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, 2015, 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 #ifdef HAVE_SERVER_SUPPORT
  88
  89 static int seq_watermark_show(struct seq_file *m, void *data)
  90 {
  91         seq_printf(m, LPU64"\n", *(__u64 *)m->private);
  92         return 0;
  93 }
  94
  95 static ssize_t seq_watermark_write(struct file *file,
  96                                    const char __user *buffer, size_t count,
  97                                    loff_t *off)
  98 {
  99         __u64 watermark;
 100         __u64 *data = ((struct seq_file *)file->private_data)->private;
 101         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 102         int rc;
 103
 104         rc = lprocfs_write_frac_u64_helper(buffer, count, &watermark, 1 << 20);
 105         if (rc) {
 106                 CERROR("Failed to set %s, rc = %d.\n",
 107                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 108                        rc);
 109                 return rc;
 110         } else if (watermark != 0 && watermark < (1 << 20)) {
 111                 CERROR("%s should be greater than 1MB.\n",
 112                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 113                 return -EINVAL;
 114         }
 115         watermark >>= 20;
 116
 117         if (wm_low) {
 118                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 119                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 120                                "lock_limit_mb.\n");
 121                         return -EINVAL;
 122                 }
 123
 124                 *data = watermark;
 125                 if (watermark != 0) {
 126                         watermark <<= 20;
 127                         do_div(watermark, sizeof(struct ldlm_lock));
 128                 }
 129                 ldlm_reclaim_threshold = watermark;
 130         } else {
 131                 if (ldlm_reclaim_threshold_mb != 0 &&
 132                     watermark < ldlm_reclaim_threshold_mb) {
 133                         CERROR("lock_limit_mb must be greater than "
 134                                "lock_reclaim_threshold_mb.\n");
 135                         return -EINVAL;
 136                 }
 137
 138                 *data = watermark;
 139                 if (watermark != 0) {
 140                         watermark <<= 20;
 141                         do_div(watermark, sizeof(struct ldlm_lock));
 142                 }
 143                 ldlm_lock_limit = watermark;
 144         }
 145
 146         return count;
 147 }
 148
 149 static int seq_watermark_open(struct inode *inode, struct file *file)
 150 {
 151         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 152 }
 153
 154 static const struct file_operations ldlm_watermark_fops = {
 155         .owner          = THIS_MODULE,
 156         .open           = seq_watermark_open,
 157         .read           = seq_read,
 158         .write          = seq_watermark_write,
 159         .llseek         = seq_lseek,
 160         .release        = lprocfs_single_release,
 161 };
 162
 163 static int seq_granted_show(struct seq_file *m, void *data)
 164 {
 165         seq_printf(m, LPU64"\n", percpu_counter_sum_positive(
 166                    (struct percpu_counter *)m->private));
 167         return 0;
 168 }
 169
 170 static int seq_granted_open(struct inode *inode, struct file *file)
 171 {
 172         return single_open(file, seq_granted_show, PDE_DATA(inode));
 173 }
 174
 175 static const struct file_operations ldlm_granted_fops = {
 176         .owner  = THIS_MODULE,
 177         .open   = seq_granted_open,
 178         .read   = seq_read,
 179         .llseek = seq_lseek,
 180         .release = seq_release,
 181 };
 182
 183 #endif /* HAVE_SERVER_SUPPORT */
 184
 185 int ldlm_proc_setup(void)
 186 {
 187         int rc;
 188         struct lprocfs_vars list[] = {
 189                 { .name =       "dump_namespaces",
 190                   .fops =       &ldlm_dump_ns_fops,
 191                   .proc_mode =  0222 },
 192                 { .name =       "dump_granted_max",
 193                   .fops =       &ldlm_rw_uint_fops,
 194                   .data =       &ldlm_dump_granted_max },
 195                 { .name =       "cancel_unused_locks_before_replay",
 196                   .fops =       &ldlm_rw_uint_fops,
 197                   .data =       &ldlm_cancel_unused_locks_before_replay },
 198 #ifdef HAVE_SERVER_SUPPORT
 199                 { .name =       "lock_reclaim_threshold_mb",
 200                   .fops =       &ldlm_watermark_fops,
 201                   .data =       &ldlm_reclaim_threshold_mb },
 202                 { .name =       "lock_limit_mb",
 203                   .fops =       &ldlm_watermark_fops,
 204                   .data =       &ldlm_lock_limit_mb },
 205                 { .name =       "lock_granted_count",
 206                   .fops =       &ldlm_granted_fops,
 207                   .data =       &ldlm_granted_total },
 208 #endif
 209                 { NULL }};
 210         ENTRY;
 211         LASSERT(ldlm_ns_proc_dir == NULL);
 212
 213         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 214                                               proc_lustre_root,
 215                                               NULL, NULL);
 216         if (IS_ERR(ldlm_type_proc_dir)) {
 217                 CERROR("LProcFS failed in ldlm-init\n");
 218                 rc = PTR_ERR(ldlm_type_proc_dir);
 219                 GOTO(err, rc);
 220         }
 221
 222         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 223                                             ldlm_type_proc_dir,
 224                                             NULL, NULL);
 225         if (IS_ERR(ldlm_ns_proc_dir)) {
 226                 CERROR("LProcFS failed in ldlm-init\n");
 227                 rc = PTR_ERR(ldlm_ns_proc_dir);
 228                 GOTO(err_type, rc);
 229         }
 230
 231         ldlm_svc_proc_dir = lprocfs_register("services",
 232                                              ldlm_type_proc_dir,
 233                                              NULL, NULL);
 234         if (IS_ERR(ldlm_svc_proc_dir)) {
 235                 CERROR("LProcFS failed in ldlm-init\n");
 236                 rc = PTR_ERR(ldlm_svc_proc_dir);
 237                 GOTO(err_ns, rc);
 238         }
 239
 240         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 241         if (rc != 0) {
 242                 CERROR("LProcFS failed in ldlm-init\n");
 243                 GOTO(err_svc, rc);
 244         }
 245
 246         RETURN(0);
 247
 248 err_svc:
 249         lprocfs_remove(&ldlm_svc_proc_dir);
 250 err_ns:
 251         lprocfs_remove(&ldlm_ns_proc_dir);
 252 err_type:
 253         lprocfs_remove(&ldlm_type_proc_dir);
