lustre/ldlm/ldlm_resource.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46
  47 int ldlm_srv_namespace_nr = 0;
  48 int ldlm_cli_namespace_nr = 0;
  49
  50 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  51 LIST_HEAD(ldlm_srv_namespace_list);
  52
  53 DEFINE_MUTEX(ldlm_cli_namespace_lock);
  54 /* Client Namespaces that have active resources in them.
  55  * Once all resources go away, ldlm_poold moves such namespaces to the
  56  * inactive list */
  57 LIST_HEAD(ldlm_cli_active_namespace_list);
  58 /* Client namespaces that don't have any locks in them */
  59 LIST_HEAD(ldlm_cli_inactive_namespace_list);
  60
  61 static struct dentry *ldlm_debugfs_dir;
  62 static struct dentry *ldlm_ns_debugfs_dir;
  63 struct dentry *ldlm_svc_debugfs_dir;
  64
  65 /* during debug dump certain amount of granted locks for one resource to avoid
  66  * DDOS. */
  67 static unsigned int ldlm_dump_granted_max = 256;
  68
  69 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
  70                                           const char __user *buffer,
  71                                           size_t count, loff_t *off)
  72 {
  73         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  75         RETURN(count);
  76 }
  77
  78 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
  79
  80 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  81 {
  82         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  83         return 0;
  84 }
  85
  86 static ssize_t
  87 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  88                        size_t count, loff_t *off)
  89 {
  90         struct seq_file *seq = file->private_data;
  91
  92         if (!count)
  93                 return 0;
  94
  95         return kstrtouint_from_user(buffer, count, 0,
  96                                     (unsigned int *)seq->private);
  97 }
  98
  99 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
 100
 101 #ifdef HAVE_SERVER_SUPPORT
 102
 103 static int seq_watermark_show(struct seq_file *m, void *data)
 104 {
 105         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 106         return 0;
 107 }
 108
 109 static ssize_t seq_watermark_write(struct file *file,
 110                                    const char __user *buffer, size_t count,
 111                                    loff_t *off)
 112 {
 113         __s64 value;
 114         __u64 watermark;
 115         __u64 *data = ((struct seq_file *)file->private_data)->private;
 116         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 117         int rc;
 118
 119         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
 120         if (rc) {
 121                 CERROR("Failed to set %s, rc = %d.\n",
 122                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 123                        rc);
 124                 return rc;
 125         } else if (value != 0 && value < (1 << 20)) {
 126                 CERROR("%s should be greater than 1MB.\n",
 127                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 128                 return -EINVAL;
 129         }
 130         watermark = value >> 20;
 131
 132         if (wm_low) {
 133                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 134                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 135                                "lock_limit_mb.\n");
 136                         return -EINVAL;
 137                 }
 138
 139                 *data = watermark;
 140                 if (watermark != 0) {
 141                         watermark <<= 20;
 142                         do_div(watermark, sizeof(struct ldlm_lock));
 143                 }
 144                 ldlm_reclaim_threshold = watermark;
 145         } else {
 146                 if (ldlm_reclaim_threshold_mb != 0 &&
 147                     watermark < ldlm_reclaim_threshold_mb) {
 148                         CERROR("lock_limit_mb must be greater than "
 149                                "lock_reclaim_threshold_mb.\n");
 150                         return -EINVAL;
 151                 }
 152
 153                 *data = watermark;
 154                 if (watermark != 0) {
 155                         watermark <<= 20;
 156                         do_div(watermark, sizeof(struct ldlm_lock));
 157                 }
 158                 ldlm_lock_limit = watermark;
 159         }
 160
 161         return count;
 162 }
 163
 164 static int seq_watermark_open(struct inode *inode, struct file *file)
 165 {
 166         return single_open(file, seq_watermark_show, inode->i_private);
 167 }
 168
 169 static const struct file_operations ldlm_watermark_fops = {
 170         .owner          = THIS_MODULE,
 171         .open           = seq_watermark_open,
 172         .read           = seq_read,
 173         .write          = seq_watermark_write,
 174         .llseek         = seq_lseek,
 175         .release        = lprocfs_single_release,
 176 };
 177
 178 static int seq_granted_show(struct seq_file *m, void *data)
 179 {
 180         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 181                    (struct percpu_counter *)m->private));
 182         return 0;
 183 }
 184
 185 static int seq_granted_open(struct inode *inode, struct file *file)
 186 {
 187         return single_open(file, seq_granted_show, inode->i_private);
 188 }
 189
 190 static const struct file_operations ldlm_granted_fops = {
 191         .owner  = THIS_MODULE,
 192         .open   = seq_granted_open,
 193         .read   = seq_read,
 194         .llseek = seq_lseek,
 195         .release = seq_release,
 196 };
 197
 198 #endif /* HAVE_SERVER_SUPPORT */
 199
 200 static struct lprocfs_vars ldlm_debugfs_list[] = {
 201         { .name =       "dump_namespaces",
 202           .fops =       &ldlm_dump_ns_fops,
 203           .proc_mode =  0222 },
 204         { .name =       "dump_granted_max",
 205           .fops =       &ldlm_rw_uint_fops,
 206           .data =       &ldlm_dump_granted_max },
 207 #ifdef HAVE_SERVER_SUPPORT
 208         { .name =       "lock_reclaim_threshold_mb",
 209           .fops =       &ldlm_watermark_fops,
 210           .data =       &ldlm_reclaim_threshold_mb },
 211         { .name =       "lock_limit_mb",
 212           .fops =       &ldlm_watermark_fops,
 213           .data =       &ldlm_lock_limit_mb },
 214         { .name =       "lock_granted_count",
 215           .fops =       &ldlm_granted_fops,
 216           .data =       &ldlm_granted_total },
 217 #endif
 218         { NULL }
 219 };
 220
 221 int ldlm_debugfs_setup(void)
 222 {
 223         int rc;
 224
 225         ENTRY;
 226         ldlm_debugfs_dir = ldebugfs_register(OBD_LDLM_DEVICENAME,
 227                                              debugfs_lustre_root,
 228                                              NULL, NULL);
 229         if (IS_ERR_OR_NULL(ldlm_debugfs_dir)) {
 230                 CERROR("LDebugFS failed in ldlm-init\n");
 231                 rc = ldlm_debugfs_dir ? PTR_ERR(ldlm_debugfs_dir) : -ENOMEM;
 232                 GOTO(err, rc);
 233         }
 234
 235         ldlm_ns_debugfs_dir = ldebugfs_register("namespaces",
