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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46
  47 int ldlm_srv_namespace_nr = 0;
  48 int ldlm_cli_namespace_nr = 0;
  49
  50 struct mutex ldlm_srv_namespace_lock;
  51 struct list_head ldlm_srv_namespace_list;
  52
  53 struct mutex ldlm_cli_namespace_lock;
  54 /* Client Namespaces that have active resources in them.
  55  * Once all resources go away, ldlm_poold moves such namespaces to the
  56  * inactive list */
  57 struct list_head ldlm_cli_active_namespace_list;
  58 /* Client namespaces that don't have any locks in them */
  59 struct list_head ldlm_cli_inactive_namespace_list;
  60
  61 static struct proc_dir_entry *ldlm_type_proc_dir;
  62 static struct proc_dir_entry *ldlm_ns_proc_dir;
  63 struct proc_dir_entry *ldlm_svc_proc_dir;
  64
  65 /* during debug dump certain amount of granted locks for one resource to avoid
  66  * DDOS. */
  67 static unsigned int ldlm_dump_granted_max = 256;
  68
  69 #ifdef CONFIG_PROC_FS
  70 static ssize_t
  71 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
  72                           size_t count, loff_t *off)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
  79
  80 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
  81 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
  82
  83 #ifdef HAVE_SERVER_SUPPORT
  84
  85 static int seq_watermark_show(struct seq_file *m, void *data)
  86 {
  87         seq_printf(m, "%llu\n", *(__u64 *)m->private);
  88         return 0;
  89 }
  90
  91 static ssize_t seq_watermark_write(struct file *file,
  92                                    const char __user *buffer, size_t count,
  93                                    loff_t *off)
  94 {
  95         __s64 value;
  96         __u64 watermark;
  97         __u64 *data = ((struct seq_file *)file->private_data)->private;
  98         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
  99         int rc;
 100
 101         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 102         if (rc) {
 103                 CERROR("Failed to set %s, rc = %d.\n",
 104                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 105                        rc);
 106                 return rc;
 107         } else if (value != 0 && value < (1 << 20)) {
 108                 CERROR("%s should be greater than 1MB.\n",
 109                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 110                 return -EINVAL;
 111         }
 112         watermark = value >> 20;
 113
 114         if (wm_low) {
 115                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 116                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 117                                "lock_limit_mb.\n");
 118                         return -EINVAL;
 119                 }
 120
 121                 *data = watermark;
 122                 if (watermark != 0) {
 123                         watermark <<= 20;
 124                         do_div(watermark, sizeof(struct ldlm_lock));
 125                 }
 126                 ldlm_reclaim_threshold = watermark;
 127         } else {
 128                 if (ldlm_reclaim_threshold_mb != 0 &&
 129                     watermark < ldlm_reclaim_threshold_mb) {
 130                         CERROR("lock_limit_mb must be greater than "
 131                                "lock_reclaim_threshold_mb.\n");
 132                         return -EINVAL;
 133                 }
 134
 135                 *data = watermark;
 136                 if (watermark != 0) {
 137                         watermark <<= 20;
 138                         do_div(watermark, sizeof(struct ldlm_lock));
 139                 }
 140                 ldlm_lock_limit = watermark;
 141         }
 142
 143         return count;
 144 }
 145
 146 static int seq_watermark_open(struct inode *inode, struct file *file)
 147 {
 148         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 149 }
 150
 151 static const struct file_operations ldlm_watermark_fops = {
 152         .owner          = THIS_MODULE,
 153         .open           = seq_watermark_open,
 154         .read           = seq_read,
 155         .write          = seq_watermark_write,
 156         .llseek         = seq_lseek,
 157         .release        = lprocfs_single_release,
 158 };
 159
 160 static int seq_granted_show(struct seq_file *m, void *data)
 161 {
 162         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 163                    (struct percpu_counter *)m->private));
 164         return 0;
 165 }
 166
 167 static int seq_granted_open(struct inode *inode, struct file *file)
 168 {
 169         return single_open(file, seq_granted_show, PDE_DATA(inode));
 170 }
 171
 172 static const struct file_operations ldlm_granted_fops = {
 173         .owner  = THIS_MODULE,
 174         .open   = seq_granted_open,
 175         .read   = seq_read,
 176         .llseek = seq_lseek,
 177         .release = seq_release,
 178 };
 179
 180 #endif /* HAVE_SERVER_SUPPORT */
 181
 182 int ldlm_proc_setup(void)
 183 {
 184         int rc;
 185         struct lprocfs_vars list[] = {
 186                 { .name =       "dump_namespaces",
 187                   .fops =       &ldlm_dump_ns_fops,
 188                   .proc_mode =  0222 },
 189                 { .name =       "dump_granted_max",
 190                   .fops =       &ldlm_rw_uint_fops,
 191                   .data =       &ldlm_dump_granted_max },
 192                 { .name =       "cancel_unused_locks_before_replay",
 193                   .fops =       &ldlm_rw_uint_fops,
 194                   .data =       &ldlm_cancel_unused_locks_before_replay },
 195 #ifdef HAVE_SERVER_SUPPORT
 196                 { .name =       "lock_reclaim_threshold_mb",
 197                   .fops =       &ldlm_watermark_fops,
 198                   .data =       &ldlm_reclaim_threshold_mb },
 199                 { .name =       "lock_limit_mb",
 200                   .fops =       &ldlm_watermark_fops,
 201                   .data =       &ldlm_lock_limit_mb },
 202                 { .name =       "lock_granted_count",
 203                   .fops =       &ldlm_granted_fops,
 204                   .data =       &ldlm_granted_total },
 205 #endif
 206                 { NULL }};
 207         ENTRY;
 208         LASSERT(ldlm_ns_proc_dir == NULL);
 209
 210         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 211                                               proc_lustre_root,
 212                                               NULL, NULL);
 213         if (IS_ERR(ldlm_type_proc_dir)) {
 214                 CERROR("LProcFS failed in ldlm-init\n");
 215                 rc = PTR_ERR(ldlm_type_proc_dir);
 216                 GOTO(err, rc);
 217         }
 218
 219         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 220                                             ldlm_type_proc_dir,
 221                                             NULL, NULL);
 222         if (IS_ERR(ldlm_ns_proc_dir)) {
 223                 CERROR("LProcFS failed in ldlm-init\n");
 224                 rc = PTR_ERR(ldlm_ns_proc_dir);
