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, 2017, 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 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  51 LIST_HEAD(ldlm_srv_namespace_list);
  52
  53 DEFINE_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 LIST_HEAD(ldlm_cli_active_namespace_list);
  58 /* Client namespaces that don't have any locks in them */
  59 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_WR_ONLY(ldlm, dump_ns);
  79
  80 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  81 {
  82         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  83         return 0;
  84 }
  85
  86 static ssize_t
  87 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  88                        size_t count, loff_t *off)
  89 {
  90         struct seq_file *seq = file->private_data;
  91
  92         if (!count)
  93                 return 0;
  94
  95         return kstrtouint_from_user(buffer, count, 0,
  96                                     (unsigned int *)seq->private);
  97 }
  98
  99 LPROC_SEQ_FOPS(ldlm_rw_uint);
 100
 101 #ifdef HAVE_SERVER_SUPPORT
 102
 103 static int seq_watermark_show(struct seq_file *m, void *data)
 104 {
 105         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 106         return 0;
 107 }
 108
 109 static ssize_t seq_watermark_write(struct file *file,
 110                                    const char __user *buffer, size_t count,
 111                                    loff_t *off)
 112 {
 113         __s64 value;
 114         __u64 watermark;
 115         __u64 *data = ((struct seq_file *)file->private_data)->private;
 116         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 117         int rc;
 118
 119         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 120         if (rc) {
 121                 CERROR("Failed to set %s, rc = %d.\n",
 122                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 123                        rc);
 124                 return rc;
 125         } else if (value != 0 && value < (1 << 20)) {
 126                 CERROR("%s should be greater than 1MB.\n",
 127                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 128                 return -EINVAL;
 129         }
 130         watermark = value >> 20;
 131
 132         if (wm_low) {
 133                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 134                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 135                                "lock_limit_mb.\n");
 136                         return -EINVAL;
 137                 }
 138
 139                 *data = watermark;
 140                 if (watermark != 0) {
 141                         watermark <<= 20;
 142                         do_div(watermark, sizeof(struct ldlm_lock));
 143                 }
 144                 ldlm_reclaim_threshold = watermark;
 145         } else {
 146                 if (ldlm_reclaim_threshold_mb != 0 &&
 147                     watermark < ldlm_reclaim_threshold_mb) {
 148                         CERROR("lock_limit_mb must be greater than "
 149                                "lock_reclaim_threshold_mb.\n");
 150                         return -EINVAL;
 151                 }
 152
 153                 *data = watermark;
 154                 if (watermark != 0) {
 155                         watermark <<= 20;
 156                         do_div(watermark, sizeof(struct ldlm_lock));
 157                 }
 158                 ldlm_lock_limit = watermark;
 159         }
 160
 161         return count;
 162 }
 163
 164 static int seq_watermark_open(struct inode *inode, struct file *file)
 165 {
 166         return single_open(file, seq_watermark_show, PDE_DATA(inode));
 167 }
 168
 169 static const struct file_operations ldlm_watermark_fops = {
 170         .owner          = THIS_MODULE,
 171         .open           = seq_watermark_open,
 172         .read           = seq_read,
 173         .write          = seq_watermark_write,
 174         .llseek         = seq_lseek,
 175         .release        = lprocfs_single_release,
 176 };
 177
 178 static int seq_granted_show(struct seq_file *m, void *data)
 179 {
 180         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 181                    (struct percpu_counter *)m->private));
 182         return 0;
 183 }
 184
 185 static int seq_granted_open(struct inode *inode, struct file *file)
 186 {
 187         return single_open(file, seq_granted_show, PDE_DATA(inode));
 188 }
 189
 190 static const struct file_operations ldlm_granted_fops = {
 191         .owner  = THIS_MODULE,
 192         .open   = seq_granted_open,
 193         .read   = seq_read,
 194         .llseek = seq_lseek,
 195         .release = seq_release,
 196 };
 197
 198 #endif /* HAVE_SERVER_SUPPORT */
 199
 200 int ldlm_proc_setup(void)
 201 {
 202         int rc;
 203         struct lprocfs_vars list[] = {
 204                 { .name =       "dump_namespaces",
 205                   .fops =       &ldlm_dump_ns_fops,
 206                   .proc_mode =  0222 },
 207                 { .name =       "dump_granted_max",
 208                   .fops =       &ldlm_rw_uint_fops,
 209                   .data =       &ldlm_dump_granted_max },
 210 #ifdef HAVE_SERVER_SUPPORT
 211                 { .name =       "lock_reclaim_threshold_mb",
 212                   .fops =       &ldlm_watermark_fops,
 213                   .data =       &ldlm_reclaim_threshold_mb },
 214                 { .name =       "lock_limit_mb",
 215                   .fops =       &ldlm_watermark_fops,
 216                   .data =       &ldlm_lock_limit_mb },
 217                 { .name =       "lock_granted_count",
 218                   .fops =       &ldlm_granted_fops,
 219                   .data =       &ldlm_granted_total },
 220 #endif
 221                 { NULL }};
 222         ENTRY;
 223         LASSERT(ldlm_ns_proc_dir == NULL);
 224
 225         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
 226                                               proc_lustre_root,
 227                                               NULL, NULL);
 228         if (IS_ERR(ldlm_type_proc_dir)) {
 229                 CERROR("LProcFS failed in ldlm-init\n");
 230                 rc = PTR_ERR(ldlm_type_proc_dir);
 231                 GOTO(err, rc);
 232         }
 233
 234         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 235                                             ldlm_type_proc_dir,
 236                                             NULL, NULL);
