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 struct kmem_cache *ldlm_inodebits_slab;
  47
  48 int ldlm_srv_namespace_nr = 0;
  49 int ldlm_cli_namespace_nr = 0;
  50
  51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  52 LIST_HEAD(ldlm_srv_namespace_list);
  53
  54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
  55 /* Client Namespaces that have active resources in them.
  56  * Once all resources go away, ldlm_poold moves such namespaces to the
  57  * inactive list */
  58 LIST_HEAD(ldlm_cli_active_namespace_list);
  59 /* Client namespaces that don't have any locks in them */
  60 LIST_HEAD(ldlm_cli_inactive_namespace_list);
  61
  62 static struct dentry *ldlm_debugfs_dir;
  63 static struct dentry *ldlm_ns_debugfs_dir;
  64 struct dentry *ldlm_svc_debugfs_dir;
  65
  66 /* during debug dump certain amount of granted locks for one resource to avoid
  67  * DDOS. */
  68 static unsigned int ldlm_dump_granted_max = 256;
  69
  70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
  71                                           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
  79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
  80
  81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  82 {
  83         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  84         return 0;
  85 }
  86
  87 static ssize_t
  88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  89                        size_t count, loff_t *off)
  90 {
  91         struct seq_file *seq = file->private_data;
  92
  93         if (!count)
  94                 return 0;
  95
  96         return kstrtouint_from_user(buffer, count, 0,
  97                                     (unsigned int *)seq->private);
  98 }
  99
 100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
 101
 102 #ifdef HAVE_SERVER_SUPPORT
 103
 104 static int seq_watermark_show(struct seq_file *m, void *data)
 105 {
 106         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 107         return 0;
 108 }
 109
 110 static ssize_t seq_watermark_write(struct file *file,
 111                                    const char __user *buffer, size_t count,
 112                                    loff_t *off)
 113 {
 114         u64 value;
 115         __u64 watermark;
 116         __u64 *data = ((struct seq_file *)file->private_data)->private;
 117         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 118         char kernbuf[22] = "";
 119         int rc;
 120
 121         if (count >= sizeof(kernbuf))
 122                 return -EINVAL;
 123
 124         if (copy_from_user(kernbuf, buffer, count))
 125                 return -EFAULT;
 126         kernbuf[count] = 0;
 127
 128         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
 129         if (rc < 0) {
 130                 CERROR("Failed to set %s, rc = %d.\n",
 131                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 132                        rc);
 133                 return rc;
 134         } else if (value != 0 && value < (1 << 20)) {
 135                 CERROR("%s should be greater than 1MB.\n",
 136                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 137                 return -EINVAL;
 138         }
 139         watermark = value >> 20;
 140
 141         if (wm_low) {
 142                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 143                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 144                                "lock_limit_mb.\n");
 145                         return -EINVAL;
 146                 }
 147
 148                 *data = watermark;
 149                 if (watermark != 0) {
 150                         watermark <<= 20;
 151                         do_div(watermark, sizeof(struct ldlm_lock));
 152                 }
 153                 ldlm_reclaim_threshold = watermark;
 154         } else {
 155                 if (ldlm_reclaim_threshold_mb != 0 &&
 156                     watermark < ldlm_reclaim_threshold_mb) {
 157                         CERROR("lock_limit_mb must be greater than "
 158                                "lock_reclaim_threshold_mb.\n");
 159                         return -EINVAL;
 160                 }
 161
 162                 *data = watermark;
 163                 if (watermark != 0) {
 164                         watermark <<= 20;
 165                         do_div(watermark, sizeof(struct ldlm_lock));
 166                 }
 167                 ldlm_lock_limit = watermark;
 168         }
 169
 170         return count;
 171 }
 172
 173 static int seq_watermark_open(struct inode *inode, struct file *file)
 174 {
 175         return single_open(file, seq_watermark_show, inode->i_private);
 176 }
 177
 178 static const struct file_operations ldlm_watermark_fops = {
 179         .owner          = THIS_MODULE,
 180         .open           = seq_watermark_open,
 181         .read           = seq_read,
 182         .write          = seq_watermark_write,
 183         .llseek         = seq_lseek,
 184         .release        = lprocfs_single_release,
 185 };
 186
 187 static int seq_granted_show(struct seq_file *m, void *data)
 188 {
 189         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 190                    (struct percpu_counter *)m->private));
 191         return 0;
 192 }
 193
 194 static int seq_granted_open(struct inode *inode, struct file *file)
 195 {
 196         return single_open(file, seq_granted_show, inode->i_private);
 197 }
 198
 199 static const struct file_operations ldlm_granted_fops = {
 200         .owner  = THIS_MODULE,
 201         .open   = seq_granted_open,
 202         .read   = seq_read,
 203         .llseek = seq_lseek,
 204         .release = seq_release,
 205 };
 206
 207 #endif /* HAVE_SERVER_SUPPORT */
 208
 209 static struct lprocfs_vars ldlm_debugfs_list[] = {
 210         { .name =       "dump_namespaces",
 211           .fops =       &ldlm_dump_ns_fops,
 212           .proc_mode =  0222 },
 213         { .name =       "dump_granted_max",
 214           .fops =       &ldlm_rw_uint_fops,
 215           .data =       &ldlm_dump_granted_max },
 216 #ifdef HAVE_SERVER_SUPPORT
 217         { .name =       "lock_reclaim_threshold_mb",
 218           .fops =       &ldlm_watermark_fops,
 219           .data =       &ldlm_reclaim_threshold_mb },
 220         { .name =       "lock_limit_mb",
 221           .fops =       &ldlm_watermark_fops,
 222           .data =       &ldlm_lock_limit_mb },
 223         { .name =       "lock_granted_count",
 224           .fops =       &ldlm_granted_fops,
 225           .data =       &ldlm_granted_total },
 226 #endif
 227         { NULL }
 228 };
 229
 230 int ldlm_debugfs_setup(void)
 231 {
 232         int rc;
 233
 234         ENTRY;
 235         ldlm_debugfs_dir = ldebugfs_register(OBD_LDLM_DEVICENAME,
 236                                              debugfs_lustre_root,
 237                                              NULL, NULL);
 238         if (IS_ERR_OR_NULL(ldlm_debugfs_dir)) {
