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         __s64 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         int rc;
 119
 120         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 121         if (rc) {
 122                 CERROR("Failed to set %s, rc = %d.\n",
 123                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 124                        rc);
 125                 return rc;
 126         } else if (value != 0 && value < (1 << 20)) {
 127                 CERROR("%s should be greater than 1MB.\n",
 128                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 129                 return -EINVAL;
 130         }
 131         watermark = value >> 20;
 132
 133         if (wm_low) {
 134                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 135                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 136                                "lock_limit_mb.\n");
 137                         return -EINVAL;
 138                 }
 139
 140                 *data = watermark;
 141                 if (watermark != 0) {
 142                         watermark <<= 20;
 143                         do_div(watermark, sizeof(struct ldlm_lock));
 144                 }
 145                 ldlm_reclaim_threshold = watermark;
 146         } else {
 147                 if (ldlm_reclaim_threshold_mb != 0 &&
 148                     watermark < ldlm_reclaim_threshold_mb) {
 149                         CERROR("lock_limit_mb must be greater than "
 150                                "lock_reclaim_threshold_mb.\n");
 151                         return -EINVAL;
 152                 }
 153
 154                 *data = watermark;
 155                 if (watermark != 0) {
 156                         watermark <<= 20;
 157                         do_div(watermark, sizeof(struct ldlm_lock));
 158                 }
 159                 ldlm_lock_limit = watermark;
 160         }
 161
 162         return count;
 163 }
 164
 165 static int seq_watermark_open(struct inode *inode, struct file *file)
 166 {
 167         return single_open(file, seq_watermark_show, inode->i_private);
 168 }
 169
 170 static const struct file_operations ldlm_watermark_fops = {
 171         .owner          = THIS_MODULE,
 172         .open           = seq_watermark_open,
 173         .read           = seq_read,
 174         .write          = seq_watermark_write,
 175         .llseek         = seq_lseek,
 176         .release        = lprocfs_single_release,
 177 };
 178
 179 static int seq_granted_show(struct seq_file *m, void *data)
 180 {
 181         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 182                    (struct percpu_counter *)m->private));
 183         return 0;
 184 }
 185
 186 static int seq_granted_open(struct inode *inode, struct file *file)
 187 {
 188         return single_open(file, seq_granted_show, inode->i_private);
 189 }
 190
 191 static const struct file_operations ldlm_granted_fops = {
 192         .owner  = THIS_MODULE,
 193         .open   = seq_granted_open,
 194         .read   = seq_read,
 195         .llseek = seq_lseek,
 196         .release = seq_release,
 197 };
 198
 199 #endif /* HAVE_SERVER_SUPPORT */
 200
 201 static struct lprocfs_vars ldlm_debugfs_list[] = {
 202         { .name =       "dump_namespaces",
 203           .fops =       &ldlm_dump_ns_fops,
 204           .proc_mode =  0222 },
 205         { .name =       "dump_granted_max",
 206           .fops =       &ldlm_rw_uint_fops,
 207           .data =       &ldlm_dump_granted_max },
 208 #ifdef HAVE_SERVER_SUPPORT
 209         { .name =       "lock_reclaim_threshold_mb",
 210           .fops =       &ldlm_watermark_fops,
 211           .data =       &ldlm_reclaim_threshold_mb },
 212         { .name =       "lock_limit_mb",
 213           .fops =       &ldlm_watermark_fops,
 214           .data =       &ldlm_lock_limit_mb },
 215         { .name =       "lock_granted_count",
 216           .fops =       &ldlm_granted_fops,
 217           .data =       &ldlm_granted_total },
 218 #endif
 219         { NULL }
 220 };
 221
 222 int ldlm_debugfs_setup(void)
 223 {
 224         int rc;
 225
 226         ENTRY;
 227         ldlm_debugfs_dir = ldebugfs_register(OBD_LDLM_DEVICENAME,
 228                                              debugfs_lustre_root,
 229                                              NULL, NULL);
 230         if (IS_ERR_OR_NULL(ldlm_debugfs_dir)) {
 231                 CERROR("LDebugFS failed in ldlm-init\n");
 232                 rc = ldlm_debugfs_dir ? PTR_ERR(ldlm_debugfs_dir) : -ENOMEM;
 233                 GOTO(err, rc);
 234         }
 235
 236         ldlm_ns_debugfs_dir = ldebugfs_register("namespaces",
 237                                                 ldlm_debugfs_dir,
 238                                                 NULL, NULL);
 239         if (IS_ERR_OR_NULL(ldlm_ns_debugfs_dir)) {
 240                 CERROR("LProcFS failed in ldlm-init\n");
 241                 rc = ldlm_ns_debugfs_dir ? PTR_ERR(ldlm_ns_debugfs_dir)
 242                                          : -ENOMEM;
 243                 GOTO(err_type, rc);
 244         }
 245
 246         ldlm_svc_debugfs_dir = ldebugfs_register("services",
 247                                                  ldlm_debugfs_dir,
 248                                                  NULL, NULL);
 249         if (IS_ERR_OR_NULL(ldlm_svc_debugfs_dir)) {
 250                 CERROR("LProcFS failed in ldlm-init\n");
 251                 rc = ldlm_svc_debugfs_dir ? PTR_ERR(ldlm_svc_debugfs_dir)
 252                                           : -ENOMEM;
 253                 GOTO(err_ns, rc);
 254         }
 255
 256         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 257         if (rc != 0) {
 258                 CERROR("LProcFS failed in ldlm-init\n");
 259                 GOTO(err_svc, rc);
 260         }
 261
 262         RETURN(0);
 263
 264 err_svc:
 265         ldebugfs_remove(&ldlm_svc_debugfs_dir);
 266 err_ns:
 267         ldebugfs_remove(&ldlm_ns_debugfs_dir);
 268 err_type:
 269         ldebugfs_remove(&ldlm_debugfs_dir);
 270 err:
 271         ldlm_svc_debugfs_dir = NULL;
 272         ldlm_ns_debugfs_dir = NULL;
 273         ldlm_debugfs_dir = NULL;
 274         RETURN(rc);
 275 }
 276
 277 void ldlm_debugfs_cleanup(void)
 278 {
 279         if (!IS_ERR_OR_NULL(ldlm_svc_debugfs_dir))
 280                 ldebugfs_remove(&ldlm_svc_debugfs_dir);
 281
 282         if (!IS_ERR_OR_NULL(ldlm_ns_debugfs_dir))
 283                 ldebugfs_remove(&ldlm_ns_debugfs_dir);
 284
 285         if (!IS_ERR_OR_NULL(ldlm_debugfs_dir))
 286                 ldebugfs_remove(&ldlm_debugfs_dir);
 287
 288         ldlm_svc_debugfs_dir = NULL;
 289         ldlm_ns_debugfs_dir = NULL;
 290         ldlm_debugfs_dir = NULL;
 291 }
 292
 293 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 294                                    char *buf)
 295 {
 296         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 297                                                  ns_kobj);
 298         __u64                   res = 0;
 299         struct cfs_hash_bd              bd;
 300         int                     i;
 301
 302         /* result is not strictly consistant */
