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 lprocfs_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 sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 462 }
 463
 464 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 465                                      struct attribute *attr,
 466                                      const char *buffer, size_t count)
 467 {
 468         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 469                                                  ns_kobj);
 470         unsigned long long tmp;
 471
 472         if (kstrtoull(buffer, 10, &tmp))
 473                 return -EINVAL;
 474
 475         ns->ns_dirty_age_limit = tmp;
 476
 477         return count;
 478 }
 479 LUSTRE_RW_ATTR(dirty_age_limit);
 480
 481 #ifdef HAVE_SERVER_SUPPORT
 482 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 483                                     struct attribute *attr, char *buf)
 484 {
 485         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 486                                                  ns_kobj);
 487
 488         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 489 }
 490
 491 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 492                                      struct attribute *attr,
 493                                      const char *buffer, size_t count)
 494 {
 495         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 496                                                  ns_kobj);
 497         unsigned long long tmp;
 498
 499         if (kstrtoull(buffer, 10, &tmp))
 500                 return -EINVAL;
 501
 502         ns->ns_ctime_age_limit = tmp;
 503
 504         return count;
 505 }
 506 LUSTRE_RW_ATTR(ctime_age_limit);
 507
 508 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 509                                   char *buf)
 510 {
 511         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 512                                                  ns_kobj);
 513
 514         return sprintf(buf, "%d\n", ns->ns_timeouts);
 515 }
 516 LUSTRE_RO_ATTR(lock_timeouts);
 517
 518 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 519                                      struct attribute *attr, char *buf)
 520 {
 521         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 522                                                  ns_kobj);
 523
 524         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 525 }
 526
 527 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 528                                       struct attribute *attr,
 529                                       const char *buffer, size_t count)
 530 {
 531         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 532                                                  ns_kobj);
 533         unsigned long tmp;
 534         int err;
 535
 536         err = kstrtoul(buffer, 10, &tmp);
 537         if (err != 0)
 538                 return -EINVAL;
 539
 540         ns->ns_max_nolock_size = tmp;
 541
 542         return count;
 543 }
 544 LUSTRE_RW_ATTR(max_nolock_bytes);
 545
 546 static ssize_t contention_seconds_show(struct kobject *kobj,
 547                                        struct attribute *attr, char *buf)
 548 {
 549         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 550                                                  ns_kobj);
 551
 552         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 553 }
 554
 555 static ssize_t contention_seconds_store(struct kobject *kobj,
 556                                         struct attribute *attr,
 557                                         const char *buffer, size_t count)
 558 {
 559         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 560                                                  ns_kobj);
 561         unsigned long long tmp;
 562
 563         if (kstrtoull(buffer, 10, &tmp))
 564                 return -EINVAL;
 565
 566         ns->ns_contention_time = tmp;
 567
 568         return count;
 569 }
 570 LUSTRE_RW_ATTR(contention_seconds);
 571
 572 static ssize_t contended_locks_show(struct kobject *kobj,
 573                                     struct attribute *attr, char *buf)
 574 {
 575         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 576                                                  ns_kobj);
 577
 578         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 579 }
 580
 581 static ssize_t contended_locks_store(struct kobject *kobj,
 582                                      struct attribute *attr,
 583                                      const char *buffer, size_t count)
 584 {
 585         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 586                                                  ns_kobj);
 587         unsigned long tmp;
 588         int err;
 589
 590         err = kstrtoul(buffer, 10, &tmp);
 591         if (err != 0)
 592                 return -EINVAL;
 593
 594         ns->ns_contended_locks = tmp;
 595
 596         return count;
 597 }
 598 LUSTRE_RW_ATTR(contended_locks);
 599
 600 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 601                                      struct attribute *attr, char *buf)
 602 {
 603         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 604                                                  ns_kobj);
 605
