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 struct kmem_cache *ldlm_interval_tree_slab;
  50
  51 int ldlm_srv_namespace_nr = 0;
  52 int ldlm_cli_namespace_nr = 0;
  53
  54 struct mutex ldlm_srv_namespace_lock;
  55 struct list_head ldlm_srv_namespace_list;
  56
  57 struct mutex ldlm_cli_namespace_lock;
  58 /* Client Namespaces that have active resources in them.
  59  * Once all resources go away, ldlm_poold moves such namespaces to the
  60  * inactive list */
  61 struct list_head ldlm_cli_active_namespace_list;
  62 /* Client namespaces that don't have any locks in them */
  63 struct list_head ldlm_cli_inactive_namespace_list;
  64
  65 static struct proc_dir_entry *ldlm_type_proc_dir;
  66 static struct proc_dir_entry *ldlm_ns_proc_dir;
  67 struct proc_dir_entry *ldlm_svc_proc_dir;
  68
  69 /* during debug dump certain amount of granted locks for one resource to avoid
  70  * DDOS. */
  71 static unsigned int ldlm_dump_granted_max = 256;
  72
  73 #ifdef CONFIG_PROC_FS
  74 static ssize_t
  75 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
  76                           size_t count, loff_t *off)
  77 {
  78         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  79         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  80         RETURN(count);
  81 }
  82 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
  83
  84 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
  85 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
  86
  87 /* Lock count is stored in the watermark, and it's display as number of MB
  88  * memory consumed by the locks */
  89 static int seq_watermark_show(struct seq_file *m, void *data)
  90 {
  91         __u64 locknr = *(__u64 *)m->private;
  92         return seq_printf(m, LPU64"\n",
  93                           (locknr * sizeof(struct ldlm_lock)) >> 20);
  94 }
  95
  96 static ssize_t seq_watermark_write(struct file *file,
  97                                    const char __user *buffer, size_t count,
  98                                    loff_t *off)
  99 {
 100         __u64 watermark;
 101         __u64 *data = ((struct seq_file *)file->private_data)->private;
 102         int rc;
 103
 104         rc = lprocfs_write_frac_u64_helper(buffer, count, &watermark, 1 << 20);
 105         if (rc) {
 106                 CERROR("Failed to set LDLM watermark, rc = %d.\n", rc);
 107                 return rc;
 108         } else if (watermark != 0 && watermark < (1 << 20)) {
 109                 CERROR("Watermark should be greater than 1MB.\n");
 110                 return -EINVAL;
 111         }
 112
 113         do_div(watermark, sizeof(struct ldlm_lock));
 114         *data = watermark;
 115
 116         if (ldlm_watermark_low != 0 && ldlm_watermark_high != 0 &&
 117             ldlm_watermark_low > ldlm_watermark_high)
 118                 ldlm_watermark_low = ldlm_watermark_high;
 119         return count;
 120 }
 121
 122 static int seq_watermark_open(struct inode *inode, struct file *file)
 123 {
 124         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 125 }
 126
 127 static const struct file_operations ldlm_watermark_fops = {
 128         .owner          = THIS_MODULE,
 129         .open           = seq_watermark_open,
 130         .read           = seq_read,
 131         .write          = seq_watermark_write,
 132         .llseek         = seq_lseek,
 133         .release        = lprocfs_single_release,
 134 };
 135
 136 int ldlm_proc_setup(void)
 137 {
 138         int rc;
 139         struct lprocfs_vars list[] = {
 140                 { .name =       "dump_namespaces",
 141                   .fops =       &ldlm_dump_ns_fops,
 142                   .proc_mode =  0222 },
 143                 { .name =       "dump_granted_max",
 144                   .fops =       &ldlm_rw_uint_fops,
 145                   .data =       &ldlm_dump_granted_max },
 146                 { .name =       "cancel_unused_locks_before_replay",
 147                   .fops =       &ldlm_rw_uint_fops,
 148                   .data =       &ldlm_cancel_unused_locks_before_replay },
 149                 { .name =       "watermark_mb_low",
 150                   .fops =       &ldlm_watermark_fops,
 151                   .data =       &ldlm_watermark_low },
 152                 { .name =       "watermark_mb_high",
 153                   .fops =       &ldlm_watermark_fops,
 154                   .data =       &ldlm_watermark_high },
 155                 { NULL }};
 156         ENTRY;
 157         LASSERT(ldlm_ns_proc_dir == NULL);
 158
 159         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 160                                               proc_lustre_root,
 161                                               NULL, NULL);
 162         if (IS_ERR(ldlm_type_proc_dir)) {
 163                 CERROR("LProcFS failed in ldlm-init\n");
 164                 rc = PTR_ERR(ldlm_type_proc_dir);
 165                 GOTO(err, rc);
 166         }
 167
 168         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 169                                             ldlm_type_proc_dir,
 170                                             NULL, NULL);
 171         if (IS_ERR(ldlm_ns_proc_dir)) {
 172                 CERROR("LProcFS failed in ldlm-init\n");
 173                 rc = PTR_ERR(ldlm_ns_proc_dir);
 174                 GOTO(err_type, rc);
 175         }
 176
 177         ldlm_svc_proc_dir = lprocfs_register("services",
 178                                              ldlm_type_proc_dir,
 179                                              NULL, NULL);
 180         if (IS_ERR(ldlm_svc_proc_dir)) {
 181                 CERROR("LProcFS failed in ldlm-init\n");
 182                 rc = PTR_ERR(ldlm_svc_proc_dir);
 183                 GOTO(err_ns, rc);
 184         }
 185
 186         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 187         if (rc != 0) {
 188                 CERROR("LProcFS failed in ldlm-init\n");
 189                 GOTO(err_svc, rc);
 190         }
 191
 192         RETURN(0);
 193
 194 err_svc:
 195         lprocfs_remove(&ldlm_svc_proc_dir);
 196 err_ns:
 197         lprocfs_remove(&ldlm_ns_proc_dir);
 198 err_type:
 199         lprocfs_remove(&ldlm_type_proc_dir);
 200 err:
 201         ldlm_svc_proc_dir = NULL;
 202         RETURN(rc);
 203 }
 204
 205 void ldlm_proc_cleanup(void)
 206 {
 207         if (ldlm_svc_proc_dir)
 208                 lprocfs_remove(&ldlm_svc_proc_dir);
 209
 210         if (ldlm_ns_proc_dir)
 211                 lprocfs_remove(&ldlm_ns_proc_dir);
 212
 213         if (ldlm_type_proc_dir)
 214                 lprocfs_remove(&ldlm_type_proc_dir);
 215 }
 216
 217 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
 218 {
 219         struct ldlm_namespace   *ns  = m->private;
 220         __u64                   res = 0;
