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