 606         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 607 }
 608
 609 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 610                                       struct attribute *attr,
 611                                       const char *buffer, size_t count)
 612 {
 613         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 614                                                  ns_kobj);
 615         unsigned long tmp;
 616         int err;
 617
 618         err = kstrtoul(buffer, 10, &tmp);
 619         if (err != 0)
 620                 return -EINVAL;
 621
 622         ns->ns_max_parallel_ast = tmp;
 623
 624         return count;
 625 }
 626 LUSTRE_RW_ATTR(max_parallel_ast);
 627
 628 #endif /* HAVE_SERVER_SUPPORT */
 629
 630 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 631 static struct attribute *ldlm_ns_attrs[] = {
 632         &lustre_attr_resource_count.attr,
 633         &lustre_attr_lock_count.attr,
 634         &lustre_attr_lock_unused_count.attr,
 635         &lustre_attr_lru_size.attr,
 636         &lustre_attr_lru_max_age.attr,
 637         &lustre_attr_early_lock_cancel.attr,
 638         &lustre_attr_dirty_age_limit.attr,
 639 #ifdef HAVE_SERVER_SUPPORT
 640         &lustre_attr_ctime_age_limit.attr,
 641         &lustre_attr_lock_timeouts.attr,
 642         &lustre_attr_max_nolock_bytes.attr,
 643         &lustre_attr_contention_seconds.attr,
 644         &lustre_attr_contended_locks.attr,
 645         &lustre_attr_max_parallel_ast.attr,
 646 #endif
 647         NULL,
 648 };
 649
 650 static void ldlm_ns_release(struct kobject *kobj)
 651 {
 652         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 653                                                  ns_kobj);
 654         complete(&ns->ns_kobj_unregister);
 655 }
 656
 657 static struct kobj_type ldlm_ns_ktype = {
 658         .default_attrs  = ldlm_ns_attrs,
 659         .sysfs_ops      = &lustre_sysfs_ops,
 660         .release        = ldlm_ns_release,
 661 };
 662
 663 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 664 {
 665         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 666                 CERROR("dlm namespace %s has no procfs dir?\n",
 667                        ldlm_ns_name(ns));
 668         else
 669                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 670
 671         if (ns->ns_stats != NULL)
 672                 lprocfs_free_stats(&ns->ns_stats);
 673 }
 674
 675 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 676 {
 677         kobject_put(&ns->ns_kobj);
 678         wait_for_completion(&ns->ns_kobj_unregister);
 679 }
 680
 681 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 682 {
 683         int err;
 684
 685         ns->ns_kobj.kset = ldlm_ns_kset;
 686         init_completion(&ns->ns_kobj_unregister);
 687         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 688                                    "%s", ldlm_ns_name(ns));
 689
 690         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 691         if (!ns->ns_stats) {
 692                 kobject_put(&ns->ns_kobj);
 693                 return -ENOMEM;
 694         }
 695
 696         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 697                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 698
 699         return err;
 700 }
 701
 702 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 703 {
 704         struct dentry *ns_entry;
 705
 706         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 707                 ns_entry = ns->ns_debugfs_entry;
 708         } else {
 709                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 710                                               ldlm_ns_debugfs_dir);
 711                 if (!ns_entry)
 712                         return -ENOMEM;
 713                 ns->ns_debugfs_entry = ns_entry;
 714         }
 715
 716         return 0;
 717 }
 718 #undef MAX_STRING_SIZE
 719
 720 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 721                                   const void *key, unsigned int mask)
 722 {
 723         const struct ldlm_res_id *id = key;
 724         unsigned int val = 0;
 725         unsigned int i;
 726
 727         for (i = 0; i < RES_NAME_SIZE; i++)
 728                 val += id->name[i];
 729         return val & mask;
 730 }
 731
 732 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 733 {
 734         struct lu_fid       fid;
 735         __u32               hash;
 736         __u32               val;
 737
 738         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 739         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 740         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 741
 742         hash = fid_flatten32(&fid);
 743         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 744         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 745                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 746         } else {
 747                 val = fid_oid(&fid);
 748         }
 749         hash += (val >> 5) + (val << 11);
 750         return cfs_hash_32(hash, bits);
 751 }
 752
 753 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 754 {
 755         struct ldlm_resource   *res;
 756
 757         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 758         return &res->lr_name;
 759 }
 760