 237         if (IS_ERR(ldlm_ns_proc_dir)) {
 238                 CERROR("LProcFS failed in ldlm-init\n");
 239                 rc = PTR_ERR(ldlm_ns_proc_dir);
 240                 GOTO(err_type, rc);
 241         }
 242
 243         ldlm_svc_proc_dir = lprocfs_register("services",
 244                                              ldlm_type_proc_dir,
 245                                              NULL, NULL);
 246         if (IS_ERR(ldlm_svc_proc_dir)) {
 247                 CERROR("LProcFS failed in ldlm-init\n");
 248                 rc = PTR_ERR(ldlm_svc_proc_dir);
 249                 GOTO(err_ns, rc);
 250         }
 251
 252         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 253         if (rc != 0) {
 254                 CERROR("LProcFS failed in ldlm-init\n");
 255                 GOTO(err_svc, rc);
 256         }
 257
 258         RETURN(0);
 259
 260 err_svc:
 261         lprocfs_remove(&ldlm_svc_proc_dir);
 262 err_ns:
 263         lprocfs_remove(&ldlm_ns_proc_dir);
 264 err_type:
 265         lprocfs_remove(&ldlm_type_proc_dir);
 266 err:
 267         ldlm_svc_proc_dir = NULL;
 268         RETURN(rc);
 269 }
 270
 271 void ldlm_proc_cleanup(void)
 272 {
 273         if (ldlm_svc_proc_dir)
 274                 lprocfs_remove(&ldlm_svc_proc_dir);
 275
 276         if (ldlm_ns_proc_dir)
 277                 lprocfs_remove(&ldlm_ns_proc_dir);
 278
 279         if (ldlm_type_proc_dir)
 280                 lprocfs_remove(&ldlm_type_proc_dir);
 281 }
 282
 283 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 284                                    char *buf)
 285 {
 286         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 287                                                  ns_kobj);
 288         __u64                   res = 0;
 289         struct cfs_hash_bd              bd;
 290         int                     i;
 291
 292         /* result is not strictly consistant */
 293         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 294                 res += cfs_hash_bd_count_get(&bd);
 295         return sprintf(buf, "%lld\n", res);
 296 }
 297 LUSTRE_RO_ATTR(resource_count);
 298
 299 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 300                                char *buf)
 301 {
 302         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 303                                                  ns_kobj);
 304         __u64                   locks;
 305
 306         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 307                                         LPROCFS_FIELDS_FLAGS_SUM);
 308         return sprintf(buf, "%lld\n", locks);
 309 }
 310 LUSTRE_RO_ATTR(lock_count);
 311
 312 static ssize_t lock_unused_count_show(struct kobject *kobj,
 313                                       struct attribute *attr,
 314                                       char *buf)
 315 {
 316         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 317                                                  ns_kobj);
 318
 319         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 320 }
 321 LUSTRE_RO_ATTR(lock_unused_count);
 322
 323 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 324                              char *buf)
 325 {
 326         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 327                                                  ns_kobj);
 328         __u32 *nr = &ns->ns_max_unused;
 329
 330         if (ns_connect_lru_resize(ns))
 331                 nr = &ns->ns_nr_unused;
 332         return sprintf(buf, "%u\n", *nr);
 333 }
 334
 335 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 336                               const char *buffer, size_t count)
 337 {
 338         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 339                                                  ns_kobj);
 340         unsigned long tmp;
 341         int lru_resize;
 342         int err;
 343
 344         if (strncmp(buffer, "clear", 5) == 0) {
 345                 CDEBUG(D_DLMTRACE,
 346                        "dropping all unused locks from namespace %s\n",
 347                        ldlm_ns_name(ns));
 348                 if (ns_connect_lru_resize(ns)) {
 349                         /* Try to cancel all @ns_nr_unused locks. */
 350                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 351                                         LDLM_LRU_FLAG_PASSED |
 352                                         LDLM_LRU_FLAG_CLEANUP);
 353                 } else {
 354                         tmp = ns->ns_max_unused;
 355                         ns->ns_max_unused = 0;
 356                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 357                                         LDLM_LRU_FLAG_CLEANUP);
 358                         ns->ns_max_unused = tmp;
 359                 }
 360                 return count;
 361         }
 362
 363         err = kstrtoul(buffer, 10, &tmp);
 364         if (err != 0) {
 365                 CERROR("lru_size: invalid value written\n");
 366                 return -EINVAL;
 367         }
 368         lru_resize = (tmp == 0);
 369
 370         if (ns_connect_lru_resize(ns)) {
 371                 if (!lru_resize)
 372                         ns->ns_max_unused = (unsigned int)tmp;
 373
 374                 if (tmp > ns->ns_nr_unused)
 375                         tmp = ns->ns_nr_unused;
 376                 tmp = ns->ns_nr_unused - tmp;
 377
 378                 CDEBUG(D_DLMTRACE,
 379                        "changing namespace %s unused locks from %u to %u\n",
 380                        ldlm_ns_name(ns), ns->ns_nr_unused,
 381                        (unsigned int)tmp);
 382                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 383
 384                 if (!lru_resize) {
 385                         CDEBUG(D_DLMTRACE,
 386                                "disable lru_resize for namespace %s\n",
 387                                ldlm_ns_name(ns));
 388                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 389                 }
 390         } else {
 391                 CDEBUG(D_DLMTRACE,
 392                        "changing namespace %s max_unused from %u to %u\n",
 393                        ldlm_ns_name(ns), ns->ns_max_unused,
 394                        (unsigned int)tmp);
 395                 ns->ns_max_unused = (unsigned int)tmp;
 396                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 397
 398                 /* Make sure that LRU resize was originally supported before
 399                  * turning it on here.
