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, 2016, 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         int err;
 413
 414         /* Did the user ask in seconds or milliseconds. Default is in ms */
 415         buf = strstr(buffer, "ms");
 416         if (!buf) {
 417                 buf = strchr(buffer, 's');
 418                 if (buf)
 419                         scale = NSEC_PER_SEC;
 420         }
 421
 422         if (buf)
 423                 *buf = '\0';
 424
 425         err = kstrtoull(buffer, 10, &tmp);
 426         if (err != 0)
 427                 return -EINVAL;
 428
 429         ns->ns_max_age = ktime_set(0, tmp * scale);
 430
 431         return count;
 432 }
 433 LUSTRE_RW_ATTR(lru_max_age);
 434
 435 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 436                                       struct attribute *attr,
 437                                       char *buf)
 438 {
 439         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 440                                                  ns_kobj);
 441
 442         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 443 }
 444
 445 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 446                                        struct attribute *attr,
 447                                        const char *buffer,
 448                                        size_t count)
 449 {
 450         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 451                                                  ns_kobj);
 452         unsigned long supp = -1;
 453         int rc;
 454
 455         rc = kstrtoul(buffer, 10, &supp);
 456         if (rc < 0)
 457                 return rc;
 458
 459         if (supp == 0)
 460                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 461         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 462                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 463         return count;
 464 }
 465 LUSTRE_RW_ATTR(early_lock_cancel);
 466
 467 #ifdef HAVE_SERVER_SUPPORT
 468 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 469                                     struct attribute *attr, char *buf)
 470 {
 471         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 472                                                  ns_kobj);
 473
 474         return sprintf(buf, "%u\n", ns->ns_ctime_age_limit);
 475 }
 476
 477 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 478                                      struct attribute *attr,
 479                                      const char *buffer, size_t count)
 480 {
 481         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 482                                                  ns_kobj);
 483         unsigned long tmp;
 484         int err;
 485
 486         err = kstrtoul(buffer, 10, &tmp);
 487         if (err != 0)
 488                 return -EINVAL;
 489
 490         ns->ns_ctime_age_limit = tmp;
 491
 492         return count;
 493 }
 494 LUSTRE_RW_ATTR(ctime_age_limit);
 495
 496 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 497                                   char *buf)
 498 {
 499         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 500                                                  ns_kobj);
 501
 502         return sprintf(buf, "%d\n", ns->ns_timeouts);
 503 }
 504 LUSTRE_RO_ATTR(lock_timeouts);
 505
 506 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 507                                      struct attribute *attr, char *buf)
 508 {
 509         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 510                                                  ns_kobj);
 511
 512         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 513 }
 514
 515 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 516                                       struct attribute *attr,
 517                                       const char *buffer, size_t count)
 518 {
 519         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 520                                                  ns_kobj);
 521         unsigned long tmp;
 522         int err;
 523
 524         err = kstrtoul(buffer, 10, &tmp);
 525         if (err != 0)
 526                 return -EINVAL;
 527
 528         ns->ns_max_nolock_size = tmp;
 529
 530         return count;
 531 }
 532 LUSTRE_RW_ATTR(max_nolock_bytes);
 533
 534 static ssize_t contention_seconds_show(struct kobject *kobj,
 535                                        struct attribute *attr, char *buf)
 536 {
 537         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 538                                                  ns_kobj);
 539
 540         return sprintf(buf, "%u\n", ns->ns_contention_time);
 541 }
 542
 543 static ssize_t contention_seconds_store(struct kobject *kobj,
 544                                         struct attribute *attr,
 545                                         const char *buffer, size_t count)
 546 {
 547         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 548                                                  ns_kobj);
 549         unsigned long tmp;
 550         int err;
 551
 552         err = kstrtoul(buffer, 10, &tmp);
 553         if (err != 0)
 554                 return -EINVAL;
 555
 556         ns->ns_contention_time = tmp;
 557
 558         return count;
 559 }
 560 LUSTRE_RW_ATTR(contention_seconds);
 561
 562 static ssize_t contended_locks_show(struct kobject *kobj,
 563                                     struct attribute *attr, char *buf)
 564 {
 565         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 566                                                  ns_kobj);
 567
 568         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 569 }
 570
 571 static ssize_t contended_locks_store(struct kobject *kobj,
 572                                      struct attribute *attr,
 573                                      const char *buffer, size_t count)
 574 {
 575         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 576                                                  ns_kobj);
 577         unsigned long tmp;
 578         int err;
 579
 580         err = kstrtoul(buffer, 10, &tmp);
 581         if (err != 0)
 582                 return -EINVAL;
 583
 584         ns->ns_contended_locks = tmp;
 585
 586         return count;
 587 }
 588 LUSTRE_RW_ATTR(contended_locks);
 589