 303         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 304                 res += cfs_hash_bd_count_get(&bd);
 305         return sprintf(buf, "%lld\n", res);
 306 }
 307 LUSTRE_RO_ATTR(resource_count);
 308
 309 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 310                                char *buf)
 311 {
 312         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 313                                                  ns_kobj);
 314         __u64                   locks;
 315
 316         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 317                                         LPROCFS_FIELDS_FLAGS_SUM);
 318         return sprintf(buf, "%lld\n", locks);
 319 }
 320 LUSTRE_RO_ATTR(lock_count);
 321
 322 static ssize_t lock_unused_count_show(struct kobject *kobj,
 323                                       struct attribute *attr,
 324                                       char *buf)
 325 {
 326         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 327                                                  ns_kobj);
 328
 329         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 330 }
 331 LUSTRE_RO_ATTR(lock_unused_count);
 332
 333 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 334                              char *buf)
 335 {
 336         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 337                                                  ns_kobj);
 338         __u32 *nr = &ns->ns_max_unused;
 339
 340         if (ns_connect_lru_resize(ns))
 341                 nr = &ns->ns_nr_unused;
 342         return sprintf(buf, "%u\n", *nr);
 343 }
 344
 345 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 346                               const char *buffer, size_t count)
 347 {
 348         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 349                                                  ns_kobj);
 350         unsigned long tmp;
 351         int lru_resize;
 352         int err;
 353
 354         if (strncmp(buffer, "clear", 5) == 0) {
 355                 CDEBUG(D_DLMTRACE,
 356                        "dropping all unused locks from namespace %s\n",
 357                        ldlm_ns_name(ns));
 358                 if (ns_connect_lru_resize(ns)) {
 359                         /* Try to cancel all @ns_nr_unused locks. */
 360                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 361                                         LDLM_LRU_FLAG_PASSED |
 362                                         LDLM_LRU_FLAG_CLEANUP);
 363                 } else {
 364                         tmp = ns->ns_max_unused;
 365                         ns->ns_max_unused = 0;
 366                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 367                                         LDLM_LRU_FLAG_CLEANUP);
 368                         ns->ns_max_unused = tmp;
 369                 }
 370                 return count;
 371         }
 372
 373         err = kstrtoul(buffer, 10, &tmp);
 374         if (err != 0) {
 375                 CERROR("lru_size: invalid value written\n");
 376                 return -EINVAL;
 377         }
 378         lru_resize = (tmp == 0);
 379
 380         if (ns_connect_lru_resize(ns)) {
 381                 if (!lru_resize)
 382                         ns->ns_max_unused = (unsigned int)tmp;
 383
 384                 if (tmp > ns->ns_nr_unused)
 385                         tmp = ns->ns_nr_unused;
 386                 tmp = ns->ns_nr_unused - tmp;
 387
 388                 CDEBUG(D_DLMTRACE,
 389                        "changing namespace %s unused locks from %u to %u\n",
 390                        ldlm_ns_name(ns), ns->ns_nr_unused,
 391                        (unsigned int)tmp);
 392                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 393
 394                 if (!lru_resize) {
 395                         CDEBUG(D_DLMTRACE,
 396                                "disable lru_resize for namespace %s\n",
 397                                ldlm_ns_name(ns));
 398                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 399                 }
 400         } else {
 401                 CDEBUG(D_DLMTRACE,
 402                        "changing namespace %s max_unused from %u to %u\n",
 403                        ldlm_ns_name(ns), ns->ns_max_unused,
 404                        (unsigned int)tmp);
 405                 ns->ns_max_unused = (unsigned int)tmp;
 406                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 407
 408                 /* Make sure that LRU resize was originally supported before
 409                  * turning it on here.
 410                  */
 411                 if (lru_resize &&
 412                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 413                         CDEBUG(D_DLMTRACE,
 414                                "enable lru_resize for namespace %s\n",
 415                                ldlm_ns_name(ns));
 416                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 417                 }
 418         }
 419
 420         return count;
 421 }
 422 LUSTRE_RW_ATTR(lru_size);
 423
 424 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 425                                 char *buf)
 426 {
 427         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 428                                                  ns_kobj);
 429
 430         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 431 }
 432
 433 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 434                                  const char *buffer, size_t count)
 435 {
 436         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 437                                                  ns_kobj);
 438         int scale = NSEC_PER_MSEC;
 439         unsigned long long tmp;
 440         char *buf;
 441
 442         /* Did the user ask in seconds or milliseconds. Default is in ms */
 443         buf = strstr(buffer, "ms");
 444         if (!buf) {
 445                 buf = strchr(buffer, 's');
 446                 if (buf)
 447                         scale = NSEC_PER_SEC;
 448         }
 449
 450         if (buf)
 451                 *buf = '\0';
 452
 453         if (kstrtoull(buffer, 10, &tmp))
 454                 return -EINVAL;
 455
 456         ns->ns_max_age = ktime_set(0, tmp * scale);
 457
 458         return count;
 459 }
 460 LUSTRE_RW_ATTR(lru_max_age);
 461
 462 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 463                                       struct attribute *attr,
 464                                       char *buf)
 465 {
 466         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 467                                                  ns_kobj);
 468
 469         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 470 }
 471
 472 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 473                                        struct attribute *attr,
 474                                        const char *buffer,
 475                                        size_t count)
 476 {
 477         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 478                                                  ns_kobj);
 479         unsigned long supp = -1;
 480         int rc;
 481
 482         rc = kstrtoul(buffer, 10, &supp);
 483         if (rc < 0)