 236                                                 ldlm_debugfs_dir,
 237                                                 NULL, NULL);
 238         if (IS_ERR_OR_NULL(ldlm_ns_debugfs_dir)) {
 239                 CERROR("LProcFS failed in ldlm-init\n");
 240                 rc = ldlm_ns_debugfs_dir ? PTR_ERR(ldlm_ns_debugfs_dir)
 241                                          : -ENOMEM;
 242                 GOTO(err_type, rc);
 243         }
 244
 245         ldlm_svc_debugfs_dir = ldebugfs_register("services",
 246                                                  ldlm_debugfs_dir,
 247                                                  NULL, NULL);
 248         if (IS_ERR_OR_NULL(ldlm_svc_debugfs_dir)) {
 249                 CERROR("LProcFS failed in ldlm-init\n");
 250                 rc = ldlm_svc_debugfs_dir ? PTR_ERR(ldlm_svc_debugfs_dir)
 251                                           : -ENOMEM;
 252                 GOTO(err_ns, rc);
 253         }
 254
 255         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 256         if (rc != 0) {
 257                 CERROR("LProcFS failed in ldlm-init\n");
 258                 GOTO(err_svc, rc);
 259         }
 260
 261         RETURN(0);
 262
 263 err_svc:
 264         ldebugfs_remove(&ldlm_svc_debugfs_dir);
 265 err_ns:
 266         ldebugfs_remove(&ldlm_ns_debugfs_dir);
 267 err_type:
 268         ldebugfs_remove(&ldlm_debugfs_dir);
 269 err:
 270         ldlm_svc_debugfs_dir = NULL;
 271         ldlm_ns_debugfs_dir = NULL;
 272         ldlm_debugfs_dir = NULL;
 273         RETURN(rc);
 274 }
 275
 276 void ldlm_debugfs_cleanup(void)
 277 {
 278         if (!IS_ERR_OR_NULL(ldlm_svc_debugfs_dir))
 279                 ldebugfs_remove(&ldlm_svc_debugfs_dir);
 280
 281         if (!IS_ERR_OR_NULL(ldlm_ns_debugfs_dir))
 282                 ldebugfs_remove(&ldlm_ns_debugfs_dir);
 283
 284         if (!IS_ERR_OR_NULL(ldlm_debugfs_dir))
 285                 ldebugfs_remove(&ldlm_debugfs_dir);
 286
 287         ldlm_svc_debugfs_dir = NULL;
 288         ldlm_ns_debugfs_dir = NULL;
 289         ldlm_debugfs_dir = NULL;
 290 }
 291
 292 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 293                                    char *buf)
 294 {
 295         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 296                                                  ns_kobj);
 297         __u64                   res = 0;
 298         struct cfs_hash_bd              bd;
 299         int                     i;
 300
 301         /* result is not strictly consistant */
 302         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 303                 res += cfs_hash_bd_count_get(&bd);
 304         return sprintf(buf, "%lld\n", res);
 305 }
 306 LUSTRE_RO_ATTR(resource_count);
 307
 308 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 309                                char *buf)
 310 {
 311         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 312                                                  ns_kobj);
 313         __u64                   locks;
 314
 315         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 316                                         LPROCFS_FIELDS_FLAGS_SUM);
 317         return sprintf(buf, "%lld\n", locks);
 318 }
 319 LUSTRE_RO_ATTR(lock_count);
 320
 321 static ssize_t lock_unused_count_show(struct kobject *kobj,
 322                                       struct attribute *attr,
 323                                       char *buf)
 324 {
 325         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 326                                                  ns_kobj);
 327
 328         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 329 }
 330 LUSTRE_RO_ATTR(lock_unused_count);
 331
 332 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 333                              char *buf)
 334 {
 335         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 336                                                  ns_kobj);
 337         __u32 *nr = &ns->ns_max_unused;
 338
 339         if (ns_connect_lru_resize(ns))
 340                 nr = &ns->ns_nr_unused;
 341         return sprintf(buf, "%u\n", *nr);
 342 }
 343
 344 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 345                               const char *buffer, size_t count)
 346 {
 347         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 348                                                  ns_kobj);
 349         unsigned long tmp;
 350         int lru_resize;
 351         int err;
 352
 353         if (strncmp(buffer, "clear", 5) == 0) {
 354                 CDEBUG(D_DLMTRACE,
 355                        "dropping all unused locks from namespace %s\n",
 356                        ldlm_ns_name(ns));
 357                 if (ns_connect_lru_resize(ns)) {
 358                         /* Try to cancel all @ns_nr_unused locks. */
 359                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 360                                         LDLM_LRU_FLAG_PASSED |
 361                                         LDLM_LRU_FLAG_CLEANUP);
 362                 } else {
 363                         tmp = ns->ns_max_unused;
 364                         ns->ns_max_unused = 0;
 365                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 366                                         LDLM_LRU_FLAG_CLEANUP);
 367                         ns->ns_max_unused = tmp;
 368                 }
 369                 return count;
 370         }
 371
 372         err = kstrtoul(buffer, 10, &tmp);
 373         if (err != 0) {
 374                 CERROR("lru_size: invalid value written\n");
 375                 return -EINVAL;
 376         }
 377         lru_resize = (tmp == 0);
 378
 379         if (ns_connect_lru_resize(ns)) {
 380                 if (!lru_resize)
 381                         ns->ns_max_unused = (unsigned int)tmp;
 382
 383                 if (tmp > ns->ns_nr_unused)
 384                         tmp = ns->ns_nr_unused;
 385                 tmp = ns->ns_nr_unused - tmp;
 386
 387                 CDEBUG(D_DLMTRACE,
 388                        "changing namespace %s unused locks from %u to %u\n",
 389                        ldlm_ns_name(ns), ns->ns_nr_unused,
 390                        (unsigned int)tmp);
 391                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 392
 393                 if (!lru_resize) {
 394                         CDEBUG(D_DLMTRACE,
 395                                "disable lru_resize for namespace %s\n",
 396                                ldlm_ns_name(ns));
 397                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 398                 }
 399         } else {
 400                 CDEBUG(D_DLMTRACE,
 401                        "changing namespace %s max_unused from %u to %u\n",
 402                        ldlm_ns_name(ns), ns->ns_max_unused,
 403                        (unsigned int)tmp);
 404                 ns->ns_max_unused = (unsigned int)tmp;
 405                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 406
 407                 /* Make sure that LRU resize was originally supported before
 408                  * turning it on here.
