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         u64 value;
 115         __u64 watermark;
 116         __u64 *data = ((struct seq_file *)file->private_data)->private;
 117         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 118         char kernbuf[22] = "";
 119         int rc;
 120
 121         if (count >= sizeof(kernbuf))
 122                 return -EINVAL;
 123
 124         if (copy_from_user(kernbuf, buffer, count))
 125                 return -EFAULT;
 126         kernbuf[count] = 0;
 127
 128         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
 129         if (rc < 0) {
 130                 CERROR("Failed to set %s, rc = %d.\n",
 131                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 132                        rc);
 133                 return rc;
 134         } else if (value != 0 && value < (1 << 20)) {
 135                 CERROR("%s should be greater than 1MB.\n",
 136                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 137                 return -EINVAL;
 138         }
 139         watermark = value >> 20;
 140
 141         if (wm_low) {
 142                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 143                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 144                                "lock_limit_mb.\n");
 145                         return -EINVAL;
 146                 }
 147
 148                 *data = watermark;
 149                 if (watermark != 0) {
 150                         watermark <<= 20;
 151                         do_div(watermark, sizeof(struct ldlm_lock));
 152                 }
 153                 ldlm_reclaim_threshold = watermark;
 154         } else {
 155                 if (ldlm_reclaim_threshold_mb != 0 &&
 156                     watermark < ldlm_reclaim_threshold_mb) {
 157                         CERROR("lock_limit_mb must be greater than "
 158                                "lock_reclaim_threshold_mb.\n");
 159                         return -EINVAL;
 160                 }
 161
 162                 *data = watermark;
 163                 if (watermark != 0) {
 164                         watermark <<= 20;
 165                         do_div(watermark, sizeof(struct ldlm_lock));
 166                 }
 167                 ldlm_lock_limit = watermark;
 168         }
 169
 170         return count;
 171 }
 172
 173 static int seq_watermark_open(struct inode *inode, struct file *file)
 174 {
 175         return single_open(file, seq_watermark_show, inode->i_private);
 176 }
 177
 178 static const struct file_operations ldlm_watermark_fops = {
 179         .owner          = THIS_MODULE,
 180         .open           = seq_watermark_open,
 181         .read           = seq_read,
 182         .write          = seq_watermark_write,
 183         .llseek         = seq_lseek,
 184         .release        = lprocfs_single_release,
 185 };
 186
 187 static int seq_granted_show(struct seq_file *m, void *data)
 188 {
 189         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 190                    (struct percpu_counter *)m->private));
 191         return 0;
 192 }
 193
 194 static int seq_granted_open(struct inode *inode, struct file *file)
 195 {
 196         return single_open(file, seq_granted_show, inode->i_private);
 197 }
 198
 199 static const struct file_operations ldlm_granted_fops = {
 200         .owner  = THIS_MODULE,
 201         .open   = seq_granted_open,
 202         .read   = seq_read,
 203         .llseek = seq_lseek,
 204         .release = seq_release,
 205 };
 206
 207 #endif /* HAVE_SERVER_SUPPORT */
 208
 209 static struct lprocfs_vars ldlm_debugfs_list[] = {
 210         { .name =       "dump_namespaces",
 211           .fops =       &ldlm_dump_ns_fops,
 212           .proc_mode =  0222 },
 213         { .name =       "dump_granted_max",
 214           .fops =       &ldlm_rw_uint_fops,
 215           .data =       &ldlm_dump_granted_max },
 216 #ifdef HAVE_SERVER_SUPPORT
 217         { .name =       "lock_reclaim_threshold_mb",
 218           .fops =       &ldlm_watermark_fops,
 219           .data =       &ldlm_reclaim_threshold_mb },
 220         { .name =       "lock_limit_mb",
 221           .fops =       &ldlm_watermark_fops,
 222           .data =       &ldlm_lock_limit_mb },
 223         { .name =       "lock_granted_count",
 224           .fops =       &ldlm_granted_fops,
 225           .data =       &ldlm_granted_total },
 226 #endif
 227         { NULL }
 228 };
 229
 230 int ldlm_debugfs_setup(void)
 231 {
 232         int rc;
 233
 234         ENTRY;
 235         ldlm_debugfs_dir = ldebugfs_register(OBD_LDLM_DEVICENAME,
 236                                              debugfs_lustre_root,
 237                                              NULL, NULL);
 238         if (IS_ERR_OR_NULL(ldlm_debugfs_dir)) {
 239                 CERROR("LDebugFS failed in ldlm-init\n");
 240                 rc = ldlm_debugfs_dir ? PTR_ERR(ldlm_debugfs_dir) : -ENOMEM;
 241                 ldlm_debugfs_dir = NULL;
 242                 GOTO(err, rc);
 243         }
 244
 245         ldlm_ns_debugfs_dir = ldebugfs_register("namespaces",
 246                                                 ldlm_debugfs_dir,