 254 err:
 255         ldlm_svc_proc_dir = NULL;
 256         RETURN(rc);
 257 }
 258
 259 void ldlm_proc_cleanup(void)
 260 {
 261         if (ldlm_svc_proc_dir)
 262                 lprocfs_remove(&ldlm_svc_proc_dir);
 263
 264         if (ldlm_ns_proc_dir)
 265                 lprocfs_remove(&ldlm_ns_proc_dir);
 266
 267         if (ldlm_type_proc_dir)
 268                 lprocfs_remove(&ldlm_type_proc_dir);
 269 }
 270
 271 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
 272 {
 273         struct ldlm_namespace   *ns  = m->private;
 274         __u64                   res = 0;
 275         struct cfs_hash_bd              bd;
 276         int                     i;
 277
 278         /* result is not strictly consistant */
 279         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 280                 res += cfs_hash_bd_count_get(&bd);
 281         return lprocfs_u64_seq_show(m, &res);
 282 }
 283 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
 284
 285 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
 286 {
 287         struct ldlm_namespace   *ns = m->private;
 288         __u64                   locks;
 289
 290         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 291                                         LPROCFS_FIELDS_FLAGS_SUM);
 292         return lprocfs_u64_seq_show(m, &locks);
 293 }
 294 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
 295
 296 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
 297 {
 298         struct ldlm_namespace *ns = m->private;
 299         __u32 *nr = &ns->ns_max_unused;
 300
 301         if (ns_connect_lru_resize(ns))
 302                 nr = &ns->ns_nr_unused;
 303         return lprocfs_uint_seq_show(m, nr);
 304 }
 305
 306 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
 307                                           const char __user *buffer,
 308                                           size_t count, loff_t *off)
 309 {
 310         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 311         char dummy[MAX_STRING_SIZE + 1], *end;
 312         unsigned long tmp;
 313         int lru_resize;
 314
 315         dummy[MAX_STRING_SIZE] = '\0';
 316         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 317                 return -EFAULT;
 318
 319         if (strncmp(dummy, "clear", 5) == 0) {
 320                 CDEBUG(D_DLMTRACE,
 321                        "dropping all unused locks from namespace %s\n",
 322                        ldlm_ns_name(ns));
 323                 if (ns_connect_lru_resize(ns)) {
 324                         int canceled, unused  = ns->ns_nr_unused;
 325
 326                         /* Try to cancel all @ns_nr_unused locks. */
 327                         canceled = ldlm_cancel_lru(ns, unused, 0,
 328                                                    LDLM_LRU_FLAG_PASSED);
 329                         if (canceled < unused) {
 330                                 CDEBUG(D_DLMTRACE,
 331                                        "not all requested locks are canceled, "
 332                                        "requested: %d, canceled: %d\n", unused,
 333                                        canceled);
 334                                 return -EINVAL;
 335                         }
 336                 } else {
 337                         tmp = ns->ns_max_unused;
 338                         ns->ns_max_unused = 0;
 339                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED);
 340                         ns->ns_max_unused = tmp;
 341                 }
 342                 return count;
 343         }
 344
 345         tmp = simple_strtoul(dummy, &end, 0);
 346         if (dummy == end) {
 347                 CERROR("invalid value written\n");
 348                 return -EINVAL;
 349         }
 350         lru_resize = (tmp == 0);
 351
 352         if (ns_connect_lru_resize(ns)) {
 353                 if (!lru_resize)
 354                         ns->ns_max_unused = tmp;
 355
 356                 if (tmp > ns->ns_nr_unused)
 357                         tmp = ns->ns_nr_unused;
 358                 tmp = ns->ns_nr_unused - tmp;
 359
 360                 CDEBUG(D_DLMTRACE,
 361                        "changing namespace %s unused locks from %u to %u\n",
 362                        ldlm_ns_name(ns), ns->ns_nr_unused,
 363                        (unsigned int)tmp);
 364                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 365
 366                 if (!lru_resize) {
 367                         CDEBUG(D_DLMTRACE,
 368                                "disable lru_resize for namespace %s\n",
 369                                ldlm_ns_name(ns));
 370                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 371                 }
 372         } else {
 373                 CDEBUG(D_DLMTRACE,
 374                        "changing namespace %s max_unused from %u to %u\n",
 375                        ldlm_ns_name(ns), ns->ns_max_unused,
 376                        (unsigned int)tmp);
 377                 ns->ns_max_unused = (unsigned int)tmp;
 378                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 379
 380                 /* Make sure that LRU resize was originally supported before
 381                  * turning it on here. */
 382                 if (lru_resize &&
 383                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 384                         CDEBUG(D_DLMTRACE,
 385                                "enable lru_resize for namespace %s\n",
 386                                ldlm_ns_name(ns));
 387                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 388                 }
 389         }
 390
 391         return count;
 392 }
 393 LPROC_SEQ_FOPS(lprocfs_lru_size);
 394
 395 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
 396 {
 397         struct ldlm_namespace *ns = m->private;
 398         unsigned int supp = ns_connect_cancelset(ns);
 399
 400         return lprocfs_uint_seq_show(m, &supp);
 401 }
 402
 403 static ssize_t lprocfs_elc_seq_write(struct file *file,
 404                                      const char __user *buffer,
 405                                      size_t count, loff_t *off)
 406 {
 407         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 408         unsigned int supp = -1;
 409         int rc;
 410
 411         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 412         if (rc < 0)
 413                 return rc;
 414
 415         if (supp == 0)
 416                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 417         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 418                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 419         return count;
 420 }
 421 LPROC_SEQ_FOPS(lprocfs_elc);