 221         struct cfs_hash_bd              bd;
 222         int                     i;
 223
 224         /* result is not strictly consistant */
 225         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 226                 res += cfs_hash_bd_count_get(&bd);
 227         return lprocfs_u64_seq_show(m, &res);
 228 }
 229 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
 230
 231 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
 232 {
 233         struct ldlm_namespace   *ns = m->private;
 234         __u64                   locks;
 235
 236         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 237                                         LPROCFS_FIELDS_FLAGS_SUM);
 238         return lprocfs_u64_seq_show(m, &locks);
 239 }
 240 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
 241
 242 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
 243 {
 244         struct ldlm_namespace *ns = m->private;
 245         __u32 *nr = &ns->ns_max_unused;
 246
 247         if (ns_connect_lru_resize(ns))
 248                 nr = &ns->ns_nr_unused;
 249         return lprocfs_uint_seq_show(m, nr);
 250 }
 251
 252 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
 253                                           const char __user *buffer,
 254                                           size_t count, loff_t *off)
 255 {
 256         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 257         char dummy[MAX_STRING_SIZE + 1], *end;
 258         unsigned long tmp;
 259         int lru_resize;
 260
 261         dummy[MAX_STRING_SIZE] = '\0';
 262         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 263                 return -EFAULT;
 264
 265         if (strncmp(dummy, "clear", 5) == 0) {
 266                 CDEBUG(D_DLMTRACE,
 267                        "dropping all unused locks from namespace %s\n",
 268                        ldlm_ns_name(ns));
 269                 if (ns_connect_lru_resize(ns)) {
 270                         int canceled, unused  = ns->ns_nr_unused;
 271
 272                         /* Try to cancel all @ns_nr_unused locks. */
 273                         canceled = ldlm_cancel_lru(ns, unused, 0,
 274                                                    LDLM_CANCEL_PASSED);
 275                         if (canceled < unused) {
 276                                 CDEBUG(D_DLMTRACE,
 277                                        "not all requested locks are canceled, "
 278                                        "requested: %d, canceled: %d\n", unused,
 279                                        canceled);
 280                                 return -EINVAL;
 281                         }
 282                 } else {
 283                         tmp = ns->ns_max_unused;
 284                         ns->ns_max_unused = 0;
 285                         ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
 286                         ns->ns_max_unused = tmp;
 287                 }
 288                 return count;
 289         }
 290
 291         tmp = simple_strtoul(dummy, &end, 0);
 292         if (dummy == end) {
 293                 CERROR("invalid value written\n");
 294                 return -EINVAL;
 295         }
 296         lru_resize = (tmp == 0);
 297
 298         if (ns_connect_lru_resize(ns)) {
 299                 if (!lru_resize)
 300                         ns->ns_max_unused = (unsigned int)tmp;
 301
 302                 if (tmp > ns->ns_nr_unused)
 303                         tmp = ns->ns_nr_unused;
 304                 tmp = ns->ns_nr_unused - tmp;
 305
 306                 CDEBUG(D_DLMTRACE,
 307                        "changing namespace %s unused locks from %u to %u\n",
 308                        ldlm_ns_name(ns), ns->ns_nr_unused,
 309                        (unsigned int)tmp);
 310                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
 311
 312                 if (!lru_resize) {
 313                         CDEBUG(D_DLMTRACE,
 314                                "disable lru_resize for namespace %s\n",
 315                                ldlm_ns_name(ns));
 316                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 317                 }
 318         } else {
 319                 CDEBUG(D_DLMTRACE,
 320                        "changing namespace %s max_unused from %u to %u\n",
 321                        ldlm_ns_name(ns), ns->ns_max_unused,
 322                        (unsigned int)tmp);
 323                 ns->ns_max_unused = (unsigned int)tmp;
 324                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
 325
 326                 /* Make sure that LRU resize was originally supported before
 327                  * turning it on here. */
 328                 if (lru_resize &&
 329                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 330                         CDEBUG(D_DLMTRACE,
 331                                "enable lru_resize for namespace %s\n",
 332                                ldlm_ns_name(ns));
 333                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 334                 }
 335         }
 336
 337         return count;
 338 }
 339 LPROC_SEQ_FOPS(lprocfs_lru_size);
 340
 341 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
 342 {
 343         struct ldlm_namespace *ns = m->private;
 344         unsigned int supp = ns_connect_cancelset(ns);
 345
 346         return lprocfs_uint_seq_show(m, &supp);
 347 }
 348
 349 static ssize_t lprocfs_elc_seq_write(struct file *file,
 350                                      const char __user *buffer,
 351                                      size_t count, loff_t *off)
 352 {
 353         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 354         unsigned int supp = -1;
 355         int rc;
 356
 357         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 358         if (rc < 0)
 359                 return rc;
 360
 361         if (supp == 0)
 362                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 363         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 364                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 365         return count;
 366 }
 367 LPROC_SEQ_FOPS(lprocfs_elc);
 368
 369 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 370 {
 371         if (ns->ns_proc_dir_entry == NULL)
 372                 CERROR("dlm namespace %s has no procfs dir?\n",
 373                        ldlm_ns_name(ns));
 374         else
 375                 lprocfs_remove(&ns->ns_proc_dir_entry);
 376
 377         if (ns->ns_stats != NULL)
 378                 lprocfs_free_stats(&ns->ns_stats);
 379 }
 380
 381 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 382 {
 383         struct lprocfs_vars lock_vars[2];
 384         char lock_name[MAX_STRING_SIZE + 1];
 385         struct proc_dir_entry *ns_pde;