 247                                                 NULL, NULL);
 248         if (IS_ERR_OR_NULL(ldlm_ns_debugfs_dir)) {
 249                 CERROR("LProcFS failed in ldlm-init\n");
 250                 rc = ldlm_ns_debugfs_dir ? PTR_ERR(ldlm_ns_debugfs_dir)
 251                                          : -ENOMEM;
 252                 GOTO(err, rc);
 253         }
 254
 255         ldlm_svc_debugfs_dir = ldebugfs_register("services",
 256                                                  ldlm_debugfs_dir,
 257                                                  NULL, NULL);
 258         if (IS_ERR_OR_NULL(ldlm_svc_debugfs_dir)) {
 259                 CERROR("LProcFS failed in ldlm-init\n");
 260                 rc = ldlm_svc_debugfs_dir ? PTR_ERR(ldlm_svc_debugfs_dir)
 261                                           : -ENOMEM;
 262                 GOTO(err, rc);
 263         }
 264
 265         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 266         if (rc != 0) {
 267                 CERROR("LProcFS failed in ldlm-init\n");
 268                 GOTO(err, rc);
 269         }
 270
 271         RETURN(0);
 272
 273 err:
 274         debugfs_remove_recursive(ldlm_debugfs_dir);
 275         ldlm_svc_debugfs_dir = NULL;
 276         ldlm_ns_debugfs_dir = NULL;
 277         ldlm_debugfs_dir = NULL;
 278         RETURN(rc);
 279 }
 280
 281 void ldlm_debugfs_cleanup(void)
 282 {
 283         debugfs_remove_recursive(ldlm_debugfs_dir);
 284
 285         ldlm_svc_debugfs_dir = NULL;
 286         ldlm_ns_debugfs_dir = NULL;
 287         ldlm_debugfs_dir = NULL;
 288 }
 289
 290 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 291                                    char *buf)
 292 {
 293         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 294                                                  ns_kobj);
 295         __u64                   res = 0;
 296         int                     i;
 297
 298         /* result is not strictly consistant */
 299         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
 300                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
 301         return sprintf(buf, "%lld\n", res);
 302 }
 303 LUSTRE_RO_ATTR(resource_count);
 304
 305 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 306                                char *buf)
 307 {
 308         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 309                                                  ns_kobj);
 310         __u64                   locks;
 311
 312         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 313                                         LPROCFS_FIELDS_FLAGS_SUM);
 314         return sprintf(buf, "%lld\n", locks);
 315 }
 316 LUSTRE_RO_ATTR(lock_count);
 317
 318 static ssize_t lock_unused_count_show(struct kobject *kobj,
 319                                       struct attribute *attr,
 320                                       char *buf)
 321 {
 322         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 323                                                  ns_kobj);
 324
 325         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 326 }
 327 LUSTRE_RO_ATTR(lock_unused_count);
 328
 329 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 330                              char *buf)
 331 {
 332         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 333                                                  ns_kobj);
 334         __u32 *nr = &ns->ns_max_unused;
 335
 336         if (ns_connect_lru_resize(ns))
 337                 nr = &ns->ns_nr_unused;
 338         return sprintf(buf, "%u\n", *nr);
 339 }
 340
 341 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 342                               const char *buffer, size_t count)
 343 {
 344         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 345                                                  ns_kobj);
 346         unsigned long tmp;
 347         int lru_resize;
 348         int err;
 349
 350         if (strncmp(buffer, "clear", 5) == 0) {
 351                 CDEBUG(D_DLMTRACE,
 352                        "dropping all unused locks from namespace %s\n",
 353                        ldlm_ns_name(ns));
 354                 if (ns_connect_lru_resize(ns)) {
 355                         /* Try to cancel all @ns_nr_unused locks. */
 356                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 357                                         LDLM_LRU_FLAG_PASSED |
 358                                         LDLM_LRU_FLAG_CLEANUP);
 359                 } else {
 360                         tmp = ns->ns_max_unused;
 361                         ns->ns_max_unused = 0;
 362                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 363                                         LDLM_LRU_FLAG_CLEANUP);
 364                         ns->ns_max_unused = tmp;
 365                 }
 366                 return count;
 367         }
 368
 369         err = kstrtoul(buffer, 10, &tmp);
 370         if (err != 0) {
 371                 CERROR("lru_size: invalid value written\n");
 372                 return -EINVAL;
 373         }
 374         lru_resize = (tmp == 0);
 375
 376         if (ns_connect_lru_resize(ns)) {
 377                 if (!lru_resize)