 400                  */
 401                 if (lru_resize &&
 402                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 403                         CDEBUG(D_DLMTRACE,
 404                                "enable lru_resize for namespace %s\n",
 405                                ldlm_ns_name(ns));
 406                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 407                 }
 408         }
 409
 410         return count;
 411 }
 412 LUSTRE_RW_ATTR(lru_size);
 413
 414 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 415                                 char *buf)
 416 {
 417         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 418                                                  ns_kobj);
 419
 420         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 421 }
 422
 423 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 424                                  const char *buffer, size_t count)
 425 {
 426         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 427                                                  ns_kobj);
 428         int scale = NSEC_PER_MSEC;
 429         unsigned long long tmp;
 430         char *buf;
 431
 432         /* Did the user ask in seconds or milliseconds. Default is in ms */
 433         buf = strstr(buffer, "ms");
 434         if (!buf) {
 435                 buf = strchr(buffer, 's');
 436                 if (buf)
 437                         scale = NSEC_PER_SEC;
 438         }
 439
 440         if (buf)
 441                 *buf = '\0';
 442
 443         if (kstrtoull(buffer, 10, &tmp))
 444                 return -EINVAL;
 445
 446         ns->ns_max_age = ktime_set(0, tmp * scale);
 447
 448         return count;
 449 }
 450 LUSTRE_RW_ATTR(lru_max_age);
 451
 452 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 453                                       struct attribute *attr,
 454                                       char *buf)
 455 {
 456         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 457                                                  ns_kobj);
 458
 459         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 460 }
 461
 462 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 463                                        struct attribute *attr,
 464                                        const char *buffer,
 465                                        size_t count)
 466 {
 467         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 468                                                  ns_kobj);
 469         unsigned long supp = -1;
 470         int rc;
 471
 472         rc = kstrtoul(buffer, 10, &supp);
 473         if (rc < 0)
 474                 return rc;
 475
 476         if (supp == 0)
 477                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 478         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 479                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 480         return count;
 481 }
 482 LUSTRE_RW_ATTR(early_lock_cancel);
 483
 484 #ifdef HAVE_SERVER_SUPPORT
 485 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 486                                     struct attribute *attr, char *buf)
 487 {
 488         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 489                                                  ns_kobj);
 490
 491         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 492 }
 493
 494 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 495                                      struct attribute *attr,
 496                                      const char *buffer, size_t count)
 497 {
 498         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 499                                                  ns_kobj);
 500         unsigned long long tmp;
 501
 502         if (kstrtoull(buffer, 10, &tmp))
 503                 return -EINVAL;
 504
 505         ns->ns_ctime_age_limit = tmp;
 506
 507         return count;
 508 }
 509 LUSTRE_RW_ATTR(ctime_age_limit);
 510
 511 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 512                                   char *buf)
 513 {
 514         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 515                                                  ns_kobj);
 516
 517         return sprintf(buf, "%d\n", ns->ns_timeouts);
 518 }
 519 LUSTRE_RO_ATTR(lock_timeouts);
 520
 521 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 522                                      struct attribute *attr, char *buf)
 523 {
 524         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 525                                                  ns_kobj);
 526
 527         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 528 }
 529
 530 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 531                                       struct attribute *attr,
 532                                       const char *buffer, size_t count)
 533 {
 534         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 535                                                  ns_kobj);
 536         unsigned long tmp;
 537         int err;
 538
 539         err = kstrtoul(buffer, 10, &tmp);
 540         if (err != 0)
 541                 return -EINVAL;
 542
 543         ns->ns_max_nolock_size = tmp;
 544
 545         return count;
 546 }
 547 LUSTRE_RW_ATTR(max_nolock_bytes);
 548
 549 static ssize_t contention_seconds_show(struct kobject *kobj,
 550                                        struct attribute *attr, char *buf)
 551 {
 552         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 553                                                  ns_kobj);
 554
 555         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 556 }
 557
 558 static ssize_t contention_seconds_store(struct kobject *kobj,
 559                                         struct attribute *attr,
 560                                         const char *buffer, size_t count)
 561 {
 562         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 563                                                  ns_kobj);
 564         unsigned long long tmp;
 565
 566         if (kstrtoull(buffer, 10, &tmp))
 567                 return -EINVAL;
 568
 569         ns->ns_contention_time = tmp;
 570
 571         return count;
 572 }
 573 LUSTRE_RW_ATTR(contention_seconds);
 574
 575 static ssize_t contended_locks_show(struct kobject *kobj,
 576                                     struct attribute *attr, char *buf)
 577 {
 578         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 579                                                  ns_kobj);
 580
 581         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 582 }
 583
 584 static ssize_t contended_locks_store(struct kobject *kobj,
 585                                      struct attribute *attr,
 586                                      const char *buffer, size_t count)
 587 {
 588         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 589                                                  ns_kobj);
 590         unsigned long tmp;
 591         int err;
 592
 593         err = kstrtoul(buffer, 10, &tmp);
 594         if (err != 0)
 595                 return -EINVAL;
 596
 597         ns->ns_contended_locks = tmp;
 598
 599         return count;
 600 }
 601 LUSTRE_RW_ATTR(contended_locks);
 602
 603 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 604                                      struct attribute *attr, char *buf)
 605 {
 606         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 607                                                  ns_kobj);
 608
 609         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 610 }
 611
 612 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 613                                       struct attribute *attr,
 614                                       const char *buffer, size_t count)
 615 {
 616         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 617                                                  ns_kobj);
 618         unsigned long tmp;
 619         int err;
 620
 621         err = kstrtoul(buffer, 10, &tmp);
 622         if (err != 0)
 623                 return -EINVAL;
 624
 625         ns->ns_max_parallel_ast = tmp;
 626
 627         return count;
 628 }
 629 LUSTRE_RW_ATTR(max_parallel_ast);
 630
 631 #endif /* HAVE_SERVER_SUPPORT */
 632
 633 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 634 static struct attribute *ldlm_ns_attrs[] = {
 635         &lustre_attr_resource_count.attr,
 636         &lustre_attr_lock_count.attr,
 637         &lustre_attr_lock_unused_count.attr,
 638         &lustre_attr_lru_size.attr,