 409                  */
 410                 if (lru_resize &&
 411                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 412                         CDEBUG(D_DLMTRACE,
 413                                "enable lru_resize for namespace %s\n",
 414                                ldlm_ns_name(ns));
 415                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 416                 }
 417         }
 418
 419         return count;
 420 }
 421 LUSTRE_RW_ATTR(lru_size);
 422
 423 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 424                                 char *buf)
 425 {
 426         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 427                                                  ns_kobj);
 428
 429         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 430 }
 431
 432 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 433                                  const char *buffer, size_t count)
 434 {
 435         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 436                                                  ns_kobj);
 437         int scale = NSEC_PER_MSEC;
 438         unsigned long long tmp;
 439         char *buf;
 440
 441         /* Did the user ask in seconds or milliseconds. Default is in ms */
 442         buf = strstr(buffer, "ms");
 443         if (!buf) {
 444                 buf = strchr(buffer, 's');
 445                 if (buf)
 446                         scale = NSEC_PER_SEC;
 447         }
 448
 449         if (buf)
 450                 *buf = '\0';
 451
 452         if (kstrtoull(buffer, 10, &tmp))
 453                 return -EINVAL;
 454
 455         ns->ns_max_age = ktime_set(0, tmp * scale);
 456
 457         return count;
 458 }
 459 LUSTRE_RW_ATTR(lru_max_age);
 460
 461 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 462                                       struct attribute *attr,
 463                                       char *buf)
 464 {
 465         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 466                                                  ns_kobj);
 467
 468         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 469 }
 470
 471 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 472                                        struct attribute *attr,
 473                                        const char *buffer,
 474                                        size_t count)
 475 {
 476         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 477                                                  ns_kobj);
 478         unsigned long supp = -1;
 479         int rc;
 480
 481         rc = kstrtoul(buffer, 10, &supp);
 482         if (rc < 0)
 483                 return rc;
 484
 485         if (supp == 0)
 486                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 487         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 488                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 489         return count;
 490 }
 491 LUSTRE_RW_ATTR(early_lock_cancel);
 492
 493 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 494                                     struct attribute *attr, char *buf)
 495 {
 496         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 497                                                  ns_kobj);
 498
 499         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 500 }
 501
 502 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 503                                      struct attribute *attr,
 504                                      const char *buffer, size_t count)
 505 {
 506         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 507                                                  ns_kobj);
 508         unsigned long long tmp;
 509
 510         if (kstrtoull(buffer, 10, &tmp))
 511                 return -EINVAL;
 512
 513         ns->ns_dirty_age_limit = tmp;
 514
 515         return count;
 516 }
 517 LUSTRE_RW_ATTR(dirty_age_limit);
 518
 519 #ifdef HAVE_SERVER_SUPPORT
 520 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 521                                     struct attribute *attr, char *buf)
 522 {
 523         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 524                                                  ns_kobj);
 525
 526         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 527 }
 528
 529 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 530                                      struct attribute *attr,
 531                                      const char *buffer, size_t count)
 532 {
 533         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 534                                                  ns_kobj);
 535         unsigned long long tmp;
 536
 537         if (kstrtoull(buffer, 10, &tmp))
 538                 return -EINVAL;
 539
 540         ns->ns_ctime_age_limit = tmp;
 541
 542         return count;
 543 }
 544 LUSTRE_RW_ATTR(ctime_age_limit);
 545
 546 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 547                                   char *buf)
 548 {
 549         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 550                                                  ns_kobj);
 551
 552         return sprintf(buf, "%d\n", ns->ns_timeouts);
 553 }
 554 LUSTRE_RO_ATTR(lock_timeouts);
 555
 556 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 557                                      struct attribute *attr, char *buf)
 558 {
 559         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 560                                                  ns_kobj);
 561
 562         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 563 }
 564
 565 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 566                                       struct attribute *attr,
 567                                       const char *buffer, size_t count)
 568 {
 569         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 570                                                  ns_kobj);
 571         unsigned long tmp;
 572         int err;
 573
 574         err = kstrtoul(buffer, 10, &tmp);
 575         if (err != 0)
 576                 return -EINVAL;
 577
 578         ns->ns_max_nolock_size = tmp;
 579
 580         return count;
 581 }
 582 LUSTRE_RW_ATTR(max_nolock_bytes);
 583
 584 static ssize_t contention_seconds_show(struct kobject *kobj,
 585                                        struct attribute *attr, char *buf)
 586 {
 587         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 588                                                  ns_kobj);
 589
 590         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 591 }
 592
 593 static ssize_t contention_seconds_store(struct kobject *kobj,
 594                                         struct attribute *attr,
 595                                         const char *buffer, size_t count)
 596 {
 597         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 598                                                  ns_kobj);
 599         unsigned long long tmp;
 600
 601         if (kstrtoull(buffer, 10, &tmp))
 602                 return -EINVAL;
 603
 604         ns->ns_contention_time = tmp;
 605
 606         return count;
 607 }
 608 LUSTRE_RW_ATTR(contention_seconds);
 609
 610 static ssize_t contended_locks_show(struct kobject *kobj,
 611                                     struct attribute *attr, char *buf)
 612 {
 613         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 614                                                  ns_kobj);
 615
 616         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 617 }
 618
 619 static ssize_t contended_locks_store(struct kobject *kobj,
 620                                      struct attribute *attr,
 621                                      const char *buffer, size_t count)
 622 {
 623         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 624                                                  ns_kobj);
 625         unsigned long tmp;
 626         int err;
 627
 628         err = kstrtoul(buffer, 10, &tmp);
 629         if (err != 0)
 630                 return -EINVAL;
 631
 632         ns->ns_contended_locks = tmp;
 633
 634         return count;
 635 }
 636 LUSTRE_RW_ATTR(contended_locks);
 637
 638 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 639                                      struct attribute *attr, char *buf)
 640 {
 641         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 642                                                  ns_kobj);
 643
 644         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 645 }
 646
 647 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 648                                       struct attribute *attr,
 649                                       const char *buffer, size_t count)
 650 {
 651         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 652                                                  ns_kobj);
 653         unsigned long tmp;
 654         int err;
 655