 386
 387         LASSERT(ns != NULL);
 388         LASSERT(ns->ns_rs_hash != NULL);
 389
 390         if (ns->ns_proc_dir_entry != NULL) {
 391                 ns_pde = ns->ns_proc_dir_entry;
 392         } else {
 393                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 394                 if (ns_pde == NULL)
 395                         return -ENOMEM;
 396                 ns->ns_proc_dir_entry = ns_pde;
 397         }
 398
 399         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 400         if (ns->ns_stats == NULL)
 401                 return -ENOMEM;
 402
 403         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 404                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 405
 406         lock_name[MAX_STRING_SIZE] = '\0';
 407
 408         memset(lock_vars, 0, sizeof(lock_vars));
 409         lock_vars[0].name = lock_name;
 410
 411         ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
 412                      &lprocfs_ns_resources_fops);
 413         ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
 414                      &lprocfs_ns_locks_fops);
 415
 416         if (ns_is_client(ns)) {
 417                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
 418                              &ns->ns_nr_unused, &ldlm_uint_fops);
 419                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
 420                              &lprocfs_lru_size_fops);
 421                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
 422                              &ns->ns_max_age, &ldlm_rw_uint_fops);
 423                 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
 424                              ns, &lprocfs_elc_fops);
 425         } else {
 426                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 427                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 428                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 429                              &ns->ns_timeouts, &ldlm_uint_fops);
 430                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 431                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 432                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 433                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 434                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 435                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 436                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 437                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 438         }
 439         return 0;
 440 }
 441 #undef MAX_STRING_SIZE
 442 #else /* CONFIG_PROC_FS */
 443
 444 #define ldlm_namespace_proc_unregister(ns)      ({;})
 445 #define ldlm_namespace_proc_register(ns)        ({0;})
 446
 447 #endif /* CONFIG_PROC_FS */
 448
 449 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 450                                   const void *key, unsigned mask)
 451 {
 452         const struct ldlm_res_id     *id  = key;
 453         unsigned                val = 0;
 454         unsigned                i;
 455
 456         for (i = 0; i < RES_NAME_SIZE; i++)
 457                 val += id->name[i];
 458         return val & mask;
 459 }
 460
 461 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 462                                       const void *key, unsigned mask)
 463 {
 464         const struct ldlm_res_id *id = key;
 465         struct lu_fid       fid;
 466         __u32               hash;
 467         __u32               val;
 468
 469         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 470         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 471         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 472
 473         hash = fid_flatten32(&fid);
 474         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 475         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 476                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 477                 hash += (val >> 5) + (val << 11);
 478         } else {
 479                 val = fid_oid(&fid);
 480         }
 481         hash = hash_long(hash, hs->hs_bkt_bits);
 482         /* give me another random factor */
 483         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 484
 485         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 486         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 487
 488         return hash & mask;
 489 }
 490
 491 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 492 {
 493         struct ldlm_resource   *res;
 494
 495         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 496         return &res->lr_name;
 497 }
 498
 499 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 500 {
 501         struct ldlm_resource   *res;
 502
 503         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 504         return ldlm_res_eq((const struct ldlm_res_id *)key,
 505                            (const struct ldlm_res_id *)&res->lr_name);
 506 }
 507
 508 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 509 {
 510         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 511 }
 512
 513 static void
 514 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 515 {
 516         struct ldlm_resource *res;
 517
 518         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 519         ldlm_resource_getref(res);
 520 }
 521
 522 static void
 523 ldlm_res_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 524 {
 525         struct ldlm_resource *res;
 526
 527         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 528         /* cfs_hash_for_each_nolock is the only chance we call it */
 529         ldlm_resource_putref_locked(res);
 530 }
 531
 532 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 533 {
 534         struct ldlm_resource *res;
 535
 536         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 537         ldlm_resource_putref(res);
 538 }
 539
 540 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 541         .hs_hash        = ldlm_res_hop_hash,
 542         .hs_key         = ldlm_res_hop_key,
 543         .hs_keycmp      = ldlm_res_hop_keycmp,
 544         .hs_keycpy      = NULL,
 545         .hs_object      = ldlm_res_hop_object,
 546         .hs_get         = ldlm_res_hop_get_locked,
 547         .hs_put_locked  = ldlm_res_hop_put_locked,
 548         .hs_put         = ldlm_res_hop_put
 549 };
 550
 551 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 552         .hs_hash        = ldlm_res_hop_fid_hash,
 553         .hs_key         = ldlm_res_hop_key,
 554         .hs_keycmp      = ldlm_res_hop_keycmp,
 555         .hs_keycpy      = NULL,
 556         .hs_object      = ldlm_res_hop_object,