 422
 423 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 424 {
 425         if (ns->ns_proc_dir_entry == NULL)
 426                 CERROR("dlm namespace %s has no procfs dir?\n",
 427                        ldlm_ns_name(ns));
 428         else
 429                 lprocfs_remove(&ns->ns_proc_dir_entry);
 430
 431         if (ns->ns_stats != NULL)
 432                 lprocfs_free_stats(&ns->ns_stats);
 433 }
 434
 435 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 436 {
 437         struct lprocfs_vars lock_vars[2];
 438         char lock_name[MAX_STRING_SIZE + 1];
 439         struct proc_dir_entry *ns_pde;
 440
 441         LASSERT(ns != NULL);
 442         LASSERT(ns->ns_rs_hash != NULL);
 443
 444         if (ns->ns_proc_dir_entry != NULL) {
 445                 ns_pde = ns->ns_proc_dir_entry;
 446         } else {
 447                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 448                 if (ns_pde == NULL)
 449                         return -ENOMEM;
 450                 ns->ns_proc_dir_entry = ns_pde;
 451         }
 452
 453         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 454         if (ns->ns_stats == NULL)
 455                 return -ENOMEM;
 456
 457         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 458                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 459
 460         lock_name[MAX_STRING_SIZE] = '\0';
 461
 462         memset(lock_vars, 0, sizeof(lock_vars));
 463         lock_vars[0].name = lock_name;
 464
 465         ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
 466                      &lprocfs_ns_resources_fops);
 467         ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
 468                      &lprocfs_ns_locks_fops);
 469
 470         if (ns_is_client(ns)) {
 471                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
 472                              &ns->ns_nr_unused, &ldlm_uint_fops);
 473                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
 474                              &lprocfs_lru_size_fops);
 475                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
 476                              &ns->ns_max_age, &ldlm_rw_uint_fops);
 477                 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
 478                              ns, &lprocfs_elc_fops);
 479         } else {
 480                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 481                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 482                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 483                              &ns->ns_timeouts, &ldlm_uint_fops);
 484                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 485                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 486                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 487                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 488                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 489                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 490                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 491                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 492         }
 493         return 0;
 494 }
 495 #undef MAX_STRING_SIZE
 496 #else /* CONFIG_PROC_FS */
 497
 498 #define ldlm_namespace_proc_unregister(ns)      ({;})
 499 #define ldlm_namespace_proc_register(ns)        ({0;})
 500
 501 #endif /* CONFIG_PROC_FS */
 502
 503 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 504                                   const void *key, unsigned mask)
 505 {
 506         const struct ldlm_res_id     *id  = key;
 507         unsigned                val = 0;
 508         unsigned                i;
 509
 510         for (i = 0; i < RES_NAME_SIZE; i++)
 511                 val += id->name[i];
 512         return val & mask;
 513 }
 514
 515 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 516                                       const void *key, unsigned mask)
 517 {
 518         const struct ldlm_res_id *id = key;
 519         struct lu_fid       fid;
 520         __u32               hash;
 521         __u32               val;
 522
 523         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 524         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 525         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 526
 527         hash = fid_flatten32(&fid);
 528         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 529         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 530                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 531                 hash += (val >> 5) + (val << 11);
 532         } else {
 533                 val = fid_oid(&fid);
 534         }
 535         hash = hash_long(hash, hs->hs_bkt_bits);
 536         /* give me another random factor */
 537         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 538
 539         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 540         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 541
 542         return hash & mask;
 543 }
 544
 545 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 546 {
 547         struct ldlm_resource   *res;
 548
 549         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 550         return &res->lr_name;
 551 }
 552
 553 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 554 {
 555         struct ldlm_resource   *res;
 556
 557         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 558         return ldlm_res_eq((const struct ldlm_res_id *)key,
 559                            (const struct ldlm_res_id *)&res->lr_name);
 560 }
 561
 562 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 563 {
 564         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 565 }
 566
 567 static void
 568 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 569 {
 570         struct ldlm_resource *res;
 571
 572         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 573         ldlm_resource_getref(res);
 574 }
 575
 576 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 577 {
 578         struct ldlm_resource *res;
 579
 580         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 581         ldlm_resource_putref(res);
 582 }
 583
 584 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 585         .hs_hash        = ldlm_res_hop_hash,
 586         .hs_key         = ldlm_res_hop_key,
 587         .hs_keycmp      = ldlm_res_hop_keycmp,
 588         .hs_keycpy      = NULL,
 589         .hs_object      = ldlm_res_hop_object,
 590         .hs_get         = ldlm_res_hop_get_locked,
 591         .hs_put         = ldlm_res_hop_put
 592 };
 593