 239                 CERROR("LDebugFS failed in ldlm-init\n");
 240                 rc = ldlm_debugfs_dir ? PTR_ERR(ldlm_debugfs_dir) : -ENOMEM;
 241                 ldlm_debugfs_dir = NULL;
 242                 GOTO(err, rc);
 243         }
 244
 245         ldlm_ns_debugfs_dir = ldebugfs_register("namespaces",
 246                                                 ldlm_debugfs_dir,
 247                                                 NULL, NULL);
 248         if (IS_ERR_OR_NULL(ldlm_ns_debugfs_dir)) {
 249                 CERROR("LProcFS failed in ldlm-init\n");
 250                 rc = ldlm_ns_debugfs_dir ? PTR_ERR(ldlm_ns_debugfs_dir)
 251                                          : -ENOMEM;
 252                 GOTO(err, rc);
 253         }
 254
 255         ldlm_svc_debugfs_dir = ldebugfs_register("services",
 256                                                  ldlm_debugfs_dir,
 257                                                  NULL, NULL);
 258         if (IS_ERR_OR_NULL(ldlm_svc_debugfs_dir)) {
 259                 CERROR("LProcFS failed in ldlm-init\n");
 260                 rc = ldlm_svc_debugfs_dir ? PTR_ERR(ldlm_svc_debugfs_dir)
 261                                           : -ENOMEM;
 262                 GOTO(err, rc);
 263         }
 264
 265         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 266         if (rc != 0) {
 267                 CERROR("LProcFS failed in ldlm-init\n");
 268                 GOTO(err, rc);
 269         }
 270
 271         RETURN(0);
 272
 273 err:
 274         debugfs_remove_recursive(ldlm_debugfs_dir);
 275         ldlm_svc_debugfs_dir = NULL;
 276         ldlm_ns_debugfs_dir = NULL;
 277         ldlm_debugfs_dir = NULL;
 278         RETURN(rc);
 279 }
 280
 281 void ldlm_debugfs_cleanup(void)
 282 {
 283         debugfs_remove_recursive(ldlm_debugfs_dir);
 284
 285         ldlm_svc_debugfs_dir = NULL;
 286         ldlm_ns_debugfs_dir = NULL;
 287         ldlm_debugfs_dir = NULL;
 288 }
 289
 290 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 291                                    char *buf)
 292 {
 293         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 294                                                  ns_kobj);
 295         __u64                   res = 0;
 296         int                     i;
 297
 298         /* result is not strictly consistant */
 299         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
 300                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
 301         return sprintf(buf, "%lld\n", res);
 302 }
 303 LUSTRE_RO_ATTR(resource_count);
 304
 305 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 306                                char *buf)
 307 {
 308         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 309                                                  ns_kobj);
 310         __u64                   locks;
 311
 312         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 313                                         LPROCFS_FIELDS_FLAGS_SUM);
 314         return sprintf(buf, "%lld\n", locks);
 315 }
 316 LUSTRE_RO_ATTR(lock_count);
 317
 318 static ssize_t lock_unused_count_show(struct kobject *kobj,
 319                                       struct attribute *attr,
 320                                       char *buf)
 321 {
 322         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 323                                                  ns_kobj);
 324
 325         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 326 }
 327 LUSTRE_RO_ATTR(lock_unused_count);
 328
 329 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 330                              char *buf)
 331 {
 332         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 333                                                  ns_kobj);
 334         __u32 *nr = &ns->ns_max_unused;
 335
 336         if (ns_connect_lru_resize(ns))
 337                 nr = &ns->ns_nr_unused;
 338         return sprintf(buf, "%u\n", *nr);
 339 }
 340
 341 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 342                               const char *buffer, size_t count)
 343 {
 344         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 345                                                  ns_kobj);
 346         unsigned long tmp;
 347         int lru_resize;
 348         int err;
 349
 350         if (strncmp(buffer, "clear", 5) == 0) {
 351                 CDEBUG(D_DLMTRACE,
 352                        "dropping all unused locks from namespace %s\n",
 353                        ldlm_ns_name(ns));
 354                 if (ns_connect_lru_resize(ns)) {
 355                         /* Try to cancel all @ns_nr_unused locks. */
 356                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 357                                         LDLM_LRU_FLAG_PASSED |
 358                                         LDLM_LRU_FLAG_CLEANUP);
 359                 } else {
 360                         tmp = ns->ns_max_unused;
 361                         ns->ns_max_unused = 0;
 362                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 363                                         LDLM_LRU_FLAG_CLEANUP);
 364                         ns->ns_max_unused = tmp;
 365                 }
 366                 return count;
 367         }
 368
 369         err = kstrtoul(buffer, 10, &tmp);
 370         if (err != 0) {
 371                 CERROR("lru_size: invalid value written\n");
 372                 return -EINVAL;
 373         }
 374         lru_resize = (tmp == 0);
 375
 376         if (ns_connect_lru_resize(ns)) {
 377                 if (!lru_resize)
 378                         ns->ns_max_unused = (unsigned int)tmp;
 379
 380                 if (tmp > ns->ns_nr_unused)
 381                         tmp = ns->ns_nr_unused;
 382                 tmp = ns->ns_nr_unused - tmp;
 383
 384                 CDEBUG(D_DLMTRACE,
 385                        "changing namespace %s unused locks from %u to %u\n",
 386                        ldlm_ns_name(ns), ns->ns_nr_unused,
 387                        (unsigned int)tmp);
 388                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 389
 390                 if (!lru_resize) {
 391                         CDEBUG(D_DLMTRACE,
 392                                "disable lru_resize for namespace %s\n",
 393                                ldlm_ns_name(ns));
 394                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 395                 }
 396         } else {
 397                 CDEBUG(D_DLMTRACE,
 398                        "changing namespace %s max_unused from %u to %u\n",
 399                        ldlm_ns_name(ns), ns->ns_max_unused,
 400                        (unsigned int)tmp);
 401                 ns->ns_max_unused = (unsigned int)tmp;
 402                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 403
 404                 /* Make sure that LRU resize was originally supported before
 405                  * turning it on here.