 557         .hs_get         = ldlm_res_hop_get_locked,
 558         .hs_put_locked  = ldlm_res_hop_put_locked,
 559         .hs_put         = ldlm_res_hop_put
 560 };
 561
 562 typedef struct {
 563         ldlm_ns_type_t  nsd_type;
 564         /** hash bucket bits */
 565         unsigned        nsd_bkt_bits;
 566         /** hash bits */
 567         unsigned        nsd_all_bits;
 568         /** hash operations */
 569         struct cfs_hash_ops *nsd_hops;
 570 } ldlm_ns_hash_def_t;
 571
 572 static ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 573 {
 574         {
 575                 .nsd_type       = LDLM_NS_TYPE_MDC,
 576                 .nsd_bkt_bits   = 11,
 577                 .nsd_all_bits   = 16,
 578                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 579         },
 580         {
 581                 .nsd_type       = LDLM_NS_TYPE_MDT,
 582                 .nsd_bkt_bits   = 14,
 583                 .nsd_all_bits   = 21,
 584                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 585         },
 586         {
 587                 .nsd_type       = LDLM_NS_TYPE_OSC,
 588                 .nsd_bkt_bits   = 8,
 589                 .nsd_all_bits   = 12,
 590                 .nsd_hops       = &ldlm_ns_hash_ops,
 591         },
 592         {
 593                 .nsd_type       = LDLM_NS_TYPE_OST,
 594                 .nsd_bkt_bits   = 11,
 595                 .nsd_all_bits   = 17,
 596                 .nsd_hops       = &ldlm_ns_hash_ops,
 597         },
 598         {
 599                 .nsd_type       = LDLM_NS_TYPE_MGC,
 600                 .nsd_bkt_bits   = 4,
 601                 .nsd_all_bits   = 4,
 602                 .nsd_hops       = &ldlm_ns_hash_ops,
 603         },
 604         {
 605                 .nsd_type       = LDLM_NS_TYPE_MGT,
 606                 .nsd_bkt_bits   = 4,
 607                 .nsd_all_bits   = 4,
 608                 .nsd_hops       = &ldlm_ns_hash_ops,
 609         },
 610         {
 611                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 612         },
 613 };
 614
 615 /**
 616  * Create and initialize new empty namespace.
 617  */
 618 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 619                                           ldlm_side_t client,
 620                                           ldlm_appetite_t apt,
 621                                           ldlm_ns_type_t ns_type)
 622 {
 623         struct ldlm_namespace *ns = NULL;
 624         struct ldlm_ns_bucket *nsb;
 625         ldlm_ns_hash_def_t    *nsd;
 626         struct cfs_hash_bd          bd;
 627         int                    idx;
 628         int                    rc;
 629         ENTRY;
 630
 631         LASSERT(obd != NULL);
 632
 633         rc = ldlm_get_ref();
 634         if (rc) {
 635                 CERROR("ldlm_get_ref failed: %d\n", rc);
 636                 RETURN(NULL);
 637         }
 638
 639         for (idx = 0;;idx++) {
 640                 nsd = &ldlm_ns_hash_defs[idx];
 641                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 642                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 643                         GOTO(out_ref, NULL);
 644                 }
 645
 646                 if (nsd->nsd_type == ns_type)
 647                         break;
 648         }
 649
 650         OBD_ALLOC_PTR(ns);
 651         if (!ns)
 652                 GOTO(out_ref, NULL);
 653
 654         ns->ns_rs_hash = cfs_hash_create(name,
 655                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 656                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 657                                          CFS_HASH_MIN_THETA,
 658                                          CFS_HASH_MAX_THETA,
 659                                          nsd->nsd_hops,
 660                                          CFS_HASH_DEPTH |
 661                                          CFS_HASH_BIGNAME |
 662                                          CFS_HASH_SPIN_BKTLOCK |
 663                                          CFS_HASH_NO_ITEMREF);
 664         if (ns->ns_rs_hash == NULL)
 665                 GOTO(out_ns, NULL);
 666
 667         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 668                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 669                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 670                 nsb->nsb_namespace = ns;
 671                 nsb->nsb_reclaim_start = 0;
 672         }
 673
 674         ns->ns_obd      = obd;
 675         ns->ns_appetite = apt;
 676         ns->ns_client   = client;
 677
 678         INIT_LIST_HEAD(&ns->ns_list_chain);
 679         INIT_LIST_HEAD(&ns->ns_unused_list);
 680         spin_lock_init(&ns->ns_lock);
 681         atomic_set(&ns->ns_bref, 0);
 682         init_waitqueue_head(&ns->ns_waitq);
 683
 684         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 685         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 686         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 687
 688         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 689         ns->ns_nr_unused          = 0;
 690         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 691         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 692         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 693         ns->ns_timeouts           = 0;
 694         ns->ns_orig_connect_flags = 0;
 695         ns->ns_connect_flags      = 0;
 696         ns->ns_stopping           = 0;
 697         ns->ns_reclaim_start      = 0;
 698         rc = ldlm_namespace_proc_register(ns);
 699         if (rc != 0) {
 700                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 701                 GOTO(out_hash, rc);
 702         }
 703
 704         idx = ldlm_namespace_nr_read(client);
 705         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 706         if (rc) {
 707                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 708                 GOTO(out_proc, rc);
 709         }
 710
 711         ldlm_namespace_register(ns, client);
 712         RETURN(ns);
 713 out_proc:
 714         ldlm_namespace_proc_unregister(ns);
 715         ldlm_namespace_cleanup(ns, 0);
 716 out_hash:
 717         cfs_hash_putref(ns->ns_rs_hash);
 718 out_ns:
 719         OBD_FREE_PTR(ns);
 720 out_ref:
 721         ldlm_put_ref();
 722         RETURN(NULL);
 723 }
 724 EXPORT_SYMBOL(ldlm_namespace_new);
 725
 726 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 727
 728 /**
 729  * Cancel and destroy all locks on a resource.
 730  *
 731  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 732  * clean up.  This is currently only used for recovery, and we make
 733  * certain assumptions as a result--notably, that we shouldn't cancel
 734  * locks with refs.