 639         &lustre_attr_lru_max_age.attr,
 640         &lustre_attr_early_lock_cancel.attr,
 641 #ifdef HAVE_SERVER_SUPPORT
 642         &lustre_attr_ctime_age_limit.attr,
 643         &lustre_attr_lock_timeouts.attr,
 644         &lustre_attr_max_nolock_bytes.attr,
 645         &lustre_attr_contention_seconds.attr,
 646         &lustre_attr_contended_locks.attr,
 647         &lustre_attr_max_parallel_ast.attr,
 648 #endif
 649         NULL,
 650 };
 651
 652 static void ldlm_ns_release(struct kobject *kobj)
 653 {
 654         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 655                                                  ns_kobj);
 656         complete(&ns->ns_kobj_unregister);
 657 }
 658
 659 static struct kobj_type ldlm_ns_ktype = {
 660         .default_attrs  = ldlm_ns_attrs,
 661         .sysfs_ops      = &lustre_sysfs_ops,
 662         .release        = ldlm_ns_release,
 663 };
 664
 665 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 666 {
 667         if (ns->ns_proc_dir_entry == NULL)
 668                 CERROR("dlm namespace %s has no procfs dir?\n",
 669                        ldlm_ns_name(ns));
 670         else
 671                 lprocfs_remove(&ns->ns_proc_dir_entry);
 672
 673         if (ns->ns_stats != NULL)
 674                 lprocfs_free_stats(&ns->ns_stats);
 675 }
 676
 677 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 678 {
 679         kobject_put(&ns->ns_kobj);
 680         wait_for_completion(&ns->ns_kobj_unregister);
 681 }
 682
 683 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 684 {
 685         int err;
 686
 687         ns->ns_kobj.kset = ldlm_ns_kset;
 688         init_completion(&ns->ns_kobj_unregister);
 689         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 690                                    "%s", ldlm_ns_name(ns));
 691
 692         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 693         if (!ns->ns_stats) {
 694                 kobject_put(&ns->ns_kobj);
 695                 return -ENOMEM;
 696         }
 697
 698         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 699                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 700
 701         return err;
 702 }
 703
 704 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 705 {
 706         struct proc_dir_entry *ns_pde;
 707
 708         LASSERT(ns != NULL);
 709         LASSERT(ns->ns_rs_hash != NULL);
 710
 711         if (ns->ns_proc_dir_entry != NULL) {
 712                 ns_pde = ns->ns_proc_dir_entry;
 713         } else {
 714                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 715                 if (ns_pde == NULL)
 716                         return -ENOMEM;
 717                 ns->ns_proc_dir_entry = ns_pde;
 718         }
 719
 720         return 0;
 721 }
 722 #undef MAX_STRING_SIZE
 723 #else /* CONFIG_PROC_FS */
 724
 725 #define ldlm_namespace_proc_unregister(ns)      ({;})
 726 #define ldlm_namespace_proc_register(ns)        ({0;})
 727
 728 #endif /* CONFIG_PROC_FS */
 729
 730 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 731                                   const void *key, unsigned mask)
 732 {
 733         const struct ldlm_res_id     *id  = key;
 734         unsigned                val = 0;
 735         unsigned                i;
 736
 737         for (i = 0; i < RES_NAME_SIZE; i++)
 738                 val += id->name[i];
 739         return val & mask;
 740 }
 741
 742 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 743                                       const void *key, unsigned mask)
 744 {
 745         const struct ldlm_res_id *id = key;
 746         struct lu_fid       fid;
 747         __u32               hash;
 748         __u32               val;
 749
 750         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 751         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 752         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 753
 754         hash = fid_flatten32(&fid);
 755         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 756         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 757                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 758                 hash += (val >> 5) + (val << 11);
 759         } else {
 760                 val = fid_oid(&fid);
 761         }
 762         hash = hash_long(hash, hs->hs_bkt_bits);
 763         /* give me another random factor */
 764         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 765
 766         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 767         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 768
 769         return hash & mask;
 770 }
 771
 772 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 773 {
 774         struct ldlm_resource   *res;
 775
 776         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 777         return &res->lr_name;
 778 }
 779
 780 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 781 {
 782         struct ldlm_resource   *res;
 783
 784         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 785         return ldlm_res_eq((const struct ldlm_res_id *)key,
 786                            (const struct ldlm_res_id *)&res->lr_name);
 787 }
 788
 789 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 790 {
 791         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 792 }
 793
 794 static void
 795 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 796 {
 797         struct ldlm_resource *res;
 798
 799         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 800         ldlm_resource_getref(res);
 801 }
 802
 803 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 804 {
 805         struct ldlm_resource *res;
 806
 807         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 808         ldlm_resource_putref(res);
 809 }
 810
 811 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 812         .hs_hash        = ldlm_res_hop_hash,
 813         .hs_key         = ldlm_res_hop_key,
 814         .hs_keycmp      = ldlm_res_hop_keycmp,
 815         .hs_keycpy      = NULL,
 816         .hs_object      = ldlm_res_hop_object,
 817         .hs_get         = ldlm_res_hop_get_locked,
 818         .hs_put         = ldlm_res_hop_put
 819 };
 820
 821 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 822         .hs_hash        = ldlm_res_hop_fid_hash,
 823         .hs_key         = ldlm_res_hop_key,
 824         .hs_keycmp      = ldlm_res_hop_keycmp,
 825         .hs_keycpy      = NULL,
 826         .hs_object      = ldlm_res_hop_object,
 827         .hs_get         = ldlm_res_hop_get_locked,
 828         .hs_put         = ldlm_res_hop_put
 829 };
 830
 831 typedef struct ldlm_ns_hash_def {
 832         enum ldlm_ns_type       nsd_type;
 833         /** hash bucket bits */
 834         unsigned                nsd_bkt_bits;
 835         /** hash bits */
 836         unsigned                nsd_all_bits;
 837         /** hash operations */
 838         struct cfs_hash_ops *nsd_hops;
 839 } ldlm_ns_hash_def_t;
 840
 841 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 842 {
 843         {
 844                 .nsd_type       = LDLM_NS_TYPE_MDC,
 845                 .nsd_bkt_bits   = 11,
 846                 .nsd_all_bits   = 16,
 847                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 848         },
 849         {
 850                 .nsd_type       = LDLM_NS_TYPE_MDT,
 851                 .nsd_bkt_bits   = 14,
 852                 .nsd_all_bits   = 21,
 853                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 854         },
 855         {
 856                 .nsd_type       = LDLM_NS_TYPE_OSC,
 857                 .nsd_bkt_bits   = 8,
 858                 .nsd_all_bits   = 12,
 859                 .nsd_hops       = &ldlm_ns_hash_ops,
 860         },
 861         {
 862                 .nsd_type       = LDLM_NS_TYPE_OST,
 863                 .nsd_bkt_bits   = 11,
 864                 .nsd_all_bits   = 17,
 865                 .nsd_hops       = &ldlm_ns_hash_ops,
 866         },
 867         {
 868                 .nsd_type       = LDLM_NS_TYPE_MGC,
 869                 .nsd_bkt_bits   = 4,
 870                 .nsd_all_bits   = 4,
 871                 .nsd_hops       = &ldlm_ns_hash_ops,
 872         },
 873         {
 874                 .nsd_type       = LDLM_NS_TYPE_MGT,
 875                 .nsd_bkt_bits   = 4,
 876                 .nsd_all_bits   = 4,
 877                 .nsd_hops       = &ldlm_ns_hash_ops,
 878         },
 879         {
 880                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 881         },
 882 };
 883
 884 /**
 885  * Create and initialize new empty namespace.