 594 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 595         .hs_hash        = ldlm_res_hop_fid_hash,
 596         .hs_key         = ldlm_res_hop_key,
 597         .hs_keycmp      = ldlm_res_hop_keycmp,
 598         .hs_keycpy      = NULL,
 599         .hs_object      = ldlm_res_hop_object,
 600         .hs_get         = ldlm_res_hop_get_locked,
 601         .hs_put         = ldlm_res_hop_put
 602 };
 603
 604 typedef struct ldlm_ns_hash_def {
 605         enum ldlm_ns_type       nsd_type;
 606         /** hash bucket bits */
 607         unsigned                nsd_bkt_bits;
 608         /** hash bits */
 609         unsigned                nsd_all_bits;
 610         /** hash operations */
 611         struct cfs_hash_ops *nsd_hops;
 612 } ldlm_ns_hash_def_t;
 613
 614 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 615 {
 616         {
 617                 .nsd_type       = LDLM_NS_TYPE_MDC,
 618                 .nsd_bkt_bits   = 11,
 619                 .nsd_all_bits   = 16,
 620                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 621         },
 622         {
 623                 .nsd_type       = LDLM_NS_TYPE_MDT,
 624                 .nsd_bkt_bits   = 14,
 625                 .nsd_all_bits   = 21,
 626                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 627         },
 628         {
 629                 .nsd_type       = LDLM_NS_TYPE_OSC,
 630                 .nsd_bkt_bits   = 8,
 631                 .nsd_all_bits   = 12,
 632                 .nsd_hops       = &ldlm_ns_hash_ops,
 633         },
 634         {
 635                 .nsd_type       = LDLM_NS_TYPE_OST,
 636                 .nsd_bkt_bits   = 11,
 637                 .nsd_all_bits   = 17,
 638                 .nsd_hops       = &ldlm_ns_hash_ops,
 639         },
 640         {
 641                 .nsd_type       = LDLM_NS_TYPE_MGC,
 642                 .nsd_bkt_bits   = 4,
 643                 .nsd_all_bits   = 4,
 644                 .nsd_hops       = &ldlm_ns_hash_ops,
 645         },
 646         {
 647                 .nsd_type       = LDLM_NS_TYPE_MGT,
 648                 .nsd_bkt_bits   = 4,
 649                 .nsd_all_bits   = 4,
 650                 .nsd_hops       = &ldlm_ns_hash_ops,
 651         },
 652         {
 653                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 654         },
 655 };
 656
 657 /**
 658  * Create and initialize new empty namespace.
 659  */
 660 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 661                                           enum ldlm_side client,
 662                                           enum ldlm_appetite apt,
 663                                           enum ldlm_ns_type ns_type)
 664 {
 665         struct ldlm_namespace *ns = NULL;
 666         struct ldlm_ns_bucket *nsb;
 667         struct ldlm_ns_hash_def *nsd;
 668         struct cfs_hash_bd bd;
 669         int idx;
 670         int rc;
 671         ENTRY;
 672
 673         LASSERT(obd != NULL);
 674
 675         rc = ldlm_get_ref();
 676         if (rc) {
 677                 CERROR("ldlm_get_ref failed: %d\n", rc);
 678                 RETURN(NULL);
 679         }
 680
 681         for (idx = 0;;idx++) {
 682                 nsd = &ldlm_ns_hash_defs[idx];
 683                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 684                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 685                         GOTO(out_ref, NULL);
 686                 }
 687
 688                 if (nsd->nsd_type == ns_type)
 689                         break;
 690         }
 691
 692         OBD_ALLOC_PTR(ns);
 693         if (!ns)
 694                 GOTO(out_ref, NULL);
 695
 696         ns->ns_rs_hash = cfs_hash_create(name,
 697                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 698                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 699                                          CFS_HASH_MIN_THETA,
 700                                          CFS_HASH_MAX_THETA,
 701                                          nsd->nsd_hops,
 702                                          CFS_HASH_DEPTH |
 703                                          CFS_HASH_BIGNAME |
 704                                          CFS_HASH_SPIN_BKTLOCK |
 705                                          CFS_HASH_NO_ITEMREF);
 706         if (ns->ns_rs_hash == NULL)
 707                 GOTO(out_ns, NULL);
 708
 709         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 710                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 711                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 712                 nsb->nsb_namespace = ns;
 713                 nsb->nsb_reclaim_start = 0;
 714         }
 715
 716         ns->ns_obd      = obd;
 717         ns->ns_appetite = apt;
 718         ns->ns_client   = client;
 719
 720         INIT_LIST_HEAD(&ns->ns_list_chain);
 721         INIT_LIST_HEAD(&ns->ns_unused_list);
 722         spin_lock_init(&ns->ns_lock);
 723         atomic_set(&ns->ns_bref, 0);
 724         init_waitqueue_head(&ns->ns_waitq);
 725
 726         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 727         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 728         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 729
 730         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 731         ns->ns_nr_unused          = 0;
 732         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 733         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 734         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 735         ns->ns_timeouts           = 0;
 736         ns->ns_orig_connect_flags = 0;
 737         ns->ns_connect_flags      = 0;
 738         ns->ns_stopping           = 0;
 739         ns->ns_reclaim_start      = 0;
 740         rc = ldlm_namespace_proc_register(ns);
 741         if (rc != 0) {
 742                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 743                 GOTO(out_hash, rc);
 744         }
 745
 746         idx = ldlm_namespace_nr_read(client);
 747         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 748         if (rc) {
 749                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 750                 GOTO(out_proc, rc);
 751         }
 752
 753         ldlm_namespace_register(ns, client);
 754         RETURN(ns);
 755 out_proc:
 756         ldlm_namespace_proc_unregister(ns);
 757         ldlm_namespace_cleanup(ns, 0);
 758 out_hash:
 759         cfs_hash_putref(ns->ns_rs_hash);
 760 out_ns:
 761         OBD_FREE_PTR(ns);
 762 out_ref:
 763         ldlm_put_ref();
 764         RETURN(NULL);
 765 }
 766 EXPORT_SYMBOL(ldlm_namespace_new);
 767
 768 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 769
 770 /**
 771  * Cancel and destroy all locks on a resource.
