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