 761 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 762 {
 763         struct ldlm_resource   *res;
 764
 765         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 766         return ldlm_res_eq((const struct ldlm_res_id *)key,
 767                            (const struct ldlm_res_id *)&res->lr_name);
 768 }
 769
 770 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 771 {
 772         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 773 }
 774
 775 static void
 776 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 777 {
 778         struct ldlm_resource *res;
 779
 780         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 781         ldlm_resource_getref(res);
 782 }
 783
 784 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 785 {
 786         struct ldlm_resource *res;
 787
 788         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 789         ldlm_resource_putref(res);
 790 }
 791
 792 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 793         .hs_hash        = ldlm_res_hop_hash,
 794         .hs_key         = ldlm_res_hop_key,
 795         .hs_keycmp      = ldlm_res_hop_keycmp,
 796         .hs_keycpy      = NULL,
 797         .hs_object      = ldlm_res_hop_object,
 798         .hs_get         = ldlm_res_hop_get_locked,
 799         .hs_put         = ldlm_res_hop_put
 800 };
 801
 802 static struct {
 803         /** hash bucket bits */
 804         unsigned                nsd_bkt_bits;
 805         /** hash bits */
 806         unsigned                nsd_all_bits;
 807 } ldlm_ns_hash_defs[] = {
 808         [LDLM_NS_TYPE_MDC] = {
 809                 .nsd_bkt_bits   = 11,
 810                 .nsd_all_bits   = 16,
 811         },
 812         [LDLM_NS_TYPE_MDT] = {
 813                 .nsd_bkt_bits   = 14,
 814                 .nsd_all_bits   = 21,
 815         },
 816         [LDLM_NS_TYPE_OSC] = {
 817                 .nsd_bkt_bits   = 8,
 818                 .nsd_all_bits   = 12,
 819         },
 820         [LDLM_NS_TYPE_OST] = {
 821                 .nsd_bkt_bits   = 11,
 822                 .nsd_all_bits   = 17,
 823         },
 824         [LDLM_NS_TYPE_MGC] = {
 825                 .nsd_bkt_bits   = 3,
 826                 .nsd_all_bits   = 4,
 827         },
 828         [LDLM_NS_TYPE_MGT] = {
 829                 .nsd_bkt_bits   = 3,
 830                 .nsd_all_bits   = 4,
 831         },
 832 };
 833
 834 /**
 835  * Create and initialize new empty namespace.
 836  */
 837 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 838                                           enum ldlm_side client,
 839                                           enum ldlm_appetite apt,
 840                                           enum ldlm_ns_type ns_type)
 841 {
 842         struct ldlm_namespace *ns = NULL;
 843         int idx;
 844         int rc;
 845
 846         ENTRY;
 847         LASSERT(obd != NULL);
 848
 849         rc = ldlm_get_ref();
 850         if (rc) {
 851                 CERROR("ldlm_get_ref failed: %d\n", rc);
 852                 RETURN(NULL);
 853         }
 854
 855         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 856             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 857                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 858                 GOTO(out_ref, NULL);
 859         }
 860
 861         OBD_ALLOC_PTR(ns);
 862         if (!ns)
 863                 GOTO(out_ref, NULL);
 864
 865         ns->ns_rs_hash = cfs_hash_create(name,
 866                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 867                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 868                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 869                                          0,
 870                                          CFS_HASH_MIN_THETA,
 871                                          CFS_HASH_MAX_THETA,
 872                                          &ldlm_ns_hash_ops,
 873                                          CFS_HASH_DEPTH |
 874                                          CFS_HASH_BIGNAME |
 875                                          CFS_HASH_SPIN_BKTLOCK |
 876                                          CFS_HASH_NO_ITEMREF);
 877         if (ns->ns_rs_hash == NULL)
 878                 GOTO(out_ns, NULL);
 879
 880         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 881                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 882
 883         OBD_ALLOC_LARGE(ns->ns_rs_buckets,
 884                         BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 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    = LDLM_DIRTY_AGE_LIMIT;
 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_LARGE(ns->ns_rs_buckets,
 955                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 956         kfree(ns->ns_name);
 957         cfs_hash_putref(ns->ns_rs_hash);
 958 out_ns:
 959         OBD_FREE_PTR(ns);
 960 out_ref:
 961         ldlm_put_ref();
 962         RETURN(NULL);
 963 }
 964 EXPORT_SYMBOL(ldlm_namespace_new);
 965
 966 /**
 967  * Cancel and destroy all locks on a resource.
 968  *
 969  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 970  * clean up.  This is currently only used for recovery, and we make
 971  * certain assumptions as a result--notably, that we shouldn't cancel
 972  * locks with refs.