 484                 return rc;
 485
 486         if (supp == 0)
 487                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 488         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 489                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 490         return count;
 491 }
 492 LUSTRE_RW_ATTR(early_lock_cancel);
 493
 494 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 495                                     struct attribute *attr, char *buf)
 496 {
 497         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 498                                                  ns_kobj);
 499
 500         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 501 }
 502
 503 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 504                                      struct attribute *attr,
 505                                      const char *buffer, size_t count)
 506 {
 507         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 508                                                  ns_kobj);
 509         unsigned long long tmp;
 510
 511         if (kstrtoull(buffer, 10, &tmp))
 512                 return -EINVAL;
 513
 514         ns->ns_dirty_age_limit = tmp;
 515
 516         return count;
 517 }
 518 LUSTRE_RW_ATTR(dirty_age_limit);
 519
 520 #ifdef HAVE_SERVER_SUPPORT
 521 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 522                                     struct attribute *attr, char *buf)
 523 {
 524         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 525                                                  ns_kobj);
 526
 527         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 528 }
 529
 530 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 531                                      struct attribute *attr,
 532                                      const char *buffer, size_t count)
 533 {
 534         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 535                                                  ns_kobj);
 536         unsigned long long tmp;
 537
 538         if (kstrtoull(buffer, 10, &tmp))
 539                 return -EINVAL;
 540
 541         ns->ns_ctime_age_limit = tmp;
 542
 543         return count;
 544 }
 545 LUSTRE_RW_ATTR(ctime_age_limit);
 546
 547 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 548                                   char *buf)
 549 {
 550         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 551                                                  ns_kobj);
 552
 553         return sprintf(buf, "%d\n", ns->ns_timeouts);
 554 }
 555 LUSTRE_RO_ATTR(lock_timeouts);
 556
 557 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 558                                      struct attribute *attr, char *buf)
 559 {
 560         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 561                                                  ns_kobj);
 562
 563         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 564 }
 565
 566 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 567                                       struct attribute *attr,
 568                                       const char *buffer, size_t count)
 569 {
 570         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 571                                                  ns_kobj);
 572         unsigned long tmp;
 573         int err;
 574
 575         err = kstrtoul(buffer, 10, &tmp);
 576         if (err != 0)
 577                 return -EINVAL;
 578
 579         ns->ns_max_nolock_size = tmp;
 580
 581         return count;
 582 }
 583 LUSTRE_RW_ATTR(max_nolock_bytes);
 584
 585 static ssize_t contention_seconds_show(struct kobject *kobj,
 586                                        struct attribute *attr, char *buf)
 587 {
 588         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 589                                                  ns_kobj);
 590
 591         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 592 }
 593
 594 static ssize_t contention_seconds_store(struct kobject *kobj,
 595                                         struct attribute *attr,
 596                                         const char *buffer, size_t count)
 597 {
 598         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 599                                                  ns_kobj);
 600         unsigned long long tmp;
 601
 602         if (kstrtoull(buffer, 10, &tmp))
 603                 return -EINVAL;
 604
 605         ns->ns_contention_time = tmp;
 606
 607         return count;
 608 }
 609 LUSTRE_RW_ATTR(contention_seconds);
 610
 611 static ssize_t contended_locks_show(struct kobject *kobj,
 612                                     struct attribute *attr, char *buf)
 613 {
 614         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 615                                                  ns_kobj);
 616
 617         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 618 }
 619
 620 static ssize_t contended_locks_store(struct kobject *kobj,
 621                                      struct attribute *attr,
 622                                      const char *buffer, size_t count)
 623 {
 624         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 625                                                  ns_kobj);
 626         unsigned long tmp;
 627         int err;
 628
 629         err = kstrtoul(buffer, 10, &tmp);
 630         if (err != 0)
 631                 return -EINVAL;
 632
 633         ns->ns_contended_locks = tmp;
 634
 635         return count;
 636 }
 637 LUSTRE_RW_ATTR(contended_locks);
 638
 639 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 640                                      struct attribute *attr, char *buf)
 641 {
 642         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 643                                                  ns_kobj);
 644
 645         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 646 }
 647
 648 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 649                                       struct attribute *attr,
 650                                       const char *buffer, size_t count)
 651 {
 652         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 653                                                  ns_kobj);
 654         unsigned long tmp;
 655         int err;
 656
 657         err = kstrtoul(buffer, 10, &tmp);
 658         if (err != 0)
 659                 return -EINVAL;
 660
 661         ns->ns_max_parallel_ast = tmp;
 662
 663         return count;
 664 }
 665 LUSTRE_RW_ATTR(max_parallel_ast);
 666
 667 #endif /* HAVE_SERVER_SUPPORT */
 668
 669 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 670 static struct attribute *ldlm_ns_attrs[] = {
 671         &lustre_attr_resource_count.attr,
 672         &lustre_attr_lock_count.attr,
 673         &lustre_attr_lock_unused_count.attr,
 674         &lustre_attr_lru_size.attr,
 675         &lustre_attr_lru_max_age.attr,
 676         &lustre_attr_early_lock_cancel.attr,
 677         &lustre_attr_dirty_age_limit.attr,
 678 #ifdef HAVE_SERVER_SUPPORT
 679         &lustre_attr_ctime_age_limit.attr,
 680         &lustre_attr_lock_timeouts.attr,
 681         &lustre_attr_max_nolock_bytes.attr,
 682         &lustre_attr_contention_seconds.attr,