 886  */
 887 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 888                                           enum ldlm_side client,
 889                                           enum ldlm_appetite apt,
 890                                           enum ldlm_ns_type ns_type)
 891 {
 892         struct ldlm_namespace *ns = NULL;
 893         struct ldlm_ns_bucket *nsb;
 894         struct ldlm_ns_hash_def *nsd;
 895         struct cfs_hash_bd bd;
 896         int idx;
 897         int rc;
 898         ENTRY;
 899
 900         LASSERT(obd != NULL);
 901
 902         rc = ldlm_get_ref();
 903         if (rc) {
 904                 CERROR("ldlm_get_ref failed: %d\n", rc);
 905                 RETURN(NULL);
 906         }
 907
 908         for (idx = 0;;idx++) {
 909                 nsd = &ldlm_ns_hash_defs[idx];
 910                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 911                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 912                         GOTO(out_ref, NULL);
 913                 }
 914
 915                 if (nsd->nsd_type == ns_type)
 916                         break;
 917         }
 918
 919         OBD_ALLOC_PTR(ns);
 920         if (!ns)
 921                 GOTO(out_ref, NULL);
 922
 923         ns->ns_rs_hash = cfs_hash_create(name,
 924                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 925                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 926                                          CFS_HASH_MIN_THETA,
 927                                          CFS_HASH_MAX_THETA,
 928                                          nsd->nsd_hops,
 929                                          CFS_HASH_DEPTH |
 930                                          CFS_HASH_BIGNAME |
 931                                          CFS_HASH_SPIN_BKTLOCK |
 932                                          CFS_HASH_NO_ITEMREF);
 933         if (ns->ns_rs_hash == NULL)
 934                 GOTO(out_ns, NULL);
 935
 936         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 937                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 938                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 939                 nsb->nsb_namespace = ns;
 940                 nsb->nsb_reclaim_start = 0;
 941         }
 942
 943         ns->ns_obd      = obd;
 944         ns->ns_appetite = apt;
 945         ns->ns_client   = client;
 946
 947         INIT_LIST_HEAD(&ns->ns_list_chain);
 948         INIT_LIST_HEAD(&ns->ns_unused_list);
 949         spin_lock_init(&ns->ns_lock);
 950         atomic_set(&ns->ns_bref, 0);
 951         init_waitqueue_head(&ns->ns_waitq);
 952
 953         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 954         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 955         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 956
 957         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 958         ns->ns_nr_unused          = 0;
 959         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 960         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 961         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 962         ns->ns_timeouts           = 0;
 963         ns->ns_orig_connect_flags = 0;
 964         ns->ns_connect_flags      = 0;
 965         ns->ns_stopping           = 0;
 966         ns->ns_reclaim_start      = 0;
 967
 968         rc = ldlm_namespace_sysfs_register(ns);
 969         if (rc) {
 970                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 971                 GOTO(out_hash, rc);
 972         }
 973
 974         rc = ldlm_namespace_proc_register(ns);
 975         if (rc) {
 976                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 977                 GOTO(out_sysfs, rc);
 978         }
 979
 980         idx = ldlm_namespace_nr_read(client);
 981         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 982         if (rc) {
 983                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 984                 GOTO(out_proc, rc);
 985         }
 986
 987         ldlm_namespace_register(ns, client);
 988         RETURN(ns);
 989 out_proc:
 990         ldlm_namespace_proc_unregister(ns);
 991 out_sysfs:
 992         ldlm_namespace_sysfs_unregister(ns);
 993         ldlm_namespace_cleanup(ns, 0);
 994 out_hash:
 995         cfs_hash_putref(ns->ns_rs_hash);
 996 out_ns:
 997         OBD_FREE_PTR(ns);
 998 out_ref:
 999         ldlm_put_ref();
1000         RETURN(NULL);
1001 }
1002 EXPORT_SYMBOL(ldlm_namespace_new);
1003
1004 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
1005
1006 /**
1007  * Cancel and destroy all locks on a resource.
1008  *
1009  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1010  * clean up.  This is currently only used for recovery, and we make
1011  * certain assumptions as a result--notably, that we shouldn't cancel
1012  * locks with refs.
1013  */
1014 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1015                              __u64 flags)
1016 {
1017         struct list_head *tmp;
1018         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1019         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1020
1021         do {
1022                 struct ldlm_lock *lock = NULL;
1023
1024                 /* First, we look for non-cleaned-yet lock
1025                  * all cleaned locks are marked by CLEANED flag. */
1026                 lock_res(res);
1027                 list_for_each(tmp, q) {
1028                         lock = list_entry(tmp, struct ldlm_lock,
1029                                           l_res_link);
1030                         if (ldlm_is_cleaned(lock)) {
1031                                 lock = NULL;
1032                                 continue;
1033                         }
1034                         LDLM_LOCK_GET(lock);
1035                         ldlm_set_cleaned(lock);
1036                         break;
1037                 }
1038
1039                 if (lock == NULL) {
1040                         unlock_res(res);
1041                         break;
1042                 }
1043
1044                 /* Set CBPENDING so nothing in the cancellation path
1045                  * can match this lock. */
1046                 ldlm_set_cbpending(lock);
1047                 ldlm_set_failed(lock);
1048                 lock->l_flags |= flags;
1049
1050                 /* ... without sending a CANCEL message for local_only. */
1051                 if (local_only)
1052                         ldlm_set_local_only(lock);
1053
1054                 if (local_only && (lock->l_readers || lock->l_writers)) {
1055                         /* This is a little bit gross, but much better than the
1056                          * alternative: pretend that we got a blocking AST from
1057                          * the server, so that when the lock is decref'd, it
1058                          * will go away ... */
1059                         unlock_res(res);
1060                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1061                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1062                                 set_current_state(TASK_UNINTERRUPTIBLE);
1063                                 schedule_timeout(cfs_time_seconds(4));
1064                                 set_current_state(TASK_RUNNING);
1065                         }
1066                         if (lock->l_completion_ast)
1067                                 lock->l_completion_ast(lock,
1068                                                        LDLM_FL_FAILED, NULL);
1069                         LDLM_LOCK_RELEASE(lock);
1070                         continue;
1071                 }
1072
1073                 if (client) {
1074                         struct lustre_handle lockh;
1075
1076                         unlock_res(res);
1077                         ldlm_lock2handle(lock, &lockh);
1078                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1079                         if (rc)
1080                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1081                 } else {
1082                         unlock_res(res);
1083                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
1084                                    "client node");
1085                         ldlm_lock_cancel(lock);
1086                 }
1087                 LDLM_LOCK_RELEASE(lock);
1088         } while (1);
1089 }
1090
1091 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1092                                struct hlist_node *hnode, void *arg)
1093 {
1094         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1095         __u64 flags = *(__u64 *)arg;
1096
1097         cleanup_resource(res, &res->lr_granted, flags);
1098         cleanup_resource(res, &res->lr_waiting, flags);
1099
1100         return 0;
1101 }
1102
1103 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1104                                   struct hlist_node *hnode, void *arg)
1105 {
1106         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1107
1108         lock_res(res);
1109         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1110                "(%d) after lock cleanup; forcing cleanup.\n",
1111                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1112                atomic_read(&res->lr_refcount) - 1);
1113
1114         ldlm_resource_dump(D_ERROR, res);
1115         unlock_res(res);
1116         return 0;
1117 }
1118
1119 /**
1120  * Cancel and destroy all locks in the namespace.