 406                  */
 407                 if (lru_resize &&
 408                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 409                         CDEBUG(D_DLMTRACE,
 410                                "enable lru_resize for namespace %s\n",
 411                                ldlm_ns_name(ns));
 412                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 413                 }
 414         }
 415
 416         return count;
 417 }
 418 LUSTRE_RW_ATTR(lru_size);
 419
 420 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 421                                 char *buf)
 422 {
 423         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 424                                                  ns_kobj);
 425
 426         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 427 }
 428
 429 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 430                                  const char *buffer, size_t count)
 431 {
 432         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 433                                                  ns_kobj);
 434         int scale = NSEC_PER_MSEC;
 435         unsigned long long tmp;
 436         char *buf;
 437
 438         /* Did the user ask in seconds or milliseconds. Default is in ms */
 439         buf = strstr(buffer, "ms");
 440         if (!buf) {
 441                 buf = strchr(buffer, 's');
 442                 if (buf)
 443                         scale = NSEC_PER_SEC;
 444         }
 445
 446         if (buf)
 447                 *buf = '\0';
 448
 449         if (kstrtoull(buffer, 10, &tmp))
 450                 return -EINVAL;
 451
 452         ns->ns_max_age = ktime_set(0, tmp * scale);
 453
 454         return count;
 455 }
 456 LUSTRE_RW_ATTR(lru_max_age);
 457
 458 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 459                                       struct attribute *attr,
 460                                       char *buf)
 461 {
 462         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 463                                                  ns_kobj);
 464
 465         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 466 }
 467
 468 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 469                                        struct attribute *attr,
 470                                        const char *buffer,
 471                                        size_t count)
 472 {
 473         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 474                                                  ns_kobj);
 475         unsigned long supp = -1;
 476         int rc;
 477
 478         rc = kstrtoul(buffer, 10, &supp);
 479         if (rc < 0)
 480                 return rc;
 481
 482         if (supp == 0)
 483                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 484         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 485                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 486         return count;
 487 }
 488 LUSTRE_RW_ATTR(early_lock_cancel);
 489
 490 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 491                                     struct attribute *attr, char *buf)
 492 {
 493         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 494                                                  ns_kobj);
 495
 496         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 497 }
 498
 499 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 500                                      struct attribute *attr,
 501                                      const char *buffer, size_t count)
 502 {
 503         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 504                                                  ns_kobj);
 505         unsigned long long tmp;
 506
 507         if (kstrtoull(buffer, 10, &tmp))
 508                 return -EINVAL;
 509
 510         ns->ns_dirty_age_limit = tmp;
 511
 512         return count;
 513 }
 514 LUSTRE_RW_ATTR(dirty_age_limit);
 515
 516 #ifdef HAVE_SERVER_SUPPORT
 517 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 518                                     struct attribute *attr, char *buf)
 519 {
 520         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 521                                                  ns_kobj);
 522
 523         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 524 }
 525
 526 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 527                                      struct attribute *attr,
 528                                      const char *buffer, size_t count)
 529 {
 530         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 531                                                  ns_kobj);
 532         unsigned long long tmp;
 533
 534         if (kstrtoull(buffer, 10, &tmp))
 535                 return -EINVAL;
 536
 537         ns->ns_ctime_age_limit = tmp;
 538
 539         return count;
 540 }
 541 LUSTRE_RW_ATTR(ctime_age_limit);
 542
 543 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 544                                   char *buf)
 545 {
 546         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 547                                                  ns_kobj);
 548
 549         return sprintf(buf, "%d\n", ns->ns_timeouts);
 550 }
 551 LUSTRE_RO_ATTR(lock_timeouts);
 552
 553 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 554                                      struct attribute *attr, char *buf)
 555 {
 556         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 557                                                  ns_kobj);
 558
 559         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 560 }
 561
 562 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 563                                       struct attribute *attr,
 564                                       const char *buffer, size_t count)
 565 {
 566         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 567                                                  ns_kobj);
 568         unsigned long tmp;
 569         int err;
 570
 571         err = kstrtoul(buffer, 10, &tmp);
 572         if (err != 0)
 573                 return -EINVAL;
 574
 575         ns->ns_max_nolock_size = tmp;
 576
 577         return count;
 578 }
 579 LUSTRE_RW_ATTR(max_nolock_bytes);
 580
 581 static ssize_t contention_seconds_show(struct kobject *kobj,
 582                                        struct attribute *attr, char *buf)
 583 {
 584         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 585                                                  ns_kobj);
 586
 587         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 588 }
 589
 590 static ssize_t contention_seconds_store(struct kobject *kobj,
 591                                         struct attribute *attr,
 592                                         const char *buffer, size_t count)
 593 {
 594         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 595                                                  ns_kobj);
 596         unsigned long long tmp;
 597
 598         if (kstrtoull(buffer, 10, &tmp))
 599                 return -EINVAL;
 600
 601         ns->ns_contention_time = tmp;
 602
 603         return count;
 604 }
 605 LUSTRE_RW_ATTR(contention_seconds);
 606
 607 static ssize_t contended_locks_show(struct kobject *kobj,
 608                                     struct attribute *attr, char *buf)
 609 {
 610         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 611                                                  ns_kobj);
 612
 613         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 614 }
 615
 616 static ssize_t contended_locks_store(struct kobject *kobj,
 617                                      struct attribute *attr,
 618                                      const char *buffer, size_t count)
 619 {
 620         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 621                                                  ns_kobj);
 622         unsigned long tmp;
 623         int err;
 624
 625         err = kstrtoul(buffer, 10, &tmp);
 626         if (err != 0)
 627                 return -EINVAL;
 628
 629         ns->ns_contended_locks = tmp;
 630