 772  *
 773  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 774  * clean up.  This is currently only used for recovery, and we make
 775  * certain assumptions as a result--notably, that we shouldn't cancel
 776  * locks with refs.
 777  */
 778 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 779                              __u64 flags)
 780 {
 781         struct list_head *tmp;
 782         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 783         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 784
 785         do {
 786                 struct ldlm_lock *lock = NULL;
 787
 788                 /* First, we look for non-cleaned-yet lock
 789                  * all cleaned locks are marked by CLEANED flag. */
 790                 lock_res(res);
 791                 list_for_each(tmp, q) {
 792                         lock = list_entry(tmp, struct ldlm_lock,
 793                                           l_res_link);
 794                         if (ldlm_is_cleaned(lock)) {
 795                                 lock = NULL;
 796                                 continue;
 797                         }
 798                         LDLM_LOCK_GET(lock);
 799                         ldlm_set_cleaned(lock);
 800                         break;
 801                 }
 802
 803                 if (lock == NULL) {
 804                         unlock_res(res);
 805                         break;
 806                 }
 807
 808                 /* Set CBPENDING so nothing in the cancellation path
 809                  * can match this lock. */
 810                 ldlm_set_cbpending(lock);
 811                 ldlm_set_failed(lock);
 812                 lock->l_flags |= flags;
 813
 814                 /* ... without sending a CANCEL message for local_only. */
 815                 if (local_only)
 816                         ldlm_set_local_only(lock);
 817
 818                 if (local_only && (lock->l_readers || lock->l_writers)) {
 819                         /* This is a little bit gross, but much better than the
 820                          * alternative: pretend that we got a blocking AST from
 821                          * the server, so that when the lock is decref'd, it
 822                          * will go away ... */
 823                         unlock_res(res);
 824                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 825                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 826                                 set_current_state(TASK_UNINTERRUPTIBLE);
 827                                 schedule_timeout(cfs_time_seconds(4));
 828                                 set_current_state(TASK_RUNNING);
 829                         }
 830                         if (lock->l_completion_ast)
 831                                 lock->l_completion_ast(lock,
 832                                                        LDLM_FL_FAILED, NULL);
 833                         LDLM_LOCK_RELEASE(lock);
 834                         continue;
 835                 }
 836
 837                 if (client) {
 838                         struct lustre_handle lockh;
 839
 840                         unlock_res(res);
 841                         ldlm_lock2handle(lock, &lockh);
 842                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
 843                         if (rc)
 844                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 845                 } else {
 846                         unlock_res(res);
 847                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 848                                    "client node");
 849                         ldlm_lock_cancel(lock);
 850                 }
 851                 LDLM_LOCK_RELEASE(lock);
 852         } while (1);
 853 }
 854
 855 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 856                                struct hlist_node *hnode, void *arg)
 857 {
 858         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 859         __u64 flags = *(__u64 *)arg;
 860
 861         cleanup_resource(res, &res->lr_granted, flags);
 862         cleanup_resource(res, &res->lr_converting, flags);
 863         cleanup_resource(res, &res->lr_waiting, flags);
 864
 865         return 0;
 866 }
 867
 868 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 869                                   struct hlist_node *hnode, void *arg)
 870 {
 871         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 872
 873         lock_res(res);
 874         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 875                "(%d) after lock cleanup; forcing cleanup.\n",
 876                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 877                atomic_read(&res->lr_refcount) - 1);
 878
 879         ldlm_resource_dump(D_ERROR, res);
 880         unlock_res(res);
 881         return 0;
 882 }
 883
 884 /**
 885  * Cancel and destroy all locks in the namespace.
 886  *
 887  * Typically used during evictions when server notified client that it was
 888  * evicted and all of its state needs to be destroyed.
 889  * Also used during shutdown.
 890  */
 891 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 892 {
 893         if (ns == NULL) {
 894                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 895                 return ELDLM_OK;
 896         }
 897
 898         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 899                                  &flags, 0);
 900         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
 901                                  NULL, 0);
 902         return ELDLM_OK;
 903 }
 904 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 905
 906 /**
 907  * Attempts to free namespace.
 908  *
 909  * Only used when namespace goes away, like during an unmount.
 910  */
 911 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 912 {
 913         ENTRY;
 914
 915         /* At shutdown time, don't call the cancellation callback */
 916         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 917
 918         if (atomic_read(&ns->ns_bref) > 0) {
 919                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 920                 int rc;
 921                 CDEBUG(D_DLMTRACE,
 922                        "dlm namespace %s free waiting on refcount %d\n",
 923                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
 924 force_wait:
 925                 if (force)
 926                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
 927                                           MSEC_PER_SEC) / 4, NULL, NULL);
 928
 929                 rc = l_wait_event(ns->ns_waitq,
 930                                   atomic_read(&ns->ns_bref) == 0, &lwi);
 931
 932                 /* Forced cleanups should be able to reclaim all references,
 933                  * so it's safe to wait forever... we can't leak locks... */
 934                 if (force && rc == -ETIMEDOUT) {
 935                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 936                                        "namespace with %d resources in use, "
 937                                        "(rc=%d)\n", ldlm_ns_name(ns),
 938                                        atomic_read(&ns->ns_bref), rc);
 939                         GOTO(force_wait, rc);
 940                 }
 941
 942                 if (atomic_read(&ns->ns_bref)) {
 943                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 944                                        "with %d resources in use, (rc=%d)\n",
 945                                        ldlm_ns_name(ns),
 946                                        atomic_read(&ns->ns_bref), rc);
 947                         RETURN(ELDLM_NAMESPACE_EXISTS);
 948                 }
 949                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 950                        ldlm_ns_name(ns));
 951         }
 952
 953         RETURN(ELDLM_OK);
 954 }
 955
 956 /**
 957  * Performs various cleanups for passed \a ns to make it drop refc and be
 958  * ready for freeing. Waits for refc == 0.