 656         err = kstrtoul(buffer, 10, &tmp);
 657         if (err != 0)
 658                 return -EINVAL;
 659
 660         ns->ns_max_parallel_ast = tmp;
 661
 662         return count;
 663 }
 664 LUSTRE_RW_ATTR(max_parallel_ast);
 665
 666 #endif /* HAVE_SERVER_SUPPORT */
 667
 668 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 669 static struct attribute *ldlm_ns_attrs[] = {
 670         &lustre_attr_resource_count.attr,
 671         &lustre_attr_lock_count.attr,
 672         &lustre_attr_lock_unused_count.attr,
 673         &lustre_attr_lru_size.attr,
 674         &lustre_attr_lru_max_age.attr,
 675         &lustre_attr_early_lock_cancel.attr,
 676         &lustre_attr_dirty_age_limit.attr,
 677 #ifdef HAVE_SERVER_SUPPORT
 678         &lustre_attr_ctime_age_limit.attr,
 679         &lustre_attr_lock_timeouts.attr,
 680         &lustre_attr_max_nolock_bytes.attr,
 681         &lustre_attr_contention_seconds.attr,
 682         &lustre_attr_contended_locks.attr,
 683         &lustre_attr_max_parallel_ast.attr,
 684 #endif
 685         NULL,
 686 };
 687
 688 static void ldlm_ns_release(struct kobject *kobj)
 689 {
 690         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 691                                                  ns_kobj);
 692         complete(&ns->ns_kobj_unregister);
 693 }
 694
 695 static struct kobj_type ldlm_ns_ktype = {
 696         .default_attrs  = ldlm_ns_attrs,
 697         .sysfs_ops      = &lustre_sysfs_ops,
 698         .release        = ldlm_ns_release,
 699 };
 700
 701 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 702 {
 703         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 704                 CERROR("dlm namespace %s has no procfs dir?\n",
 705                        ldlm_ns_name(ns));
 706         else
 707                 ldebugfs_remove(&ns->ns_debugfs_entry);
 708
 709         if (ns->ns_stats != NULL)
 710                 lprocfs_free_stats(&ns->ns_stats);
 711 }
 712
 713 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 714 {
 715         kobject_put(&ns->ns_kobj);
 716         wait_for_completion(&ns->ns_kobj_unregister);
 717 }
 718
 719 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 720 {
 721         int err;
 722
 723         ns->ns_kobj.kset = ldlm_ns_kset;
 724         init_completion(&ns->ns_kobj_unregister);
 725         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 726                                    "%s", ldlm_ns_name(ns));
 727
 728         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 729         if (!ns->ns_stats) {
 730                 kobject_put(&ns->ns_kobj);
 731                 return -ENOMEM;
 732         }
 733
 734         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 735                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 736
 737         return err;
 738 }
 739
 740 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 741 {
 742         struct dentry *ns_entry;
 743
 744         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 745                 ns_entry = ns->ns_debugfs_entry;
 746         } else {
 747                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 748                                               ldlm_ns_debugfs_dir);
 749                 if (!ns_entry)
 750                         return -ENOMEM;
 751                 ns->ns_debugfs_entry = ns_entry;
 752         }
 753
 754         return 0;
 755 }
 756 #undef MAX_STRING_SIZE
 757
 758 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 759                                   const void *key, unsigned mask)
 760 {
 761         const struct ldlm_res_id     *id  = key;
 762         unsigned                val = 0;
 763         unsigned                i;
 764
 765         for (i = 0; i < RES_NAME_SIZE; i++)
 766                 val += id->name[i];
 767         return val & mask;
 768 }
 769
 770 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
 771                                       const void *key, unsigned mask)
 772 {
 773         const struct ldlm_res_id *id = key;
 774         struct lu_fid       fid;
 775         __u32               hash;
 776         __u32               val;
 777
 778         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 779         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 780         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 781
 782         hash = fid_flatten32(&fid);
 783         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 784         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 785                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 786                 hash += (val >> 5) + (val << 11);
 787         } else {
 788                 val = fid_oid(&fid);
 789         }
 790         hash = hash_long(hash, hs->hs_bkt_bits);
 791         /* give me another random factor */
 792         hash -= hash_long((unsigned long)hs, val % 11 + 3);
 793
 794         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 795         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 796
 797         return hash & mask;
 798 }
 799
 800 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 801 {
 802         struct ldlm_resource   *res;
 803
 804         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 805         return &res->lr_name;
 806 }
 807
 808 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 809 {
 810         struct ldlm_resource   *res;
 811
 812         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 813         return ldlm_res_eq((const struct ldlm_res_id *)key,
 814                            (const struct ldlm_res_id *)&res->lr_name);
 815 }
 816
 817 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 818 {
 819         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 820 }
 821
 822 static void
 823 ldlm_res_hop_get_locked(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_getref(res);
 829 }
 830
 831 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 832 {
 833         struct ldlm_resource *res;
 834
 835         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 836         ldlm_resource_putref(res);
 837 }
 838
 839 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 840         .hs_hash        = ldlm_res_hop_hash,
 841         .hs_key         = ldlm_res_hop_key,
 842         .hs_keycmp      = ldlm_res_hop_keycmp,
 843         .hs_keycpy      = NULL,
 844         .hs_object      = ldlm_res_hop_object,
 845         .hs_get         = ldlm_res_hop_get_locked,
 846         .hs_put         = ldlm_res_hop_put
 847 };
 848
 849 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
 850         .hs_hash        = ldlm_res_hop_fid_hash,
 851         .hs_key         = ldlm_res_hop_key,
 852         .hs_keycmp      = ldlm_res_hop_keycmp,
 853         .hs_keycpy      = NULL,
 854         .hs_object      = ldlm_res_hop_object,
 855         .hs_get         = ldlm_res_hop_get_locked,
 856         .hs_put         = ldlm_res_hop_put
 857 };
 858
 859 typedef struct ldlm_ns_hash_def {
 860         enum ldlm_ns_type       nsd_type;
 861         /** hash bucket bits */
 862         unsigned                nsd_bkt_bits;
 863         /** hash bits */
 864         unsigned                nsd_all_bits;
 865         /** hash operations */
 866         struct cfs_hash_ops *nsd_hops;
 867 } ldlm_ns_hash_def_t;
 868
 869 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 870 {
 871         {
 872                 .nsd_type       = LDLM_NS_TYPE_MDC,
 873                 .nsd_bkt_bits   = 11,
 874                 .nsd_all_bits   = 16,
 875                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 876         },
 877         {
 878                 .nsd_type       = LDLM_NS_TYPE_MDT,
 879                 .nsd_bkt_bits   = 14,
 880                 .nsd_all_bits   = 21,
 881                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 882         },
 883         {
 884                 .nsd_type       = LDLM_NS_TYPE_OSC,
 885                 .nsd_bkt_bits   = 8,
 886                 .nsd_all_bits   = 12,
 887                 .nsd_hops       = &ldlm_ns_hash_ops,
 888         },
 889         {
 890                 .nsd_type       = LDLM_NS_TYPE_OST,
 891                 .nsd_bkt_bits   = 11,
 892                 .nsd_all_bits   = 17,
 893                 .nsd_hops       = &ldlm_ns_hash_ops,
 894         },
 895         {
 896                 .nsd_type       = LDLM_NS_TYPE_MGC,
 897                 .nsd_bkt_bits   = 4,
 898                 .nsd_all_bits   = 4,
 899                 .nsd_hops       = &ldlm_ns_hash_ops,
 900         },
 901         {
 902                 .nsd_type       = LDLM_NS_TYPE_MGT,
 903                 .nsd_bkt_bits   = 4,
 904                 .nsd_all_bits   = 4,
 905                 .nsd_hops       = &ldlm_ns_hash_ops,
 906         },
 907         {
 908                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 909         },
 910 };
 911
 912 /**
 913  * Create and initialize new empty namespace.