 225                 GOTO(err_type, rc);
 226         }
 227
 228         ldlm_svc_proc_dir = lprocfs_register("services",
 229                                              ldlm_type_proc_dir,
 230                                              NULL, NULL);
 231         if (IS_ERR(ldlm_svc_proc_dir)) {
 232                 CERROR("LProcFS failed in ldlm-init\n");
 233                 rc = PTR_ERR(ldlm_svc_proc_dir);
 234                 GOTO(err_ns, rc);
 235         }
 236
 237         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 238         if (rc != 0) {
 239                 CERROR("LProcFS failed in ldlm-init\n");
 240                 GOTO(err_svc, rc);
 241         }
 242
 243         RETURN(0);
 244
 245 err_svc:
 246         lprocfs_remove(&ldlm_svc_proc_dir);
 247 err_ns:
 248         lprocfs_remove(&ldlm_ns_proc_dir);
 249 err_type:
 250         lprocfs_remove(&ldlm_type_proc_dir);
 251 err:
 252         ldlm_svc_proc_dir = NULL;
 253         RETURN(rc);
 254 }
 255
 256 void ldlm_proc_cleanup(void)
 257 {
 258         if (ldlm_svc_proc_dir)
 259                 lprocfs_remove(&ldlm_svc_proc_dir);
 260
 261         if (ldlm_ns_proc_dir)
 262                 lprocfs_remove(&ldlm_ns_proc_dir);
 263
 264         if (ldlm_type_proc_dir)
 265                 lprocfs_remove(&ldlm_type_proc_dir);
 266 }
 267
 268 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
 269 {
 270         struct ldlm_namespace   *ns  = m->private;
 271         __u64                   res = 0;
 272         struct cfs_hash_bd              bd;
 273         int                     i;
 274
 275         /* result is not strictly consistant */
 276         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 277                 res += cfs_hash_bd_count_get(&bd);
 278         return lprocfs_u64_seq_show(m, &res);
 279 }
 280 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
 281
 282 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
 283 {
 284         struct ldlm_namespace   *ns = m->private;
 285         __u64                   locks;
 286
 287         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 288                                         LPROCFS_FIELDS_FLAGS_SUM);
 289         return lprocfs_u64_seq_show(m, &locks);
 290 }
 291 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
 292
 293 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
 294 {
 295         struct ldlm_namespace *ns = m->private;
 296         __u32 *nr = &ns->ns_max_unused;
 297
 298         if (ns_connect_lru_resize(ns))
 299                 nr = &ns->ns_nr_unused;
 300         return lprocfs_uint_seq_show(m, nr);
 301 }
 302
 303 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
 304                                           const char __user *buffer,
 305                                           size_t count, loff_t *off)
 306 {
 307         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 308         char dummy[MAX_STRING_SIZE + 1], *end;
 309         unsigned long tmp;
 310         int lru_resize;
 311
 312         dummy[MAX_STRING_SIZE] = '\0';
 313         if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 314                 return -EFAULT;
 315
 316         if (strncmp(dummy, "clear", 5) == 0) {
 317                 CDEBUG(D_DLMTRACE,
 318                        "dropping all unused locks from namespace %s\n",
 319                        ldlm_ns_name(ns));
 320                 if (ns_connect_lru_resize(ns)) {
 321                         int canceled, unused  = ns->ns_nr_unused;
 322
 323                         /* Try to cancel all @ns_nr_unused locks. */
 324                         canceled = ldlm_cancel_lru(ns, unused, 0,
 325                                                    LDLM_LRU_FLAG_PASSED);
 326                         if (canceled < unused) {
 327                                 CDEBUG(D_DLMTRACE,
 328                                        "not all requested locks are canceled, "
 329                                        "requested: %d, canceled: %d\n", unused,
 330                                        canceled);
 331                                 return -EINVAL;
 332                         }
 333                 } else {
 334                         tmp = ns->ns_max_unused;
 335                         ns->ns_max_unused = 0;
 336                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED);
 337                         ns->ns_max_unused = tmp;
 338                 }
 339                 return count;
 340         }
 341
 342         tmp = simple_strtoul(dummy, &end, 0);
 343         if (dummy == end) {
 344                 CERROR("invalid value written\n");
 345                 return -EINVAL;
 346         }
 347         lru_resize = (tmp == 0);
 348
 349         if (ns_connect_lru_resize(ns)) {
 350                 if (!lru_resize)
 351                         ns->ns_max_unused = tmp;
 352
 353                 if (tmp > ns->ns_nr_unused)
 354                         tmp = ns->ns_nr_unused;
 355                 tmp = ns->ns_nr_unused - tmp;
 356
 357                 CDEBUG(D_DLMTRACE,
 358                        "changing namespace %s unused locks from %u to %u\n",
 359                        ldlm_ns_name(ns), ns->ns_nr_unused,
 360                        (unsigned int)tmp);
 361                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 362
 363                 if (!lru_resize) {
 364                         CDEBUG(D_DLMTRACE,
 365                                "disable lru_resize for namespace %s\n",
 366                                ldlm_ns_name(ns));
 367                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 368                 }
 369         } else {
 370                 CDEBUG(D_DLMTRACE,
 371                        "changing namespace %s max_unused from %u to %u\n",
 372                        ldlm_ns_name(ns), ns->ns_max_unused,
 373                        (unsigned int)tmp);
 374                 ns->ns_max_unused = (unsigned int)tmp;
 375                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 376
 377                 /* Make sure that LRU resize was originally supported before
 378                  * turning it on here. */
 379                 if (lru_resize &&
 380                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 381                         CDEBUG(D_DLMTRACE,
 382                                "enable lru_resize for namespace %s\n",
 383                                ldlm_ns_name(ns));
 384                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 385                 }
 386         }
 387
 388         return count;
 389 }
 390 LPROC_SEQ_FOPS(lprocfs_lru_size);
 391
 392 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
 393 {
 394         struct ldlm_namespace *ns = m->private;
 395         unsigned int supp = ns_connect_cancelset(ns);
 396
 397         return lprocfs_uint_seq_show(m, &supp);
 398 }
 399
 400 static ssize_t lprocfs_elc_seq_write(struct file *file,
 401                                      const char __user *buffer,
 402                                      size_t count, loff_t *off)