 590 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 591                                      struct attribute *attr, char *buf)
 592 {
 593         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 594                                                  ns_kobj);
 595
 596         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 597 }
 598
 599 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 600                                       struct attribute *attr,
 601                                       const char *buffer, size_t count)
 602 {
 603         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 604                                                  ns_kobj);
 605         unsigned long tmp;
 606         int err;
 607
 608         err = kstrtoul(buffer, 10, &tmp);
 609         if (err != 0)
 610                 return -EINVAL;
 611
 612         ns->ns_max_parallel_ast = tmp;
 613
 614         return count;
 615 }
 616 LUSTRE_RW_ATTR(max_parallel_ast);
 617
 618 #endif /* HAVE_SERVER_SUPPORT */
 619
 620 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 621 static struct attribute *ldlm_ns_attrs[] = {
 622         &lustre_attr_resource_count.attr,
 623         &lustre_attr_lock_count.attr,
 624         &lustre_attr_lock_unused_count.attr,
 625         &lustre_attr_lru_size.attr,
 626         &lustre_attr_lru_max_age.attr,
 627         &lustre_attr_early_lock_cancel.attr,
 628 #ifdef HAVE_SERVER_SUPPORT
 629         &lustre_attr_ctime_age_limit.attr,
 630         &lustre_attr_lock_timeouts.attr,
 631         &lustre_attr_max_nolock_bytes.attr,
 632         &lustre_attr_contention_seconds.attr,
 633         &lustre_attr_contended_locks.attr,
 634         &lustre_attr_max_parallel_ast.attr,
 635 #endif
 636         NULL,
 637 };
 638
 639 static void ldlm_ns_release(struct kobject *kobj)
 640 {
 641         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 642                                                  ns_kobj);
 643         complete(&ns->ns_kobj_unregister);
 644 }
 645
 646 static struct kobj_type ldlm_ns_ktype = {
 647         .default_attrs  = ldlm_ns_attrs,
 648         .sysfs_ops      = &lustre_sysfs_ops,
 649         .release        = ldlm_ns_release,
 650 };
 651
 652 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 653 {
 654         if (ns->ns_proc_dir_entry == NULL)
 655                 CERROR("dlm namespace %s has no procfs dir?\n",
 656                        ldlm_ns_name(ns));
 657         else
 658                 lprocfs_remove(&ns->ns_proc_dir_entry);
 659
 660         if (ns->ns_stats != NULL)
 661                 lprocfs_free_stats(&ns->ns_stats);
 662 }
 663
 664 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 665 {
 666         kobject_put(&ns->ns_kobj);
 667         wait_for_completion(&ns->ns_kobj_unregister);
 668 }
 669
 670 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 671 {
 672         int err;
 673
 674         ns->ns_kobj.kset = ldlm_ns_kset;
 675         init_completion(&ns->ns_kobj_unregister);
 676         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 677                                    "%s", ldlm_ns_name(ns));
 678
 679         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 680         if (!ns->ns_stats) {
 681                 kobject_put(&ns->ns_kobj);
 682                 return -ENOMEM;
 683         }
 684
 685         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 686                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 687
 688         return err;
 689 }
 690
 691 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 692 {
 693         struct proc_dir_entry *ns_pde;
 694
 695         LASSERT(ns != NULL);
 696         LASSERT(ns->ns_rs_hash != NULL);
 697
 698         if (ns->ns_proc_dir_entry != NULL) {
 699                 ns_pde = ns->ns_proc_dir_entry;
 700         } else {
 701                 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
 702                 if (ns_pde == NULL)
 703                         return -ENOMEM;
 704                 ns->ns_proc_dir_entry = ns_pde;
 705         }
 706
 707         return 0;
 708 }
 709 #undef MAX_STRING_SIZE
 710 #else /* CONFIG_PROC_FS */
 711
 712 #define ldlm_namespace_proc_unregister(ns)      ({;})
 713 #define ldlm_namespace_proc_register(ns)        ({0;})
 714
 715 #endif /* CONFIG_PROC_FS */
 716
 717 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 718                                   const void *key, unsigned mask)
 719 {
 720         const struct ldlm_res_id     *id  = key;
 721         unsigned                val = 0;
 722         unsigned                i;
 723
 724         for (i = 0; i < RES_NAME_SIZE; i++)
 725                 val += id->name[i];
 726         return val & mask;
 727 }
 728
 729 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 730                                       const void *key, unsigned mask)
 731 {
 732         const struct ldlm_res_id *id = key;
 733         struct lu_fid       fid;
 734         __u32               hash;
 735         __u32               val;
 736
 737         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 738         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 739         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 740
 741         hash = fid_flatten32(&fid);
 742         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 743         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 744                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 745                 hash += (val >> 5) + (val << 11);
 746         } else {
 747                 val = fid_oid(&fid);
 748         }
 749         hash = hash_long(hash, hs->hs_bkt_bits);
 750         /* give me another random factor */
 751         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 752
 753         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 754         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 755
 756         return hash & mask;
 757 }
 758
 759 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 760 {
 761         struct ldlm_resource   *res;
 762
 763         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 764         return &res->lr_name;
 765 }
 766
 767 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 768 {
 769         struct ldlm_resource   *res;
 770
 771         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 772         return ldlm_res_eq((const struct ldlm_res_id *)key,