 914  */
 915 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 916                                           enum ldlm_side client,
 917                                           enum ldlm_appetite apt,
 918                                           enum ldlm_ns_type ns_type)
 919 {
 920         struct ldlm_namespace *ns = NULL;
 921         struct ldlm_ns_bucket *nsb;
 922         struct ldlm_ns_hash_def *nsd;
 923         struct cfs_hash_bd bd;
 924         int idx;
 925         int rc;
 926         ENTRY;
 927
 928         LASSERT(obd != NULL);
 929
 930         rc = ldlm_get_ref();
 931         if (rc) {
 932                 CERROR("ldlm_get_ref failed: %d\n", rc);
 933                 RETURN(NULL);
 934         }
 935
 936         for (idx = 0;;idx++) {
 937                 nsd = &ldlm_ns_hash_defs[idx];
 938                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 939                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 940                         GOTO(out_ref, NULL);
 941                 }
 942
 943                 if (nsd->nsd_type == ns_type)
 944                         break;
 945         }
 946
 947         OBD_ALLOC_PTR(ns);
 948         if (!ns)
 949                 GOTO(out_ref, NULL);
 950
 951         ns->ns_rs_hash = cfs_hash_create(name,
 952                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 953                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 954                                          CFS_HASH_MIN_THETA,
 955                                          CFS_HASH_MAX_THETA,
 956                                          nsd->nsd_hops,
 957                                          CFS_HASH_DEPTH |
 958                                          CFS_HASH_BIGNAME |
 959                                          CFS_HASH_SPIN_BKTLOCK |
 960                                          CFS_HASH_NO_ITEMREF);
 961         if (ns->ns_rs_hash == NULL)
 962                 GOTO(out_ns, NULL);
 963
 964         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 965                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 966                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 967                 nsb->nsb_namespace = ns;
 968                 nsb->nsb_reclaim_start = 0;
 969         }
 970
 971         ns->ns_obd = obd;
 972         ns->ns_appetite = apt;
 973         ns->ns_client = client;
 974         ns->ns_name = kstrdup(name, GFP_KERNEL);
 975         if (!ns->ns_name)
 976                 goto out_hash;
 977
 978         INIT_LIST_HEAD(&ns->ns_list_chain);
 979         INIT_LIST_HEAD(&ns->ns_unused_list);
 980         spin_lock_init(&ns->ns_lock);
 981         atomic_set(&ns->ns_bref, 0);
 982         init_waitqueue_head(&ns->ns_waitq);
 983
 984         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 985         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 986         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 987
 988         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 989         ns->ns_nr_unused          = 0;
 990         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 991         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 992         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 993         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 994         ns->ns_timeouts           = 0;
 995         ns->ns_orig_connect_flags = 0;
 996         ns->ns_connect_flags      = 0;
 997         ns->ns_stopping           = 0;
 998         ns->ns_reclaim_start      = 0;
 999         ns->ns_last_pos           = &ns->ns_unused_list;
1000
1001         rc = ldlm_namespace_sysfs_register(ns);
1002         if (rc) {
1003                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
1004                 GOTO(out_hash, rc);
1005         }
1006
1007         rc = ldlm_namespace_debugfs_register(ns);
1008         if (rc) {
1009                 CERROR("Can't initialize ns proc, rc %d\n", rc);
1010                 GOTO(out_sysfs, rc);
1011         }
1012
1013         idx = ldlm_namespace_nr_read(client);
1014         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
1015         if (rc) {
1016                 CERROR("Can't initialize lock pool, rc %d\n", rc);
1017                 GOTO(out_proc, rc);
1018         }
1019
1020         ldlm_namespace_register(ns, client);
1021         RETURN(ns);
1022 out_proc:
1023         ldlm_namespace_debugfs_unregister(ns);
1024 out_sysfs:
1025         ldlm_namespace_sysfs_unregister(ns);
1026         ldlm_namespace_cleanup(ns, 0);
1027 out_hash:
1028         kfree(ns->ns_name);
1029         cfs_hash_putref(ns->ns_rs_hash);
1030 out_ns:
1031         OBD_FREE_PTR(ns);
1032 out_ref:
1033         ldlm_put_ref();
1034         RETURN(NULL);
1035 }
1036 EXPORT_SYMBOL(ldlm_namespace_new);
1037
1038 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
1039
1040 /**
1041  * Cancel and destroy all locks on a resource.
1042  *
1043  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1044  * clean up.  This is currently only used for recovery, and we make
1045  * certain assumptions as a result--notably, that we shouldn't cancel
1046  * locks with refs.