 378                         ns->ns_max_unused = (unsigned int)tmp;
 379
 380                 if (tmp > ns->ns_nr_unused)
 381                         tmp = ns->ns_nr_unused;
 382                 tmp = ns->ns_nr_unused - tmp;
 383
 384                 CDEBUG(D_DLMTRACE,
 385                        "changing namespace %s unused locks from %u to %u\n",
 386                        ldlm_ns_name(ns), ns->ns_nr_unused,
 387                        (unsigned int)tmp);
 388                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 389
 390                 if (!lru_resize) {
 391                         CDEBUG(D_DLMTRACE,
 392                                "disable lru_resize for namespace %s\n",
 393                                ldlm_ns_name(ns));
 394                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 395                 }
 396         } else {
 397                 CDEBUG(D_DLMTRACE,
 398                        "changing namespace %s max_unused from %u to %u\n",
 399                        ldlm_ns_name(ns), ns->ns_max_unused,
 400                        (unsigned int)tmp);
 401                 ns->ns_max_unused = (unsigned int)tmp;
 402                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 403
 404                 /* Make sure that LRU resize was originally supported before
 405                  * turning it on here.
 406                  */
 407                 if (lru_resize &&
 408                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 409                         CDEBUG(D_DLMTRACE,
 410                                "enable lru_resize for namespace %s\n",
 411                                ldlm_ns_name(ns));
 412                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 413                 }
 414         }
 415
 416         return count;
 417 }
 418 LUSTRE_RW_ATTR(lru_size);
 419
 420 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 421                                 char *buf)
 422 {
 423         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 424                                                  ns_kobj);
 425
 426         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 427 }
 428
 429 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 430                                  const char *buffer, size_t count)
 431 {
 432         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 433                                                  ns_kobj);
 434         int scale = NSEC_PER_MSEC;
 435         unsigned long long tmp;
 436         char *buf;
 437
 438         /* Did the user ask in seconds or milliseconds. Default is in ms */
 439         buf = strstr(buffer, "ms");
 440         if (!buf) {
 441                 buf = strchr(buffer, 's');
 442                 if (buf)
 443                         scale = NSEC_PER_SEC;
 444         }
 445
 446         if (buf)
 447                 *buf = '\0';
 448
 449         if (kstrtoull(buffer, 10, &tmp))
 450                 return -EINVAL;
 451
 452         ns->ns_max_age = ktime_set(0, tmp * scale);
 453
 454         return count;
 455 }
 456 LUSTRE_RW_ATTR(lru_max_age);
 457
 458 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 459                                       struct attribute *attr,
 460                                       char *buf)
 461 {
 462         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 463                                                  ns_kobj);
 464
 465         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 466 }
 467
 468 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 469                                        struct attribute *attr,
 470                                        const char *buffer,
 471                                        size_t count)
 472 {
 473         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 474                                                  ns_kobj);
 475         unsigned long supp = -1;
 476         int rc;
 477
 478         rc = kstrtoul(buffer, 10, &supp);
 479         if (rc < 0)
 480                 return rc;
 481
 482         if (supp == 0)
 483                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 484         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 485                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 486         return count;
 487 }
 488 LUSTRE_RW_ATTR(early_lock_cancel);
 489
 490 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 491                                     struct attribute *attr, char *buf)
 492 {
 493         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 494                                                  ns_kobj);
 495
 496         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 497 }
 498
 499 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 500                                      struct attribute *attr,
 501                                      const char *buffer, size_t count)
 502 {
 503         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 504                                                  ns_kobj);
 505         unsigned long long tmp;
 506
 507         if (kstrtoull(buffer, 10, &tmp))
 508                 return -EINVAL;
 509
 510         ns->ns_dirty_age_limit = tmp;
