace_list(client)) {
1312                 struct ldlm_namespace *ns;
1313                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1314                 ldlm_namespace_dump(level, ns);
1315         }
1316
1317         mutex_unlock(ldlm_namespace_lock(client));
1318 }
1319
1320 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1321                               struct hlist_node *hnode, void *arg)
1322 {
1323         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1324         int    level = (int)(unsigned long)arg;
1325
1326         lock_res(res);
1327         ldlm_resource_dump(level, res);
1328         unlock_res(res);
1329
1330         return 0;
1331 }
1332
1333 /**
1334  * Print information about all locks in this namespace on this node to debug
1335  * log.
1336  */
1337 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1338 {
1339         if (!((libcfs_debug | D_ERROR) & level))
1340                 return;
1341
1342         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1343                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1344                ns_is_client(ns) ? "client" : "server");
1345
1346         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1347                 return;
1348
1349         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1350                                  ldlm_res_hash_dump,
1351                                  (void *)(unsigned long)level);
1352         spin_lock(&ns->ns_lock);
1353         ns->ns_next_dump = cfs_time_shift(10);
1354         spin_unlock(&ns->ns_lock);
1355 }
1356
1357 /**
1358  * Print information about all locks in this resource to debug log.
1359  */
1360 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1361 {
1362         struct ldlm_lock *lock;
1363         unsigned int granted = 0;
1364
1365         CLASSERT(RES_NAME_SIZE == 4);
1366
1367         if (!((libcfs_debug | D_ERROR) & level))
1368                 return;
1369
1370         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1371                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1372
1373         if (!list_empty(&res->lr_granted)) {
1374                 CDEBUG(level, "Granted locks (in reverse order):\n");
1375                 list_for_each_entry_reverse(lock, &res->lr_granted,
1376                                                 l_res_link) {
1377                         LDLM_DEBUG_LIMIT(level, lock, "###");
1378                         if (!(level & D_CANTMASK) &&
1379                             ++granted > ldlm_dump_granted_max) {
1380                                 CDEBUG(level, "only dump %d granted locks to "
1381                                        "avoid DDOS.\n", granted);
1382                                 break;
1383                         }
1384                 }
1385         }
1386         if (!list_empty(&res->lr_converting)) {
1387                 CDEBUG(level, "Converting locks:\n");
1388                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1389                         LDLM_DEBUG_LIMIT(level, lock, "###");
1390         }
1391         if (!list_empty(&res->lr_waiting)) {
1392                 CDEBUG(level, "Waiting locks:\n");
1393                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1394                         LDLM_DEBUG_LIMIT(level, lock, "###");
1395         }
1396 }
1397 EXPORT_SYMBOL(ldlm_resource_dump);