1047  */
1048 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1049                              __u64 flags)
1050 {
1051         struct list_head *tmp;
1052         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1053         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1054
1055         do {
1056                 struct ldlm_lock *lock = NULL;
1057
1058                 /* First, we look for non-cleaned-yet lock
1059                  * all cleaned locks are marked by CLEANED flag. */
1060                 lock_res(res);
1061                 list_for_each(tmp, q) {
1062                         lock = list_entry(tmp, struct ldlm_lock,
1063                                           l_res_link);
1064                         if (ldlm_is_cleaned(lock)) {
1065                                 lock = NULL;
1066                                 continue;
1067                         }
1068                         LDLM_LOCK_GET(lock);
1069                         ldlm_set_cleaned(lock);
1070                         break;
1071                 }
1072
1073                 if (lock == NULL) {
1074                         unlock_res(res);
1075                         break;
1076                 }
1077
1078                 /* Set CBPENDING so nothing in the cancellation path
1079                  * can match this lock. */
1080                 ldlm_set_cbpending(lock);
1081                 ldlm_set_failed(lock);
1082                 lock->l_flags |= flags;
1083
1084                 /* ... without sending a CANCEL message for local_only. */
1085                 if (local_only)
1086                         ldlm_set_local_only(lock);
1087
1088                 if (local_only && (lock->l_readers || lock->l_writers)) {
1089                         /* This is a little bit gross, but much better than the
1090                          * alternative: pretend that we got a blocking AST from
1091                          * the server, so that when the lock is decref'd, it
1092                          * will go away ... */
1093                         unlock_res(res);
1094                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1095                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1096                                 set_current_state(TASK_UNINTERRUPTIBLE);
1097                                 schedule_timeout(cfs_time_seconds(4));
1098                                 set_current_state(TASK_RUNNING);
1099                         }
1100                         if (lock->l_completion_ast)
1101                                 lock->l_completion_ast(lock,
1102                                                        LDLM_FL_FAILED, NULL);
1103                         LDLM_LOCK_RELEASE(lock);
1104                         continue;
1105                 }
1106
1107                 if (client) {
1108                         struct lustre_handle lockh;
1109
1110                         unlock_res(res);
1111                         ldlm_lock2handle(lock, &lockh);
1112                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1113                         if (rc)
1114                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1115                 } else {
1116                         unlock_res(res);
1117                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
1118                                    "client node");
1119                         ldlm_lock_cancel(lock);
1120                 }
1121                 LDLM_LOCK_RELEASE(lock);
1122         } while (1);
1123 }
1124
1125 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1126                                struct hlist_node *hnode, void *arg)
1127 {
1128         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1129         __u64 flags = *(__u64 *)arg;
1130
1131         cleanup_resource(res, &res->lr_granted, flags);
1132         cleanup_resource(res, &res->lr_waiting, flags);
1133
1134         return 0;
1135 }
1136
1137 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1138                                   struct hlist_node *hnode, void *arg)
1139 {
1140         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1141
1142         lock_res(res);
1143         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1144                "(%d) after lock cleanup; forcing cleanup.\n",
1145                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1146                atomic_read(&res->lr_refcount) - 1);
1147
1148         ldlm_resource_dump(D_ERROR, res);
1149         unlock_res(res);
1150         return 0;
1151 }
1152
1153 /**
1154  * Cancel and destroy all locks in the namespace.
1155  *
1156  * Typically used during evictions when server notified client that it was
1157  * evicted and all of its state needs to be destroyed.
1158  * Also used during shutdown.
1159  */
1160 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1161 {
1162         if (ns == NULL) {
1163                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1164                 return ELDLM_OK;
1165         }
1166
1167         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1168                                  &flags, 0);
1169         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1170                                  NULL, 0);
1171         return ELDLM_OK;
1172 }
1173 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1174
1175 /**
1176  * Attempts to free namespace.
1177  *
1178  * Only used when namespace goes away, like during an unmount.
1179  */
1180 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1181 {
1182         ENTRY;
1183
1184         /* At shutdown time, don't call the cancellation callback */
1185         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1186
1187         if (atomic_read(&ns->ns_bref) > 0) {
1188                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1189                 int rc;
1190                 CDEBUG(D_DLMTRACE,
1191                        "dlm namespace %s free waiting on refcount %d\n",
1192                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1193 force_wait:
1194                 if (force)
1195                         lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
1196                                           MSEC_PER_SEC) / 4, NULL, NULL);
1197
1198                 rc = l_wait_event(ns->ns_waitq,
1199                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1200
1201                 /* Forced cleanups should be able to reclaim all references,
1202                  * so it's safe to wait forever... we can't leak locks... */
1203                 if (force && rc == -ETIMEDOUT) {
1204                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1205                                        "namespace with %d resources in use, "
1206                                        "(rc=%d)\n", ldlm_ns_name(ns),
1207                                        atomic_read(&ns->ns_bref), rc);
1208                         GOTO(force_wait, rc);
1209                 }
1210
1211                 if (atomic_read(&ns->ns_bref)) {
1212                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1213                                        "with %d resources in use, (rc=%d)\n",
1214                                        ldlm_ns_name(ns),
1215                                        atomic_read(&ns->ns_bref), rc);
1216                         RETURN(ELDLM_NAMESPACE_EXISTS);
1217                 }
1218                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1219                        ldlm_ns_name(ns));
1220         }
1221
1222         RETURN(ELDLM_OK);
1223 }
1224
1225 /**
1226  * Performs various cleanups for passed \a ns to make it drop refc and be
1227  * ready for freeing. Waits for refc == 0.
1228  *
1229  * The following is done:
1230  * (0) Unregister \a ns from its list to make inaccessible for potential
1231  * users like pools thread and others;
1232  * (1) Clear all locks in \a ns.
1233  */
1234 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1235                                struct obd_import *imp,
1236                                int force)
1237 {
1238         int rc;
1239         ENTRY;
1240         if (!ns) {
1241                 EXIT;
1242                 return;
1243         }
1244
1245         spin_lock(&ns->ns_lock);
1246         ns->ns_stopping = 1;
1247         spin_unlock(&ns->ns_lock);
1248
1249         /*
1250          * Can fail with -EINTR when force == 0 in which case try harder.
1251          */
1252         rc = __ldlm_namespace_free(ns, force);
1253         if (rc != ELDLM_OK) {
1254                 if (imp) {
1255                         ptlrpc_disconnect_import(imp, 0);
1256                         ptlrpc_invalidate_import(imp);
1257                 }
1258
1259                 /*
1260                  * With all requests dropped and the import inactive
1261                  * we are gaurenteed all reference will be dropped.
1262                  */
1263                 rc = __ldlm_namespace_free(ns, 1);
1264                 LASSERT(rc == 0);
1265         }
1266         EXIT;
1267 }
1268 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1269
1270 /**
1271  * Performs freeing memory structures related to \a ns. This is only done
1272  * when ldlm_namespce_free_prior() successfully removed all resources
1273  * referencing \a ns and its refc == 0.
1274  */
1275 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1276 {
1277         ENTRY;
1278         if (!ns) {
1279                 EXIT;
1280                 return;
1281         }
1282
1283         /* Make sure that nobody can find this ns in its list. */
1284         ldlm_namespace_unregister(ns, ns->ns_client);
1285         /* Fini pool _before_ parent proc dir is removed. This is important as
1286          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1287          * Removing it after @dir may cause oops. */
1288         ldlm_pool_fini(&ns->ns_pool);
1289
1290         ldlm_namespace_debugfs_unregister(ns);
1291         ldlm_namespace_sysfs_unregister(ns);
1292         cfs_hash_putref(ns->ns_rs_hash);
1293         kfree(ns->ns_name);
1294         /* Namespace \a ns should be not on list at this time, otherwise
1295          * this will cause issues related to using freed \a ns in poold
1296          * thread.
1297          */
1298         LASSERT(list_empty(&ns->ns_list_chain));
1299         OBD_FREE_PTR(ns);
1300         ldlm_put_ref();
1301         EXIT;
1302 }
1303 EXPORT_SYMBOL(ldlm_namespace_free_post);
1304
1305 /**
1306  * Cleanup the resource, and free namespace.