 403 {
 404         struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
 405         unsigned int supp = -1;
 406         int rc;
 407
 408         rc = lprocfs_wr_uint(file, buffer, count, &supp);
 409         if (rc < 0)
 410                 return rc;
 411
 412         if (supp == 0)
 413                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 414         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 415                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 416         return count;
 417 }
 418 LPROC_SEQ_FOPS(lprocfs_elc);
 419
 420 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 421 {
 422         if (ns->ns_proc_dir_entry == NULL)
 423                 CERROR("dlm namespace %s has no procfs dir?\n",
 424                        ldlm_ns_name(ns));
 425         else
 426                 lprocfs_remove(&ns->ns_proc_dir_entry);
 427
 428         if (ns->ns_stats != NULL)
 429                 lprocfs_free_stats(&ns->ns_stats);
 430 }
 431
 432 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 433 {
 434         struct lprocfs_vars lock_vars[2];
 435         char lock_name[MAX_STRING_SIZE + 1];
 436         struct proc_dir_entry *ns_pde;
 437
 438         LASSERT(ns != NULL);
 439         LASSERT(ns->ns_rs_hash != NULL);
 440
 441         if (ns->ns_proc_dir_entry != NULL) {
 442                 ns_pde = ns->ns_proc_dir_entry;
 443         } else {
 444                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 445                 if (ns_pde == NULL)
 446                         return -ENOMEM;
 447                 ns->ns_proc_dir_entry = ns_pde;
 448         }
 449
 450         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 451         if (ns->ns_stats == NULL)
 452                 return -ENOMEM;
 453
 454         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 455                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 456
 457         lock_name[MAX_STRING_SIZE] = '\0';
 458
 459         memset(lock_vars, 0, sizeof(lock_vars));
 460         lock_vars[0].name = lock_name;
 461
 462         ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
 463                      &lprocfs_ns_resources_fops);
 464         ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
 465                      &lprocfs_ns_locks_fops);
 466
 467         if (ns_is_client(ns)) {
 468                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
 469                              &ns->ns_nr_unused, &ldlm_uint_fops);
 470                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
 471                              &lprocfs_lru_size_fops);
 472                 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
 473                              &ns->ns_max_age, &ldlm_rw_uint_fops);
 474                 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
 475                              ns, &lprocfs_elc_fops);
 476         } else {
 477                 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
 478                              &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
 479                 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
 480                              &ns->ns_timeouts, &ldlm_uint_fops);
 481                 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
 482                              &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
 483                 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
 484                              &ns->ns_contention_time, &ldlm_rw_uint_fops);
 485                 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
 486                              &ns->ns_contended_locks, &ldlm_rw_uint_fops);
 487                 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
 488                              &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
 489         }
 490         return 0;
 491 }
 492 #undef MAX_STRING_SIZE
 493 #else /* CONFIG_PROC_FS */
 494
 495 #define ldlm_namespace_proc_unregister(ns)      ({;})
 496 #define ldlm_namespace_proc_register(ns)        ({0;})
 497
 498 #endif /* CONFIG_PROC_FS */
 499
 500 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 501                                   const void *key, unsigned mask)
 502 {
 503         const struct ldlm_res_id     *id  = key;
 504         unsigned                val = 0;
 505         unsigned                i;
 506
 507         for (i = 0; i < RES_NAME_SIZE; i++)
 508                 val += id->name[i];
 509         return val & mask;
 510 }
 511
 512 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 513                                       const void *key, unsigned mask)
 514 {
 515         const struct ldlm_res_id *id = key;
 516         struct lu_fid       fid;
 517         __u32               hash;
 518         __u32               val;
 519
 520         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 521         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 522         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 523
 524         hash = fid_flatten32(&fid);
 525         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 526         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 527                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 528                 hash += (val >> 5) + (val << 11);
 529         } else {
 530                 val = fid_oid(&fid);
 531         }
 532         hash = hash_long(hash, hs->hs_bkt_bits);
 533         /* give me another random factor */
 534         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 535
 536         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 537         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 538
 539         return hash & mask;
 540 }
 541
 542 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 543 {
 544         struct ldlm_resource   *res;
 545
 546         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 547         return &res->lr_name;
 548 }
 549
 550 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 551 {
 552         struct ldlm_resource   *res;
 553
 554         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 555         return ldlm_res_eq((const struct ldlm_res_id *)key,
 556                            (const struct ldlm_res_id *)&res->lr_name);
 557 }
 558
 559 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 560 {
 561         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 562 }
 563
 564 static void
 565 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 566 {
 567         struct ldlm_resource *res;
 568
 569         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 570         ldlm_resource_getref(res);
 571 }
 572
 573 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 574 {
 575         struct ldlm_resource *res;
 576
 577         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 578         ldlm_resource_putref(res);
 579 }
 580
 581 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 582         .hs_hash        = ldlm_res_hop_hash,
 583         .hs_key         = ldlm_res_hop_key,