 631         return count;
 632 }
 633 LUSTRE_RW_ATTR(contended_locks);
 634
 635 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 636                                      struct attribute *attr, char *buf)
 637 {
 638         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 639                                                  ns_kobj);
 640
 641         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 642 }
 643
 644 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 645                                       struct attribute *attr,
 646                                       const char *buffer, size_t count)
 647 {
 648         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 649                                                  ns_kobj);
 650         unsigned long tmp;
 651         int err;
 652
 653         err = kstrtoul(buffer, 10, &tmp);
 654         if (err != 0)
 655                 return -EINVAL;
 656
 657         ns->ns_max_parallel_ast = tmp;
 658
 659         return count;
 660 }
 661 LUSTRE_RW_ATTR(max_parallel_ast);
 662
 663 #endif /* HAVE_SERVER_SUPPORT */
 664
 665 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 666 static struct attribute *ldlm_ns_attrs[] = {
 667         &lustre_attr_resource_count.attr,
 668         &lustre_attr_lock_count.attr,
 669         &lustre_attr_lock_unused_count.attr,
 670         &lustre_attr_lru_size.attr,
 671         &lustre_attr_lru_max_age.attr,
 672         &lustre_attr_early_lock_cancel.attr,
 673         &lustre_attr_dirty_age_limit.attr,
 674 #ifdef HAVE_SERVER_SUPPORT
 675         &lustre_attr_ctime_age_limit.attr,
 676         &lustre_attr_lock_timeouts.attr,
 677         &lustre_attr_max_nolock_bytes.attr,
 678         &lustre_attr_contention_seconds.attr,
 679         &lustre_attr_contended_locks.attr,
 680         &lustre_attr_max_parallel_ast.attr,
 681 #endif
 682         NULL,
 683 };
 684
 685 static void ldlm_ns_release(struct kobject *kobj)
 686 {
 687         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 688                                                  ns_kobj);
 689         complete(&ns->ns_kobj_unregister);
 690 }
 691
 692 static struct kobj_type ldlm_ns_ktype = {
 693         .default_attrs  = ldlm_ns_attrs,
 694         .sysfs_ops      = &lustre_sysfs_ops,
 695         .release        = ldlm_ns_release,
 696 };
 697
 698 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 699 {
 700         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 701                 CERROR("dlm namespace %s has no procfs dir?\n",
 702                        ldlm_ns_name(ns));
 703         else
 704                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 705
 706         if (ns->ns_stats != NULL)
 707                 lprocfs_free_stats(&ns->ns_stats);
 708 }
 709
 710 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 711 {
 712         kobject_put(&ns->ns_kobj);
 713         wait_for_completion(&ns->ns_kobj_unregister);
 714 }
 715
 716 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 717 {
 718         int err;
 719
 720         ns->ns_kobj.kset = ldlm_ns_kset;
 721         init_completion(&ns->ns_kobj_unregister);
 722         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 723                                    "%s", ldlm_ns_name(ns));
 724
 725         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 726         if (!ns->ns_stats) {
 727                 kobject_put(&ns->ns_kobj);
 728                 return -ENOMEM;
 729         }
 730
 731         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 732                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 733
 734         return err;
 735 }
 736
 737 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 738 {
 739         struct dentry *ns_entry;
 740
 741         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 742                 ns_entry = ns->ns_debugfs_entry;
 743         } else {
 744                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 745                                               ldlm_ns_debugfs_dir);
 746                 if (!ns_entry)
 747                         return -ENOMEM;
 748                 ns->ns_debugfs_entry = ns_entry;
 749         }
 750
 751         return 0;
 752 }
 753 #undef MAX_STRING_SIZE
 754
 755 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 756                                   const void *key, unsigned int mask)
 757 {
 758         const struct ldlm_res_id *id = key;
 759         unsigned int val = 0;
 760         unsigned int i;
 761
 762         for (i = 0; i < RES_NAME_SIZE; i++)
 763                 val += id->name[i];
 764         return val & mask;
 765 }
 766
 767 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 768 {
 769         struct lu_fid       fid;
 770         __u32               hash;
 771         __u32               val;
 772
 773         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 774         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 775         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 776
 777         hash = fid_flatten32(&fid);
 778         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 779         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 780                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 781         } else {
 782                 val = fid_oid(&fid);
 783         }
 784         hash += (val >> 5) + (val << 11);
 785         return cfs_hash_32(hash, bits);
 786 }
 787
 788 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 789 {
 790         struct ldlm_resource   *res;
 791
 792         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 793         return &res->lr_name;
 794 }
 795
 796 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 797 {
 798         struct ldlm_resource   *res;
 799
 800         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 801         return ldlm_res_eq((const struct ldlm_res_id *)key,
 802                            (const struct ldlm_res_id *)&res->lr_name);
 803 }
 804
 805 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 806 {
 807         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 808 }
 809
 810 static void
 811 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 812 {
 813         struct ldlm_resource *res;
 814
 815         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 816         ldlm_resource_getref(res);
 817 }
 818
 819 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 820 {
 821         struct ldlm_resource *res;
 822
 823         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 824         ldlm_resource_putref(res);
 825 }
 826
 827 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 828         .hs_hash        = ldlm_res_hop_hash,
 829         .hs_key         = ldlm_res_hop_key,
 830         .hs_keycmp      = ldlm_res_hop_keycmp,
 831         .hs_keycpy      = NULL,
 832         .hs_object      = ldlm_res_hop_object,
 833         .hs_get         = ldlm_res_hop_get_locked,
 834         .hs_put         = ldlm_res_hop_put
 835 };
 836
 837 static struct {
 838         /** hash bucket bits */
 839         unsigned                nsd_bkt_bits;
 840         /** hash bits */
 841         unsigned                nsd_all_bits;
 842 } ldlm_ns_hash_defs[] = {
 843         [LDLM_NS_TYPE_MDC] = {
 844                 .nsd_bkt_bits   = 11,
 845                 .nsd_all_bits   = 16,
 846         },
 847         [LDLM_NS_TYPE_MDT] = {
 848                 .nsd_bkt_bits   = 14,
 849                 .nsd_all_bits   = 21,
 850         },
 851         [LDLM_NS_TYPE_OSC] = {
 852                 .nsd_bkt_bits   = 8,
 853                 .nsd_all_bits   = 12,
 854         },
 855         [LDLM_NS_TYPE_OST] = {
 856                 .nsd_bkt_bits   = 11,
 857                 .nsd_all_bits   = 17,
 858         },
 859         [LDLM_NS_TYPE_MGC] = {
 860                 .nsd_bkt_bits   = 3,
 861                 .nsd_all_bits   = 4,
 862         },
 863         [LDLM_NS_TYPE_MGT] = {
 864                 .nsd_bkt_bits   = 3,
 865                 .nsd_all_bits   = 4,
 866         },
 867 };
 868
 869 /**
 870  * Create and initialize new empty namespace.