 773                            (const struct ldlm_res_id *)&res->lr_name);
 774 }
 775
 776 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 777 {
 778         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 779 }
 780
 781 static void
 782 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 783 {
 784         struct ldlm_resource *res;
 785
 786         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 787         ldlm_resource_getref(res);
 788 }
 789
 790 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 791 {
 792         struct ldlm_resource *res;
 793
 794         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 795         ldlm_resource_putref(res);
 796 }
 797
 798 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 799         .hs_hash        = ldlm_res_hop_hash,
 800         .hs_key         = ldlm_res_hop_key,
 801         .hs_keycmp      = ldlm_res_hop_keycmp,
 802         .hs_keycpy      = NULL,
 803         .hs_object      = ldlm_res_hop_object,
 804         .hs_get         = ldlm_res_hop_get_locked,
 805         .hs_put         = ldlm_res_hop_put
 806 };
 807
 808 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 809         .hs_hash        = ldlm_res_hop_fid_hash,
 810         .hs_key         = ldlm_res_hop_key,
 811         .hs_keycmp      = ldlm_res_hop_keycmp,
 812         .hs_keycpy      = NULL,
 813         .hs_object      = ldlm_res_hop_object,
 814         .hs_get         = ldlm_res_hop_get_locked,
 815         .hs_put         = ldlm_res_hop_put
 816 };
 817
 818 typedef struct ldlm_ns_hash_def {
 819         enum ldlm_ns_type       nsd_type;
 820         /** hash bucket bits */
 821         unsigned                nsd_bkt_bits;
 822         /** hash bits */
 823         unsigned                nsd_all_bits;
 824         /** hash operations */
 825         struct cfs_hash_ops *nsd_hops;
 826 } ldlm_ns_hash_def_t;
 827
 828 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 829 {
 830         {
 831                 .nsd_type       = LDLM_NS_TYPE_MDC,
 832                 .nsd_bkt_bits   = 11,
 833                 .nsd_all_bits   = 16,
 834                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 835         },
 836         {
 837                 .nsd_type       = LDLM_NS_TYPE_MDT,
 838                 .nsd_bkt_bits   = 14,
 839                 .nsd_all_bits   = 21,
 840                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 841         },
 842         {
 843                 .nsd_type       = LDLM_NS_TYPE_OSC,
 844                 .nsd_bkt_bits   = 8,
 845                 .nsd_all_bits   = 12,
 846                 .nsd_hops       = &ldlm_ns_hash_ops,
 847         },
 848         {
 849                 .nsd_type       = LDLM_NS_TYPE_OST,
 850                 .nsd_bkt_bits   = 11,
 851                 .nsd_all_bits   = 17,
 852                 .nsd_hops       = &ldlm_ns_hash_ops,
 853         },
 854         {
 855                 .nsd_type       = LDLM_NS_TYPE_MGC,
 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_MGT,
 862                 .nsd_bkt_bits   = 4,
 863                 .nsd_all_bits   = 4,
 864                 .nsd_hops       = &ldlm_ns_hash_ops,
 865         },
 866         {
 867                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 868         },
 869 };
 870
 871 /**
 872  * Create and initialize new empty namespace.
 873  */
 874 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 875                                           enum ldlm_side client,
 876                                           enum ldlm_appetite apt,
 877                                           enum ldlm_ns_type ns_type)
 878 {
 879         struct ldlm_namespace *ns = NULL;
 880         struct ldlm_ns_bucket *nsb;
 881         struct ldlm_ns_hash_def *nsd;
 882         struct cfs_hash_bd bd;
 883         int idx;
 884         int rc;
 885         ENTRY;
 886
 887         LASSERT(obd != NULL);
 888
 889         rc = ldlm_get_ref();
 890         if (rc) {
 891                 CERROR("ldlm_get_ref failed: %d\n", rc);
 892                 RETURN(NULL);
 893         }
 894
 895         for (idx = 0;;idx++) {
 896                 nsd = &ldlm_ns_hash_defs[idx];
 897                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 898                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 899                         GOTO(out_ref, NULL);
 900                 }
 901
 902                 if (nsd->nsd_type == ns_type)
 903                         break;
 904         }
 905
 906         OBD_ALLOC_PTR(ns);
 907         if (!ns)
 908                 GOTO(out_ref, NULL);
 909
 910         ns->ns_rs_hash = cfs_hash_create(name,
 911                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 912                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 913                                          CFS_HASH_MIN_THETA,
 914                                          CFS_HASH_MAX_THETA,
 915                                          nsd->nsd_hops,
 916                                          CFS_HASH_DEPTH |
 917                                          CFS_HASH_BIGNAME |
 918                                          CFS_HASH_SPIN_BKTLOCK |
 919                                          CFS_HASH_NO_ITEMREF);
 920         if (ns->ns_rs_hash == NULL)
 921                 GOTO(out_ns, NULL);
 922
 923         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 924                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 925                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 926                 nsb->nsb_namespace = ns;
 927                 nsb->nsb_reclaim_start = 0;
 928         }
 929
 930         ns->ns_obd      = obd;
 931         ns->ns_appetite = apt;
 932         ns->ns_client   = client;
 933
 934         INIT_LIST_HEAD(&ns->ns_list_chain);
 935         INIT_LIST_HEAD(&ns->ns_unused_list);
 936         spin_lock_init(&ns->ns_lock);
 937         atomic_set(&ns->ns_bref, 0);
 938         init_waitqueue_head(&ns->ns_waitq);
 939
 940         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 941         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 942         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 943
 944         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 945         ns->ns_nr_unused          = 0;