 511
 512         return count;
 513 }
 514 LUSTRE_RW_ATTR(dirty_age_limit);
 515
 516 #ifdef HAVE_SERVER_SUPPORT
 517 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 518                                     struct attribute *attr, char *buf)
 519 {
 520         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 521                                                  ns_kobj);
 522
 523         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 524 }
 525
 526 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 527                                      struct attribute *attr,
 528                                      const char *buffer, size_t count)
 529 {
 530         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 531                                                  ns_kobj);
 532         unsigned long long tmp;
 533
 534         if (kstrtoull(buffer, 10, &tmp))
 535                 return -EINVAL;
 536
 537         ns->ns_ctime_age_limit = tmp;
 538
 539         return count;
 540 }
 541 LUSTRE_RW_ATTR(ctime_age_limit);
 542
 543 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 544                                   char *buf)
 545 {
 546         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 547                                                  ns_kobj);
 548
 549         return sprintf(buf, "%d\n", ns->ns_timeouts);
 550 }
 551 LUSTRE_RO_ATTR(lock_timeouts);
 552
 553 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 554                                      struct attribute *attr, char *buf)
 555 {
 556         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 557                                                  ns_kobj);
 558
 559         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 560 }
 561
 562 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 563                                       struct attribute *attr,
 564                                       const char *buffer, size_t count)
 565 {
 566         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 567                                                  ns_kobj);
 568         unsigned long tmp;
 569         int err;
 570
 571         err = kstrtoul(buffer, 10, &tmp);
 572         if (err != 0)
 573                 return -EINVAL;
 574
 575         ns->ns_max_nolock_size = tmp;
 576
 577         return count;
 578 }
 579 LUSTRE_RW_ATTR(max_nolock_bytes);
 580
 581 static ssize_t contention_seconds_show(struct kobject *kobj,
 582                                        struct attribute *attr, char *buf)
 583 {
 584         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 585                                                  ns_kobj);
 586
 587         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 588 }
 589
 590 static ssize_t contention_seconds_store(struct kobject *kobj,
 591                                         struct attribute *attr,
 592                                         const char *buffer, size_t count)
 593 {
 594         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 595                                                  ns_kobj);
 596         unsigned long long tmp;
 597
 598         if (kstrtoull(buffer, 10, &tmp))
 599                 return -EINVAL;
 600
 601         ns->ns_contention_time = tmp;
 602
 603         return count;
 604 }
 605 LUSTRE_RW_ATTR(contention_seconds);
 606
 607 static ssize_t contended_locks_show(struct kobject *kobj,
 608                                     struct attribute *attr, char *buf)
 609 {
 610         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 611                                                  ns_kobj);
 612
 613         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 614 }
 615
 616 static ssize_t contended_locks_store(struct kobject *kobj,
 617                                      struct attribute *attr,
 618                                      const char *buffer, size_t count)
 619 {
 620         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 621                                                  ns_kobj);
 622         unsigned long tmp;
 623         int err;
 624
 625         err = kstrtoul(buffer, 10, &tmp);
 626         if (err != 0)
 627                 return -EINVAL;
 628
 629         ns->ns_contended_locks = tmp;
 630
 631         return count;
 632 }
 633 LUSTRE_RW_ATTR(contended_locks);
 634
 635 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 636                                      struct attribute *attr, char *buf)
 637 {
 638         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 639                                                  ns_kobj);
 640
 641         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 642 }
 643
 644 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 645                                       struct attribute *attr,
 646                                       const char *buffer, size_t count)
 647 {
 648         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 649                                                  ns_kobj);
 650         unsigned long tmp;
 651         int err;
 652
 653         err = kstrtoul(buffer, 10, &tmp);
 654         if (err != 0)
 655                 return -EINVAL;