 584         .hs_keycmp      = ldlm_res_hop_keycmp,
 585         .hs_keycpy      = NULL,
 586         .hs_object      = ldlm_res_hop_object,
 587         .hs_get         = ldlm_res_hop_get_locked,
 588         .hs_put         = ldlm_res_hop_put
 589 };
 590
 591 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 592         .hs_hash        = ldlm_res_hop_fid_hash,
 593         .hs_key         = ldlm_res_hop_key,
 594         .hs_keycmp      = ldlm_res_hop_keycmp,
 595         .hs_keycpy      = NULL,
 596         .hs_object      = ldlm_res_hop_object,
 597         .hs_get         = ldlm_res_hop_get_locked,
 598         .hs_put         = ldlm_res_hop_put
 599 };
 600
 601 typedef struct ldlm_ns_hash_def {
 602         enum ldlm_ns_type       nsd_type;
 603         /** hash bucket bits */
 604         unsigned                nsd_bkt_bits;
 605         /** hash bits */
 606         unsigned                nsd_all_bits;
 607         /** hash operations */
 608         struct cfs_hash_ops *nsd_hops;
 609 } ldlm_ns_hash_def_t;
 610
 611 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 612 {
 613         {
 614                 .nsd_type       = LDLM_NS_TYPE_MDC,
 615                 .nsd_bkt_bits   = 11,
 616                 .nsd_all_bits   = 16,
 617                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 618         },
 619         {
 620                 .nsd_type       = LDLM_NS_TYPE_MDT,
 621                 .nsd_bkt_bits   = 14,
 622                 .nsd_all_bits   = 21,
 623                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 624         },
 625         {
 626                 .nsd_type       = LDLM_NS_TYPE_OSC,
 627                 .nsd_bkt_bits   = 8,
 628                 .nsd_all_bits   = 12,
 629                 .nsd_hops       = &ldlm_ns_hash_ops,
 630         },
 631         {
 632                 .nsd_type       = LDLM_NS_TYPE_OST,
 633                 .nsd_bkt_bits   = 11,
 634                 .nsd_all_bits   = 17,
 635                 .nsd_hops       = &ldlm_ns_hash_ops,
 636         },
 637         {
 638                 .nsd_type       = LDLM_NS_TYPE_MGC,
 639                 .nsd_bkt_bits   = 4,
 640                 .nsd_all_bits   = 4,
 641                 .nsd_hops       = &ldlm_ns_hash_ops,
 642         },
 643         {
 644                 .nsd_type       = LDLM_NS_TYPE_MGT,
 645                 .nsd_bkt_bits   = 4,
 646                 .nsd_all_bits   = 4,
 647                 .nsd_hops       = &ldlm_ns_hash_ops,
 648         },
 649         {
 650                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 651         },
 652 };
 653
 654 /**
 655  * Create and initialize new empty namespace.
 656  */
 657 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 658                                           enum ldlm_side client,
 659                                           enum ldlm_appetite apt,
 660                                           enum ldlm_ns_type ns_type)
 661 {
 662         struct ldlm_namespace *ns = NULL;
 663         struct ldlm_ns_bucket *nsb;
 664         struct ldlm_ns_hash_def *nsd;
 665         struct cfs_hash_bd bd;
 666         int idx;
 667         int rc;
 668         ENTRY;
 669
 670         LASSERT(obd != NULL);
 671
 672         rc = ldlm_get_ref();
 673         if (rc) {
 674                 CERROR("ldlm_get_ref failed: %d\n", rc);
 675                 RETURN(NULL);
 676         }
 677
 678         for (idx = 0;;idx++) {
 679                 nsd = &ldlm_ns_hash_defs[idx];
 680                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 681                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 682                         GOTO(out_ref, NULL);
 683                 }
 684
 685                 if (nsd->nsd_type == ns_type)
 686                         break;
 687         }
 688
 689         OBD_ALLOC_PTR(ns);
 690         if (!ns)
 691                 GOTO(out_ref, NULL);
 692
 693         ns->ns_rs_hash = cfs_hash_create(name,
 694                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 695                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 696                                          CFS_HASH_MIN_THETA,
 697                                          CFS_HASH_MAX_THETA,
 698                                          nsd->nsd_hops,
 699                                          CFS_HASH_DEPTH |
 700                                          CFS_HASH_BIGNAME |
 701                                          CFS_HASH_SPIN_BKTLOCK |
 702                                          CFS_HASH_NO_ITEMREF);
 703         if (ns->ns_rs_hash == NULL)
 704                 GOTO(out_ns, NULL);
 705
 706         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 707                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 708                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 709                 nsb->nsb_namespace = ns;
 710                 nsb->nsb_reclaim_start = 0;
 711         }
 712
 713         ns->ns_obd      = obd;
 714         ns->ns_appetite = apt;
 715         ns->ns_client   = client;
 716
 717         INIT_LIST_HEAD(&ns->ns_list_chain);
 718         INIT_LIST_HEAD(&ns->ns_unused_list);
 719         spin_lock_init(&ns->ns_lock);
 720         atomic_set(&ns->ns_bref, 0);
 721         init_waitqueue_head(&ns->ns_waitq);
 722
 723         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 724         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 725         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 726
 727         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 728         ns->ns_nr_unused          = 0;
 729         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 730         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 731         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 732         ns->ns_timeouts           = 0;
 733         ns->ns_orig_connect_flags = 0;
 734         ns->ns_connect_flags      = 0;
 735         ns->ns_stopping           = 0;
 736         ns->ns_reclaim_start      = 0;
 737         rc = ldlm_namespace_proc_register(ns);
 738         if (rc != 0) {
 739                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 740                 GOTO(out_hash, rc);
 741         }
 742
 743         idx = ldlm_namespace_nr_read(client);
 744         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 745         if (rc) {
 746                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 747                 GOTO(out_proc, rc);
 748         }
 749
 750         ldlm_namespace_register(ns, client);
 751         RETURN(ns);
 752 out_proc:
 753         ldlm_namespace_proc_unregister(ns);
 754         ldlm_namespace_cleanup(ns, 0);
 755 out_hash:
 756         cfs_hash_putref(ns->ns_rs_hash);
 757 out_ns:
 758         OBD_FREE_PTR(ns);
 759 out_ref:
 760         ldlm_put_ref();
 761         RETURN(NULL);
 762 }
 763 EXPORT_SYMBOL(ldlm_namespace_new);
 764
 765 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 766
 767 /**
 768  * Cancel and destroy all locks on a resource.