 946         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 947         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 948         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 949         ns->ns_timeouts           = 0;
 950         ns->ns_orig_connect_flags = 0;
 951         ns->ns_connect_flags      = 0;
 952         ns->ns_stopping           = 0;
 953         ns->ns_reclaim_start      = 0;
 954
 955         rc = ldlm_namespace_sysfs_register(ns);
 956         if (rc) {
 957                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 958                 GOTO(out_hash, rc);
 959         }
 960
 961         rc = ldlm_namespace_proc_register(ns);
 962         if (rc) {
 963                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 964                 GOTO(out_sysfs, rc);
 965         }
 966
 967         idx = ldlm_namespace_nr_read(client);
 968         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 969         if (rc) {
 970                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 971                 GOTO(out_proc, rc);
 972         }
 973
 974         ldlm_namespace_register(ns, client);
 975         RETURN(ns);
 976 out_proc:
 977         ldlm_namespace_proc_unregister(ns);
 978 out_sysfs:
 979         ldlm_namespace_sysfs_unregister(ns);
 980         ldlm_namespace_cleanup(ns, 0);
 981 out_hash:
 982         cfs_hash_putref(ns->ns_rs_hash);
 983 out_ns:
 984         OBD_FREE_PTR(ns);
 985 out_ref:
 986         ldlm_put_ref();
 987         RETURN(NULL);
 988 }
 989 EXPORT_SYMBOL(ldlm_namespace_new);
 990
 991 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 992
 993 /**
 994  * Cancel and destroy all locks on a resource.
 995  *
 996  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 997  * clean up.  This is currently only used for recovery, and we make
 998  * certain assumptions as a result--notably, that we shouldn't cancel
 999  * locks with refs.
1000  */
1001 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1002                              __u64 flags)
1003 {
1004         struct list_head *tmp;
1005         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1006         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1007
1008         do {
1009                 struct ldlm_lock *lock = NULL;
1010
1011                 /* First, we look for non-cleaned-yet lock
1012                  * all cleaned locks are marked by CLEANED flag. */
1013                 lock_res(res);
1014                 list_for_each(tmp, q) {
1015                         lock = list_entry(tmp, struct ldlm_lock,
1016                                           l_res_link);
1017                         if (ldlm_is_cleaned(lock)) {
1018                                 lock = NULL;
1019                                 continue;
1020                         }
1021                         LDLM_LOCK_GET(lock);
1022                         ldlm_set_cleaned(lock);
1023                         break;
1024                 }
1025
1026                 if (lock == NULL) {
1027                         unlock_res(res);
1028                         break;
1029                 }
1030
1031                 /* Set CBPENDING so nothing in the cancellation path
1032                  * can match this lock. */
1033                 ldlm_set_cbpending(lock);
1034                 ldlm_set_failed(lock);
1035                 lock->l_flags |= flags;
1036
1037                 /* ... without sending a CANCEL message for local_only. */
1038                 if (local_only)
1039                         ldlm_set_local_only(lock);
1040
1041                 if (local_only && (lock->l_readers || lock->l_writers)) {
1042                         /* This is a little bit gross, but much better than the
1043                          * alternative: pretend that we got a blocking AST from
1044                          * the server, so that when the lock is decref'd, it
1045                          * will go away ... */
1046                         unlock_res(res);
1047                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1048                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1049                                 set_current_state(TASK_UNINTERRUPTIBLE);
1050                                 schedule_timeout(cfs_time_seconds(4));
1051                                 set_current_state(TASK_RUNNING);
1052                         }
1053                         if (lock->l_completion_ast)
1054                                 lock->l_completion_ast(lock,
1055                                                        LDLM_FL_FAILED, NULL);
1056                         LDLM_LOCK_RELEASE(lock);
1057                         continue;
1058                 }
1059
1060                 if (client) {
1061                         struct lustre_handle lockh;
1062
1063                         unlock_res(res);
1064                         ldlm_lock2handle(lock, &lockh);
1065                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1066                         if (rc)
1067                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1068                 } else {
1069                         unlock_res(res);
1070                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
1071                                    "client node");
1072                         ldlm_lock_cancel(lock);
1073                 }
1074                 LDLM_LOCK_RELEASE(lock);
1075         } while (1);
1076 }
1077
1078 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1079                                struct hlist_node *hnode, void *arg)
1080 {
1081         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1082         __u64 flags = *(__u64 *)arg;
1083
1084         cleanup_resource(res, &res->lr_granted, flags);
1085         cleanup_resource(res, &res->lr_converting, flags);
1086         cleanup_resource(res, &res->lr_waiting, flags);
1087
1088         return 0;
1089 }
1090
1091 static int ldlm_resource_complain(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
1096         lock_res(res);
1097         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1098                "(%d) after lock cleanup; forcing cleanup.\n",
1099                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1100                atomic_read(&res->lr_refcount) - 1);
1101
1102         ldlm_resource_dump(D_ERROR, res);
1103         unlock_res(res);
1104         return 0;
1105 }
1106
1107 /**
1108  * Cancel and destroy all locks in the namespace.