 973  */
 974 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 975                              __u64 flags)
 976 {
 977         struct list_head *tmp;
 978         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 979         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 980
 981         do {
 982                 struct ldlm_lock *lock = NULL;
 983
 984                 /* First, we look for non-cleaned-yet lock
 985                  * all cleaned locks are marked by CLEANED flag. */
 986                 lock_res(res);
 987                 list_for_each(tmp, q) {
 988                         lock = list_entry(tmp, struct ldlm_lock,
 989                                           l_res_link);
 990                         if (ldlm_is_cleaned(lock)) {
 991                                 lock = NULL;
 992                                 continue;
 993                         }
 994                         LDLM_LOCK_GET(lock);
 995                         ldlm_set_cleaned(lock);
 996                         break;
 997                 }
 998
 999                 if (lock == NULL) {
1000                         unlock_res(res);
1001                         break;
1002                 }
1003
1004                 /* Set CBPENDING so nothing in the cancellation path
1005                  * can match this lock. */
1006                 ldlm_set_cbpending(lock);
1007                 ldlm_set_failed(lock);
1008                 lock->l_flags |= flags;
1009
1010                 /* ... without sending a CANCEL message for local_only. */
1011                 if (local_only)
1012                         ldlm_set_local_only(lock);
1013
1014                 if (local_only && (lock->l_readers || lock->l_writers)) {
1015                         /*
1016                          * This is a little bit gross, but much better than the
1017                          * alternative: pretend that we got a blocking AST from
1018                          * the server, so that when the lock is decref'd, it
1019                          * will go away ...
1020                          */
1021                         unlock_res(res);
1022                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1023                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1024                                 schedule_timeout_uninterruptible(
1025                                         cfs_time_seconds(4));
1026
1027                         if (lock->l_completion_ast)
1028                                 lock->l_completion_ast(lock,
1029                                                        LDLM_FL_FAILED, NULL);
1030                         LDLM_LOCK_RELEASE(lock);
1031                         continue;
1032                 }
1033
1034                 if (client) {
1035                         struct lustre_handle lockh;
1036
1037                         unlock_res(res);
1038                         ldlm_lock2handle(lock, &lockh);
1039                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1040                         if (rc)
1041                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1042                 } else {
1043                         unlock_res(res);
1044                         LDLM_DEBUG(lock,
1045                                    "Freeing a lock still held by a client node");
1046                         ldlm_lock_cancel(lock);
1047                 }
1048                 LDLM_LOCK_RELEASE(lock);
1049         } while (1);
1050 }
1051
1052 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1053                                struct hlist_node *hnode, void *arg)
1054 {
1055         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1056         __u64 flags = *(__u64 *)arg;
1057
1058         cleanup_resource(res, &res->lr_granted, flags);
1059         cleanup_resource(res, &res->lr_waiting, flags);
1060
1061         return 0;
1062 }
1063
1064 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1065                                   struct hlist_node *hnode, void *arg)
1066 {
1067         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1068
1069         lock_res(res);
1070         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1071                "(%d) after lock cleanup; forcing cleanup.\n",
1072                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1073                atomic_read(&res->lr_refcount) - 1);
1074
1075         /* Use D_NETERROR since it is in the default mask */
1076         ldlm_resource_dump(D_NETERROR, res);
1077         unlock_res(res);
1078         return 0;
1079 }
1080
1081 /**
1082  * Cancel and destroy all locks in the namespace.
1083  *
1084  * Typically used during evictions when server notified client that it was
1085  * evicted and all of its state needs to be destroyed.
1086  * Also used during shutdown.
1087  */
1088 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1089 {
1090         if (ns == NULL) {
1091                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1092                 return ELDLM_OK;
1093         }
1094
1095         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1096                                  &flags, 0);
1097         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1098                                  NULL, 0);
1099         return ELDLM_OK;
1100 }
1101 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1102
1103 /**
1104  * Attempts to free namespace.
1105  *
1106  * Only used when namespace goes away, like during an unmount.
1107  */
1108 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1109 {
1110         ENTRY;
1111
1112         /* At shutdown time, don't call the cancellation callback */
1113         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1114
1115         if (atomic_read(&ns->ns_bref) > 0) {
1116                 int rc;
1117                 CDEBUG(D_DLMTRACE,
1118                        "dlm namespace %s free waiting on refcount %d\n",
1119                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1120 force_wait:
1121                 if (force)
1122                         rc = wait_event_idle_timeout(
1123                                 ns->ns_waitq,
1124                                 atomic_read(&ns->ns_bref) == 0,
1125                                 cfs_time_seconds(1) / 4);
1126                 else
1127                         rc = l_wait_event_abortable(
1128                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1129
1130                 /* Forced cleanups should be able to reclaim all references,
1131                  * so it's safe to wait forever... we can't leak locks... */
1132                 if (force && rc == 0) {
1133                         rc = -ETIMEDOUT;
1134                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1135                                        "namespace with %d resources in use, "
1136                                        "(rc=%d)\n", ldlm_ns_name(ns),
1137                                        atomic_read(&ns->ns_bref), rc);
1138                         GOTO(force_wait, rc);
1139                 }
1140
1141                 if (atomic_read(&ns->ns_bref)) {
1142                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1143                                        "with %d resources in use, (rc=%d)\n",
1144                                        ldlm_ns_name(ns),
1145                                        atomic_read(&ns->ns_bref), rc);
1146                         RETURN(ELDLM_NAMESPACE_EXISTS);
1147                 }
1148                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1149                        ldlm_ns_name(ns));
1150         }
1151
1152         RETURN(ELDLM_OK);
1153 }
1154
1155 /**
1156  * Performs various cleanups for passed \a ns to make it drop refc and be
1157  * ready for freeing. Waits for refc == 0.
