lustre/ldlm/ldlm_resource.c

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