1109  *
1110  * Typically used during evictions when server notified client that it was
1111  * evicted and all of its state needs to be destroyed.
1112  * Also used during shutdown.
1113  */
1114 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1115 {
1116         if (ns == NULL) {
1117                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1118                 return ELDLM_OK;
1119         }
1120
1121         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1122                                  &flags, 0);
1123         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1124                                  NULL, 0);
1125         return ELDLM_OK;
1126 }
1127 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1128
1129 /**
1130  * Attempts to free namespace.
1131  *
1132  * Only used when namespace goes away, like during an unmount.
1133  */
1134 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1135 {
1136         ENTRY;
1137
1138         /* At shutdown time, don't call the cancellation callback */
1139         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1140
1141         if (atomic_read(&ns->ns_bref) > 0) {
1142                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1143                 int rc;
1144                 CDEBUG(D_DLMTRACE,
1145                        "dlm namespace %s free waiting on refcount %d\n",
1146                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1147 force_wait:
1148                 if (force)
1149                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
1150                                           MSEC_PER_SEC) / 4, NULL, NULL);
1151
1152                 rc = l_wait_event(ns->ns_waitq,
1153                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1154
1155                 /* Forced cleanups should be able to reclaim all references,
1156                  * so it's safe to wait forever... we can't leak locks... */
1157                 if (force && rc == -ETIMEDOUT) {
1158                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1159                                        "namespace with %d resources in use, "
1160                                        "(rc=%d)\n", ldlm_ns_name(ns),
1161                                        atomic_read(&ns->ns_bref), rc);
1162                         GOTO(force_wait, rc);
1163                 }
1164
1165                 if (atomic_read(&ns->ns_bref)) {
1166                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1167                                        "with %d resources in use, (rc=%d)\n",
1168                                        ldlm_ns_name(ns),
1169                                        atomic_read(&ns->ns_bref), rc);
1170                         RETURN(ELDLM_NAMESPACE_EXISTS);
1171                 }
1172                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1173                        ldlm_ns_name(ns));
1174         }
1175
1176         RETURN(ELDLM_OK);
1177 }
1178
1179 /**
1180  * Performs various cleanups for passed \a ns to make it drop refc and be
1181  * ready for freeing. Waits for refc == 0.
1182  *
1183  * The following is done:
1184  * (0) Unregister \a ns from its list to make inaccessible for potential
1185  * users like pools thread and others;
1186  * (1) Clear all locks in \a ns.
1187  */
1188 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1189                                struct obd_import *imp,
1190                                int force)
1191 {
1192         int rc;
1193         ENTRY;
1194         if (!ns) {
1195                 EXIT;
1196                 return;
1197         }
1198
1199         spin_lock(&ns->ns_lock);
1200         ns->ns_stopping = 1;
1201         spin_unlock(&ns->ns_lock);
1202
1203         /*
1204          * Can fail with -EINTR when force == 0 in which case try harder.
1205          */
1206         rc = __ldlm_namespace_free(ns, force);
1207         if (rc != ELDLM_OK) {
1208                 if (imp) {
1209                         ptlrpc_disconnect_import(imp, 0);
1210                         ptlrpc_invalidate_import(imp);
1211                 }
1212
1213                 /*
1214                  * With all requests dropped and the import inactive
1215                  * we are gaurenteed all reference will be dropped.
1216                  */
1217                 rc = __ldlm_namespace_free(ns, 1);
1218                 LASSERT(rc == 0);
1219         }
1220         EXIT;
1221 }
1222 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1223
1224 /**
1225  * Performs freeing memory structures related to \a ns. This is only done
1226  * when ldlm_namespce_free_prior() successfully removed all resources
1227  * referencing \a ns and its refc == 0.