 769  *
 770  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 771  * clean up.  This is currently only used for recovery, and we make
 772  * certain assumptions as a result--notably, that we shouldn't cancel
 773  * locks with refs.
 774  */
 775 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 776                              __u64 flags)
 777 {
 778         struct list_head *tmp;
 779         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 780         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 781
 782         do {
 783                 struct ldlm_lock *lock = NULL;
 784
 785                 /* First, we look for non-cleaned-yet lock
 786                  * all cleaned locks are marked by CLEANED flag. */
 787                 lock_res(res);
 788                 list_for_each(tmp, q) {
 789                         lock = list_entry(tmp, struct ldlm_lock,
 790                                           l_res_link);
 791                         if (ldlm_is_cleaned(lock)) {
 792                                 lock = NULL;
 793                                 continue;
 794                         }
 795                         LDLM_LOCK_GET(lock);
 796                         ldlm_set_cleaned(lock);
 797                         break;
 798                 }
 799
 800                 if (lock == NULL) {
 801                         unlock_res(res);
 802                         break;
 803                 }
 804
 805                 /* Set CBPENDING so nothing in the cancellation path
 806                  * can match this lock. */
 807                 ldlm_set_cbpending(lock);
 808                 ldlm_set_failed(lock);
 809                 lock->l_flags |= flags;
 810
 811                 /* ... without sending a CANCEL message for local_only. */
 812                 if (local_only)
 813                         ldlm_set_local_only(lock);
 814
 815                 if (local_only && (lock->l_readers || lock->l_writers)) {
 816                         /* This is a little bit gross, but much better than the
 817                          * alternative: pretend that we got a blocking AST from
 818                          * the server, so that when the lock is decref'd, it
 819                          * will go away ... */
 820                         unlock_res(res);
 821                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 822                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
 823                                 set_current_state(TASK_UNINTERRUPTIBLE);
 824                                 schedule_timeout(cfs_time_seconds(4));
 825                                 set_current_state(TASK_RUNNING);
 826                         }
 827                         if (lock->l_completion_ast)
 828                                 lock->l_completion_ast(lock,
 829                                                        LDLM_FL_FAILED, NULL);
 830                         LDLM_LOCK_RELEASE(lock);
 831                         continue;
 832                 }
 833
 834                 if (client) {
 835                         struct lustre_handle lockh;
 836
 837                         unlock_res(res);
 838                         ldlm_lock2handle(lock, &lockh);
 839                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
 840                         if (rc)
 841                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 842                 } else {
 843                         unlock_res(res);
 844                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 845                                    "client node");
 846                         ldlm_lock_cancel(lock);
 847                 }
 848                 LDLM_LOCK_RELEASE(lock);
 849         } while (1);
 850 }
 851
 852 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 853                                struct hlist_node *hnode, void *arg)
 854 {
 855         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 856         __u64 flags = *(__u64 *)arg;
 857
 858         cleanup_resource(res, &res->lr_granted, flags);
 859         cleanup_resource(res, &res->lr_converting, flags);
 860         cleanup_resource(res, &res->lr_waiting, flags);
 861
 862         return 0;
 863 }
 864
 865 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 866                                   struct hlist_node *hnode, void *arg)
 867 {
 868         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 869
 870         lock_res(res);
 871         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
 872                "(%d) after lock cleanup; forcing cleanup.\n",
 873                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
 874                atomic_read(&res->lr_refcount) - 1);
 875
 876         ldlm_resource_dump(D_ERROR, res);
 877         unlock_res(res);
 878         return 0;
 879 }
 880
 881 /**
 882  * Cancel and destroy all locks in the namespace.
 883  *
 884  * Typically used during evictions when server notified client that it was
 885  * evicted and all of its state needs to be destroyed.
 886  * Also used during shutdown.
 887  */
 888 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
 889 {
 890         if (ns == NULL) {
 891                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 892                 return ELDLM_OK;
 893         }
 894
 895         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 896                                  &flags, 0);
 897         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
 898                                  NULL, 0);
 899         return ELDLM_OK;
 900 }
 901 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 902
 903 /**
 904  * Attempts to free namespace.
 905  *
 906  * Only used when namespace goes away, like during an unmount.