 683         &lustre_attr_contended_locks.attr,
 684         &lustre_attr_max_parallel_ast.attr,
 685 #endif
 686         NULL,
 687 };
 688
 689 static void ldlm_ns_release(struct kobject *kobj)
 690 {
 691         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 692                                                  ns_kobj);
 693         complete(&ns->ns_kobj_unregister);
 694 }
 695
 696 static struct kobj_type ldlm_ns_ktype = {
 697         .default_attrs  = ldlm_ns_attrs,
 698         .sysfs_ops      = &lustre_sysfs_ops,
 699         .release        = ldlm_ns_release,
 700 };
 701
 702 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 703 {
 704         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 705                 CERROR("dlm namespace %s has no procfs dir?\n",
 706                        ldlm_ns_name(ns));
 707         else
 708                 ldebugfs_remove(&ns->ns_debugfs_entry);
 709
 710         if (ns->ns_stats != NULL)
 711                 lprocfs_free_stats(&ns->ns_stats);
 712 }
 713
 714 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 715 {
 716         kobject_put(&ns->ns_kobj);
 717         wait_for_completion(&ns->ns_kobj_unregister);
 718 }
 719
 720 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 721 {
 722         int err;
 723
 724         ns->ns_kobj.kset = ldlm_ns_kset;
 725         init_completion(&ns->ns_kobj_unregister);
 726         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 727                                    "%s", ldlm_ns_name(ns));
 728
 729         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 730         if (!ns->ns_stats) {
 731                 kobject_put(&ns->ns_kobj);
 732                 return -ENOMEM;
 733         }
 734
 735         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 736                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 737
 738         return err;
 739 }
 740
 741 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 742 {
 743         struct dentry *ns_entry;
 744
 745         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 746                 ns_entry = ns->ns_debugfs_entry;
 747         } else {
 748                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 749                                               ldlm_ns_debugfs_dir);
 750                 if (!ns_entry)
 751                         return -ENOMEM;
 752                 ns->ns_debugfs_entry = ns_entry;
 753         }
 754
 755         return 0;
 756 }
 757 #undef MAX_STRING_SIZE
 758
 759 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 760                                   const void *key, unsigned int mask)
 761 {
 762         const struct ldlm_res_id *id = key;
 763         unsigned int val = 0;
 764         unsigned int i;
 765
 766         for (i = 0; i < RES_NAME_SIZE; i++)
 767                 val += id->name[i];
 768         return val & mask;
 769 }
 770
 771 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 772 {
 773         struct lu_fid       fid;
 774         __u32               hash;
 775         __u32               val;
 776
 777         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 778         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 779         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 780
 781         hash = fid_flatten32(&fid);
 782         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 783         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 784                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 785         } else {
 786                 val = fid_oid(&fid);
 787         }
 788         hash += (val >> 5) + (val << 11);
 789         return cfs_hash_32(hash, bits);
 790 }
 791
 792 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 793 {
 794         struct ldlm_resource   *res;
 795
 796         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 797         return &res->lr_name;
 798 }
 799
 800 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 801 {
 802         struct ldlm_resource   *res;
 803
 804         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 805         return ldlm_res_eq((const struct ldlm_res_id *)key,
 806                            (const struct ldlm_res_id *)&res->lr_name);
 807 }
 808
 809 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 810 {
 811         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 812 }
 813
 814 static void
 815 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 816 {
 817         struct ldlm_resource *res;
 818
 819         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 820         ldlm_resource_getref(res);
 821 }
 822
 823 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 824 {
 825         struct ldlm_resource *res;
 826
 827         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 828         ldlm_resource_putref(res);
 829 }
 830
 831 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 832         .hs_hash        = ldlm_res_hop_hash,
 833         .hs_key         = ldlm_res_hop_key,
 834         .hs_keycmp      = ldlm_res_hop_keycmp,
 835         .hs_keycpy      = NULL,
 836         .hs_object      = ldlm_res_hop_object,
 837         .hs_get         = ldlm_res_hop_get_locked,
 838         .hs_put         = ldlm_res_hop_put
 839 };
 840
 841 typedef struct ldlm_ns_hash_def {
 842         enum ldlm_ns_type       nsd_type;
 843         /** hash bucket bits */
 844         unsigned                nsd_bkt_bits;
 845         /** hash bits */
 846         unsigned                nsd_all_bits;
 847         /** hash operations */
 848         struct cfs_hash_ops *nsd_hops;
 849 } ldlm_ns_hash_def_t;
 850
 851 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 852 {
 853         {
 854                 .nsd_type       = LDLM_NS_TYPE_MDC,
 855                 .nsd_bkt_bits   = 11,
 856                 .nsd_all_bits   = 16,
 857                 .nsd_hops       = &ldlm_ns_hash_ops,
 858         },
 859         {
 860                 .nsd_type       = LDLM_NS_TYPE_MDT,
 861                 .nsd_bkt_bits   = 14,
 862                 .nsd_all_bits   = 21,
 863                 .nsd_hops       = &ldlm_ns_hash_ops,
 864         },
 865         {
 866                 .nsd_type       = LDLM_NS_TYPE_OSC,
 867                 .nsd_bkt_bits   = 8,
 868                 .nsd_all_bits   = 12,
 869                 .nsd_hops       = &ldlm_ns_hash_ops,
 870         },
 871         {
 872                 .nsd_type       = LDLM_NS_TYPE_OST,
 873                 .nsd_bkt_bits   = 11,
 874                 .nsd_all_bits   = 17,
 875                 .nsd_hops       = &ldlm_ns_hash_ops,
 876         },
 877         {
 878                 .nsd_type       = LDLM_NS_TYPE_MGC,
 879                 .nsd_bkt_bits   = 3,
 880                 .nsd_all_bits   = 4,
 881                 .nsd_hops       = &ldlm_ns_hash_ops,
 882         },
 883         {
 884                 .nsd_type       = LDLM_NS_TYPE_MGT,
 885                 .nsd_bkt_bits   = 3,
 886                 .nsd_all_bits   = 4,
 887                 .nsd_hops       = &ldlm_ns_hash_ops,
 888         },
 889         {
 890                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 891         },
 892 };
 893
 894 /**
 895  * Create and initialize new empty namespace.