1158  *
1159  * The following is done:
1160  * (0) Unregister \a ns from its list to make inaccessible for potential
1161  * users like pools thread and others;
1162  * (1) Clear all locks in \a ns.
1163  */
1164 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1165                                struct obd_import *imp,
1166                                int force)
1167 {
1168         int rc;
1169
1170         ENTRY;
1171         if (!ns) {
1172                 EXIT;
1173                 return;
1174         }
1175
1176         spin_lock(&ns->ns_lock);
1177         ns->ns_stopping = 1;
1178         spin_unlock(&ns->ns_lock);
1179
1180         /*
1181          * Can fail with -EINTR when force == 0 in which case try harder.
1182          */
1183         rc = __ldlm_namespace_free(ns, force);
1184         if (rc != ELDLM_OK) {
1185                 if (imp) {
1186                         ptlrpc_disconnect_import(imp, 0);
1187                         ptlrpc_invalidate_import(imp);
1188                 }
1189
1190                 /*
1191                  * With all requests dropped and the import inactive
1192                  * we are gaurenteed all reference will be dropped.
1193                  */
1194                 rc = __ldlm_namespace_free(ns, 1);
1195                 LASSERT(rc == 0);
1196         }
1197         EXIT;
1198 }
1199 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1200
1201 /**
1202  * Performs freeing memory structures related to \a ns. This is only done
1203  * when ldlm_namespce_free_prior() successfully removed all resources
1204  * referencing \a ns and its refc == 0.
1205  */
1206 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1207 {
1208         ENTRY;
1209         if (!ns) {
1210                 EXIT;
1211                 return;
1212         }
1213
1214         /* Make sure that nobody can find this ns in its list. */
1215         ldlm_namespace_unregister(ns, ns->ns_client);
1216         /* Fini pool _before_ parent proc dir is removed. This is important as
1217          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1218          * Removing it after @dir may cause oops. */
1219         ldlm_pool_fini(&ns->ns_pool);
1220
1221         ldlm_namespace_debugfs_unregister(ns);
1222         ldlm_namespace_sysfs_unregister(ns);
1223         cfs_hash_putref(ns->ns_rs_hash);
1224         OBD_FREE_LARGE(ns->ns_rs_buckets,
1225                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1226         kfree(ns->ns_name);
1227         /* Namespace \a ns should be not on list at this time, otherwise
1228          * this will cause issues related to using freed \a ns in poold
1229          * thread.
1230          */
1231         LASSERT(list_empty(&ns->ns_list_chain));
1232         OBD_FREE_PTR(ns);
1233         ldlm_put_ref();
1234         EXIT;
1235 }
1236 EXPORT_SYMBOL(ldlm_namespace_free_post);
1237
1238 /**
1239  * Cleanup the resource, and free namespace.
1240  * bug 12864:
1241  * Deadlock issue:
1242  * proc1: destroy import
1243  *        class_disconnect_export(grab cl_sem) ->
1244  *              -> ldlm_namespace_free ->
1245  *              -> lprocfs_remove(grab _lprocfs_lock).
1246  * proc2: read proc info
1247  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1248  *              -> osc_rd_active, etc(grab cl_sem).
1249  *
1250  * So that I have to split the ldlm_namespace_free into two parts - the first
1251  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1252  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1253  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1254  * held.
