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