1228  */
1229 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1230 {
1231         ENTRY;
1232         if (!ns) {
1233                 EXIT;
1234                 return;
1235         }
1236
1237         /* Make sure that nobody can find this ns in its list. */
1238         ldlm_namespace_unregister(ns, ns->ns_client);
1239         /* Fini pool _before_ parent proc dir is removed. This is important as
1240          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1241          * Removing it after @dir may cause oops. */
1242         ldlm_pool_fini(&ns->ns_pool);
1243
1244         ldlm_namespace_proc_unregister(ns);
1245         ldlm_namespace_sysfs_unregister(ns);
1246         cfs_hash_putref(ns->ns_rs_hash);
1247         /* Namespace \a ns should be not on list at this time, otherwise
1248          * this will cause issues related to using freed \a ns in poold
1249          * thread. */
1250         LASSERT(list_empty(&ns->ns_list_chain));
1251         OBD_FREE_PTR(ns);
1252         ldlm_put_ref();
1253         EXIT;
1254 }
1255 EXPORT_SYMBOL(ldlm_namespace_free_post);
1256
1257 /**
1258  * Cleanup the resource, and free namespace.
1259  * bug 12864:
1260  * Deadlock issue:
1261  * proc1: destroy import
1262  *        class_disconnect_export(grab cl_sem) ->
1263  *              -> ldlm_namespace_free ->
1264  *              -> lprocfs_remove(grab _lprocfs_lock).
1265  * proc2: read proc info
1266  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1267  *              -> osc_rd_active, etc(grab cl_sem).
1268  *
1269  * So that I have to split the ldlm_namespace_free into two parts - the first
1270  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1271  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1272  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1273  * held.
1274  */
1275 void ldlm_namespace_free(struct ldlm_namespace *ns,
1276                          struct obd_import *imp,
1277                          int force)
1278 {
1279         ldlm_namespace_free_prior(ns, imp, force);
1280         ldlm_namespace_free_post(ns);
1281 }
1282 EXPORT_SYMBOL(ldlm_namespace_free);
1283
1284 void ldlm_namespace_get(struct ldlm_namespace *ns)
1285 {
1286         atomic_inc(&ns->ns_bref);
1287 }
1288
1289 /* This is only for callers that care about refcount */
1290 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1291 {
1292         return atomic_inc_return(&ns->ns_bref);
1293 }
1294
1295 void ldlm_namespace_put(struct ldlm_namespace *ns)
1296 {
1297         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1298                 wake_up(&ns->ns_waitq);
1299                 spin_unlock(&ns->ns_lock);
1300         }
1301 }
1302
1303 /** Register \a ns in the list of namespaces */
1304 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1305 {
1306         mutex_lock(ldlm_namespace_lock(client));
1307         LASSERT(list_empty(&ns->ns_list_chain));
1308         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1309         ldlm_namespace_nr_inc(client);
1310         mutex_unlock(ldlm_namespace_lock(client));
1311 }
1312
1313 /** Unregister \a ns from the list of namespaces. */
1314 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1315 {
1316         mutex_lock(ldlm_namespace_lock(client));
1317         LASSERT(!list_empty(&ns->ns_list_chain));
1318         /* Some asserts and possibly other parts of the code are still
1319          * using list_empty(&ns->ns_list_chain). This is why it is
1320          * important to use list_del_init() here. */
1321         list_del_init(&ns->ns_list_chain);
1322         ldlm_namespace_nr_dec(client);
1323         mutex_unlock(ldlm_namespace_lock(client));
1324 }
1325
1326 /** Should be called with ldlm_namespace_lock(client) taken. */
1327 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1328                                           enum ldlm_side client)
1329 {
1330         LASSERT(!list_empty(&ns->ns_list_chain));
1331         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1332         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1333 }
1334
1335 /** Should be called with ldlm_namespace_lock(client) taken. */
1336 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1337                                             enum ldlm_side client)
1338 {
1339         LASSERT(!list_empty(&ns->ns_list_chain));
1340         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1341         list_move_tail(&ns->ns_list_chain,
1342                        ldlm_namespace_inactive_list(client));
1343 }
1344
1345 /** Should be called with ldlm_namespace_lock(client) taken. */
1346 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1347 {
1348         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1349         LASSERT(!list_empty(ldlm_namespace_list(client)));
1350         return container_of(ldlm_namespace_list(client)->next,
1351                             struct ldlm_namespace, ns_list_chain);
1352 }
1353
1354 /** Create and initialize new resource. */
1355 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1356 {
1357         struct ldlm_resource *res;
1358         int idx;
1359
1360         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1361         if (res == NULL)
1362                 return NULL;
1363
1364         if (ldlm_type == LDLM_EXTENT) {
1365                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1366                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1367                 if (res->lr_itree == NULL) {
1368                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1369                         return NULL;
1370                 }
1371                 /* Initialize interval trees for each lock mode. */
1372                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1373                         res->lr_itree[idx].lit_size = 0;
1374                         res->lr_itree[idx].lit_mode = 1 << idx;
1375                         res->lr_itree[idx].lit_root = NULL;
1376                 }
1377         }
1378
1379         INIT_LIST_HEAD(&res->lr_granted);
1380         INIT_LIST_HEAD(&res->lr_converting);
1381         INIT_LIST_HEAD(&res->lr_waiting);
1382
1383         atomic_set(&res->lr_refcount, 1);
1384         spin_lock_init(&res->lr_lock);
1385         lu_ref_init(&res->lr_reference);
1386
1387         /* Since LVB init can be delayed now, there is no longer need to
1388          * immediatelly acquire mutex here. */
1389         mutex_init(&res->lr_lvb_mutex);
1390         res->lr_lvb_initialized = false;
1391
1392         return res;
1393 }
1394
1395 /**
1396  * Return a reference to resource with given name, creating it if necessary.