 735  */
 736 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 737                              __u64 flags)
 738 {
 739         struct list_head *tmp;
 740         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 741         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 742
 743         do {
 744                 struct ldlm_lock *lock = NULL;
 745
 746                 /* First, we look for non-cleaned-yet lock
 747                  * all cleaned locks are marked by CLEANED flag. */
 748                 lock_res(res);
 749                 list_for_each(tmp, q) {
 750                         lock = list_entry(tmp, struct ldlm_lock,
 751                                           l_res_link);
 752                         if (ldlm_is_cleaned(lock)) {
 753                                 lock = NULL;
 754                                 continue;
 755                         }
 756                         LDLM_LOCK_GET(lock);
 757                         ldlm_set_cleaned(lock);
 758                         break;
 759                 }
 760
 761                 if (lock == NULL) {
 762                         unlock_res(res);
 763                         break;
 764                 }
 765
 766                 /* Set CBPENDING so nothing in the cancellation path
 767                  * can match this lock. */
 768                 ldlm_set_cbpending(lock);
 769                 ldlm_set_failed(lock);
 770                 lock->l_flags |= flags;
 771
 772                 /* ... without sending a CANCEL message for local_only. */
 773                 if (local_only)
 774                         ldlm_set_local_only(lock);
 775
 776                 if (local_only && (lock->l_readers || lock->l_writers)) {
 777                         /* This is a little bit gross, but much better than the
 778                          * alternative: pretend that we got a blocking AST from
 779                          * the server, so that when the lock is decref'd, it
 780                          * will go away ... */
 781                         unlock_res(res);
 782                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 783                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 784                                 set_current_state(TASK_UNINTERRUPTIBLE);
 785                                 schedule_timeout(cfs_time_seconds(4));
 786                                 set_current_state(TASK_RUNNING);
 787                         }
 788                         if (lock->l_completion_ast)
 789                                 lock->l_completion_ast(lock,
 790                                                        LDLM_FL_FAILED, NULL);
 791                         LDLM_LOCK_RELEASE(lock);
 792                         continue;
 793                 }
 794
 795                 if (client) {
 796                         struct lustre_handle lockh;
 797
 798                         unlock_res(res);
 799                         ldlm_lock2handle(lock, &lockh);
 800                         rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 801                         if (rc)
 802                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 803                 } else {
 804                         ldlm_resource_unlink_lock(lock);
 805                         unlock_res(res);
 806                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 807                                    "client node");
 808                         ldlm_lock_destroy(lock);
 809                 }
 810                 LDLM_LOCK_RELEASE(lock);
 811         } while (1);
 812 }
 813
 814 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 815                                struct hlist_node *hnode, void *arg)
 816 {
 817         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 818         __u64 flags = *(__u64 *)arg;
 819
 820         cleanup_resource(res, &res->lr_granted, flags);
 821         cleanup_resource(res, &res->lr_converting, flags);
 822         cleanup_resource(res, &res->lr_waiting, flags);
 823
 824         return 0;
 825 }
 826
 827 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 828                                   struct hlist_node *hnode, void *arg)
 829 {
 830         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 831
 832         lock_res(res);
 833         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 834                "(%d) after lock cleanup; forcing cleanup.\n",
 835                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 836                atomic_read(&res->lr_refcount) - 1);
 837
 838         ldlm_resource_dump(D_ERROR, res);
 839         unlock_res(res);
 840         return 0;
 841 }
 842
 843 /**
 844  * Cancel and destroy all locks in the namespace.
 845  *
 846  * Typically used during evictions when server notified client that it was
 847  * evicted and all of its state needs to be destroyed.
 848  * Also used during shutdown.
 849  */
 850 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 851 {
 852         if (ns == NULL) {
 853                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 854                 return ELDLM_OK;
 855         }
 856
 857         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 858                                  &flags, 0);
 859         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
 860                                  NULL, 0);
 861         return ELDLM_OK;
 862 }
 863 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 864
 865 /**
 866  * Attempts to free namespace.
 867  *
 868  * Only used when namespace goes away, like during an unmount.
 869  */
 870 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 871 {
 872         ENTRY;
 873
 874         /* At shutdown time, don't call the cancellation callback */
 875         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 876
 877         if (atomic_read(&ns->ns_bref) > 0) {
 878                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 879                 int rc;
 880                 CDEBUG(D_DLMTRACE,
 881                        "dlm namespace %s free waiting on refcount %d\n",
 882                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
 883 force_wait:
 884                 if (force)
 885                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
 886                                           MSEC_PER_SEC) / 4, NULL, NULL);
 887
 888                 rc = l_wait_event(ns->ns_waitq,
 889                                   atomic_read(&ns->ns_bref) == 0, &lwi);
 890
 891                 /* Forced cleanups should be able to reclaim all references,
 892                  * so it's safe to wait forever... we can't leak locks... */
 893                 if (force && rc == -ETIMEDOUT) {
 894                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 895                                        "namespace with %d resources in use, "
 896                                        "(rc=%d)\n", ldlm_ns_name(ns),
 897                                        atomic_read(&ns->ns_bref), rc);
 898                         GOTO(force_wait, rc);
 899                 }
 900
 901                 if (atomic_read(&ns->ns_bref)) {
 902                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 903                                        "with %d resources in use, (rc=%d)\n",
 904                                        ldlm_ns_name(ns),
 905                                        atomic_read(&ns->ns_bref), rc);
 906                         RETURN(ELDLM_NAMESPACE_EXISTS);
 907                 }
 908                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 909                        ldlm_ns_name(ns));
 910         }
 911
 912         RETURN(ELDLM_OK);
 913 }
 914
 915 /**
 916  * Performs various cleanups for passed \a ns to make it drop refc and be
 917  * ready for freeing. Waits for refc == 0.