 907  */
 908 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 909 {
 910         ENTRY;
 911
 912         /* At shutdown time, don't call the cancellation callback */
 913         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 914
 915         if (atomic_read(&ns->ns_bref) > 0) {
 916                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 917                 int rc;
 918                 CDEBUG(D_DLMTRACE,
 919                        "dlm namespace %s free waiting on refcount %d\n",
 920                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
 921 force_wait:
 922                 if (force)
 923                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
 924                                           MSEC_PER_SEC) / 4, NULL, NULL);
 925
 926                 rc = l_wait_event(ns->ns_waitq,
 927                                   atomic_read(&ns->ns_bref) == 0, &lwi);
 928
 929                 /* Forced cleanups should be able to reclaim all references,
 930                  * so it's safe to wait forever... we can't leak locks... */
 931                 if (force && rc == -ETIMEDOUT) {
 932                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 933                                        "namespace with %d resources in use, "
 934                                        "(rc=%d)\n", ldlm_ns_name(ns),
 935                                        atomic_read(&ns->ns_bref), rc);
 936                         GOTO(force_wait, rc);
 937                 }
 938
 939                 if (atomic_read(&ns->ns_bref)) {
 940                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 941                                        "with %d resources in use, (rc=%d)\n",
 942                                        ldlm_ns_name(ns),
 943                                        atomic_read(&ns->ns_bref), rc);
 944                         RETURN(ELDLM_NAMESPACE_EXISTS);
 945                 }
 946                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 947                        ldlm_ns_name(ns));
 948         }
 949
 950         RETURN(ELDLM_OK);
 951 }
 952
 953 /**
 954  * Performs various cleanups for passed \a ns to make it drop refc and be
 955  * ready for freeing. Waits for refc == 0.
 956  *
 957  * The following is done:
 958  * (0) Unregister \a ns from its list to make inaccessible for potential
 959  * users like pools thread and others;
 960  * (1) Clear all locks in \a ns.
 961  */
 962 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 963                                struct obd_import *imp,
 964                                int force)
 965 {
 966         int rc;
 967         ENTRY;
 968         if (!ns) {
 969                 EXIT;
 970                 return;
 971         }
 972
 973         spin_lock(&ns->ns_lock);
 974         ns->ns_stopping = 1;
 975         spin_unlock(&ns->ns_lock);
 976
 977         /*
 978          * Can fail with -EINTR when force == 0 in which case try harder.
 979          */
 980         rc = __ldlm_namespace_free(ns, force);
 981         if (rc != ELDLM_OK) {
 982                 if (imp) {
 983                         ptlrpc_disconnect_import(imp, 0);
 984                         ptlrpc_invalidate_import(imp);
 985                 }
 986
 987                 /*
 988                  * With all requests dropped and the import inactive
 989                  * we are gaurenteed all reference will be dropped.
 990                  */
 991                 rc = __ldlm_namespace_free(ns, 1);
 992                 LASSERT(rc == 0);
 993         }
 994         EXIT;
 995 }
 996 EXPORT_SYMBOL(ldlm_namespace_free_prior);
 997
 998 /**
 999  * Performs freeing memory structures related to \a ns. This is only done
1000  * when ldlm_namespce_free_prior() successfully removed all resources
1001  * referencing \a ns and its refc == 0.
1002  */
1003 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1004 {
1005         ENTRY;
1006         if (!ns) {
1007                 EXIT;
1008                 return;
1009         }
1010
1011         /* Make sure that nobody can find this ns in its list. */
1012         ldlm_namespace_unregister(ns, ns->ns_client);
1013         /* Fini pool _before_ parent proc dir is removed. This is important as
1014          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1015          * Removing it after @dir may cause oops. */
1016         ldlm_pool_fini(&ns->ns_pool);
1017
1018         ldlm_namespace_proc_unregister(ns);
1019         cfs_hash_putref(ns->ns_rs_hash);
1020         /* Namespace \a ns should be not on list at this time, otherwise
1021          * this will cause issues related to using freed \a ns in poold
1022          * thread. */
1023         LASSERT(list_empty(&ns->ns_list_chain));
1024         OBD_FREE_PTR(ns);
1025         ldlm_put_ref();
1026         EXIT;
1027 }
1028 EXPORT_SYMBOL(ldlm_namespace_free_post);
1029
1030 /**
1031  * Cleanup the resource, and free namespace.
1032  * bug 12864:
1033  * Deadlock issue:
1034  * proc1: destroy import
1035  *        class_disconnect_export(grab cl_sem) ->
1036  *              -> ldlm_namespace_free ->
1037  *              -> lprocfs_remove(grab _lprocfs_lock).
1038  * proc2: read proc info
1039  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1040  *              -> osc_rd_active, etc(grab cl_sem).
1041  *
1042  * So that I have to split the ldlm_namespace_free into two parts - the first
1043  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1044  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1045  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1046  * held.