 871  */
 872 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 873                                           enum ldlm_side client,
 874                                           enum ldlm_appetite apt,
 875                                           enum ldlm_ns_type ns_type)
 876 {
 877         struct ldlm_namespace *ns = NULL;
 878         int idx;
 879         int rc;
 880
 881         ENTRY;
 882         LASSERT(obd != NULL);
 883
 884         rc = ldlm_get_ref();
 885         if (rc) {
 886                 CERROR("ldlm_get_ref failed: %d\n", rc);
 887                 RETURN(NULL);
 888         }
 889
 890         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 891             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 892                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 893                 GOTO(out_ref, NULL);
 894         }
 895
 896         OBD_ALLOC_PTR(ns);
 897         if (!ns)
 898                 GOTO(out_ref, NULL);
 899
 900         ns->ns_rs_hash = cfs_hash_create(name,
 901                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 902                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 903                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 904                                          0,
 905                                          CFS_HASH_MIN_THETA,
 906                                          CFS_HASH_MAX_THETA,
 907                                          &ldlm_ns_hash_ops,
 908                                          CFS_HASH_DEPTH |
 909                                          CFS_HASH_BIGNAME |
 910                                          CFS_HASH_SPIN_BKTLOCK |
 911                                          CFS_HASH_NO_ITEMREF);
 912         if (ns->ns_rs_hash == NULL)
 913                 GOTO(out_ns, NULL);
 914
 915         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 916                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 917
 918         OBD_ALLOC_LARGE(ns->ns_rs_buckets,
 919                         BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 920         if (!ns->ns_rs_buckets)
 921                 goto out_hash;
 922
 923         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 924                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 925
 926                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 927                 nsb->nsb_namespace = ns;
 928                 nsb->nsb_reclaim_start = 0;
 929                 atomic_set(&nsb->nsb_count, 0);
 930         }
 931
 932         ns->ns_obd = obd;
 933         ns->ns_appetite = apt;
 934         ns->ns_client = client;
 935         ns->ns_name = kstrdup(name, GFP_KERNEL);
 936         if (!ns->ns_name)
 937                 goto out_hash;
 938
 939         INIT_LIST_HEAD(&ns->ns_list_chain);
 940         INIT_LIST_HEAD(&ns->ns_unused_list);
 941         spin_lock_init(&ns->ns_lock);
 942         atomic_set(&ns->ns_bref, 0);
 943         init_waitqueue_head(&ns->ns_waitq);
 944
 945         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 946         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 947         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 948
 949         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 950         ns->ns_nr_unused          = 0;
 951         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 952         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 953         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 954         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 955         ns->ns_timeouts           = 0;
 956         ns->ns_orig_connect_flags = 0;
 957         ns->ns_connect_flags      = 0;
 958         ns->ns_stopping           = 0;
 959         ns->ns_reclaim_start      = 0;
 960         ns->ns_last_pos           = &ns->ns_unused_list;
 961
 962         rc = ldlm_namespace_sysfs_register(ns);
 963         if (rc) {
 964                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 965                 GOTO(out_hash, rc);
 966         }
 967
 968         rc = ldlm_namespace_debugfs_register(ns);
 969         if (rc) {
 970                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 971                 GOTO(out_sysfs, rc);
 972         }
 973
 974         idx = ldlm_namespace_nr_read(client);
 975         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 976         if (rc) {
 977                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 978                 GOTO(out_proc, rc);
 979         }
 980
 981         ldlm_namespace_register(ns, client);
 982         RETURN(ns);
 983 out_proc:
 984         ldlm_namespace_debugfs_unregister(ns);
 985 out_sysfs:
 986         ldlm_namespace_sysfs_unregister(ns);
 987         ldlm_namespace_cleanup(ns, 0);
 988 out_hash:
 989         OBD_FREE_LARGE(ns->ns_rs_buckets,
 990                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 991         kfree(ns->ns_name);
 992         cfs_hash_putref(ns->ns_rs_hash);
 993 out_ns:
 994         OBD_FREE_PTR(ns);
 995 out_ref:
 996         ldlm_put_ref();
 997         RETURN(NULL);
 998 }
 999 EXPORT_SYMBOL(ldlm_namespace_new);
1000
1001 /**
1002  * Cancel and destroy all locks on a resource.
1003  *
1004  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1005  * clean up.  This is currently only used for recovery, and we make
1006  * certain assumptions as a result--notably, that we shouldn't cancel
1007  * locks with refs.
1008  */
1009 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1010                              __u64 flags)
1011 {
1012         struct list_head *tmp;
1013         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1014         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1015
1016         do {
1017                 struct ldlm_lock *lock = NULL;
1018
1019                 /* First, we look for non-cleaned-yet lock
1020                  * all cleaned locks are marked by CLEANED flag. */
1021                 lock_res(res);
1022                 list_for_each(tmp, q) {
1023                         lock = list_entry(tmp, struct ldlm_lock,
1024                                           l_res_link);
1025                         if (ldlm_is_cleaned(lock)) {
1026                                 lock = NULL;
1027                                 continue;
1028                         }
1029                         LDLM_LOCK_GET(lock);
1030                         ldlm_set_cleaned(lock);
1031                         break;
1032                 }
1033
1034                 if (lock == NULL) {
1035                         unlock_res(res);
1036                         break;
1037                 }
1038
1039                 /* Set CBPENDING so nothing in the cancellation path
1040                  * can match this lock. */
1041                 ldlm_set_cbpending(lock);
1042                 ldlm_set_failed(lock);
1043                 lock->l_flags |= flags;
1044
1045                 /* ... without sending a CANCEL message for local_only. */
1046                 if (local_only)
1047                         ldlm_set_local_only(lock);
1048
1049                 if (local_only && (lock->l_readers || lock->l_writers)) {
1050                         /*
1051                          * This is a little bit gross, but much better than the
1052                          * alternative: pretend that we got a blocking AST from
1053                          * the server, so that when the lock is decref'd, it
1054                          * will go away ...
1055                          */
1056                         unlock_res(res);
1057                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1058                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1059                                 set_current_state(TASK_UNINTERRUPTIBLE);
1060                                 schedule_timeout(cfs_time_seconds(4));
1061                                 set_current_state(TASK_RUNNING);
1062                         }
1063                         if (lock->l_completion_ast)
1064                                 lock->l_completion_ast(lock,
1065                                                        LDLM_FL_FAILED, NULL);
1066                         LDLM_LOCK_RELEASE(lock);
1067                         continue;
1068                 }
1069
1070                 if (client) {
1071                         struct lustre_handle lockh;
1072
1073                         unlock_res(res);
1074                         ldlm_lock2handle(lock, &lockh);
1075                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1076                         if (rc)
1077                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1078                 } else {
1079                         unlock_res(res);
1080                         LDLM_DEBUG(lock,
1081                                    "Freeing a lock still held by a client node");
1082                         ldlm_lock_cancel(lock);
1083                 }
1084                 LDLM_LOCK_RELEASE(lock);
1085         } while (1);
1086 }
1087
1088 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1089                                struct hlist_node *hnode, void *arg)
1090 {
1091         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1092         __u64 flags = *(__u64 *)arg;
1093
1094         cleanup_resource(res, &res->lr_granted, flags);
1095         cleanup_resource(res, &res->lr_waiting, flags);
1096
1097         return 0;
1098 }
1099
1100 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1101                                   struct hlist_node *hnode, void *arg)
1102 {
1103         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1104
1105         lock_res(res);
1106         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1107                "(%d) after lock cleanup; forcing cleanup.\n",
1108                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1109                atomic_read(&res->lr_refcount) - 1);
1110
1111         /* Use D_NETERROR since it is in the default mask */
1112         ldlm_resource_dump(D_NETERROR, res);
1113         unlock_res(res);
1114         return 0;
1115 }
1116
1117 /**
1118  * Cancel and destroy all locks in the namespace.