1255  */
1256 void ldlm_namespace_free(struct ldlm_namespace *ns,
1257                          struct obd_import *imp,
1258                          int force)
1259 {
1260         ldlm_namespace_free_prior(ns, imp, force);
1261         ldlm_namespace_free_post(ns);
1262 }
1263 EXPORT_SYMBOL(ldlm_namespace_free);
1264
1265 void ldlm_namespace_get(struct ldlm_namespace *ns)
1266 {
1267         atomic_inc(&ns->ns_bref);
1268 }
1269
1270 /* This is only for callers that care about refcount */
1271 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1272 {
1273         return atomic_inc_return(&ns->ns_bref);
1274 }
1275
1276 void ldlm_namespace_put(struct ldlm_namespace *ns)
1277 {
1278         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1279                 wake_up(&ns->ns_waitq);
1280                 spin_unlock(&ns->ns_lock);
1281         }
1282 }
1283
1284 /** Register \a ns in the list of namespaces */
1285 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1286 {
1287         mutex_lock(ldlm_namespace_lock(client));
1288         LASSERT(list_empty(&ns->ns_list_chain));
1289         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1290         ldlm_namespace_nr_inc(client);
1291         mutex_unlock(ldlm_namespace_lock(client));
1292 }
1293
1294 /** Unregister \a ns from the list of namespaces. */
1295 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1296 {
1297         mutex_lock(ldlm_namespace_lock(client));
1298         LASSERT(!list_empty(&ns->ns_list_chain));
1299         /* Some asserts and possibly other parts of the code are still
1300          * using list_empty(&ns->ns_list_chain). This is why it is
1301          * important to use list_del_init() here. */
1302         list_del_init(&ns->ns_list_chain);
1303         ldlm_namespace_nr_dec(client);
1304         mutex_unlock(ldlm_namespace_lock(client));
1305 }
1306
1307 /** Should be called with ldlm_namespace_lock(client) taken. */
1308 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1309                                           enum ldlm_side client)
1310 {
1311         LASSERT(!list_empty(&ns->ns_list_chain));
1312         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1313         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1314 }
1315
1316 /** Should be called with ldlm_namespace_lock(client) taken. */
1317 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1318                                             enum ldlm_side client)
1319 {
1320         LASSERT(!list_empty(&ns->ns_list_chain));
1321         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1322         list_move_tail(&ns->ns_list_chain,
1323                        ldlm_namespace_inactive_list(client));
1324 }
1325
1326 /** Should be called with ldlm_namespace_lock(client) taken. */
1327 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1328 {
1329         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1330         LASSERT(!list_empty(ldlm_namespace_list(client)));
1331         return container_of(ldlm_namespace_list(client)->next,
1332                             struct ldlm_namespace, ns_list_chain);
1333 }
1334
1335 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1336 {
1337         int idx;
1338
1339         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1340                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1341         if (res->lr_itree == NULL)
1342                 return false;
1343         /* Initialize interval trees for each lock mode. */
1344         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1345                 res->lr_itree[idx].lit_size = 0;
1346                 res->lr_itree[idx].lit_mode = 1 << idx;
1347                 res->lr_itree[idx].lit_root = NULL;
1348         }
1349         return true;
1350 }
1351
1352 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1353 {
1354         int i;
1355
1356         OBD_ALLOC_PTR(res->lr_ibits_queues);
1357         if (res->lr_ibits_queues == NULL)
1358                 return false;
1359         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1360                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1361         return true;
1362 }
1363
1364 /** Create and initialize new resource. */
1365 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1366 {
1367         struct ldlm_resource *res;
1368         bool rc;
1369
1370         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1371         if (res == NULL)
1372                 return NULL;
1373
1374         switch (ldlm_type) {
1375         case LDLM_EXTENT:
1376                 rc = ldlm_resource_extent_new(res);
1377                 break;
1378         case LDLM_IBITS:
1379                 rc = ldlm_resource_inodebits_new(res);
1380                 break;
1381         default:
1382                 rc = true;
1383                 break;
1384         }
1385         if (!rc) {
1386                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1387                 return NULL;
1388         }
1389
1390         INIT_LIST_HEAD(&res->lr_granted);
1391         INIT_LIST_HEAD(&res->lr_waiting);
1392
1393         atomic_set(&res->lr_refcount, 1);
1394         spin_lock_init(&res->lr_lock);
1395         lu_ref_init(&res->lr_reference);
1396
1397         /* Since LVB init can be delayed now, there is no longer need to
1398          * immediatelly acquire mutex here. */
1399         mutex_init(&res->lr_lvb_mutex);
1400         res->lr_lvb_initialized = false;
1401
1402         return res;
1403 }
1404
1405 static void ldlm_resource_free(struct ldlm_resource *res)
1406 {
1407         if (res->lr_type == LDLM_EXTENT) {
1408                 if (res->lr_itree != NULL)
1409                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1410                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1411         } else if (res->lr_type == LDLM_IBITS) {
1412                 if (res->lr_ibits_queues != NULL)
1413                         OBD_FREE_PTR(res->lr_ibits_queues);
1414         }
1415
1416         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1417 }
1418
1419 /**
1420  * Return a reference to resource with given name, creating it if necessary.