1047  */
1048 void ldlm_namespace_free(struct ldlm_namespace *ns,
1049                          struct obd_import *imp,
1050                          int force)
1051 {
1052         ldlm_namespace_free_prior(ns, imp, force);
1053         ldlm_namespace_free_post(ns);
1054 }
1055 EXPORT_SYMBOL(ldlm_namespace_free);
1056
1057 void ldlm_namespace_get(struct ldlm_namespace *ns)
1058 {
1059         atomic_inc(&ns->ns_bref);
1060 }
1061
1062 /* This is only for callers that care about refcount */
1063 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1064 {
1065         return atomic_inc_return(&ns->ns_bref);
1066 }
1067
1068 void ldlm_namespace_put(struct ldlm_namespace *ns)
1069 {
1070         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1071                 wake_up(&ns->ns_waitq);
1072                 spin_unlock(&ns->ns_lock);
1073         }
1074 }
1075
1076 /** Register \a ns in the list of namespaces */
1077 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1078 {
1079         mutex_lock(ldlm_namespace_lock(client));
1080         LASSERT(list_empty(&ns->ns_list_chain));
1081         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1082         ldlm_namespace_nr_inc(client);
1083         mutex_unlock(ldlm_namespace_lock(client));
1084 }
1085
1086 /** Unregister \a ns from the list of namespaces. */
1087 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1088 {
1089         mutex_lock(ldlm_namespace_lock(client));
1090         LASSERT(!list_empty(&ns->ns_list_chain));
1091         /* Some asserts and possibly other parts of the code are still
1092          * using list_empty(&ns->ns_list_chain). This is why it is
1093          * important to use list_del_init() here. */
1094         list_del_init(&ns->ns_list_chain);
1095         ldlm_namespace_nr_dec(client);
1096         mutex_unlock(ldlm_namespace_lock(client));
1097 }
1098
1099 /** Should be called with ldlm_namespace_lock(client) taken. */
1100 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1101                                           enum ldlm_side client)
1102 {
1103         LASSERT(!list_empty(&ns->ns_list_chain));
1104         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1105         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1106 }
1107
1108 /** Should be called with ldlm_namespace_lock(client) taken. */
1109 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1110                                             enum ldlm_side client)
1111 {
1112         LASSERT(!list_empty(&ns->ns_list_chain));
1113         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1114         list_move_tail(&ns->ns_list_chain,
1115                        ldlm_namespace_inactive_list(client));
1116 }
1117
1118 /** Should be called with ldlm_namespace_lock(client) taken. */
1119 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1120 {
1121         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1122         LASSERT(!list_empty(ldlm_namespace_list(client)));
1123         return container_of(ldlm_namespace_list(client)->next,
1124                             struct ldlm_namespace, ns_list_chain);
1125 }
1126
1127 /** Create and initialize new resource. */
1128 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1129 {
1130         struct ldlm_resource *res;
1131         int idx;
1132
1133         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1134         if (res == NULL)
1135                 return NULL;
1136
1137         if (ldlm_type == LDLM_EXTENT) {
1138                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1139                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1140                 if (res->lr_itree == NULL) {
1141                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1142                         return NULL;
1143                 }
1144                 /* Initialize interval trees for each lock mode. */
1145                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1146                         res->lr_itree[idx].lit_size = 0;
1147                         res->lr_itree[idx].lit_mode = 1 << idx;
1148                         res->lr_itree[idx].lit_root = NULL;
1149                 }
1150         }
1151
1152         INIT_LIST_HEAD(&res->lr_granted);
1153         INIT_LIST_HEAD(&res->lr_converting);
1154         INIT_LIST_HEAD(&res->lr_waiting);
1155
1156         atomic_set(&res->lr_refcount, 1);
1157         spin_lock_init(&res->lr_lock);
1158         lu_ref_init(&res->lr_reference);
1159
1160         /* Since LVB init can be delayed now, there is no longer need to
1161          * immediatelly acquire mutex here. */
1162         mutex_init(&res->lr_lvb_mutex);
1163         res->lr_lvb_initialized = false;
1164
1165         return res;
1166 }
1167
1168 /**
1169  * Return a reference to resource with given name, creating it if necessary.
1170  * Args: namespace with ns_lock unlocked
1171  * Locks: takes and releases NS hash-lock and res->lr_lock
1172  * Returns: referenced, unlocked ldlm_resource or NULL
1173  */
1174 struct ldlm_resource *
1175 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1176                   const struct ldlm_res_id *name, enum ldlm_type type,
1177                   int create)
1178 {
1179         struct hlist_node       *hnode;
1180         struct ldlm_resource    *res = NULL;
1181         struct cfs_hash_bd              bd;
1182         __u64                   version;
1183         int                     ns_refcount = 0;
1184
1185         LASSERT(ns != NULL);
1186         LASSERT(parent == NULL);
1187         LASSERT(ns->ns_rs_hash != NULL);
1188         LASSERT(name->name[0] != 0);
1189
1190         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1191         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1192         if (hnode != NULL) {
1193                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1194                 GOTO(found, res);
1195         }
1196
1197         version = cfs_hash_bd_version_get(&bd);
1198         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1199
1200         if (create == 0)
1201                 return ERR_PTR(-ENOENT);
1202
1203         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1204                  "type: %d\n", type);
1205         res = ldlm_resource_new(type);
1206         if (res == NULL)
1207                 return ERR_PTR(-ENOMEM);
1208
1209         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1210         res->lr_name = *name;
1211         res->lr_type = type;
1212
1213         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1214         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1215                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1216
1217         if (hnode != NULL) {
1218                 /* Someone won the race and already added the resource. */
1219                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1220                 /* Clean lu_ref for failed resource. */
1221                 lu_ref_fini(&res->lr_reference);
1222                 if (res->lr_itree != NULL)
1223                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1224                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1225                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1226 found:
1227                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1228                 return res;
1229         }
1230         /* We won! Let's add the resource. */
1231         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1232         if (cfs_hash_bd_count_get(&bd) == 1)
1233                 ns_refcount = ldlm_namespace_get_return(ns);
1234
1235         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1236
1237         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1238
1239         /* Let's see if we happened to be the very first resource in this
1240          * namespace. If so, and this is a client namespace, we need to move
1241          * the namespace into the active namespaces list to be patrolled by
1242          * the ldlm_poold. */
1243         if (ns_is_client(ns) && ns_refcount == 1) {
1244                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1245                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1246                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1247         }
1248
1249         return res;
1250 }
1251 EXPORT_SYMBOL(ldlm_resource_get);
1252
1253 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1254 {
1255         LASSERT(res != NULL);
1256         LASSERT(res != LP_POISON);
1257         atomic_inc(&res->lr_refcount);
1258         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1259                atomic_read(&res->lr_refcount));
1260         return res;
1261 }
1262
1263 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1264                                          struct ldlm_resource *res)
1265 {
1266         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1267
1268         if (!list_empty(&res->lr_granted)) {
1269                 ldlm_resource_dump(D_ERROR, res);
1270                 LBUG();
1271         }
1272
1273         if (!list_empty(&res->lr_converting)) {
1274                 ldlm_resource_dump(D_ERROR, res);
1275                 LBUG();
1276         }
1277
1278         if (!list_empty(&res->lr_waiting)) {
1279                 ldlm_resource_dump(D_ERROR, res);
1280                 LBUG();
1281         }
1282
1283         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1284                                bd, &res->lr_hash);
1285         lu_ref_fini(&res->lr_reference);
1286         if (cfs_hash_bd_count_get(bd) == 0)
1287                 ldlm_namespace_put(nsb->nsb_namespace);
1288 }
1289
1290 /* Returns 1 if the resource was freed, 0 if it remains. */
1291 int ldlm_resource_putref(struct ldlm_resource *res)
1292 {
1293         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1294         struct cfs_hash_bd   bd;
1295
1296         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1297         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1298                res, atomic_read(&res->lr_refcount) - 1);
1299
1300         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1301         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1302                 __ldlm_resource_putref_final(&bd, res);
1303                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1304                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1305                         ns->ns_lvbo->lvbo_free(res);
1306                 if (res->lr_itree != NULL)
1307                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1308                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1309                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1310                 return 1;
1311         }
1312         return 0;
1313 }
1314 EXPORT_SYMBOL(ldlm_resource_putref);
1315
1316 /**
1317  * Add a lock into a given resource into specified lock list.