1119  *
1120  * Typically used during evictions when server notified client that it was
1121  * evicted and all of its state needs to be destroyed.
1122  * Also used during shutdown.
1123  */
1124 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1125 {
1126         if (ns == NULL) {
1127                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1128                 return ELDLM_OK;
1129         }
1130
1131         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1132                                  &flags, 0);
1133         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1134                                  NULL, 0);
1135         return ELDLM_OK;
1136 }
1137 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1138
1139 /**
1140  * Attempts to free namespace.
1141  *
1142  * Only used when namespace goes away, like during an unmount.
1143  */
1144 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1145 {
1146         ENTRY;
1147
1148         /* At shutdown time, don't call the cancellation callback */
1149         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1150
1151         if (atomic_read(&ns->ns_bref) > 0) {
1152                 int rc;
1153                 CDEBUG(D_DLMTRACE,
1154                        "dlm namespace %s free waiting on refcount %d\n",
1155                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1156 force_wait:
1157                 if (force)
1158                         rc = wait_event_idle_timeout(
1159                                 ns->ns_waitq,
1160                                 atomic_read(&ns->ns_bref) == 0,
1161                                 cfs_time_seconds(1) / 4);
1162                 else
1163                         rc = l_wait_event_abortable(
1164                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1165
1166                 /* Forced cleanups should be able to reclaim all references,
1167                  * so it's safe to wait forever... we can't leak locks... */
1168                 if (force && rc == 0) {
1169                         rc = -ETIMEDOUT;
1170                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1171                                        "namespace with %d resources in use, "
1172                                        "(rc=%d)\n", ldlm_ns_name(ns),
1173                                        atomic_read(&ns->ns_bref), rc);
1174                         GOTO(force_wait, rc);
1175                 }
1176
1177                 if (atomic_read(&ns->ns_bref)) {
1178                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1179                                        "with %d resources in use, (rc=%d)\n",
1180                                        ldlm_ns_name(ns),
1181                                        atomic_read(&ns->ns_bref), rc);
1182                         RETURN(ELDLM_NAMESPACE_EXISTS);
1183                 }
1184                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1185                        ldlm_ns_name(ns));
1186         }
1187
1188         RETURN(ELDLM_OK);
1189 }
1190
1191 /**
1192  * Performs various cleanups for passed \a ns to make it drop refc and be
1193  * ready for freeing. Waits for refc == 0.
1194  *
1195  * The following is done:
1196  * (0) Unregister \a ns from its list to make inaccessible for potential
1197  * users like pools thread and others;
1198  * (1) Clear all locks in \a ns.
1199  */
1200 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1201                                struct obd_import *imp,
1202                                int force)
1203 {
1204         int rc;
1205
1206         ENTRY;
1207         if (!ns) {
1208                 EXIT;
1209                 return;
1210         }
1211
1212         spin_lock(&ns->ns_lock);
1213         ns->ns_stopping = 1;
1214         spin_unlock(&ns->ns_lock);
1215
1216         /*
1217          * Can fail with -EINTR when force == 0 in which case try harder.
1218          */
1219         rc = __ldlm_namespace_free(ns, force);
1220         if (rc != ELDLM_OK) {
1221                 if (imp) {
1222                         ptlrpc_disconnect_import(imp, 0);
1223                         ptlrpc_invalidate_import(imp);
1224                 }
1225
1226                 /*
1227                  * With all requests dropped and the import inactive
1228                  * we are gaurenteed all reference will be dropped.
1229                  */
1230                 rc = __ldlm_namespace_free(ns, 1);
1231                 LASSERT(rc == 0);
1232         }
1233         EXIT;
1234 }
1235 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1236
1237 /**
1238  * Performs freeing memory structures related to \a ns. This is only done
1239  * when ldlm_namespce_free_prior() successfully removed all resources
1240  * referencing \a ns and its refc == 0.
1241  */
1242 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1243 {
1244         ENTRY;
1245         if (!ns) {
1246                 EXIT;
1247                 return;
1248         }
1249
1250         /* Make sure that nobody can find this ns in its list. */
1251         ldlm_namespace_unregister(ns, ns->ns_client);
1252         /* Fini pool _before_ parent proc dir is removed. This is important as
1253          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1254          * Removing it after @dir may cause oops. */
1255         ldlm_pool_fini(&ns->ns_pool);
1256
1257         ldlm_namespace_debugfs_unregister(ns);
1258         ldlm_namespace_sysfs_unregister(ns);
1259         cfs_hash_putref(ns->ns_rs_hash);
1260         OBD_FREE_LARGE(ns->ns_rs_buckets,
1261                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1262         kfree(ns->ns_name);
1263         /* Namespace \a ns should be not on list at this time, otherwise
1264          * this will cause issues related to using freed \a ns in poold
1265          * thread.
1266          */
1267         LASSERT(list_empty(&ns->ns_list_chain));
1268         OBD_FREE_PTR(ns);
1269         ldlm_put_ref();
1270         EXIT;
1271 }
1272 EXPORT_SYMBOL(ldlm_namespace_free_post);
1273
1274 /**
1275  * Cleanup the resource, and free namespace.
1276  * bug 12864:
1277  * Deadlock issue:
1278  * proc1: destroy import
1279  *        class_disconnect_export(grab cl_sem) ->
1280  *              -> ldlm_namespace_free ->
1281  *              -> lprocfs_remove(grab _lprocfs_lock).
1282  * proc2: read proc info
1283  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1284  *              -> osc_rd_active, etc(grab cl_sem).
1285  *
1286  * So that I have to split the ldlm_namespace_free into two parts - the first
1287  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1288  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1289  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1290  * held.