1121  *
1122  * Typically used during evictions when server notified client that it was
1123  * evicted and all of its state needs to be destroyed.
1124  * Also used during shutdown.
1125  */
1126 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1127 {
1128         if (ns == NULL) {
1129                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1130                 return ELDLM_OK;
1131         }
1132
1133         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1134                                  &flags, 0);
1135         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1136                                  NULL, 0);
1137         return ELDLM_OK;
1138 }
1139 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1140
1141 /**
1142  * Attempts to free namespace.
1143  *
1144  * Only used when namespace goes away, like during an unmount.
1145  */
1146 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1147 {
1148         ENTRY;
1149
1150         /* At shutdown time, don't call the cancellation callback */
1151         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1152
1153         if (atomic_read(&ns->ns_bref) > 0) {
1154                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1155                 int rc;
1156                 CDEBUG(D_DLMTRACE,
1157                        "dlm namespace %s free waiting on refcount %d\n",
1158                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1159 force_wait:
1160                 if (force)
1161                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
1162                                           MSEC_PER_SEC) / 4, NULL, NULL);
1163
1164                 rc = l_wait_event(ns->ns_waitq,
1165                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1166
1167                 /* Forced cleanups should be able to reclaim all references,
1168                  * so it's safe to wait forever... we can't leak locks... */
1169                 if (force && rc == -ETIMEDOUT) {
1170                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1171                                        "namespace with %d resources in use, "
1172                                        "(rc=%d)\n", ldlm_ns_name(ns),
1173                                        atomic_read(&ns->ns_bref), rc);
1174                         GOTO(force_wait, rc);
1175                 }
1176
1177                 if (atomic_read(&ns->ns_bref)) {
1178                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1179                                        "with %d resources in use, (rc=%d)\n",
1180                                        ldlm_ns_name(ns),
1181                                        atomic_read(&ns->ns_bref), rc);
1182                         RETURN(ELDLM_NAMESPACE_EXISTS);
1183                 }
1184                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1185                        ldlm_ns_name(ns));
1186         }
1187
1188         RETURN(ELDLM_OK);
1189 }
1190
1191 /**
1192  * Performs various cleanups for passed \a ns to make it drop refc and be
1193  * ready for freeing. Waits for refc == 0.
1194  *
1195  * The following is done:
1196  * (0) Unregister \a ns from its list to make inaccessible for potential
1197  * users like pools thread and others;
1198  * (1) Clear all locks in \a ns.
1199  */
1200 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1201                                struct obd_import *imp,
1202                                int force)
1203 {
1204         int rc;
1205         ENTRY;
1206         if (!ns) {
1207                 EXIT;
1208                 return;
1209         }
1210
1211         spin_lock(&ns->ns_lock);
1212         ns->ns_stopping = 1;
1213         spin_unlock(&ns->ns_lock);
1214
1215         /*
1216          * Can fail with -EINTR when force == 0 in which case try harder.
1217          */
1218         rc = __ldlm_namespace_free(ns, force);
1219         if (rc != ELDLM_OK) {
1220                 if (imp) {
1221                         ptlrpc_disconnect_import(imp, 0);
1222                         ptlrpc_invalidate_import(imp);
1223                 }
1224
1225                 /*
1226                  * With all requests dropped and the import inactive
1227                  * we are gaurenteed all reference will be dropped.
1228                  */
1229                 rc = __ldlm_namespace_free(ns, 1);
1230                 LASSERT(rc == 0);
1231         }
1232         EXIT;
1233 }
1234 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1235
1236 /**
1237  * Performs freeing memory structures related to \a ns. This is only done
1238  * when ldlm_namespce_free_prior() successfully removed all resources
1239  * referencing \a ns and its refc == 0.
1240  */
1241 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1242 {
1243         ENTRY;
1244         if (!ns) {
1245                 EXIT;
1246                 return;
1247         }
1248
1249         /* Make sure that nobody can find this ns in its list. */
1250         ldlm_namespace_unregister(ns, ns->ns_client);
1251         /* Fini pool _before_ parent proc dir is removed. This is important as
1252          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1253          * Removing it after @dir may cause oops. */
1254         ldlm_pool_fini(&ns->ns_pool);
1255
1256         ldlm_namespace_proc_unregister(ns);
1257         ldlm_namespace_sysfs_unregister(ns);
1258         cfs_hash_putref(ns->ns_rs_hash);
1259         /* Namespace \a ns should be not on list at this time, otherwise
1260          * this will cause issues related to using freed \a ns in poold
1261          * thread. */
1262         LASSERT(list_empty(&ns->ns_list_chain));
1263         OBD_FREE_PTR(ns);
1264         ldlm_put_ref();
1265         EXIT;
1266 }
1267 EXPORT_SYMBOL(ldlm_namespace_free_post);
1268
1269 /**
1270  * Cleanup the resource, and free namespace.
1271  * bug 12864:
1272  * Deadlock issue:
1273  * proc1: destroy import
1274  *        class_disconnect_export(grab cl_sem) ->
1275  *              -> ldlm_namespace_free ->
1276  *              -> lprocfs_remove(grab _lprocfs_lock).
1277  * proc2: read proc info
1278  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1279  *              -> osc_rd_active, etc(grab cl_sem).
1280  *
1281  * So that I have to split the ldlm_namespace_free into two parts - the first
1282  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1283  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1284  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1285  * held.