 896  */
 897 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 898                                           enum ldlm_side client,
 899                                           enum ldlm_appetite apt,
 900                                           enum ldlm_ns_type ns_type)
 901 {
 902         struct ldlm_namespace *ns = NULL;
 903         struct ldlm_ns_hash_def *nsd;
 904         int idx;
 905         int rc;
 906
 907         ENTRY;
 908         LASSERT(obd != NULL);
 909
 910         rc = ldlm_get_ref();
 911         if (rc) {
 912                 CERROR("ldlm_get_ref failed: %d\n", rc);
 913                 RETURN(NULL);
 914         }
 915
 916         for (idx = 0; ; idx++) {
 917                 nsd = &ldlm_ns_hash_defs[idx];
 918                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 919                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 920                         GOTO(out_ref, NULL);
 921                 }
 922
 923                 if (nsd->nsd_type == ns_type)
 924                         break;
 925         }
 926
 927         OBD_ALLOC_PTR(ns);
 928         if (!ns)
 929                 GOTO(out_ref, NULL);
 930
 931         ns->ns_rs_hash = cfs_hash_create(name,
 932                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 933                                          nsd->nsd_bkt_bits, 0,
 934                                          CFS_HASH_MIN_THETA,
 935                                          CFS_HASH_MAX_THETA,
 936                                          nsd->nsd_hops,
 937                                          CFS_HASH_DEPTH |
 938                                          CFS_HASH_BIGNAME |
 939                                          CFS_HASH_SPIN_BKTLOCK |
 940                                          CFS_HASH_NO_ITEMREF);
 941         if (ns->ns_rs_hash == NULL)
 942                 GOTO(out_ns, NULL);
 943
 944         ns->ns_bucket_bits = nsd->nsd_all_bits - nsd->nsd_bkt_bits;
 945         OBD_ALLOC_LARGE(ns->ns_rs_buckets,
 946                         BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 947         if (!ns->ns_rs_buckets)
 948                 goto out_hash;
 949
 950         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 951                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 952
 953                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 954                 nsb->nsb_namespace = ns;
 955                 nsb->nsb_reclaim_start = 0;
 956         }
 957
 958         ns->ns_obd = obd;
 959         ns->ns_appetite = apt;
 960         ns->ns_client = client;
 961         ns->ns_name = kstrdup(name, GFP_KERNEL);
 962         if (!ns->ns_name)
 963                 goto out_hash;
 964
 965         INIT_LIST_HEAD(&ns->ns_list_chain);
 966         INIT_LIST_HEAD(&ns->ns_unused_list);
 967         spin_lock_init(&ns->ns_lock);
 968         atomic_set(&ns->ns_bref, 0);
 969         init_waitqueue_head(&ns->ns_waitq);
 970
 971         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 972         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 973         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 974
 975         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 976         ns->ns_nr_unused          = 0;
 977         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 978         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 979         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 980         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 981         ns->ns_timeouts           = 0;
 982         ns->ns_orig_connect_flags = 0;
 983         ns->ns_connect_flags      = 0;
 984         ns->ns_stopping           = 0;
 985         ns->ns_reclaim_start      = 0;
 986         ns->ns_last_pos           = &ns->ns_unused_list;
 987
 988         rc = ldlm_namespace_sysfs_register(ns);
 989         if (rc) {
 990                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 991                 GOTO(out_hash, rc);
 992         }
 993
 994         rc = ldlm_namespace_debugfs_register(ns);
 995         if (rc) {
 996                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 997                 GOTO(out_sysfs, rc);
 998         }
 999
1000         idx = ldlm_namespace_nr_read(client);
1001         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
1002         if (rc) {
1003                 CERROR("Can't initialize lock pool, rc %d\n", rc);
1004                 GOTO(out_proc, rc);
1005         }
1006
1007         ldlm_namespace_register(ns, client);
1008         RETURN(ns);
1009 out_proc:
1010         ldlm_namespace_debugfs_unregister(ns);
1011 out_sysfs:
1012         ldlm_namespace_sysfs_unregister(ns);
1013         ldlm_namespace_cleanup(ns, 0);
1014 out_hash:
1015         OBD_FREE_LARGE(ns->ns_rs_buckets,
1016                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1017         kfree(ns->ns_name);
1018         cfs_hash_putref(ns->ns_rs_hash);
1019 out_ns:
1020         OBD_FREE_PTR(ns);
1021 out_ref:
1022         ldlm_put_ref();
1023         RETURN(NULL);
1024 }
1025 EXPORT_SYMBOL(ldlm_namespace_new);
1026
1027 /**
1028  * Cancel and destroy all locks on a resource.
1029  *
1030  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1031  * clean up.  This is currently only used for recovery, and we make
1032  * certain assumptions as a result--notably, that we shouldn't cancel
1033  * locks with refs.
1034  */
1035 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1036                              __u64 flags)
1037 {
1038         struct list_head *tmp;
1039         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1040         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1041
1042         do {
1043                 struct ldlm_lock *lock = NULL;
1044
1045                 /* First, we look for non-cleaned-yet lock
1046                  * all cleaned locks are marked by CLEANED flag. */
1047                 lock_res(res);
1048                 list_for_each(tmp, q) {
1049                         lock = list_entry(tmp, struct ldlm_lock,
1050                                           l_res_link);
1051                         if (ldlm_is_cleaned(lock)) {
1052                                 lock = NULL;
1053                                 continue;
1054                         }
1055                         LDLM_LOCK_GET(lock);
1056                         ldlm_set_cleaned(lock);
1057                         break;
1058                 }
1059
1060                 if (lock == NULL) {
1061                         unlock_res(res);
1062                         break;
1063                 }
1064
1065                 /* Set CBPENDING so nothing in the cancellation path
1066                  * can match this lock. */
1067                 ldlm_set_cbpending(lock);
1068                 ldlm_set_failed(lock);
1069                 lock->l_flags |= flags;
1070
1071                 /* ... without sending a CANCEL message for local_only. */
1072                 if (local_only)
1073                         ldlm_set_local_only(lock);
1074
1075                 if (local_only && (lock->l_readers || lock->l_writers)) {
1076                         /*
1077                          * This is a little bit gross, but much better than the
1078                          * alternative: pretend that we got a blocking AST from
1079                          * the server, so that when the lock is decref'd, it
1080                          * will go away ...
1081                          */
1082                         unlock_res(res);
1083                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1084                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1085                                 set_current_state(TASK_UNINTERRUPTIBLE);
1086                                 schedule_timeout(cfs_time_seconds(4));
1087                                 set_current_state(TASK_RUNNING);
1088                         }
1089                         if (lock->l_completion_ast)
1090                                 lock->l_completion_ast(lock,
1091                                                        LDLM_FL_FAILED, NULL);
1092                         LDLM_LOCK_RELEASE(lock);
1093                         continue;
1094                 }
1095
1096                 if (client) {
1097                         struct lustre_handle lockh;
1098
1099                         unlock_res(res);
1100                         ldlm_lock2handle(lock, &lockh);
1101                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1102                         if (rc)
1103                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1104                 } else {
1105                         unlock_res(res);
1106                         LDLM_DEBUG(lock,
1107                                    "Freeing a lock still held by a client node");
1108                         ldlm_lock_cancel(lock);
1109                 }
1110                 LDLM_LOCK_RELEASE(lock);
1111         } while (1);
1112 }
1113
1114 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1115                                struct hlist_node *hnode, void *arg)
1116 {
1117         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1118         __u64 flags = *(__u64 *)arg;
1119
1120         cleanup_resource(res, &res->lr_granted, flags);
1121         cleanup_resource(res, &res->lr_waiting, flags);
1122
1123         return 0;
1124 }
1125
1126 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1127                                   struct hlist_node *hnode, void *arg)
1128 {
1129         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1130
1131         lock_res(res);
1132         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1133                "(%d) after lock cleanup; forcing cleanup.\n",
1134                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1135                atomic_read(&res->lr_refcount) - 1);
1136
1137         /* Use D_NETERROR since it is in the default mask */
1138         ldlm_resource_dump(D_NETERROR, res);
1139         unlock_res(res);
1140         return 0;
1141 }
1142
1143 /**
1144  * Cancel and destroy all locks in the namespace.
1145  *
1146  * Typically used during evictions when server notified client that it was
1147  * evicted and all of its state needs to be destroyed.