 959  *
 960  * The following is done:
 961  * (0) Unregister \a ns from its list to make inaccessible for potential
 962  * users like pools thread and others;
 963  * (1) Clear all locks in \a ns.
 964  */
 965 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 966                                struct obd_import *imp,
 967                                int force)
 968 {
 969         int rc;
 970         ENTRY;
 971         if (!ns) {
 972                 EXIT;
 973                 return;
 974         }
 975
 976         spin_lock(&ns->ns_lock);
 977         ns->ns_stopping = 1;
 978         spin_unlock(&ns->ns_lock);
 979
 980         /*
 981          * Can fail with -EINTR when force == 0 in which case try harder.
 982          */
 983         rc = __ldlm_namespace_free(ns, force);
 984         if (rc != ELDLM_OK) {
 985                 if (imp) {
 986                         ptlrpc_disconnect_import(imp, 0);
 987                         ptlrpc_invalidate_import(imp);
 988                 }
 989
 990                 /*
 991                  * With all requests dropped and the import inactive
 992                  * we are gaurenteed all reference will be dropped.
 993                  */
 994                 rc = __ldlm_namespace_free(ns, 1);
 995                 LASSERT(rc == 0);
 996         }
 997         EXIT;
 998 }
 999 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1000
1001 /**
1002  * Performs freeing memory structures related to \a ns. This is only done
1003  * when ldlm_namespce_free_prior() successfully removed all resources
1004  * referencing \a ns and its refc == 0.
1005  */
1006 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1007 {
1008         ENTRY;
1009         if (!ns) {
1010                 EXIT;
1011                 return;
1012         }
1013
1014         /* Make sure that nobody can find this ns in its list. */
1015         ldlm_namespace_unregister(ns, ns->ns_client);
1016         /* Fini pool _before_ parent proc dir is removed. This is important as
1017          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1018          * Removing it after @dir may cause oops. */
1019         ldlm_pool_fini(&ns->ns_pool);
1020
1021         ldlm_namespace_proc_unregister(ns);
1022         cfs_hash_putref(ns->ns_rs_hash);
1023         /* Namespace \a ns should be not on list at this time, otherwise
1024          * this will cause issues related to using freed \a ns in poold
1025          * thread. */
1026         LASSERT(list_empty(&ns->ns_list_chain));
1027         OBD_FREE_PTR(ns);
1028         ldlm_put_ref();
1029         EXIT;
1030 }
1031 EXPORT_SYMBOL(ldlm_namespace_free_post);
1032
1033 /**
1034  * Cleanup the resource, and free namespace.
1035  * bug 12864:
1036  * Deadlock issue:
1037  * proc1: destroy import
1038  *        class_disconnect_export(grab cl_sem) ->
1039  *              -> ldlm_namespace_free ->
1040  *              -> lprocfs_remove(grab _lprocfs_lock).
1041  * proc2: read proc info
1042  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1043  *              -> osc_rd_active, etc(grab cl_sem).
1044  *
1045  * So that I have to split the ldlm_namespace_free into two parts - the first
1046  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1047  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1048  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1049  * held.
1050  */
1051 void ldlm_namespace_free(struct ldlm_namespace *ns,
1052                          struct obd_import *imp,
1053                          int force)
1054 {
1055         ldlm_namespace_free_prior(ns, imp, force);
1056         ldlm_namespace_free_post(ns);
1057 }
1058 EXPORT_SYMBOL(ldlm_namespace_free);
1059
1060 void ldlm_namespace_get(struct ldlm_namespace *ns)
1061 {
1062         atomic_inc(&ns->ns_bref);
1063 }
1064
1065 /* This is only for callers that care about refcount */
1066 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1067 {
1068         return atomic_inc_return(&ns->ns_bref);
1069 }
1070
1071 void ldlm_namespace_put(struct ldlm_namespace *ns)
1072 {
1073         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1074                 wake_up(&ns->ns_waitq);
1075                 spin_unlock(&ns->ns_lock);
1076         }
1077 }
1078
1079 /** Register \a ns in the list of namespaces */
1080 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1081 {
1082         mutex_lock(ldlm_namespace_lock(client));
1083         LASSERT(list_empty(&ns->ns_list_chain));
1084         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1085         ldlm_namespace_nr_inc(client);
1086         mutex_unlock(ldlm_namespace_lock(client));
1087 }
1088
1089 /** Unregister \a ns from the list of namespaces. */
1090 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1091 {
1092         mutex_lock(ldlm_namespace_lock(client));
1093         LASSERT(!list_empty(&ns->ns_list_chain));
1094         /* Some asserts and possibly other parts of the code are still
1095          * using list_empty(&ns->ns_list_chain). This is why it is
1096          * important to use list_del_init() here. */
1097         list_del_init(&ns->ns_list_chain);
1098         ldlm_namespace_nr_dec(client);
1099         mutex_unlock(ldlm_namespace_lock(client));
1100 }
1101
1102 /** Should be called with ldlm_namespace_lock(client) taken. */
1103 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1104                                           enum ldlm_side client)
1105 {
1106         LASSERT(!list_empty(&ns->ns_list_chain));
1107         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1108         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1109 }
1110
1111 /** Should be called with ldlm_namespace_lock(client) taken. */
1112 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1113                                             enum ldlm_side client)
1114 {
1115         LASSERT(!list_empty(&ns->ns_list_chain));
1116         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1117         list_move_tail(&ns->ns_list_chain,
1118                        ldlm_namespace_inactive_list(client));
1119 }
1120
1121 /** Should be called with ldlm_namespace_lock(client) taken. */
1122 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1123 {
1124         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1125         LASSERT(!list_empty(ldlm_namespace_list(client)));
1126         return container_of(ldlm_namespace_list(client)->next,
1127                             struct ldlm_namespace, ns_list_chain);
1128 }
1129
1130 /** Create and initialize new resource. */
1131 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1132 {
1133         struct ldlm_resource *res;
1134         int idx;
1135
1136         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1137         if (res == NULL)
1138                 return NULL;
1139
1140         if (ldlm_type == LDLM_EXTENT) {
1141                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1142                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1143                 if (res->lr_itree == NULL) {
1144                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1145                         return NULL;
1146                 }