 918  *
 919  * The following is done:
 920  * (0) Unregister \a ns from its list to make inaccessible for potential
 921  * users like pools thread and others;
 922  * (1) Clear all locks in \a ns.
 923  */
 924 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 925                                struct obd_import *imp,
 926                                int force)
 927 {
 928         int rc;
 929         ENTRY;
 930         if (!ns) {
 931                 EXIT;
 932                 return;
 933         }
 934
 935         spin_lock(&ns->ns_lock);
 936         ns->ns_stopping = 1;
 937         spin_unlock(&ns->ns_lock);
 938
 939         /*
 940          * Can fail with -EINTR when force == 0 in which case try harder.
 941          */
 942         rc = __ldlm_namespace_free(ns, force);
 943         if (rc != ELDLM_OK) {
 944                 if (imp) {
 945                         ptlrpc_disconnect_import(imp, 0);
 946                         ptlrpc_invalidate_import(imp);
 947                 }
 948
 949                 /*
 950                  * With all requests dropped and the import inactive
 951                  * we are gaurenteed all reference will be dropped.
 952                  */
 953                 rc = __ldlm_namespace_free(ns, 1);
 954                 LASSERT(rc == 0);
 955         }
 956         EXIT;
 957 }
 958
 959 /**
 960  * Performs freeing memory structures related to \a ns. This is only done
 961  * when ldlm_namespce_free_prior() successfully removed all resources
 962  * referencing \a ns and its refc == 0.
 963  */
 964 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 965 {
 966         ENTRY;
 967         if (!ns) {
 968                 EXIT;
 969                 return;
 970         }
 971
 972         /* Make sure that nobody can find this ns in its list. */
 973         ldlm_namespace_unregister(ns, ns->ns_client);
 974         /* Fini pool _before_ parent proc dir is removed. This is important as
 975          * ldlm_pool_fini() removes own proc dir which is child to @dir.
 976          * Removing it after @dir may cause oops. */
 977         ldlm_pool_fini(&ns->ns_pool);
 978
 979         ldlm_namespace_proc_unregister(ns);
 980         cfs_hash_putref(ns->ns_rs_hash);
 981         /* Namespace \a ns should be not on list at this time, otherwise
 982          * this will cause issues related to using freed \a ns in poold
 983          * thread. */
 984         LASSERT(list_empty(&ns->ns_list_chain));
 985         OBD_FREE_PTR(ns);
 986         ldlm_put_ref();
 987         EXIT;
 988 }
 989
 990 /**
 991  * Cleanup the resource, and free namespace.
 992  * bug 12864:
 993  * Deadlock issue:
 994  * proc1: destroy import
 995  *        class_disconnect_export(grab cl_sem) ->
 996  *              -> ldlm_namespace_free ->
 997  *              -> lprocfs_remove(grab _lprocfs_lock).
 998  * proc2: read proc info
 999  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1000  *              -> osc_rd_active, etc(grab cl_sem).
1001  *
1002  * So that I have to split the ldlm_namespace_free into two parts - the first
1003  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1004  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1005  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1006  * held.
1007  */
1008 void ldlm_namespace_free(struct ldlm_namespace *ns,
1009                          struct obd_import *imp,
1010                          int force)
1011 {
1012         ldlm_namespace_free_prior(ns, imp, force);
1013         ldlm_namespace_free_post(ns);
1014 }
1015 EXPORT_SYMBOL(ldlm_namespace_free);
1016
1017 void ldlm_namespace_get(struct ldlm_namespace *ns)
1018 {
1019         atomic_inc(&ns->ns_bref);
1020 }
1021
1022 /* This is only for callers that care about refcount */
1023 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1024 {
1025         return atomic_inc_return(&ns->ns_bref);
1026 }
1027
1028 void ldlm_namespace_put(struct ldlm_namespace *ns)
1029 {
1030         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1031                 wake_up(&ns->ns_waitq);
1032                 spin_unlock(&ns->ns_lock);
1033         }
1034 }
1035
1036 /** Register \a ns in the list of namespaces */
1037 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
1038 {
1039         mutex_lock(ldlm_namespace_lock(client));
1040         LASSERT(list_empty(&ns->ns_list_chain));
1041         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1042         ldlm_namespace_nr_inc(client);
1043         mutex_unlock(ldlm_namespace_lock(client));
1044 }
1045
1046 /** Unregister \a ns from the list of namespaces. */
1047 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1048 {
1049         mutex_lock(ldlm_namespace_lock(client));
1050         LASSERT(!list_empty(&ns->ns_list_chain));
1051         /* Some asserts and possibly other parts of the code are still
1052          * using list_empty(&ns->ns_list_chain). This is why it is
1053          * important to use list_del_init() here. */
1054         list_del_init(&ns->ns_list_chain);
1055         ldlm_namespace_nr_dec(client);
1056         mutex_unlock(ldlm_namespace_lock(client));
1057 }
1058
1059 /** Should be called with ldlm_namespace_lock(client) taken. */
1060 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1061                                        ldlm_side_t client)
1062 {
1063         LASSERT(!list_empty(&ns->ns_list_chain));
1064         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1065         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1066 }
1067
1068 /** Should be called with ldlm_namespace_lock(client) taken. */
1069 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1070                                          ldlm_side_t client)
1071 {
1072         LASSERT(!list_empty(&ns->ns_list_chain));
1073         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1074         list_move_tail(&ns->ns_list_chain,
1075                        ldlm_namespace_inactive_list(client));
1076 }
1077
1078 /** Should be called with ldlm_namespace_lock(client) taken. */
1079 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1080 {
1081         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1082         LASSERT(!list_empty(ldlm_namespace_list(client)));
1083         return container_of(ldlm_namespace_list(client)->next,
1084                             struct ldlm_namespace, ns_list_chain);
1085 }
1086
1087 /** Create and initialize new resource. */
1088 static struct ldlm_resource *ldlm_resource_new(ldlm_type_t type)
1089 {
1090         struct ldlm_resource *res;
1091         int idx;
1092
1093         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1094         if (res == NULL)
1095                 return NULL;
1096
1097         if (type == LDLM_EXTENT) {
1098                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1099                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1100                 if (res->lr_itree == NULL) {
1101                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1102                         return NULL;
1103                 }
1104                 /* Initialize interval trees for each lock mode. */
1105                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1106                         res->lr_itree[idx].lit_size = 0;
1107                         res->lr_itree[idx].lit_mode = 1 << idx;
1108                         res->lr_itree[idx].lit_root = NULL;
1109                 }
1110         }
1111
1112         INIT_LIST_HEAD(&res->lr_granted);
1113         INIT_LIST_HEAD(&res->lr_converting);
1114         INIT_LIST_HEAD(&res->lr_waiting);
1115
1116         atomic_set(&res->lr_refcount, 1);
1117         spin_lock_init(&res->lr_lock);
1118         lu_ref_init(&res->lr_reference);
1119
1120         /* Since LVB init can be delayed now, there is no longer need to
1121          * immediatelly acquire mutex here. */
1122         mutex_init(&res->lr_lvb_mutex);
1123         res->lr_lvb_initialized = false;
1124
1125         return res;
1126 }
1127
1128 /**
1129  * Return a reference to resource with given name, creating it if necessary.