1148  * Also used during shutdown.
1149  */
1150 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1151 {
1152         if (ns == NULL) {
1153                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1154                 return ELDLM_OK;
1155         }
1156
1157         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1158                                  &flags, 0);
1159         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1160                                  NULL, 0);
1161         return ELDLM_OK;
1162 }
1163 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1164
1165 /**
1166  * Attempts to free namespace.
1167  *
1168  * Only used when namespace goes away, like during an unmount.
1169  */
1170 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1171 {
1172         ENTRY;
1173
1174         /* At shutdown time, don't call the cancellation callback */
1175         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1176
1177         if (atomic_read(&ns->ns_bref) > 0) {
1178                 int rc;
1179                 CDEBUG(D_DLMTRACE,
1180                        "dlm namespace %s free waiting on refcount %d\n",
1181                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1182 force_wait:
1183                 if (force)
1184                         rc = wait_event_idle_timeout(
1185                                 ns->ns_waitq,
1186                                 atomic_read(&ns->ns_bref) == 0,
1187                                 cfs_time_seconds(1) / 4);
1188                 else
1189                         rc = l_wait_event_abortable(
1190                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1191
1192                 /* Forced cleanups should be able to reclaim all references,
1193                  * so it's safe to wait forever... we can't leak locks... */
1194                 if (force && rc == 0) {
1195                         rc = -ETIMEDOUT;
1196                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1197                                        "namespace with %d resources in use, "
1198                                        "(rc=%d)\n", ldlm_ns_name(ns),
1199                                        atomic_read(&ns->ns_bref), rc);
1200                         GOTO(force_wait, rc);
1201                 }
1202
1203                 if (atomic_read(&ns->ns_bref)) {
1204                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1205                                        "with %d resources in use, (rc=%d)\n",
1206                                        ldlm_ns_name(ns),
1207                                        atomic_read(&ns->ns_bref), rc);
1208                         RETURN(ELDLM_NAMESPACE_EXISTS);
1209                 }
1210                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1211                        ldlm_ns_name(ns));
1212         }
1213
1214         RETURN(ELDLM_OK);
1215 }
1216
1217 /**
1218  * Performs various cleanups for passed \a ns to make it drop refc and be
1219  * ready for freeing. Waits for refc == 0.
1220  *
1221  * The following is done:
1222  * (0) Unregister \a ns from its list to make inaccessible for potential
1223  * users like pools thread and others;
1224  * (1) Clear all locks in \a ns.
1225  */
1226 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1227                                struct obd_import *imp,
1228                                int force)
1229 {
1230         int rc;
1231
1232         ENTRY;
1233         if (!ns) {
1234                 EXIT;
1235                 return;
1236         }
1237
1238         spin_lock(&ns->ns_lock);
1239         ns->ns_stopping = 1;
1240         spin_unlock(&ns->ns_lock);
1241
1242         /*
1243          * Can fail with -EINTR when force == 0 in which case try harder.
1244          */
1245         rc = __ldlm_namespace_free(ns, force);
1246         if (rc != ELDLM_OK) {
1247                 if (imp) {
1248                         ptlrpc_disconnect_import(imp, 0);
1249                         ptlrpc_invalidate_import(imp);
1250                 }
1251
1252                 /*
1253                  * With all requests dropped and the import inactive
1254                  * we are gaurenteed all reference will be dropped.
1255                  */
1256                 rc = __ldlm_namespace_free(ns, 1);
1257                 LASSERT(rc == 0);
1258         }
1259         EXIT;
1260 }
1261 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1262
1263 /**
1264  * Performs freeing memory structures related to \a ns. This is only done
1265  * when ldlm_namespce_free_prior() successfully removed all resources
1266  * referencing \a ns and its refc == 0.
1267  */
1268 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1269 {
1270         ENTRY;
1271         if (!ns) {
1272                 EXIT;
1273                 return;
1274         }
1275
1276         /* Make sure that nobody can find this ns in its list. */
1277         ldlm_namespace_unregister(ns, ns->ns_client);
1278         /* Fini pool _before_ parent proc dir is removed. This is important as
1279          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1280          * Removing it after @dir may cause oops. */
1281         ldlm_pool_fini(&ns->ns_pool);
1282
1283         ldlm_namespace_debugfs_unregister(ns);
1284         ldlm_namespace_sysfs_unregister(ns);
1285         cfs_hash_putref(ns->ns_rs_hash);
1286         OBD_FREE_LARGE(ns->ns_rs_buckets,
1287                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1288         kfree(ns->ns_name);
1289         /* Namespace \a ns should be not on list at this time, otherwise
1290          * this will cause issues related to using freed \a ns in poold
1291          * thread.
1292          */
1293         LASSERT(list_empty(&ns->ns_list_chain));
1294         OBD_FREE_PTR(ns);
1295         ldlm_put_ref();
1296         EXIT;
1297 }
1298 EXPORT_SYMBOL(ldlm_namespace_free_post);
1299
1300 /**
1301  * Cleanup the resource, and free namespace.
1302  * bug 12864:
1303  * Deadlock issue:
1304  * proc1: destroy import
1305  *        class_disconnect_export(grab cl_sem) ->
1306  *              -> ldlm_namespace_free ->
1307  *              -> lprocfs_remove(grab _lprocfs_lock).
1308  * proc2: read proc info
1309  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1310  *              -> osc_rd_active, etc(grab cl_sem).
1311  *
1312  * So that I have to split the ldlm_namespace_free into two parts - the first
1313  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1314  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1315  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1316  * held.