1147                 /* Initialize interval trees for each lock mode. */
1148                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1149                         res->lr_itree[idx].lit_size = 0;
1150                         res->lr_itree[idx].lit_mode = 1 << idx;
1151                         res->lr_itree[idx].lit_root = NULL;
1152                 }
1153         }
1154
1155         INIT_LIST_HEAD(&res->lr_granted);
1156         INIT_LIST_HEAD(&res->lr_converting);
1157         INIT_LIST_HEAD(&res->lr_waiting);
1158
1159         atomic_set(&res->lr_refcount, 1);
1160         spin_lock_init(&res->lr_lock);
1161         lu_ref_init(&res->lr_reference);
1162
1163         /* Since LVB init can be delayed now, there is no longer need to
1164          * immediatelly acquire mutex here. */
1165         mutex_init(&res->lr_lvb_mutex);
1166         res->lr_lvb_initialized = false;
1167
1168         return res;
1169 }
1170
1171 /**
1172  * Return a reference to resource with given name, creating it if necessary.
1173  * Args: namespace with ns_lock unlocked
1174  * Locks: takes and releases NS hash-lock and res->lr_lock
1175  * Returns: referenced, unlocked ldlm_resource or NULL
1176  */
1177 struct ldlm_resource *
1178 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1179                   const struct ldlm_res_id *name, enum ldlm_type type,
1180                   int create)
1181 {
1182         struct hlist_node       *hnode;
1183         struct ldlm_resource    *res = NULL;
1184         struct cfs_hash_bd              bd;
1185         __u64                   version;
1186         int                     ns_refcount = 0;
1187
1188         LASSERT(ns != NULL);
1189         LASSERT(parent == NULL);
1190         LASSERT(ns->ns_rs_hash != NULL);
1191         LASSERT(name->name[0] != 0);
1192
1193         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1194         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1195         if (hnode != NULL) {
1196                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1197                 GOTO(found, res);
1198         }
1199
1200         version = cfs_hash_bd_version_get(&bd);
1201         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1202
1203         if (create == 0)
1204                 return ERR_PTR(-ENOENT);
1205
1206         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1207                  "type: %d\n", type);
1208         res = ldlm_resource_new(type);
1209         if (res == NULL)
1210                 return ERR_PTR(-ENOMEM);
1211
1212         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1213         res->lr_name = *name;
1214         res->lr_type = type;
1215
1216         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1217         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1218                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1219
1220         if (hnode != NULL) {
1221                 /* Someone won the race and already added the resource. */
1222                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1223                 /* Clean lu_ref for failed resource. */
1224                 lu_ref_fini(&res->lr_reference);
1225                 if (res->lr_itree != NULL)
1226                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1227                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1228                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1229 found:
1230                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1231                 return res;
1232         }
1233         /* We won! Let's add the resource. */
1234         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1235         if (cfs_hash_bd_count_get(&bd) == 1)
1236                 ns_refcount = ldlm_namespace_get_return(ns);
1237
1238         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1239
1240         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1241
1242         /* Let's see if we happened to be the very first resource in this
1243          * namespace. If so, and this is a client namespace, we need to move
1244          * the namespace into the active namespaces list to be patrolled by
1245          * the ldlm_poold. */
1246         if (ns_is_client(ns) && ns_refcount == 1) {
1247                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1248                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1249                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1250         }
1251
1252         return res;
1253 }
1254 EXPORT_SYMBOL(ldlm_resource_get);
1255
1256 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1257 {
1258         LASSERT(res != NULL);
1259         LASSERT(res != LP_POISON);
1260         atomic_inc(&res->lr_refcount);
1261         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1262                atomic_read(&res->lr_refcount));
1263         return res;
1264 }
1265
1266 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1267                                          struct ldlm_resource *res)
1268 {
1269         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1270
1271         if (!list_empty(&res->lr_granted)) {
1272                 ldlm_resource_dump(D_ERROR, res);
1273                 LBUG();
1274         }
1275
1276         if (!list_empty(&res->lr_converting)) {
1277                 ldlm_resource_dump(D_ERROR, res);
1278                 LBUG();
1279         }
1280
1281         if (!list_empty(&res->lr_waiting)) {
1282                 ldlm_resource_dump(D_ERROR, res);
1283                 LBUG();
1284         }
1285
1286         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1287                                bd, &res->lr_hash);
1288         lu_ref_fini(&res->lr_reference);
1289         if (cfs_hash_bd_count_get(bd) == 0)
1290                 ldlm_namespace_put(nsb->nsb_namespace);
1291 }
1292
1293 /* Returns 1 if the resource was freed, 0 if it remains. */
1294 int ldlm_resource_putref(struct ldlm_resource *res)
1295 {
1296         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1297         struct cfs_hash_bd   bd;
1298
1299         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1300         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1301                res, atomic_read(&res->lr_refcount) - 1);
1302
1303         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1304         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1305                 __ldlm_resource_putref_final(&bd, res);
1306                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1307                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1308                         ns->ns_lvbo->lvbo_free(res);
1309                 if (res->lr_itree != NULL)
1310                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1311                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1312                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1313                 return 1;
1314         }
1315         return 0;
1316 }
1317 EXPORT_SYMBOL(ldlm_resource_putref);
1318
1319 /**
1320  * Add a lock into a given resource into specified lock list.