1130  * Args: namespace with ns_lock unlocked
1131  * Locks: takes and releases NS hash-lock and res->lr_lock
1132  * Returns: referenced, unlocked ldlm_resource or NULL
1133  */
1134 struct ldlm_resource *
1135 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1136                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1137 {
1138         struct hlist_node       *hnode;
1139         struct ldlm_resource    *res = NULL;
1140         struct cfs_hash_bd              bd;
1141         __u64                   version;
1142         int                     ns_refcount = 0;
1143
1144         LASSERT(ns != NULL);
1145         LASSERT(parent == NULL);
1146         LASSERT(ns->ns_rs_hash != NULL);
1147         LASSERT(name->name[0] != 0);
1148
1149         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1150         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1151         if (hnode != NULL) {
1152                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1153                 GOTO(found, res);
1154         }
1155
1156         version = cfs_hash_bd_version_get(&bd);
1157         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1158
1159         if (create == 0)
1160                 return ERR_PTR(-ENOENT);
1161
1162         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1163                  "type: %d\n", type);
1164         res = ldlm_resource_new(type);
1165         if (res == NULL)
1166                 return ERR_PTR(-ENOMEM);
1167
1168         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1169         res->lr_name       = *name;
1170         res->lr_type       = type;
1171
1172         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1173         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1174                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1175
1176         if (hnode != NULL) {
1177                 /* Someone won the race and already added the resource. */
1178                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1179                 /* Clean lu_ref for failed resource. */
1180                 lu_ref_fini(&res->lr_reference);
1181                 if (res->lr_itree != NULL)
1182                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1183                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1184                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1185 found:
1186                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1187                 return res;
1188         }
1189         /* We won! Let's add the resource. */
1190         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1191         if (cfs_hash_bd_count_get(&bd) == 1)
1192                 ns_refcount = ldlm_namespace_get_return(ns);
1193
1194         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1195
1196         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1197
1198         /* Let's see if we happened to be the very first resource in this
1199          * namespace. If so, and this is a client namespace, we need to move
1200          * the namespace into the active namespaces list to be patrolled by
1201          * the ldlm_poold. */
1202         if (ns_is_client(ns) && ns_refcount == 1) {
1203                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1204                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1205                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1206         }
1207
1208         return res;
1209 }
1210 EXPORT_SYMBOL(ldlm_resource_get);
1211
1212 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1213 {
1214         LASSERT(res != NULL);
1215         LASSERT(res != LP_POISON);
1216         atomic_inc(&res->lr_refcount);
1217         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1218                atomic_read(&res->lr_refcount));
1219         return res;
1220 }
1221
1222 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1223                                          struct ldlm_resource *res)
1224 {
1225         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1226
1227         if (!list_empty(&res->lr_granted)) {
1228                 ldlm_resource_dump(D_ERROR, res);
1229                 LBUG();
1230         }
1231
1232         if (!list_empty(&res->lr_converting)) {
1233                 ldlm_resource_dump(D_ERROR, res);
1234                 LBUG();
1235         }
1236
1237         if (!list_empty(&res->lr_waiting)) {
1238                 ldlm_resource_dump(D_ERROR, res);
1239                 LBUG();
1240         }
1241
1242         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1243                                bd, &res->lr_hash);
1244         lu_ref_fini(&res->lr_reference);
1245         if (cfs_hash_bd_count_get(bd) == 0)
1246                 ldlm_namespace_put(nsb->nsb_namespace);
1247 }
1248
1249 /* Returns 1 if the resource was freed, 0 if it remains. */
1250 int ldlm_resource_putref(struct ldlm_resource *res)
1251 {
1252         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1253         struct cfs_hash_bd   bd;
1254
1255         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1256         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1257                res, atomic_read(&res->lr_refcount) - 1);
1258
1259         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1260         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1261                 __ldlm_resource_putref_final(&bd, res);
1262                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1263                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1264                         ns->ns_lvbo->lvbo_free(res);
1265                 if (res->lr_itree != NULL)
1266                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1267                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1268                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1269                 return 1;
1270         }
1271         return 0;
1272 }
1273 EXPORT_SYMBOL(ldlm_resource_putref);
1274
1275 /* Returns 1 if the resource was freed, 0 if it remains. */
1276 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1277 {
1278         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1279
1280         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1281         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1282                res, atomic_read(&res->lr_refcount) - 1);
1283
1284         if (atomic_dec_and_test(&res->lr_refcount)) {
1285                 struct cfs_hash_bd bd;
1286
1287                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1288                                 &res->lr_name, &bd);
1289                 __ldlm_resource_putref_final(&bd, res);
1290                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1291                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1292                  * so we should never be here while calling cfs_hash_del,
1293                  * cfs_hash_for_each_nolock is the only case we can get
1294                  * here, which is safe to release cfs_hash_bd_lock.