1317  */
1318 void ldlm_namespace_free(struct ldlm_namespace *ns,
1319                          struct obd_import *imp,
1320                          int force)
1321 {
1322         ldlm_namespace_free_prior(ns, imp, force);
1323         ldlm_namespace_free_post(ns);
1324 }
1325 EXPORT_SYMBOL(ldlm_namespace_free);
1326
1327 void ldlm_namespace_get(struct ldlm_namespace *ns)
1328 {
1329         atomic_inc(&ns->ns_bref);
1330 }
1331
1332 /* This is only for callers that care about refcount */
1333 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1334 {
1335         return atomic_inc_return(&ns->ns_bref);
1336 }
1337
1338 void ldlm_namespace_put(struct ldlm_namespace *ns)
1339 {
1340         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1341                 wake_up(&ns->ns_waitq);
1342                 spin_unlock(&ns->ns_lock);
1343         }
1344 }
1345
1346 /** Register \a ns in the list of namespaces */
1347 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1348 {
1349         mutex_lock(ldlm_namespace_lock(client));
1350         LASSERT(list_empty(&ns->ns_list_chain));
1351         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1352         ldlm_namespace_nr_inc(client);
1353         mutex_unlock(ldlm_namespace_lock(client));
1354 }
1355
1356 /** Unregister \a ns from the list of namespaces. */
1357 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1358 {
1359         mutex_lock(ldlm_namespace_lock(client));
1360         LASSERT(!list_empty(&ns->ns_list_chain));
1361         /* Some asserts and possibly other parts of the code are still
1362          * using list_empty(&ns->ns_list_chain). This is why it is
1363          * important to use list_del_init() here. */
1364         list_del_init(&ns->ns_list_chain);
1365         ldlm_namespace_nr_dec(client);
1366         mutex_unlock(ldlm_namespace_lock(client));
1367 }
1368
1369 /** Should be called with ldlm_namespace_lock(client) taken. */
1370 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1371                                           enum ldlm_side client)
1372 {
1373         LASSERT(!list_empty(&ns->ns_list_chain));
1374         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1375         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1376 }
1377
1378 /** Should be called with ldlm_namespace_lock(client) taken. */
1379 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1380                                             enum ldlm_side client)
1381 {
1382         LASSERT(!list_empty(&ns->ns_list_chain));
1383         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1384         list_move_tail(&ns->ns_list_chain,
1385                        ldlm_namespace_inactive_list(client));
1386 }
1387
1388 /** Should be called with ldlm_namespace_lock(client) taken. */
1389 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1390 {
1391         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1392         LASSERT(!list_empty(ldlm_namespace_list(client)));
1393         return container_of(ldlm_namespace_list(client)->next,
1394                             struct ldlm_namespace, ns_list_chain);
1395 }
1396
1397 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1398 {
1399         int idx;
1400
1401         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1402                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1403         if (res->lr_itree == NULL)
1404                 return false;
1405         /* Initialize interval trees for each lock mode. */
1406         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1407                 res->lr_itree[idx].lit_size = 0;
1408                 res->lr_itree[idx].lit_mode = 1 << idx;
1409                 res->lr_itree[idx].lit_root = NULL;
1410         }
1411         return true;
1412 }
1413
1414 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1415 {
1416         int i;
1417
1418         OBD_ALLOC_PTR(res->lr_ibits_queues);
1419         if (res->lr_ibits_queues == NULL)
1420                 return false;
1421         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1422                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1423         return true;
1424 }
1425
1426 /** Create and initialize new resource. */
1427 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1428 {
1429         struct ldlm_resource *res;
1430         bool rc;
1431
1432         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1433         if (res == NULL)
1434                 return NULL;
1435
1436         switch (ldlm_type) {
1437         case LDLM_EXTENT:
1438                 rc = ldlm_resource_extent_new(res);
1439                 break;
1440         case LDLM_IBITS:
1441                 rc = ldlm_resource_inodebits_new(res);
1442                 break;
1443         default:
1444                 rc = true;
1445                 break;
1446         }
1447         if (!rc) {
1448                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1449                 return NULL;
1450         }
1451
1452         INIT_LIST_HEAD(&res->lr_granted);
1453         INIT_LIST_HEAD(&res->lr_waiting);
1454
1455         atomic_set(&res->lr_refcount, 1);
1456         spin_lock_init(&res->lr_lock);
1457         lu_ref_init(&res->lr_reference);
1458
1459         /* Since LVB init can be delayed now, there is no longer need to
1460          * immediatelly acquire mutex here. */
1461         mutex_init(&res->lr_lvb_mutex);
1462         res->lr_lvb_initialized = false;
1463
1464         return res;
1465 }
1466
1467 static void ldlm_resource_free(struct ldlm_resource *res)
1468 {
1469         if (res->lr_type == LDLM_EXTENT) {
1470                 if (res->lr_itree != NULL)
1471                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1472                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1473         } else if (res->lr_type == LDLM_IBITS) {
1474                 if (res->lr_ibits_queues != NULL)
1475                         OBD_FREE_PTR(res->lr_ibits_queues);
1476         }
1477
1478         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1479 }
1480
1481 /**
1482  * Return a reference to resource with given name, creating it if necessary.
1483  * Args: namespace with ns_lock unlocked
1484  * Locks: takes and releases NS hash-lock and res->lr_lock
1485  * Returns: referenced, unlocked ldlm_resource or NULL
1486  */
1487 struct ldlm_resource *
1488 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1489                   const struct ldlm_res_id *name, enum ldlm_type type,
1490                   int create)
1491 {
1492         struct hlist_node       *hnode;
1493         struct ldlm_resource    *res = NULL;
1494         struct cfs_hash_bd              bd;
1495         __u64                   version;
1496         int                     ns_refcount = 0;
1497         int hash;
1498
1499         LASSERT(ns != NULL);
1500         LASSERT(parent == NULL);
1501         LASSERT(ns->ns_rs_hash != NULL);
1502         LASSERT(name->name[0] != 0);
1503
1504         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1505         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1506         if (hnode != NULL) {
1507                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1508                 GOTO(found, res);
1509         }
1510
1511         version = cfs_hash_bd_version_get(&bd);
1512         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1513
1514         if (create == 0)
1515                 return ERR_PTR(-ENOENT);
1516
1517         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1518                  "type: %d\n", type);
1519         res = ldlm_resource_new(type);
1520         if (res == NULL)
1521                 return ERR_PTR(-ENOMEM);
1522
1523         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1524         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1525         res->lr_name = *name;
1526         res->lr_type = type;
1527
1528         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1529         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1530                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1531
1532         if (hnode != NULL) {
1533                 /* Someone won the race and already added the resource. */
1534                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1535                 /* Clean lu_ref for failed resource. */
1536                 lu_ref_fini(&res->lr_reference);
1537                 ldlm_resource_free(res);
1538 found:
1539                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1540                 return res;
1541         }
1542         /* We won! Let's add the resource. */
1543         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1544         if (cfs_hash_bd_count_get(&bd) == 1)
1545                 ns_refcount = ldlm_namespace_get_return(ns);
1546
1547         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1548
1549         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1550
1551         /* Let's see if we happened to be the very first resource in this
1552          * namespace. If so, and this is a client namespace, we need to move
1553          * the namespace into the active namespaces list to be patrolled by
1554          * the ldlm_poold. */
1555         if (ns_is_client(ns) && ns_refcount == 1) {
1556                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1557                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1558                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1559         }
1560
1561         return res;
1562 }
1563 EXPORT_SYMBOL(ldlm_resource_get);
1564
1565 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1566 {
1567         LASSERT(res != NULL);
1568         LASSERT(res != LP_POISON);
1569         atomic_inc(&res->lr_refcount);
1570         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1571                atomic_read(&res->lr_refcount));
1572         return res;
1573 }
1574
1575 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1576                                          struct ldlm_resource *res)
1577 {
1578         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1579
1580         if (!list_empty(&res->lr_granted)) {
1581                 ldlm_resource_dump(D_ERROR, res);
1582                 LBUG();
1583         }
1584
1585         if (!list_empty(&res->lr_waiting)) {
1586                 ldlm_resource_dump(D_ERROR, res);
1587                 LBUG();
1588         }
1589
1590         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1591                                bd, &res->lr_hash);
1592         lu_ref_fini(&res->lr_reference);
1593         if (cfs_hash_bd_count_get(bd) == 0)
1594                 ldlm_namespace_put(nsb->nsb_namespace);
1595 }
1596
1597 /* Returns 1 if the resource was freed, 0 if it remains. */
1598 int ldlm_resource_putref(struct ldlm_resource *res)
1599 {
1600         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1601         struct cfs_hash_bd   bd;
1602
1603         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1604         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1605                res, atomic_read(&res->lr_refcount) - 1);
1606
1607         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1608         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1609                 __ldlm_resource_putref_final(&bd, res);
1610                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1611                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1612                         ns->ns_lvbo->lvbo_free(res);
1613                 ldlm_resource_free(res);
1614                 return 1;
1615         }
1616         return 0;
1617 }
1618 EXPORT_SYMBOL(ldlm_resource_putref);
1619
1620 /**
1621  * Add a lock into a given resource into specified lock list.