1291  */
1292 void ldlm_namespace_free(struct ldlm_namespace *ns,
1293                          struct obd_import *imp,
1294                          int force)
1295 {
1296         ldlm_namespace_free_prior(ns, imp, force);
1297         ldlm_namespace_free_post(ns);
1298 }
1299 EXPORT_SYMBOL(ldlm_namespace_free);
1300
1301 void ldlm_namespace_get(struct ldlm_namespace *ns)
1302 {
1303         atomic_inc(&ns->ns_bref);
1304 }
1305
1306 /* This is only for callers that care about refcount */
1307 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1308 {
1309         return atomic_inc_return(&ns->ns_bref);
1310 }
1311
1312 void ldlm_namespace_put(struct ldlm_namespace *ns)
1313 {
1314         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1315                 wake_up(&ns->ns_waitq);
1316                 spin_unlock(&ns->ns_lock);
1317         }
1318 }
1319
1320 /** Register \a ns in the list of namespaces */
1321 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1322 {
1323         mutex_lock(ldlm_namespace_lock(client));
1324         LASSERT(list_empty(&ns->ns_list_chain));
1325         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1326         ldlm_namespace_nr_inc(client);
1327         mutex_unlock(ldlm_namespace_lock(client));
1328 }
1329
1330 /** Unregister \a ns from the list of namespaces. */
1331 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1332 {
1333         mutex_lock(ldlm_namespace_lock(client));
1334         LASSERT(!list_empty(&ns->ns_list_chain));
1335         /* Some asserts and possibly other parts of the code are still
1336          * using list_empty(&ns->ns_list_chain). This is why it is
1337          * important to use list_del_init() here. */
1338         list_del_init(&ns->ns_list_chain);
1339         ldlm_namespace_nr_dec(client);
1340         mutex_unlock(ldlm_namespace_lock(client));
1341 }
1342
1343 /** Should be called with ldlm_namespace_lock(client) taken. */
1344 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1345                                           enum ldlm_side client)
1346 {
1347         LASSERT(!list_empty(&ns->ns_list_chain));
1348         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1349         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1350 }
1351
1352 /** Should be called with ldlm_namespace_lock(client) taken. */
1353 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1354                                             enum ldlm_side client)
1355 {
1356         LASSERT(!list_empty(&ns->ns_list_chain));
1357         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1358         list_move_tail(&ns->ns_list_chain,
1359                        ldlm_namespace_inactive_list(client));
1360 }
1361
1362 /** Should be called with ldlm_namespace_lock(client) taken. */
1363 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1364 {
1365         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1366         LASSERT(!list_empty(ldlm_namespace_list(client)));
1367         return container_of(ldlm_namespace_list(client)->next,
1368                             struct ldlm_namespace, ns_list_chain);
1369 }
1370
1371 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1372 {
1373         int idx;
1374
1375         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1376                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1377         if (res->lr_itree == NULL)
1378                 return false;
1379         /* Initialize interval trees for each lock mode. */
1380         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1381                 res->lr_itree[idx].lit_size = 0;
1382                 res->lr_itree[idx].lit_mode = 1 << idx;
1383                 res->lr_itree[idx].lit_root = NULL;
1384         }
1385         return true;
1386 }
1387
1388 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1389 {
1390         int i;
1391
1392         OBD_ALLOC_PTR(res->lr_ibits_queues);
1393         if (res->lr_ibits_queues == NULL)
1394                 return false;
1395         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1396                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1397         return true;
1398 }
1399
1400 /** Create and initialize new resource. */
1401 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1402 {
1403         struct ldlm_resource *res;
1404         bool rc;
1405
1406         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1407         if (res == NULL)
1408                 return NULL;
1409
1410         switch (ldlm_type) {
1411         case LDLM_EXTENT:
1412                 rc = ldlm_resource_extent_new(res);
1413                 break;
1414         case LDLM_IBITS:
1415                 rc = ldlm_resource_inodebits_new(res);
1416                 break;
1417         default:
1418                 rc = true;
1419                 break;
1420         }
1421         if (!rc) {
1422                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1423                 return NULL;
1424         }
1425
1426         INIT_LIST_HEAD(&res->lr_granted);
1427         INIT_LIST_HEAD(&res->lr_waiting);
1428
1429         atomic_set(&res->lr_refcount, 1);
1430         spin_lock_init(&res->lr_lock);
1431         lu_ref_init(&res->lr_reference);
1432
1433         /* Since LVB init can be delayed now, there is no longer need to
1434          * immediatelly acquire mutex here. */
1435         mutex_init(&res->lr_lvb_mutex);
1436         res->lr_lvb_initialized = false;
1437
1438         return res;
1439 }
1440
1441 static void ldlm_resource_free(struct ldlm_resource *res)
1442 {
1443         if (res->lr_type == LDLM_EXTENT) {
1444                 if (res->lr_itree != NULL)
1445                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1446                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1447         } else if (res->lr_type == LDLM_IBITS) {
1448                 if (res->lr_ibits_queues != NULL)
1449                         OBD_FREE_PTR(res->lr_ibits_queues);
1450         }
1451
1452         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1453 }
1454
1455 /**
1456  * Return a reference to resource with given name, creating it if necessary.
1457  * Args: namespace with ns_lock unlocked
1458  * Locks: takes and releases NS hash-lock and res->lr_lock
1459  * Returns: referenced, unlocked ldlm_resource or NULL
1460  */
1461 struct ldlm_resource *
1462 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1463                   const struct ldlm_res_id *name, enum ldlm_type type,
1464                   int create)
1465 {
1466         struct hlist_node       *hnode;
1467         struct ldlm_resource    *res = NULL;
1468         struct cfs_hash_bd              bd;
1469         __u64                   version;
1470         int                     ns_refcount = 0;
1471         int hash;
1472
1473         LASSERT(ns != NULL);
1474         LASSERT(parent == NULL);
1475         LASSERT(ns->ns_rs_hash != NULL);
1476         LASSERT(name->name[0] != 0);
1477
1478         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1479         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1480         if (hnode != NULL) {
1481                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1482                 GOTO(found, res);
1483         }
1484
1485         version = cfs_hash_bd_version_get(&bd);
1486         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1487
1488         if (create == 0)
1489                 return ERR_PTR(-ENOENT);
1490
1491         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1492                  "type: %d\n", type);
1493         res = ldlm_resource_new(type);
1494         if (res == NULL)
1495                 return ERR_PTR(-ENOMEM);
1496
1497         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1498         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1499         res->lr_name = *name;
1500         res->lr_type = type;
1501
1502         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1503         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1504                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1505
1506         if (hnode != NULL) {
1507                 /* Someone won the race and already added the resource. */
1508                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1509                 /* Clean lu_ref for failed resource. */
1510                 lu_ref_fini(&res->lr_reference);
1511                 ldlm_resource_free(res);
1512 found:
1513                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1514                 return res;
1515         }
1516         /* We won! Let's add the resource. */
1517         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1518         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1519                 ns_refcount = ldlm_namespace_get_return(ns);
1520
1521         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1522
1523         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1524
1525         /* Let's see if we happened to be the very first resource in this
1526          * namespace. If so, and this is a client namespace, we need to move
1527          * the namespace into the active namespaces list to be patrolled by
1528          * the ldlm_poold. */
1529         if (ns_is_client(ns) && ns_refcount == 1) {
1530                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1531                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1532                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1533         }
1534
1535         return res;
1536 }
1537 EXPORT_SYMBOL(ldlm_resource_get);
1538
1539 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1540 {
1541         LASSERT(res != NULL);
1542         LASSERT(res != LP_POISON);
1543         atomic_inc(&res->lr_refcount);
1544         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1545                atomic_read(&res->lr_refcount));
1546         return res;
1547 }
1548
1549 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1550                                          struct ldlm_resource *res)
1551 {
1552         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1553
1554         if (!list_empty(&res->lr_granted)) {
1555                 ldlm_resource_dump(D_ERROR, res);
1556                 LBUG();
1557         }
1558
1559         if (!list_empty(&res->lr_waiting)) {
1560                 ldlm_resource_dump(D_ERROR, res);
1561                 LBUG();
1562         }
1563
1564         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1565                                bd, &res->lr_hash);
1566         lu_ref_fini(&res->lr_reference);
1567         if (atomic_dec_and_test(&nsb->nsb_count))
1568                 ldlm_namespace_put(nsb->nsb_namespace);
1569 }
1570
1571 /* Returns 1 if the resource was freed, 0 if it remains. */
1572 int ldlm_resource_putref(struct ldlm_resource *res)
1573 {
1574         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1575         struct cfs_hash_bd   bd;
1576
1577         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1578         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1579                res, atomic_read(&res->lr_refcount) - 1);
1580
1581         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1582         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1583                 __ldlm_resource_putref_final(&bd, res);
1584                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1585                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1586                         ns->ns_lvbo->lvbo_free(res);
1587                 ldlm_resource_free(res);
1588                 return 1;
1589         }
1590         return 0;
1591 }
1592 EXPORT_SYMBOL(ldlm_resource_putref);
1593
1594 /**
1595  * Add a lock into a given resource into specified lock list.