 656
 657         ns->ns_max_parallel_ast = tmp;
 658
 659         return count;
 660 }
 661 LUSTRE_RW_ATTR(max_parallel_ast);
 662
 663 #endif /* HAVE_SERVER_SUPPORT */
 664
 665 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 666 static struct attribute *ldlm_ns_attrs[] = {
 667         &lustre_attr_resource_count.attr,
 668         &lustre_attr_lock_count.attr,
 669         &lustre_attr_lock_unused_count.attr,
 670         &lustre_attr_lru_size.attr,
 671         &lustre_attr_lru_max_age.attr,
 672         &lustre_attr_early_lock_cancel.attr,
 673         &lustre_attr_dirty_age_limit.attr,
 674 #ifdef HAVE_SERVER_SUPPORT
 675         &lustre_attr_ctime_age_limit.attr,
 676         &lustre_attr_lock_timeouts.attr,
 677         &lustre_attr_max_nolock_bytes.attr,
 678         &lustre_attr_contention_seconds.attr,
 679         &lustre_attr_contended_locks.attr,
 680         &lustre_attr_max_parallel_ast.attr,
 681 #endif
 682         NULL,
 683 };
 684
 685 static void ldlm_ns_release(struct kobject *kobj)
 686 {
 687         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 688                                                  ns_kobj);
 689         complete(&ns->ns_kobj_unregister);
 690 }
 691
 692 static struct kobj_type ldlm_ns_ktype = {
 693         .default_attrs  = ldlm_ns_attrs,
 694         .sysfs_ops      = &lustre_sysfs_ops,
 695         .release        = ldlm_ns_release,
 696 };
 697
 698 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 699 {
 700         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 701                 CERROR("dlm namespace %s has no procfs dir?\n",
 702                        ldlm_ns_name(ns));
 703         else
 704                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 705
 706         if (ns->ns_stats != NULL)
 707                 lprocfs_free_stats(&ns->ns_stats);
 708 }
 709
 710 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 711 {
 712         kobject_put(&ns->ns_kobj);
 713         wait_for_completion(&ns->ns_kobj_unregister);
 714 }
 715
 716 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 717 {
 718         int err;
 719
 720         ns->ns_kobj.kset = ldlm_ns_kset;
 721         init_completion(&ns->ns_kobj_unregister);
 722         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 723                                    "%s", ldlm_ns_name(ns));
 724
 725         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 726         if (!ns->ns_stats) {
 727                 kobject_put(&ns->ns_kobj);
 728                 return -ENOMEM;
 729         }
 730
 731         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 732                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 733
 734         return err;
 735 }
 736
 737 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 738 {
 739         struct dentry *ns_entry;
 740
 741         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 742                 ns_entry = ns->ns_debugfs_entry;
 743         } else {
 744                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 745                                               ldlm_ns_debugfs_dir);
 746                 if (!ns_entry)
 747                         return -ENOMEM;
 748                 ns->ns_debugfs_entry = ns_entry;
 749         }
 750
 751         return 0;
 752 }
 753 #undef MAX_STRING_SIZE
 754
 755 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 756                                   const void *key, unsigned int mask)
 757 {
 758         const struct ldlm_res_id *id = key;
 759         unsigned int val = 0;
 760         unsigned int i;
 761
 762         for (i = 0; i < RES_NAME_SIZE; i++)
 763                 val += id->name[i];
 764         return val & mask;
 765 }
 766
 767 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 768 {
 769         struct lu_fid       fid;
 770         __u32               hash;
 771         __u32               val;
 772
 773         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 774         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 775         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 776
 777         hash = fid_flatten32(&fid);
 778         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 779         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 780                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 781         } else {
 782                 val = fid_oid(&fid);
 783         }
 784         hash += (val >> 5) + (val << 11);
 785         return cfs_hash_32(hash, bits);
 786 }
 787
 788 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 789 {
 790         struct ldlm_resource   *res;
 791
 792         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 793         return &res->lr_name;
 794 }
 795
 796 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 797 {
 798         struct ldlm_resource   *res;
 799
 800         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 801         return ldlm_res_eq((const struct ldlm_res_id *)key,
 802                            (const struct ldlm_res_id *)&res->lr_name);
 803 }
 804
 805 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 806 {
 807         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 808 }