1622  */
1623 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1624                             struct ldlm_lock *lock)
1625 {
1626         check_res_locked(res);
1627
1628         LDLM_DEBUG(lock, "About to add this lock");
1629
1630         if (ldlm_is_destroyed(lock)) {
1631                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1632                 return;
1633         }
1634
1635         LASSERT(list_empty(&lock->l_res_link));
1636
1637         list_add_tail(&lock->l_res_link, head);
1638
1639         if (res->lr_type == LDLM_IBITS)
1640                 ldlm_inodebits_add_lock(res, head, lock);
1641 }
1642
1643 /**
1644  * Insert a lock into resource after specified lock.
1645  *
1646  * Obtain resource description from the lock we are inserting after.
1647  */
1648 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1649                                      struct ldlm_lock *new)
1650 {
1651         struct ldlm_resource *res = original->l_resource;
1652
1653         check_res_locked(res);
1654
1655         ldlm_resource_dump(D_INFO, res);
1656         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1657
1658         if (ldlm_is_destroyed(new)) {
1659                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1660                 goto out;
1661         }
1662
1663         LASSERT(list_empty(&new->l_res_link));
1664
1665         list_add(&new->l_res_link, &original->l_res_link);
1666  out:;
1667 }
1668
1669 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1670 {
1671         int type = lock->l_resource->lr_type;
1672
1673         check_res_locked(lock->l_resource);
1674         switch (type) {
1675         case LDLM_PLAIN:
1676                 ldlm_unlink_lock_skiplist(lock);
1677                 break;
1678         case LDLM_EXTENT:
1679                 ldlm_extent_unlink_lock(lock);
1680                 break;
1681         case LDLM_IBITS:
1682                 ldlm_inodebits_unlink_lock(lock);
1683                 break;
1684         }
1685         list_del_init(&lock->l_res_link);
1686 }
1687 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1688
1689 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1690 {
1691         desc->lr_type = res->lr_type;
1692         desc->lr_name = res->lr_name;
1693 }
1694
1695 /**
1696  * Print information about all locks in all namespaces on this node to debug
1697  * log.
1698  */
1699 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1700 {
1701         struct list_head *tmp;
1702
1703         if (!((libcfs_debug | D_ERROR) & level))
1704                 return;
1705
1706         mutex_lock(ldlm_namespace_lock(client));
1707
1708         list_for_each(tmp, ldlm_namespace_list(client)) {
1709                 struct ldlm_namespace *ns;
1710
1711                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1712                 ldlm_namespace_dump(level, ns);
1713         }
1714
1715         mutex_unlock(ldlm_namespace_lock(client));
1716 }
1717
1718 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1719                               struct hlist_node *hnode, void *arg)
1720 {
1721         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1722         int    level = (int)(unsigned long)arg;
1723
1724         lock_res(res);
1725         ldlm_resource_dump(level, res);
1726         unlock_res(res);
1727
1728         return 0;
1729 }
1730
1731 /**
1732  * Print information about all locks in this namespace on this node to debug
1733  * log.
1734  */
1735 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1736 {
1737         if (!((libcfs_debug | D_ERROR) & level))
1738                 return;
1739
1740         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1741                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1742                ns_is_client(ns) ? "client" : "server");
1743
1744         if (ktime_get_seconds() < ns->ns_next_dump)
1745                 return;
1746
1747         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1748                                  ldlm_res_hash_dump,
1749                                  (void *)(unsigned long)level, 0);
1750         spin_lock(&ns->ns_lock);
1751         ns->ns_next_dump = ktime_get_seconds() + 10;
1752         spin_unlock(&ns->ns_lock);
1753 }
1754
1755 /**
1756  * Print information about all locks in this resource to debug log.
1757  */
1758 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1759 {
1760         struct ldlm_lock *lock;
1761         unsigned int granted = 0;
1762
1763         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1764
1765         if (!((libcfs_debug | D_ERROR) & level))
1766                 return;
1767
1768         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1769                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1770
1771         if (!list_empty(&res->lr_granted)) {
1772                 CDEBUG(level, "Granted locks (in reverse order):\n");
1773                 list_for_each_entry_reverse(lock, &res->lr_granted,
1774                                                 l_res_link) {
1775                         LDLM_DEBUG_LIMIT(level, lock, "###");
1776                         if (!(level & D_CANTMASK) &&
1777                             ++granted > ldlm_dump_granted_max) {
1778                                 CDEBUG(level,
1779                                        "only dump %d granted locks to avoid DDOS.\n",
1780                                        granted);
1781                                 break;
1782                         }
1783                 }
1784         }
1785
1786         if (!list_empty(&res->lr_waiting)) {
1787                 CDEBUG(level, "Waiting locks:\n");
1788                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1789                         LDLM_DEBUG_LIMIT(level, lock, "###");
1790         }
1791 }
1792 EXPORT_SYMBOL(ldlm_resource_dump);