1596  */
1597 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1598                             struct ldlm_lock *lock)
1599 {
1600         check_res_locked(res);
1601
1602         LDLM_DEBUG(lock, "About to add this lock");
1603
1604         if (ldlm_is_destroyed(lock)) {
1605                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1606                 return;
1607         }
1608
1609         LASSERT(list_empty(&lock->l_res_link));
1610
1611         list_add_tail(&lock->l_res_link, head);
1612
1613         if (res->lr_type == LDLM_IBITS)
1614                 ldlm_inodebits_add_lock(res, head, lock);
1615 }
1616
1617 /**
1618  * Insert a lock into resource after specified lock.
1619  *
1620  * Obtain resource description from the lock we are inserting after.
1621  */
1622 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1623                                      struct ldlm_lock *new)
1624 {
1625         struct ldlm_resource *res = original->l_resource;
1626
1627         check_res_locked(res);
1628
1629         ldlm_resource_dump(D_INFO, res);
1630         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1631
1632         if (ldlm_is_destroyed(new)) {
1633                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1634                 goto out;
1635         }
1636
1637         LASSERT(list_empty(&new->l_res_link));
1638
1639         list_add(&new->l_res_link, &original->l_res_link);
1640  out:;
1641 }
1642
1643 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1644 {
1645         int type = lock->l_resource->lr_type;
1646
1647         check_res_locked(lock->l_resource);
1648         switch (type) {
1649         case LDLM_PLAIN:
1650                 ldlm_unlink_lock_skiplist(lock);
1651                 break;
1652         case LDLM_EXTENT:
1653                 ldlm_extent_unlink_lock(lock);
1654                 break;
1655         case LDLM_IBITS:
1656                 ldlm_inodebits_unlink_lock(lock);
1657                 break;
1658         }
1659         list_del_init(&lock->l_res_link);
1660 }
1661 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1662
1663 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1664 {
1665         desc->lr_type = res->lr_type;
1666         desc->lr_name = res->lr_name;
1667 }
1668
1669 /**
1670  * Print information about all locks in all namespaces on this node to debug
1671  * log.
1672  */
1673 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1674 {
1675         struct list_head *tmp;
1676
1677         if (!((libcfs_debug | D_ERROR) & level))
1678                 return;
1679
1680         mutex_lock(ldlm_namespace_lock(client));
1681
1682         list_for_each(tmp, ldlm_namespace_list(client)) {
1683                 struct ldlm_namespace *ns;
1684
1685                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1686                 ldlm_namespace_dump(level, ns);
1687         }
1688
1689         mutex_unlock(ldlm_namespace_lock(client));
1690 }
1691
1692 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1693                               struct hlist_node *hnode, void *arg)
1694 {
1695         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1696         int    level = (int)(unsigned long)arg;
1697
1698         lock_res(res);
1699         ldlm_resource_dump(level, res);
1700         unlock_res(res);
1701
1702         return 0;
1703 }
1704
1705 /**
1706  * Print information about all locks in this namespace on this node to debug
1707  * log.
1708  */
1709 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1710 {
1711         if (!((libcfs_debug | D_ERROR) & level))
1712                 return;
1713
1714         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1715                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1716                ns_is_client(ns) ? "client" : "server");
1717
1718         if (ktime_get_seconds() < ns->ns_next_dump)
1719                 return;
1720
1721         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1722                                  ldlm_res_hash_dump,
1723                                  (void *)(unsigned long)level, 0);
1724         spin_lock(&ns->ns_lock);
1725         ns->ns_next_dump = ktime_get_seconds() + 10;
1726         spin_unlock(&ns->ns_lock);
1727 }
1728
1729 /**
1730  * Print information about all locks in this resource to debug log.
1731  */
1732 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1733 {
1734         struct ldlm_lock *lock;
1735         unsigned int granted = 0;
1736
1737         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1738
1739         if (!((libcfs_debug | D_ERROR) & level))
1740                 return;
1741
1742         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1743                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1744
1745         if (!list_empty(&res->lr_granted)) {
1746                 CDEBUG(level, "Granted locks (in reverse order):\n");
1747                 list_for_each_entry_reverse(lock, &res->lr_granted,
1748                                                 l_res_link) {
1749                         LDLM_DEBUG_LIMIT(level, lock, "###");
1750                         if (!(level & D_CANTMASK) &&
1751                             ++granted > ldlm_dump_granted_max) {
1752                                 CDEBUG(level,
1753                                        "only dump %d granted locks to avoid DDOS.\n",
1754                                        granted);
1755                                 break;
1756                         }
1757                 }
1758         }
1759
1760         if (!list_empty(&res->lr_waiting)) {
1761                 CDEBUG(level, "Waiting locks:\n");
1762                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1763                         LDLM_DEBUG_LIMIT(level, lock, "###");
1764         }
1765 }
1766 EXPORT_SYMBOL(ldlm_resource_dump);