 809
 810 static void
 811 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 812 {
 813         struct ldlm_resource *res;
 814
 815         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 816         ldlm_resource_getref(res);
 817 }
 818
 819 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 820 {
 821         struct ldlm_resource *res;
 822
 823         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 824         ldlm_resource_putref(res);
 825 }
 826
 827 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 828         .hs_hash        = ldlm_res_hop_hash,
 829         .hs_key         = ldlm_res_hop_key,
 830         .hs_keycmp      = ldlm_res_hop_keycmp,
 831         .hs_keycpy      = NULL,
 832         .hs_object      = ldlm_res_hop_object,
 833         .hs_get         = ldlm_res_hop_get_locked,
 834         .hs_put         = ldlm_res_hop_put
 835 };
 836
 837 typedef struct ldlm_ns_hash_def {
 838         enum ldlm_ns_type       nsd_type;
 839         /** hash bucket bits */
 840         unsigned                nsd_bkt_bits;
 841         /** hash bits */
 842         unsigned                nsd_all_bits;
 843         /** hash operations */
 844         struct cfs_hash_ops *nsd_hops;
 845 } ldlm_ns_hash_def_t;
 846
 847 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
 848 {
 849         {
 850                 .nsd_type       = LDLM_NS_TYPE_MDC,
 851                 .nsd_bkt_bits   = 11,
 852                 .nsd_all_bits   = 16,
 853                 .nsd_hops       = &ldlm_ns_hash_ops,
 854         },
 855         {
 856                 .nsd_type       = LDLM_NS_TYPE_MDT,
 857                 .nsd_bkt_bits   = 14,
 858                 .nsd_all_bits   = 21,
 859                 .nsd_hops       = &ldlm_ns_hash_ops,
 860         },
 861         {
 862                 .nsd_type       = LDLM_NS_TYPE_OSC,
 863                 .nsd_bkt_bits   = 8,
 864                 .nsd_all_bits   = 12,
 865                 .nsd_hops       = &ldlm_ns_hash_ops,
 866         },
 867         {
 868                 .nsd_type       = LDLM_NS_TYPE_OST,
 869                 .nsd_bkt_bits   = 11,
 870                 .nsd_all_bits   = 17,
 871                 .nsd_hops       = &ldlm_ns_hash_ops,
 872         },
 873         {
 874                 .nsd_type       = LDLM_NS_TYPE_MGC,
 875                 .nsd_bkt_bits   = 3,
 876                 .nsd_all_bits   = 4,
 877                 .nsd_hops       = &ldlm_ns_hash_ops,
 878         },
 879         {
 880                 .nsd_type       = LDLM_NS_TYPE_MGT,
 881                 .nsd_bkt_bits   = 3,
 882                 .nsd_all_bits   = 4,
 883                 .nsd_hops       = &ldlm_ns_hash_ops,
 884         },
 885         {
 886                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 887         },
 888 };
 889
 890 /**
 891  * Create and initialize new empty namespace.
 892  */
 893 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 894                                           enum ldlm_side client,
 895                                           enum ldlm_appetite apt,
 896                                           enum ldlm_ns_type ns_type)
 897 {
 898         struct ldlm_namespace *ns = NULL;
 899         struct ldlm_ns_hash_def *nsd;
 900         int idx;
 901         int rc;
 902
 903         ENTRY;
 904         LASSERT(obd != NULL);
 905
 906         rc = ldlm_get_ref();
 907         if (rc) {
 908                 CERROR("ldlm_get_ref failed: %d\n", rc);
 909                 RETURN(NULL);
 910         }
 911
 912         for (idx = 0; ; idx++) {
 913                 nsd = &ldlm_ns_hash_defs[idx];
 914                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 915                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 916                         GOTO(out_ref, NULL);
 917                 }
 918
 919                 if (nsd->nsd_type == ns_type)
 920                         break;
 921         }
 922
 923         OBD_ALLOC_PTR(ns);
 924         if (!ns)
 925                 GOTO(out_ref, NULL);
 926
 927         ns->ns_rs_hash = cfs_hash_create(name,
 928                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 929                                          nsd->nsd_bkt_bits, 0,
 930                                          CFS_HASH_MIN_THETA,
 931                                          CFS_HASH_MAX_THETA,
 932                                          nsd->nsd_hops,
 933                                          CFS_HASH_DEPTH |
 934                                          CFS_HASH_BIGNAME |
 935                                          CFS_HASH_SPIN_BKTLOCK |
 936                                          CFS_HASH_NO_ITEMREF);
 937         if (ns->ns_rs_hash == NULL)
 938                 GOTO(out_ns, NULL);
 939
 940         ns->ns_bucket_bits = nsd->nsd_all_bits - nsd->nsd_bkt_bits;
 941         OBD_ALLOC_LARGE(ns->ns_rs_buckets,
 942                         BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 943         if (!ns->ns_rs_buckets)
 944                 goto out_hash;
 945
 946         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 947                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 948