1286  */
1287 void ldlm_namespace_free(struct ldlm_namespace *ns,
1288                          struct obd_import *imp,
1289                          int force)
1290 {
1291         ldlm_namespace_free_prior(ns, imp, force);
1292         ldlm_namespace_free_post(ns);
1293 }
1294 EXPORT_SYMBOL(ldlm_namespace_free);
1295
1296 void ldlm_namespace_get(struct ldlm_namespace *ns)
1297 {
1298         atomic_inc(&ns->ns_bref);
1299 }
1300
1301 /* This is only for callers that care about refcount */
1302 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1303 {
1304         return atomic_inc_return(&ns->ns_bref);
1305 }
1306
1307 void ldlm_namespace_put(struct ldlm_namespace *ns)
1308 {
1309         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1310                 wake_up(&ns->ns_waitq);
1311                 spin_unlock(&ns->ns_lock);
1312         }
1313 }
1314
1315 /** Register \a ns in the list of namespaces */
1316 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1317 {
1318         mutex_lock(ldlm_namespace_lock(client));
1319         LASSERT(list_empty(&ns->ns_list_chain));
1320         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1321         ldlm_namespace_nr_inc(client);
1322         mutex_unlock(ldlm_namespace_lock(client));
1323 }
1324
1325 /** Unregister \a ns from the list of namespaces. */
1326 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1327 {
1328         mutex_lock(ldlm_namespace_lock(client));
1329         LASSERT(!list_empty(&ns->ns_list_chain));
1330         /* Some asserts and possibly other parts of the code are still
1331          * using list_empty(&ns->ns_list_chain). This is why it is
1332          * important to use list_del_init() here. */
1333         list_del_init(&ns->ns_list_chain);
1334         ldlm_namespace_nr_dec(client);
1335         mutex_unlock(ldlm_namespace_lock(client));
1336 }
1337
1338 /** Should be called with ldlm_namespace_lock(client) taken. */
1339 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1340                                           enum ldlm_side client)
1341 {
1342         LASSERT(!list_empty(&ns->ns_list_chain));
1343         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1344         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1345 }
1346
1347 /** Should be called with ldlm_namespace_lock(client) taken. */
1348 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1349                                             enum ldlm_side client)
1350 {
1351         LASSERT(!list_empty(&ns->ns_list_chain));
1352         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1353         list_move_tail(&ns->ns_list_chain,
1354                        ldlm_namespace_inactive_list(client));
1355 }
1356
1357 /** Should be called with ldlm_namespace_lock(client) taken. */
1358 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1359 {
1360         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1361         LASSERT(!list_empty(ldlm_namespace_list(client)));
1362         return container_of(ldlm_namespace_list(client)->next,
1363                             struct ldlm_namespace, ns_list_chain);
1364 }
1365
1366 /** Create and initialize new resource. */
1367 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1368 {
1369         struct ldlm_resource *res;
1370         int idx;
1371
1372         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1373         if (res == NULL)
1374                 return NULL;
1375
1376         if (ldlm_type == LDLM_EXTENT) {
1377                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1378                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1379                 if (res->lr_itree == NULL) {
1380                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1381                         return NULL;
1382                 }
1383                 /* Initialize interval trees for each lock mode. */
1384                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1385                         res->lr_itree[idx].lit_size = 0;
1386                         res->lr_itree[idx].lit_mode = 1 << idx;
1387                         res->lr_itree[idx].lit_root = NULL;
1388                 }
1389         }
1390
1391         INIT_LIST_HEAD(&res->lr_granted);
1392         INIT_LIST_HEAD(&res->lr_waiting);
1393
1394         atomic_set(&res->lr_refcount, 1);
1395         spin_lock_init(&res->lr_lock);
1396         lu_ref_init(&res->lr_reference);
1397
1398         /* Since LVB init can be delayed now, there is no longer need to
1399          * immediatelly acquire mutex here. */
1400         mutex_init(&res->lr_lvb_mutex);
1401         res->lr_lvb_initialized = false;
1402
1403         return res;
1404 }
1405
1406 /**
1407  * Return a reference to resource with given name, creating it if necessary.
1408  * Args: namespace with ns_lock unlocked
1409  * Locks: takes and releases NS hash-lock and res->lr_lock
1410  * Returns: referenced, unlocked ldlm_resource or NULL
1411  */
1412 struct ldlm_resource *
1413 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1414                   const struct ldlm_res_id *name, enum ldlm_type type,
1415                   int create)
1416 {
1417         struct hlist_node       *hnode;
1418         struct ldlm_resource    *res = NULL;
1419         struct cfs_hash_bd              bd;
1420         __u64                   version;
1421         int                     ns_refcount = 0;
1422
1423         LASSERT(ns != NULL);
1424         LASSERT(parent == NULL);
1425         LASSERT(ns->ns_rs_hash != NULL);
1426         LASSERT(name->name[0] != 0);
1427
1428         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1429         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1430         if (hnode != NULL) {
1431                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1432                 GOTO(found, res);
1433         }
1434
1435         version = cfs_hash_bd_version_get(&bd);
1436         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1437
1438         if (create == 0)
1439                 return ERR_PTR(-ENOENT);
1440
1441         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1442                  "type: %d\n", type);
1443         res = ldlm_resource_new(type);
1444         if (res == NULL)
1445                 return ERR_PTR(-ENOMEM);
1446
1447         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1448         res->lr_name = *name;
1449         res->lr_type = type;
1450
1451         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1452         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1453                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1454
1455         if (hnode != NULL) {
1456                 /* Someone won the race and already added the resource. */
1457                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1458                 /* Clean lu_ref for failed resource. */
1459                 lu_ref_fini(&res->lr_reference);
1460                 if (res->lr_itree != NULL)
1461                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1462                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1463                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1464 found:
1465                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1466                 return res;
1467         }
1468         /* We won! Let's add the resource. */
1469         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1470         if (cfs_hash_bd_count_get(&bd) == 1)
1471                 ns_refcount = ldlm_namespace_get_return(ns);
1472
1473         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1474
1475         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1476
1477         /* Let's see if we happened to be the very first resource in this
1478          * namespace. If so, and this is a client namespace, we need to move
1479          * the namespace into the active namespaces list to be patrolled by
1480          * the ldlm_poold. */
1481         if (ns_is_client(ns) && ns_refcount == 1) {
1482                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1483                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1484                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1485         }
1486
1487         return res;
1488 }
1489 EXPORT_SYMBOL(ldlm_resource_get);
1490
1491 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1492 {
1493         LASSERT(res != NULL);
1494         LASSERT(res != LP_POISON);
1495         atomic_inc(&res->lr_refcount);
1496         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1497                atomic_read(&res->lr_refcount));
1498         return res;
1499 }
1500
1501 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1502                                          struct ldlm_resource *res)
1503 {
1504         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1505
1506         if (!list_empty(&res->lr_granted)) {
1507                 ldlm_resource_dump(D_ERROR, res);
1508                 LBUG();
1509         }
1510
1511         if (!list_empty(&res->lr_waiting)) {
1512                 ldlm_resource_dump(D_ERROR, res);
1513                 LBUG();
1514         }
1515
1516         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1517                                bd, &res->lr_hash);
1518         lu_ref_fini(&res->lr_reference);
1519         if (cfs_hash_bd_count_get(bd) == 0)
1520                 ldlm_namespace_put(nsb->nsb_namespace);
1521 }
1522
1523 /* Returns 1 if the resource was freed, 0 if it remains. */
1524 int ldlm_resource_putref(struct ldlm_resource *res)
1525 {
1526         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1527         struct cfs_hash_bd   bd;
1528
1529         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1530         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1531                res, atomic_read(&res->lr_refcount) - 1);
1532
1533         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1534         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1535                 __ldlm_resource_putref_final(&bd, res);
1536                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1537                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1538                         ns->ns_lvbo->lvbo_free(res);
1539                 if (res->lr_itree != NULL)
1540                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1541                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1542                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1543                 return 1;
1544         }
1545         return 0;
1546 }
1547 EXPORT_SYMBOL(ldlm_resource_putref);
1548
1549 /**
1550  * Add a lock into a given resource into specified lock list.