1421  * Args: namespace with ns_lock unlocked
1422  * Locks: takes and releases NS hash-lock and res->lr_lock
1423  * Returns: referenced, unlocked ldlm_resource or NULL
1424  */
1425 struct ldlm_resource *
1426 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1427                   const struct ldlm_res_id *name, enum ldlm_type type,
1428                   int create)
1429 {
1430         struct hlist_node       *hnode;
1431         struct ldlm_resource    *res = NULL;
1432         struct cfs_hash_bd              bd;
1433         __u64                   version;
1434         int                     ns_refcount = 0;
1435         int hash;
1436
1437         LASSERT(ns != NULL);
1438         LASSERT(parent == NULL);
1439         LASSERT(ns->ns_rs_hash != NULL);
1440         LASSERT(name->name[0] != 0);
1441
1442         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1443         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1444         if (hnode != NULL) {
1445                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1446                 GOTO(found, res);
1447         }
1448
1449         version = cfs_hash_bd_version_get(&bd);
1450         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1451
1452         if (create == 0)
1453                 return ERR_PTR(-ENOENT);
1454
1455         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1456                  "type: %d\n", type);
1457         res = ldlm_resource_new(type);
1458         if (res == NULL)
1459                 return ERR_PTR(-ENOMEM);
1460
1461         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1462         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1463         res->lr_name = *name;
1464         res->lr_type = type;
1465
1466         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1467         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1468                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1469
1470         if (hnode != NULL) {
1471                 /* Someone won the race and already added the resource. */
1472                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1473                 /* Clean lu_ref for failed resource. */
1474                 lu_ref_fini(&res->lr_reference);
1475                 ldlm_resource_free(res);
1476 found:
1477                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1478                 return res;
1479         }
1480         /* We won! Let's add the resource. */
1481         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1482         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1483                 ns_refcount = ldlm_namespace_get_return(ns);
1484
1485         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1486
1487         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1488
1489         /* Let's see if we happened to be the very first resource in this
1490          * namespace. If so, and this is a client namespace, we need to move
1491          * the namespace into the active namespaces list to be patrolled by
1492          * the ldlm_poold. */
1493         if (ns_is_client(ns) && ns_refcount == 1) {
1494                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1495                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1496                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1497         }
1498
1499         return res;
1500 }
1501 EXPORT_SYMBOL(ldlm_resource_get);
1502
1503 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1504 {
1505         LASSERT(res != NULL);
1506         LASSERT(res != LP_POISON);
1507         atomic_inc(&res->lr_refcount);
1508         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1509                atomic_read(&res->lr_refcount));
1510         return res;
1511 }
1512
1513 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1514                                          struct ldlm_resource *res)
1515 {
1516         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1517
1518         if (!list_empty(&res->lr_granted)) {
1519                 ldlm_resource_dump(D_ERROR, res);
1520                 LBUG();
1521         }
1522
1523         if (!list_empty(&res->lr_waiting)) {
1524                 ldlm_resource_dump(D_ERROR, res);
1525                 LBUG();
1526         }
1527
1528         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1529                                bd, &res->lr_hash);
1530         lu_ref_fini(&res->lr_reference);
1531         if (atomic_dec_and_test(&nsb->nsb_count))
1532                 ldlm_namespace_put(nsb->nsb_namespace);
1533 }
1534
1535 /* Returns 1 if the resource was freed, 0 if it remains. */
1536 int ldlm_resource_putref(struct ldlm_resource *res)
1537 {
1538         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1539         struct cfs_hash_bd   bd;
1540
1541         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1542         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1543                res, atomic_read(&res->lr_refcount) - 1);
1544
1545         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1546         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1547                 __ldlm_resource_putref_final(&bd, res);
1548                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1549                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1550                         ns->ns_lvbo->lvbo_free(res);
1551                 ldlm_resource_free(res);
1552                 return 1;
1553         }
1554         return 0;
1555 }
1556 EXPORT_SYMBOL(ldlm_resource_putref);
1557
1558 /**
1559  * Add a lock into a given resource into specified lock list.
1560  */
1561 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1562                             struct ldlm_lock *lock)
1563 {
1564         check_res_locked(res);
1565
1566         LDLM_DEBUG(lock, "About to add this lock");
1567
1568         if (ldlm_is_destroyed(lock)) {
1569                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1570                 return;
1571         }
1572
1573         LASSERT(list_empty(&lock->l_res_link));
1574
1575         list_add_tail(&lock->l_res_link, head);
1576
1577         if (res->lr_type == LDLM_IBITS)
1578                 ldlm_inodebits_add_lock(res, head, lock);
1579 }
1580
1581 /**
1582  * Insert a lock into resource after specified lock.