1295                  */
1296                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1297                         ns->ns_lvbo->lvbo_free(res);
1298                 if (res->lr_itree != NULL)
1299                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1300                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1301                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1302
1303                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1304                 return 1;
1305         }
1306         return 0;
1307 }
1308
1309 /**
1310  * Add a lock into a given resource into specified lock list.
1311  */
1312 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1313                             struct ldlm_lock *lock)
1314 {
1315         check_res_locked(res);
1316
1317         LDLM_DEBUG(lock, "About to add this lock:\n");
1318
1319         if (ldlm_is_destroyed(lock)) {
1320                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1321                 return;
1322         }
1323
1324         LASSERT(list_empty(&lock->l_res_link));
1325
1326         list_add_tail(&lock->l_res_link, head);
1327 }
1328
1329 /**
1330  * Insert a lock into resource after specified lock.
1331  *
1332  * Obtain resource description from the lock we are inserting after.
1333  */
1334 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1335                                      struct ldlm_lock *new)
1336 {
1337         struct ldlm_resource *res = original->l_resource;
1338
1339         check_res_locked(res);
1340
1341         ldlm_resource_dump(D_INFO, res);
1342         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1343
1344         if (ldlm_is_destroyed(new)) {
1345                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1346                 goto out;
1347         }
1348
1349         LASSERT(list_empty(&new->l_res_link));
1350
1351         list_add(&new->l_res_link, &original->l_res_link);
1352  out:;
1353 }
1354
1355 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1356 {
1357         int type = lock->l_resource->lr_type;
1358
1359         check_res_locked(lock->l_resource);
1360         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1361                 ldlm_unlink_lock_skiplist(lock);
1362         else if (type == LDLM_EXTENT)
1363                 ldlm_extent_unlink_lock(lock);
1364         list_del_init(&lock->l_res_link);
1365 }
1366 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1367
1368 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1369 {
1370         desc->lr_type = res->lr_type;
1371         desc->lr_name = res->lr_name;
1372 }
1373
1374 /**
1375  * Print information about all locks in all namespaces on this node to debug
1376  * log.
1377  */
1378 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1379 {
1380         struct list_head *tmp;
1381
1382         if (!((libcfs_debug | D_ERROR) & level))
1383                 return;
1384
1385         mutex_lock(ldlm_namespace_lock(client));
1386
1387         list_for_each(tmp, ldlm_namespace_list(client)) {
1388                 struct ldlm_namespace *ns;
1389                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1390                 ldlm_namespace_dump(level, ns);
1391         }
1392
1393         mutex_unlock(ldlm_namespace_lock(client));
1394 }
1395
1396 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1397                               struct hlist_node *hnode, void *arg)
1398 {
1399         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1400         int    level = (int)(unsigned long)arg;
1401
1402         lock_res(res);
1403         ldlm_resource_dump(level, res);
1404         unlock_res(res);
1405
1406         return 0;
1407 }
1408
1409 /**
1410  * Print information about all locks in this namespace on this node to debug
1411  * log.
1412  */
1413 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1414 {
1415         if (!((libcfs_debug | D_ERROR) & level))
1416                 return;
1417
1418         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1419                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1420                ns_is_client(ns) ? "client" : "server");
1421
1422         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1423                 return;
1424
1425         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1426                                  ldlm_res_hash_dump,
1427                                  (void *)(unsigned long)level, 0);
1428         spin_lock(&ns->ns_lock);
1429         ns->ns_next_dump = cfs_time_shift(10);
1430         spin_unlock(&ns->ns_lock);
1431 }
1432
1433 /**
1434  * Print information about all locks in this resource to debug log.
1435  */
1436 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1437 {
1438         struct ldlm_lock *lock;
1439         unsigned int granted = 0;
1440
1441         CLASSERT(RES_NAME_SIZE == 4);
1442
1443         if (!((libcfs_debug | D_ERROR) & level))
1444                 return;
1445
1446         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1447                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1448
1449         if (!list_empty(&res->lr_granted)) {
1450                 CDEBUG(level, "Granted locks (in reverse order):\n");
1451                 list_for_each_entry_reverse(lock, &res->lr_granted,
1452                                                 l_res_link) {
1453                         LDLM_DEBUG_LIMIT(level, lock, "###");
1454                         if (!(level & D_CANTMASK) &&
1455                             ++granted > ldlm_dump_granted_max) {
1456                                 CDEBUG(level, "only dump %d granted locks to "
1457                                        "avoid DDOS.\n", granted);
1458                                 break;
1459                         }
1460                 }
1461         }
1462         if (!list_empty(&res->lr_converting)) {
1463                 CDEBUG(level, "Converting locks:\n");
1464                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1465                         LDLM_DEBUG_LIMIT(level, lock, "###");
1466         }
1467         if (!list_empty(&res->lr_waiting)) {
1468                 CDEBUG(level, "Waiting locks:\n");
1469                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1470                         LDLM_DEBUG_LIMIT(level, lock, "###");
1471         }
1472 }
1473 EXPORT_SYMBOL(ldlm_resource_dump);