1318  */
1319 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1320                             struct ldlm_lock *lock)
1321 {
1322         check_res_locked(res);
1323
1324         LDLM_DEBUG(lock, "About to add this lock");
1325
1326         if (ldlm_is_destroyed(lock)) {
1327                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1328                 return;
1329         }
1330
1331         LASSERT(list_empty(&lock->l_res_link));
1332
1333         list_add_tail(&lock->l_res_link, head);
1334 }
1335
1336 /**
1337  * Insert a lock into resource after specified lock.
1338  *
1339  * Obtain resource description from the lock we are inserting after.
1340  */
1341 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1342                                      struct ldlm_lock *new)
1343 {
1344         struct ldlm_resource *res = original->l_resource;
1345
1346         check_res_locked(res);
1347
1348         ldlm_resource_dump(D_INFO, res);
1349         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1350
1351         if (ldlm_is_destroyed(new)) {
1352                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1353                 goto out;
1354         }
1355
1356         LASSERT(list_empty(&new->l_res_link));
1357
1358         list_add(&new->l_res_link, &original->l_res_link);
1359  out:;
1360 }
1361
1362 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1363 {
1364         int type = lock->l_resource->lr_type;
1365
1366         check_res_locked(lock->l_resource);
1367         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1368                 ldlm_unlink_lock_skiplist(lock);
1369         else if (type == LDLM_EXTENT)
1370                 ldlm_extent_unlink_lock(lock);
1371         list_del_init(&lock->l_res_link);
1372 }
1373 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1374
1375 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1376 {
1377         desc->lr_type = res->lr_type;
1378         desc->lr_name = res->lr_name;
1379 }
1380
1381 /**
1382  * Print information about all locks in all namespaces on this node to debug
1383  * log.
1384  */
1385 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1386 {
1387         struct list_head *tmp;
1388
1389         if (!((libcfs_debug | D_ERROR) & level))
1390                 return;
1391
1392         mutex_lock(ldlm_namespace_lock(client));
1393
1394         list_for_each(tmp, ldlm_namespace_list(client)) {
1395                 struct ldlm_namespace *ns;
1396
1397                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1398                 ldlm_namespace_dump(level, ns);
1399         }
1400
1401         mutex_unlock(ldlm_namespace_lock(client));
1402 }
1403
1404 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1405                               struct hlist_node *hnode, void *arg)
1406 {
1407         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1408         int    level = (int)(unsigned long)arg;
1409
1410         lock_res(res);
1411         ldlm_resource_dump(level, res);
1412         unlock_res(res);
1413
1414         return 0;
1415 }
1416
1417 /**
1418  * Print information about all locks in this namespace on this node to debug
1419  * log.
1420  */
1421 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1422 {
1423         if (!((libcfs_debug | D_ERROR) & level))
1424                 return;
1425
1426         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1427                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1428                ns_is_client(ns) ? "client" : "server");
1429
1430         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1431                 return;
1432
1433         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1434                                  ldlm_res_hash_dump,
1435                                  (void *)(unsigned long)level, 0);
1436         spin_lock(&ns->ns_lock);
1437         ns->ns_next_dump = cfs_time_shift(10);
1438         spin_unlock(&ns->ns_lock);
1439 }
1440
1441 /**
1442  * Print information about all locks in this resource to debug log.
1443  */
1444 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1445 {
1446         struct ldlm_lock *lock;
1447         unsigned int granted = 0;
1448
1449         CLASSERT(RES_NAME_SIZE == 4);
1450
1451         if (!((libcfs_debug | D_ERROR) & level))
1452                 return;
1453
1454         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1455                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1456
1457         if (!list_empty(&res->lr_granted)) {
1458                 CDEBUG(level, "Granted locks (in reverse order):\n");
1459                 list_for_each_entry_reverse(lock, &res->lr_granted,
1460                                                 l_res_link) {
1461                         LDLM_DEBUG_LIMIT(level, lock, "###");
1462                         if (!(level & D_CANTMASK) &&
1463                             ++granted > ldlm_dump_granted_max) {
1464                                 CDEBUG(level, "only dump %d granted locks to "
1465                                        "avoid DDOS.\n", granted);
1466                                 break;
1467                         }
1468                 }
1469         }
1470         if (!list_empty(&res->lr_converting)) {
1471                 CDEBUG(level, "Converting locks:\n");
1472                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1473                         LDLM_DEBUG_LIMIT(level, lock, "###");
1474         }
1475         if (!list_empty(&res->lr_waiting)) {
1476                 CDEBUG(level, "Waiting locks:\n");
1477                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1478                         LDLM_DEBUG_LIMIT(level, lock, "###");
1479         }
1480 }
1481 EXPORT_SYMBOL(ldlm_resource_dump);