1583  *
1584  * Obtain resource description from the lock we are inserting after.
1585  */
1586 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1587                                      struct ldlm_lock *new)
1588 {
1589         struct ldlm_resource *res = original->l_resource;
1590
1591         check_res_locked(res);
1592
1593         ldlm_resource_dump(D_INFO, res);
1594         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1595
1596         if (ldlm_is_destroyed(new)) {
1597                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1598                 goto out;
1599         }
1600
1601         LASSERT(list_empty(&new->l_res_link));
1602
1603         list_add(&new->l_res_link, &original->l_res_link);
1604  out:;
1605 }
1606
1607 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1608 {
1609         int type = lock->l_resource->lr_type;
1610
1611         check_res_locked(lock->l_resource);
1612         switch (type) {
1613         case LDLM_PLAIN:
1614                 ldlm_unlink_lock_skiplist(lock);
1615                 break;
1616         case LDLM_EXTENT:
1617                 ldlm_extent_unlink_lock(lock);
1618                 break;
1619         case LDLM_IBITS:
1620                 ldlm_inodebits_unlink_lock(lock);
1621                 break;
1622         }
1623         list_del_init(&lock->l_res_link);
1624 }
1625 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1626
1627 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1628 {
1629         desc->lr_type = res->lr_type;
1630         desc->lr_name = res->lr_name;
1631 }
1632
1633 /**
1634  * Print information about all locks in all namespaces on this node to debug
1635  * log.
1636  */
1637 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1638 {
1639         struct list_head *tmp;
1640
1641         if (!((libcfs_debug | D_ERROR) & level))
1642                 return;
1643
1644         mutex_lock(ldlm_namespace_lock(client));
1645
1646         list_for_each(tmp, ldlm_namespace_list(client)) {
1647                 struct ldlm_namespace *ns;
1648
1649                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1650                 ldlm_namespace_dump(level, ns);
1651         }
1652
1653         mutex_unlock(ldlm_namespace_lock(client));
1654 }
1655
1656 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1657                               struct hlist_node *hnode, void *arg)
1658 {
1659         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1660         int    level = (int)(unsigned long)arg;
1661
1662         lock_res(res);
1663         ldlm_resource_dump(level, res);
1664         unlock_res(res);
1665
1666         return 0;
1667 }
1668
1669 /**
1670  * Print information about all locks in this namespace on this node to debug
1671  * log.
1672  */
1673 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1674 {
1675         if (!((libcfs_debug | D_ERROR) & level))
1676                 return;
1677
1678         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1679                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1680                ns_is_client(ns) ? "client" : "server");
1681
1682         if (ktime_get_seconds() < ns->ns_next_dump)
1683                 return;
1684
1685         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1686                                  ldlm_res_hash_dump,
1687                                  (void *)(unsigned long)level, 0);
1688         spin_lock(&ns->ns_lock);
1689         ns->ns_next_dump = ktime_get_seconds() + 10;
1690         spin_unlock(&ns->ns_lock);
1691 }
1692
1693 /**
1694  * Print information about all locks in this resource to debug log.
1695  */
1696 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1697 {
1698         struct ldlm_lock *lock;
1699         unsigned int granted = 0;
1700
1701         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1702
1703         if (!((libcfs_debug | D_ERROR) & level))
1704                 return;
1705
1706         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1707                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1708
1709         if (!list_empty(&res->lr_granted)) {
1710                 CDEBUG(level, "Granted locks (in reverse order):\n");
1711                 list_for_each_entry_reverse(lock, &res->lr_granted,
1712                                                 l_res_link) {
1713                         LDLM_DEBUG_LIMIT(level, lock, "###");
1714                         if (!(level & D_CANTMASK) &&
1715                             ++granted > ldlm_dump_granted_max) {
1716                                 CDEBUG(level,
1717                                        "only dump %d granted locks to avoid DDOS.\n",
1718                                        granted);
1719                                 break;
1720                         }
1721                 }
1722         }
1723
1724         if (!list_empty(&res->lr_waiting)) {
1725                 CDEBUG(level, "Waiting locks:\n");
1726                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1727                         LDLM_DEBUG_LIMIT(level, lock, "###");
1728         }
1729 }
1730 EXPORT_SYMBOL(ldlm_resource_dump);