1307  * bug 12864:
1308  * Deadlock issue:
1309  * proc1: destroy import
1310  *        class_disconnect_export(grab cl_sem) ->
1311  *              -> ldlm_namespace_free ->
1312  *              -> lprocfs_remove(grab _lprocfs_lock).
1313  * proc2: read proc info
1314  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1315  *              -> osc_rd_active, etc(grab cl_sem).
1316  *
1317  * So that I have to split the ldlm_namespace_free into two parts - the first
1318  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1319  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1320  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1321  * held.
1322  */
1323 void ldlm_namespace_free(struct ldlm_namespace *ns,
1324                          struct obd_import *imp,
1325                          int force)
1326 {
1327         ldlm_namespace_free_prior(ns, imp, force);
1328         ldlm_namespace_free_post(ns);
1329 }
1330 EXPORT_SYMBOL(ldlm_namespace_free);
1331
1332 void ldlm_namespace_get(struct ldlm_namespace *ns)
1333 {
1334         atomic_inc(&ns->ns_bref);
1335 }
1336
1337 /* This is only for callers that care about refcount */
1338 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1339 {
1340         return atomic_inc_return(&ns->ns_bref);
1341 }
1342
1343 void ldlm_namespace_put(struct ldlm_namespace *ns)
1344 {
1345         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1346                 wake_up(&ns->ns_waitq);
1347                 spin_unlock(&ns->ns_lock);
1348         }
1349 }
1350
1351 /** Register \a ns in the list of namespaces */
1352 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1353 {
1354         mutex_lock(ldlm_namespace_lock(client));
1355         LASSERT(list_empty(&ns->ns_list_chain));
1356         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1357         ldlm_namespace_nr_inc(client);
1358         mutex_unlock(ldlm_namespace_lock(client));
1359 }
1360
1361 /** Unregister \a ns from the list of namespaces. */
1362 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1363 {
1364         mutex_lock(ldlm_namespace_lock(client));
1365         LASSERT(!list_empty(&ns->ns_list_chain));
1366         /* Some asserts and possibly other parts of the code are still
1367          * using list_empty(&ns->ns_list_chain). This is why it is
1368          * important to use list_del_init() here. */
1369         list_del_init(&ns->ns_list_chain);
1370         ldlm_namespace_nr_dec(client);
1371         mutex_unlock(ldlm_namespace_lock(client));
1372 }
1373
1374 /** Should be called with ldlm_namespace_lock(client) taken. */
1375 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1376                                           enum ldlm_side client)
1377 {
1378         LASSERT(!list_empty(&ns->ns_list_chain));
1379         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1380         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1381 }
1382
1383 /** Should be called with ldlm_namespace_lock(client) taken. */
1384 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1385                                             enum ldlm_side client)
1386 {
1387         LASSERT(!list_empty(&ns->ns_list_chain));
1388         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1389         list_move_tail(&ns->ns_list_chain,
1390                        ldlm_namespace_inactive_list(client));
1391 }
1392
1393 /** Should be called with ldlm_namespace_lock(client) taken. */
1394 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1395 {
1396         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1397         LASSERT(!list_empty(ldlm_namespace_list(client)));
1398         return container_of(ldlm_namespace_list(client)->next,
1399                             struct ldlm_namespace, ns_list_chain);
1400 }
1401
1402 /** Create and initialize new resource. */
1403 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1404 {
1405         struct ldlm_resource *res;
1406         int idx;
1407
1408         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1409         if (res == NULL)
1410                 return NULL;
1411
1412         if (ldlm_type == LDLM_EXTENT) {
1413                 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1414                                sizeof(*res->lr_itree) * LCK_MODE_NUM);
1415                 if (res->lr_itree == NULL) {
1416                         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1417                         return NULL;
1418                 }
1419                 /* Initialize interval trees for each lock mode. */
1420                 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1421                         res->lr_itree[idx].lit_size = 0;
1422                         res->lr_itree[idx].lit_mode = 1 << idx;
1423                         res->lr_itree[idx].lit_root = NULL;
1424                 }
1425         }
1426
1427         INIT_LIST_HEAD(&res->lr_granted);
1428         INIT_LIST_HEAD(&res->lr_waiting);
1429
1430         atomic_set(&res->lr_refcount, 1);
1431         spin_lock_init(&res->lr_lock);
1432         lu_ref_init(&res->lr_reference);
1433
1434         /* Since LVB init can be delayed now, there is no longer need to
1435          * immediatelly acquire mutex here. */
1436         mutex_init(&res->lr_lvb_mutex);
1437         res->lr_lvb_initialized = false;
1438
1439         return res;
1440 }
1441
1442 /**
1443  * Return a reference to resource with given name, creating it if necessary.
1444  * Args: namespace with ns_lock unlocked
1445  * Locks: takes and releases NS hash-lock and res->lr_lock
1446  * Returns: referenced, unlocked ldlm_resource or NULL
1447  */
1448 struct ldlm_resource *
1449 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1450                   const struct ldlm_res_id *name, enum ldlm_type type,
1451                   int create)
1452 {
1453         struct hlist_node       *hnode;
1454         struct ldlm_resource    *res = NULL;
1455         struct cfs_hash_bd              bd;
1456         __u64                   version;
1457         int                     ns_refcount = 0;
1458
1459         LASSERT(ns != NULL);
1460         LASSERT(parent == NULL);
1461         LASSERT(ns->ns_rs_hash != NULL);
1462         LASSERT(name->name[0] != 0);
1463
1464         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1465         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1466         if (hnode != NULL) {
1467                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1468                 GOTO(found, res);
1469         }
1470
1471         version = cfs_hash_bd_version_get(&bd);
1472         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1473
1474         if (create == 0)
1475                 return ERR_PTR(-ENOENT);
1476
1477         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1478                  "type: %d\n", type);
1479         res = ldlm_resource_new(type);
1480         if (res == NULL)
1481                 return ERR_PTR(-ENOMEM);
1482
1483         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1484         res->lr_name = *name;
1485         res->lr_type = type;
1486
1487         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1488         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1489                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1490
1491         if (hnode != NULL) {
1492                 /* Someone won the race and already added the resource. */
1493                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1494                 /* Clean lu_ref for failed resource. */
1495                 lu_ref_fini(&res->lr_reference);
1496                 if (res->lr_itree != NULL)
1497                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1498                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1499                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1500 found:
1501                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1502                 return res;
1503         }
1504         /* We won! Let's add the resource. */
1505         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1506         if (cfs_hash_bd_count_get(&bd) == 1)
1507                 ns_refcount = ldlm_namespace_get_return(ns);
1508
1509         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1510
1511         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1512
1513         /* Let's see if we happened to be the very first resource in this