 949                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 950                 nsb->nsb_namespace = ns;
 951                 nsb->nsb_reclaim_start = 0;
 952                 atomic_set(&nsb->nsb_count, 0);
 953         }
 954
 955         ns->ns_obd = obd;
 956         ns->ns_appetite = apt;
 957         ns->ns_client = client;
 958         ns->ns_name = kstrdup(name, GFP_KERNEL);
 959         if (!ns->ns_name)
 960                 goto out_hash;
 961
 962         INIT_LIST_HEAD(&ns->ns_list_chain);
 963         INIT_LIST_HEAD(&ns->ns_unused_list);
 964         spin_lock_init(&ns->ns_lock);
 965         atomic_set(&ns->ns_bref, 0);
 966         init_waitqueue_head(&ns->ns_waitq);
 967
 968         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 969         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 970         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 971
 972         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 973         ns->ns_nr_unused          = 0;
 974         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 975         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 976         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 977         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 978         ns->ns_timeouts           = 0;
 979         ns->ns_orig_connect_flags = 0;
 980         ns->ns_connect_flags      = 0;
 981         ns->ns_stopping           = 0;
 982         ns->ns_reclaim_start      = 0;
 983         ns->ns_last_pos           = &ns->ns_unused_list;
 984
 985         rc = ldlm_namespace_sysfs_register(ns);
 986         if (rc) {
 987                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 988                 GOTO(out_hash, rc);
 989         }
 990
 991         rc = ldlm_namespace_debugfs_register(ns);
 992         if (rc) {
 993                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 994                 GOTO(out_sysfs, rc);
 995         }
 996
 997         idx = ldlm_namespace_nr_read(client);
 998         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 999         if (rc) {
1000                 CERROR("Can't initialize lock pool, rc %d\n", rc);
1001                 GOTO(out_proc, rc);
1002         }
1003
1004         ldlm_namespace_register(ns, client);
1005         RETURN(ns);
1006 out_proc:
1007         ldlm_namespace_debugfs_unregister(ns);
1008 out_sysfs:
1009         ldlm_namespace_sysfs_unregister(ns);
1010         ldlm_namespace_cleanup(ns, 0);
1011 out_hash:
1012         OBD_FREE_LARGE(ns->ns_rs_buckets,
1013                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1014         kfree(ns->ns_name);
1015         cfs_hash_putref(ns->ns_rs_hash);
1016 out_ns:
1017         OBD_FREE_PTR(ns);
1018 out_ref:
1019         ldlm_put_ref();
1020         RETURN(NULL);
1021 }
1022 EXPORT_SYMBOL(ldlm_namespace_new);
1023
1024 /**
1025  * Cancel and destroy all locks on a resource.
1026  *
1027  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1028  * clean up.  This is currently only used for recovery, and we make
1029  * certain assumptions as a result--notably, that we shouldn't cancel
1030  * locks with refs.
1031  */
1032 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1033                              __u64 flags)
1034 {
1035         struct list_head *tmp;
1036         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1037         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1038
1039         do {
1040                 struct ldlm_lock *lock = NULL;
1041
1042                 /* First, we look for non-cleaned-yet lock
1043                  * all cleaned locks are marked by CLEANED flag. */
1044                 lock_res(res);
1045                 list_for_each(tmp, q) {
1046                         lock = list_entry(tmp, struct ldlm_lock,
1047                                           l_res_link);
1048                         if (ldlm_is_cleaned(lock)) {
1049                                 lock = NULL;
1050                                 continue;
1051                         }
1052                         LDLM_LOCK_GET(lock);
1053                         ldlm_set_cleaned(lock);
1054                         break;
1055                 }
1056
1057                 if (lock == NULL) {
1058                         unlock_res(res);
1059                         break;
1060                 }
1061
1062                 /* Set CBPENDING so nothing in the cancellation path
1063                  * can match this lock. */
1064                 ldlm_set_cbpending(lock);
1065                 ldlm_set_failed(lock);
1066                 lock->l_flags |= flags;
1067
1068                 /* ... without sending a CANCEL message for local_only. */
1069                 if (local_only)
1070                         ldlm_set_local_only(lock);
1071
1072                 if (local_only && (lock->l_readers || lock->l_writers)) {
1073                         /*
1074                          * This is a little bit gross, but much better than the
1075                          * alternative: pretend that we got a blocking AST from
1076                          * the server, so that when the lock is decref'd, it
1077                          * will go away ...