1551  */
1552 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1553                             struct ldlm_lock *lock)
1554 {
1555         check_res_locked(res);
1556
1557         LDLM_DEBUG(lock, "About to add this lock");
1558
1559         if (ldlm_is_destroyed(lock)) {
1560                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1561                 return;
1562         }
1563
1564         LASSERT(list_empty(&lock->l_res_link));
1565
1566         list_add_tail(&lock->l_res_link, head);
1567 }
1568
1569 /**
1570  * Insert a lock into resource after specified lock.
1571  *
1572  * Obtain resource description from the lock we are inserting after.
1573  */
1574 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1575                                      struct ldlm_lock *new)
1576 {
1577         struct ldlm_resource *res = original->l_resource;
1578
1579         check_res_locked(res);
1580
1581         ldlm_resource_dump(D_INFO, res);
1582         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1583
1584         if (ldlm_is_destroyed(new)) {
1585                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1586                 goto out;
1587         }
1588
1589         LASSERT(list_empty(&new->l_res_link));
1590
1591         list_add(&new->l_res_link, &original->l_res_link);
1592  out:;
1593 }
1594
1595 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1596 {
1597         int type = lock->l_resource->lr_type;
1598
1599         check_res_locked(lock->l_resource);
1600         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1601                 ldlm_unlink_lock_skiplist(lock);
1602         else if (type == LDLM_EXTENT)
1603                 ldlm_extent_unlink_lock(lock);
1604         list_del_init(&lock->l_res_link);
1605 }
1606 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1607
1608 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1609 {
1610         desc->lr_type = res->lr_type;
1611         desc->lr_name = res->lr_name;
1612 }
1613
1614 /**
1615  * Print information about all locks in all namespaces on this node to debug
1616  * log.
1617  */
1618 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1619 {
1620         struct list_head *tmp;
1621
1622         if (!((libcfs_debug | D_ERROR) & level))
1623                 return;
1624
1625         mutex_lock(ldlm_namespace_lock(client));
1626
1627         list_for_each(tmp, ldlm_namespace_list(client)) {
1628                 struct ldlm_namespace *ns;
1629
1630                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1631                 ldlm_namespace_dump(level, ns);
1632         }
1633
1634         mutex_unlock(ldlm_namespace_lock(client));
1635 }
1636
1637 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1638                               struct hlist_node *hnode, void *arg)
1639 {
1640         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1641         int    level = (int)(unsigned long)arg;
1642
1643         lock_res(res);
1644         ldlm_resource_dump(level, res);
1645         unlock_res(res);
1646
1647         return 0;
1648 }
1649
1650 /**
1651  * Print information about all locks in this namespace on this node to debug
1652  * log.
1653  */
1654 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1655 {
1656         if (!((libcfs_debug | D_ERROR) & level))
1657                 return;
1658
1659         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1660                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1661                ns_is_client(ns) ? "client" : "server");
1662
1663         if (ktime_get_seconds() < ns->ns_next_dump)
1664                 return;
1665
1666         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1667                                  ldlm_res_hash_dump,
1668                                  (void *)(unsigned long)level, 0);
1669         spin_lock(&ns->ns_lock);
1670         ns->ns_next_dump = ktime_get_seconds() + 10;
1671         spin_unlock(&ns->ns_lock);
1672 }
1673
1674 /**
1675  * Print information about all locks in this resource to debug log.
1676  */
1677 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1678 {
1679         struct ldlm_lock *lock;
1680         unsigned int granted = 0;
1681
1682         CLASSERT(RES_NAME_SIZE == 4);
1683
1684         if (!((libcfs_debug | D_ERROR) & level))
1685                 return;
1686
1687         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1688                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1689
1690         if (!list_empty(&res->lr_granted)) {
1691                 CDEBUG(level, "Granted locks (in reverse order):\n");
1692                 list_for_each_entry_reverse(lock, &res->lr_granted,
1693                                                 l_res_link) {
1694                         LDLM_DEBUG_LIMIT(level, lock, "###");
1695                         if (!(level & D_CANTMASK) &&
1696                             ++granted > ldlm_dump_granted_max) {
1697                                 CDEBUG(level, "only dump %d granted locks to "
1698                                        "avoid DDOS.\n", granted);
1699                                 break;
1700                         }
1701                 }
1702         }
1703
1704         if (!list_empty(&res->lr_waiting)) {
1705                 CDEBUG(level, "Waiting locks:\n");
1706                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1707                         LDLM_DEBUG_LIMIT(level, lock, "###");
1708         }
1709 }
1710 EXPORT_SYMBOL(ldlm_resource_dump);