1397  * Args: namespace with ns_lock unlocked
1398  * Locks: takes and releases NS hash-lock and res->lr_lock
1399  * Returns: referenced, unlocked ldlm_resource or NULL
1400  */
1401 struct ldlm_resource *
1402 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1403                   const struct ldlm_res_id *name, enum ldlm_type type,
1404                   int create)
1405 {
1406         struct hlist_node       *hnode;
1407         struct ldlm_resource    *res = NULL;
1408         struct cfs_hash_bd              bd;
1409         __u64                   version;
1410         int                     ns_refcount = 0;
1411
1412         LASSERT(ns != NULL);
1413         LASSERT(parent == NULL);
1414         LASSERT(ns->ns_rs_hash != NULL);
1415         LASSERT(name->name[0] != 0);
1416
1417         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1418         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1419         if (hnode != NULL) {
1420                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1421                 GOTO(found, res);
1422         }
1423
1424         version = cfs_hash_bd_version_get(&bd);
1425         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1426
1427         if (create == 0)
1428                 return ERR_PTR(-ENOENT);
1429
1430         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1431                  "type: %d\n", type);
1432         res = ldlm_resource_new(type);
1433         if (res == NULL)
1434                 return ERR_PTR(-ENOMEM);
1435
1436         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1437         res->lr_name = *name;
1438         res->lr_type = type;
1439
1440         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1441         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1442                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1443
1444         if (hnode != NULL) {
1445                 /* Someone won the race and already added the resource. */
1446                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1447                 /* Clean lu_ref for failed resource. */
1448                 lu_ref_fini(&res->lr_reference);
1449                 if (res->lr_itree != NULL)
1450                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1451                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1452                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1453 found:
1454                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1455                 return res;
1456         }
1457         /* We won! Let's add the resource. */
1458         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1459         if (cfs_hash_bd_count_get(&bd) == 1)
1460                 ns_refcount = ldlm_namespace_get_return(ns);
1461
1462         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1463
1464         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1465
1466         /* Let's see if we happened to be the very first resource in this
1467          * namespace. If so, and this is a client namespace, we need to move
1468          * the namespace into the active namespaces list to be patrolled by
1469          * the ldlm_poold. */
1470         if (ns_is_client(ns) && ns_refcount == 1) {
1471                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1472                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1473                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1474         }
1475
1476         return res;
1477 }
1478 EXPORT_SYMBOL(ldlm_resource_get);
1479
1480 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1481 {
1482         LASSERT(res != NULL);
1483         LASSERT(res != LP_POISON);
1484         atomic_inc(&res->lr_refcount);
1485         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1486                atomic_read(&res->lr_refcount));
1487         return res;
1488 }
1489
1490 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1491                                          struct ldlm_resource *res)
1492 {
1493         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1494
1495         if (!list_empty(&res->lr_granted)) {
1496                 ldlm_resource_dump(D_ERROR, res);
1497                 LBUG();
1498         }
1499
1500         if (!list_empty(&res->lr_converting)) {
1501                 ldlm_resource_dump(D_ERROR, res);
1502                 LBUG();
1503         }
1504
1505         if (!list_empty(&res->lr_waiting)) {
1506                 ldlm_resource_dump(D_ERROR, res);
1507                 LBUG();
1508         }
1509
1510         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1511                                bd, &res->lr_hash);
1512         lu_ref_fini(&res->lr_reference);
1513         if (cfs_hash_bd_count_get(bd) == 0)
1514                 ldlm_namespace_put(nsb->nsb_namespace);
1515 }
1516
1517 /* Returns 1 if the resource was freed, 0 if it remains. */
1518 int ldlm_resource_putref(struct ldlm_resource *res)
1519 {
1520         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1521         struct cfs_hash_bd   bd;
1522
1523         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1524         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1525                res, atomic_read(&res->lr_refcount) - 1);
1526
1527         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1528         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1529                 __ldlm_resource_putref_final(&bd, res);
1530                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1531                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1532                         ns->ns_lvbo->lvbo_free(res);
1533                 if (res->lr_itree != NULL)
1534                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1535                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1536                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1537                 return 1;
1538         }
1539         return 0;
1540 }
1541 EXPORT_SYMBOL(ldlm_resource_putref);
1542
1543 /**
1544  * Add a lock into a given resource into specified lock list.