1321  */
1322 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1323                             struct ldlm_lock *lock)
1324 {
1325         check_res_locked(res);
1326
1327         LDLM_DEBUG(lock, "About to add this lock");
1328
1329         if (ldlm_is_destroyed(lock)) {
1330                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1331                 return;
1332         }
1333
1334         LASSERT(list_empty(&lock->l_res_link));
1335
1336         list_add_tail(&lock->l_res_link, head);
1337 }
1338
1339 /**
1340  * Insert a lock into resource after specified lock.
1341  *
1342  * Obtain resource description from the lock we are inserting after.
1343  */
1344 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1345                                      struct ldlm_lock *new)
1346 {
1347         struct ldlm_resource *res = original->l_resource;
1348
1349         check_res_locked(res);
1350
1351         ldlm_resource_dump(D_INFO, res);
1352         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1353
1354         if (ldlm_is_destroyed(new)) {
1355                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1356                 goto out;
1357         }
1358
1359         LASSERT(list_empty(&new->l_res_link));
1360
1361         list_add(&new->l_res_link, &original->l_res_link);
1362  out:;
1363 }
1364
1365 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1366 {
1367         int type = lock->l_resource->lr_type;
1368
1369         check_res_locked(lock->l_resource);
1370         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1371                 ldlm_unlink_lock_skiplist(lock);
1372         else if (type == LDLM_EXTENT)
1373                 ldlm_extent_unlink_lock(lock);
1374         list_del_init(&lock->l_res_link);
1375 }
1376 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1377
1378 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1379 {
1380         desc->lr_type = res->lr_type;
1381         desc->lr_name = res->lr_name;
1382 }
1383
1384 /**
1385  * Print information about all locks in all namespaces on this node to debug
1386  * log.
1387  */
1388 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1389 {
1390         struct list_head *tmp;
1391
1392         if (!((libcfs_debug | D_ERROR) & level))
1393                 return;
1394
1395         mutex_lock(ldlm_namespace_lock(client));
1396
1397         list_for_each(tmp, ldlm_namespace_list(client)) {
1398                 struct ldlm_namespace *ns;
1399
1400                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1401                 ldlm_namespace_dump(level, ns);
1402         }
1403
1404         mutex_unlock(ldlm_namespace_lock(client));
1405 }
1406
1407 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1408                               struct hlist_node *hnode, void *arg)
1409 {
1410         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1411         int    level = (int)(unsigned long)arg;
1412
1413         lock_res(res);
1414         ldlm_resource_dump(level, res);
1415         unlock_res(res);
1416
1417         return 0;
1418 }
1419
1420 /**
1421  * Print information about all locks in this namespace on this node to debug
1422  * log.
1423  */
1424 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1425 {
1426         if (!((libcfs_debug | D_ERROR) & level))
1427                 return;
1428
1429         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1430                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1431                ns_is_client(ns) ? "client" : "server");
1432
1433         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1434                 return;
1435
1436         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1437                                  ldlm_res_hash_dump,
1438                                  (void *)(unsigned long)level, 0);
1439         spin_lock(&ns->ns_lock);
1440         ns->ns_next_dump = cfs_time_shift(10);
1441         spin_unlock(&ns->ns_lock);
1442 }
1443
1444 /**
1445  * Print information about all locks in this resource to debug log.
1446  */
1447 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1448 {
1449         struct ldlm_lock *lock;
1450         unsigned int granted = 0;
1451
1452         CLASSERT(RES_NAME_SIZE == 4);
1453
1454         if (!((libcfs_debug | D_ERROR) & level))
1455                 return;
1456
1457         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1458                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1459
1460         if (!list_empty(&res->lr_granted)) {
1461                 CDEBUG(level, "Granted locks (in reverse order):\n");
1462                 list_for_each_entry_reverse(lock, &res->lr_granted,
1463                                                 l_res_link) {
1464                         LDLM_DEBUG_LIMIT(level, lock, "###");
1465                         if (!(level & D_CANTMASK) &&
1466                             ++granted > ldlm_dump_granted_max) {
1467                                 CDEBUG(level, "only dump %d granted locks to "
1468                                        "avoid DDOS.\n", granted);
1469                                 break;
1470                         }
1471                 }
1472         }
1473         if (!list_empty(&res->lr_converting)) {
1474                 CDEBUG(level, "Converting locks:\n");
1475                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1476                         LDLM_DEBUG_LIMIT(level, lock, "###");
1477         }
1478         if (!list_empty(&res->lr_waiting)) {
1479                 CDEBUG(level, "Waiting locks:\n");
1480                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1481                         LDLM_DEBUG_LIMIT(level, lock, "###");
1482         }
1483 }
1484 EXPORT_SYMBOL(ldlm_resource_dump);