1514          * namespace. If so, and this is a client namespace, we need to move
1515          * the namespace into the active namespaces list to be patrolled by
1516          * the ldlm_poold. */
1517         if (ns_is_client(ns) && ns_refcount == 1) {
1518                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1519                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1520                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1521         }
1522
1523         return res;
1524 }
1525 EXPORT_SYMBOL(ldlm_resource_get);
1526
1527 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1528 {
1529         LASSERT(res != NULL);
1530         LASSERT(res != LP_POISON);
1531         atomic_inc(&res->lr_refcount);
1532         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1533                atomic_read(&res->lr_refcount));
1534         return res;
1535 }
1536
1537 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1538                                          struct ldlm_resource *res)
1539 {
1540         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1541
1542         if (!list_empty(&res->lr_granted)) {
1543                 ldlm_resource_dump(D_ERROR, res);
1544                 LBUG();
1545         }
1546
1547         if (!list_empty(&res->lr_waiting)) {
1548                 ldlm_resource_dump(D_ERROR, res);
1549                 LBUG();
1550         }
1551
1552         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1553                                bd, &res->lr_hash);
1554         lu_ref_fini(&res->lr_reference);
1555         if (cfs_hash_bd_count_get(bd) == 0)
1556                 ldlm_namespace_put(nsb->nsb_namespace);
1557 }
1558
1559 /* Returns 1 if the resource was freed, 0 if it remains. */
1560 int ldlm_resource_putref(struct ldlm_resource *res)
1561 {
1562         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1563         struct cfs_hash_bd   bd;
1564
1565         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1566         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1567                res, atomic_read(&res->lr_refcount) - 1);
1568
1569         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1570         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1571                 __ldlm_resource_putref_final(&bd, res);
1572                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1573                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1574                         ns->ns_lvbo->lvbo_free(res);
1575                 if (res->lr_itree != NULL)
1576                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1577                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1578                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1579                 return 1;
1580         }
1581         return 0;
1582 }
1583 EXPORT_SYMBOL(ldlm_resource_putref);
1584
1585 /**
1586  * Add a lock into a given resource into specified lock list.
1587  */
1588 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1589                             struct ldlm_lock *lock)
1590 {
1591         check_res_locked(res);
1592
1593         LDLM_DEBUG(lock, "About to add this lock");
1594
1595         if (ldlm_is_destroyed(lock)) {
1596                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1597                 return;
1598         }
1599
1600         LASSERT(list_empty(&lock->l_res_link));
1601
1602         list_add_tail(&lock->l_res_link, head);
1603 }
1604
1605 /**
1606  * Insert a lock into resource after specified lock.
1607  *
1608  * Obtain resource description from the lock we are inserting after.
1609  */
1610 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1611                                      struct ldlm_lock *new)
1612 {
1613         struct ldlm_resource *res = original->l_resource;
1614
1615         check_res_locked(res);
1616
1617         ldlm_resource_dump(D_INFO, res);
1618         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1619
1620         if (ldlm_is_destroyed(new)) {
1621                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1622                 goto out;
1623         }
1624
1625         LASSERT(list_empty(&new->l_res_link));
1626
1627         list_add(&new->l_res_link, &original->l_res_link);
1628  out:;
1629 }
1630
1631 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1632 {
1633         int type = lock->l_resource->lr_type;
1634
1635         check_res_locked(lock->l_resource);
1636         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1637                 ldlm_unlink_lock_skiplist(lock);
1638         else if (type == LDLM_EXTENT)
1639                 ldlm_extent_unlink_lock(lock);
1640         list_del_init(&lock->l_res_link);
1641 }
1642 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1643
1644 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1645 {
1646         desc->lr_type = res->lr_type;
1647         desc->lr_name = res->lr_name;
1648 }
1649
1650 /**
1651  * Print information about all locks in all namespaces on this node to debug
1652  * log.
1653  */
1654 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1655 {
1656         struct list_head *tmp;
1657
1658         if (!((libcfs_debug | D_ERROR) & level))
1659                 return;
1660
1661         mutex_lock(ldlm_namespace_lock(client));
1662
1663         list_for_each(tmp, ldlm_namespace_list(client)) {
1664                 struct ldlm_namespace *ns;
1665
1666                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1667                 ldlm_namespace_dump(level, ns);
1668         }
1669
1670         mutex_unlock(ldlm_namespace_lock(client));
1671 }
1672
1673 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1674                               struct hlist_node *hnode, void *arg)
1675 {
1676         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1677         int    level = (int)(unsigned long)arg;
1678
1679         lock_res(res);
1680         ldlm_resource_dump(level, res);
1681         unlock_res(res);
1682
1683         return 0;
1684 }
1685
1686 /**
1687  * Print information about all locks in this namespace on this node to debug
1688  * log.
1689  */
1690 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1691 {
1692         if (!((libcfs_debug | D_ERROR) & level))
1693                 return;
1694
1695         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1696                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1697                ns_is_client(ns) ? "client" : "server");
1698
1699         if (ktime_get_seconds() < ns->ns_next_dump)
1700                 return;
1701
1702         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1703                                  ldlm_res_hash_dump,
1704                                  (void *)(unsigned long)level, 0);
1705         spin_lock(&ns->ns_lock);
1706         ns->ns_next_dump = ktime_get_seconds() + 10;
1707         spin_unlock(&ns->ns_lock);
1708 }
1709
1710 /**
1711  * Print information about all locks in this resource to debug log.
1712  */
1713 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1714 {
1715         struct ldlm_lock *lock;
1716         unsigned int granted = 0;
1717
1718         CLASSERT(RES_NAME_SIZE == 4);
1719
1720         if (!((libcfs_debug | D_ERROR) & level))
1721                 return;
1722
1723         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1724                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1725
1726         if (!list_empty(&res->lr_granted)) {
1727                 CDEBUG(level, "Granted locks (in reverse order):\n");
1728                 list_for_each_entry_reverse(lock, &res->lr_granted,
1729                                                 l_res_link) {
1730                         LDLM_DEBUG_LIMIT(level, lock, "###");
1731                         if (!(level & D_CANTMASK) &&
1732                             ++granted > ldlm_dump_granted_max) {
1733                                 CDEBUG(level, "only dump %d granted locks to "
1734                                        "avoid DDOS.\n", granted);
1735                                 break;
1736                         }
1737                 }
1738         }
1739
1740         if (!list_empty(&res->lr_waiting)) {
1741                 CDEBUG(level, "Waiting locks:\n");
1742                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1743                         LDLM_DEBUG_LIMIT(level, lock, "###");
1744         }
1745 }
1746 EXPORT_SYMBOL(ldlm_resource_dump);