1545  */
1546 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1547                             struct ldlm_lock *lock)
1548 {
1549         check_res_locked(res);
1550
1551         LDLM_DEBUG(lock, "About to add this lock");
1552
1553         if (ldlm_is_destroyed(lock)) {
1554                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1555                 return;
1556         }
1557
1558         LASSERT(list_empty(&lock->l_res_link));
1559
1560         list_add_tail(&lock->l_res_link, head);
1561 }
1562
1563 /**
1564  * Insert a lock into resource after specified lock.
1565  *
1566  * Obtain resource description from the lock we are inserting after.
1567  */
1568 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1569                                      struct ldlm_lock *new)
1570 {
1571         struct ldlm_resource *res = original->l_resource;
1572
1573         check_res_locked(res);
1574
1575         ldlm_resource_dump(D_INFO, res);
1576         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1577
1578         if (ldlm_is_destroyed(new)) {
1579                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1580                 goto out;
1581         }
1582
1583         LASSERT(list_empty(&new->l_res_link));
1584
1585         list_add(&new->l_res_link, &original->l_res_link);
1586  out:;
1587 }
1588
1589 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1590 {
1591         int type = lock->l_resource->lr_type;
1592
1593         check_res_locked(lock->l_resource);
1594         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1595                 ldlm_unlink_lock_skiplist(lock);
1596         else if (type == LDLM_EXTENT)
1597                 ldlm_extent_unlink_lock(lock);
1598         list_del_init(&lock->l_res_link);
1599 }
1600 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1601
1602 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1603 {
1604         desc->lr_type = res->lr_type;
1605         desc->lr_name = res->lr_name;
1606 }
1607
1608 /**
1609  * Print information about all locks in all namespaces on this node to debug
1610  * log.
1611  */
1612 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1613 {
1614         struct list_head *tmp;
1615
1616         if (!((libcfs_debug | D_ERROR) & level))
1617                 return;
1618
1619         mutex_lock(ldlm_namespace_lock(client));
1620
1621         list_for_each(tmp, ldlm_namespace_list(client)) {
1622                 struct ldlm_namespace *ns;
1623
1624                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1625                 ldlm_namespace_dump(level, ns);
1626         }
1627
1628         mutex_unlock(ldlm_namespace_lock(client));
1629 }
1630
1631 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1632                               struct hlist_node *hnode, void *arg)
1633 {
1634         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1635         int    level = (int)(unsigned long)arg;
1636
1637         lock_res(res);
1638         ldlm_resource_dump(level, res);
1639         unlock_res(res);
1640
1641         return 0;
1642 }
1643
1644 /**
1645  * Print information about all locks in this namespace on this node to debug
1646  * log.
1647  */
1648 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1649 {
1650         if (!((libcfs_debug | D_ERROR) & level))
1651                 return;
1652
1653         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1654                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1655                ns_is_client(ns) ? "client" : "server");
1656
1657         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1658                 return;
1659
1660         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1661                                  ldlm_res_hash_dump,
1662                                  (void *)(unsigned long)level, 0);
1663         spin_lock(&ns->ns_lock);
1664         ns->ns_next_dump = cfs_time_shift(10);
1665         spin_unlock(&ns->ns_lock);
1666 }
1667
1668 /**
1669  * Print information about all locks in this resource to debug log.
1670  */
1671 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1672 {
1673         struct ldlm_lock *lock;
1674         unsigned int granted = 0;
1675
1676         CLASSERT(RES_NAME_SIZE == 4);
1677
1678         if (!((libcfs_debug | D_ERROR) & level))
1679                 return;
1680
1681         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1682                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1683
1684         if (!list_empty(&res->lr_granted)) {
1685                 CDEBUG(level, "Granted locks (in reverse order):\n");
1686                 list_for_each_entry_reverse(lock, &res->lr_granted,
1687                                                 l_res_link) {
1688                         LDLM_DEBUG_LIMIT(level, lock, "###");
1689                         if (!(level & D_CANTMASK) &&
1690                             ++granted > ldlm_dump_granted_max) {
1691                                 CDEBUG(level, "only dump %d granted locks to "
1692                                        "avoid DDOS.\n", granted);
1693                                 break;
1694                         }
1695                 }
1696         }
1697         if (!list_empty(&res->lr_converting)) {
1698                 CDEBUG(level, "Converting locks:\n");
1699                 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1700                         LDLM_DEBUG_LIMIT(level, lock, "###");
1701         }
1702         if (!list_empty(&res->lr_waiting)) {
1703                 CDEBUG(level, "Waiting locks:\n");
1704                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1705                         LDLM_DEBUG_LIMIT(level, lock, "###");
1706         }
1707 }
1708 EXPORT_SYMBOL(ldlm_resource_dump);