1078                          */
1079                         unlock_res(res);
1080                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1081                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1082                                 set_current_state(TASK_UNINTERRUPTIBLE);
1083                                 schedule_timeout(cfs_time_seconds(4));
1084                                 set_current_state(TASK_RUNNING);
1085                         }
1086                         if (lock->l_completion_ast)
1087                                 lock->l_completion_ast(lock,
1088                                                        LDLM_FL_FAILED, NULL);
1089                         LDLM_LOCK_RELEASE(lock);
1090                         continue;
1091                 }
1092
1093                 if (client) {
1094                         struct lustre_handle lockh;
1095
1096                         unlock_res(res);
1097                         ldlm_lock2handle(lock, &lockh);
1098                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1099                         if (rc)
1100                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1101                 } else {
1102                         unlock_res(res);
1103                         LDLM_DEBUG(lock,
1104                                    "Freeing a lock still held by a client node");
1105                         ldlm_lock_cancel(lock);
1106                 }
1107                 LDLM_LOCK_RELEASE(lock);
1108         } while (1);
1109 }
1110
1111 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1112                                struct hlist_node *hnode, void *arg)
1113 {
1114         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1115         __u64 flags = *(__u64 *)arg;
1116
1117         cleanup_resource(res, &res->lr_granted, flags);
1118         cleanup_resource(res, &res->lr_waiting, flags);
1119
1120         return 0;
1121 }
1122
1123 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1124                                   struct hlist_node *hnode, void *arg)
1125 {
1126         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1127
1128         lock_res(res);
1129         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1130                "(%d) after lock cleanup; forcing cleanup.\n",
1131                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1132                atomic_read(&res->lr_refcount) - 1);
1133
1134         /* Use D_NETERROR since it is in the default mask */
1135         ldlm_resource_dump(D_NETERROR, res);
1136         unlock_res(res);
1137         return 0;
1138 }
1139
1140 /**
1141  * Cancel and destroy all locks in the namespace.
1142  *
1143  * Typically used during evictions when server notified client that it was
1144  * evicted and all of its state needs to be destroyed.
1145  * Also used during shutdown.
1146  */
1147 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1148 {
1149         if (ns == NULL) {
1150                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1151                 return ELDLM_OK;
1152         }
1153
1154         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1155                                  &flags, 0);
1156         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1157                                  NULL, 0);
1158         return ELDLM_OK;
1159 }
1160 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1161
1162 /**
1163  * Attempts to free namespace.
1164  *
1165  * Only used when namespace goes away, like during an unmount.
1166  */
1167 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1168 {
1169         ENTRY;
1170
1171         /* At shutdown time, don't call the cancellation callback */
1172         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1173
1174         if (atomic_read(&ns->ns_bref) > 0) {
1175                 int rc;
1176                 CDEBUG(D_DLMTRACE,
1177                        "dlm namespace %s free waiting on refcount %d\n",
1178                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1179 force_wait:
1180                 if (force)
1181                         rc = wait_event_idle_timeout(
1182                                 ns->ns_waitq,
1183                                 atomic_read(&ns->ns_bref) == 0,
1184                                 cfs_time_seconds(1) / 4);
1185                 else
1186                         rc = l_wait_event_abortable(
1187                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1188
1189                 /* Forced cleanups should be able to reclaim all references,
1190                  * so it's safe to wait forever... we can't leak locks... */
1191                 if (force && rc == 0) {
1192                         rc = -ETIMEDOUT;
1193                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1194                                        "namespace with %d resources in use, "
1195                                        "(rc=%d)\n", ldlm_ns_name(ns),
1196                                        atomic_read(&ns->ns_bref), rc);
1197                         GOTO(force_wait, rc);
1198                 }
1199
1200                 if (atomic_read(&ns->ns_bref)) {
1201                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1202                                        "with %d resources in use, (rc=%d)\n",
1203                                        ldlm_ns_name(ns),
1204                                        atomic_read(&ns->ns_bref), rc);
1205                         RETURN(ELDLM_NAMESPACE_EXISTS);
1206                 }
1207                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1208                        ldlm_ns_name(ns));
1209         }
1210
1211         RETURN(ELDLM_OK);
1212 }
1213
1214 /**
1215  * Performs various cleanups for passed \a ns to make it drop refc and be
1216  * ready for freeing. Waits for refc == 0.
1217  *
1218  * The following is done:
1219  * (0) Unregister \a ns from its list to make inaccessible for potential
1220  * users like pools thread and others;
1221  * (1) Clear all locks in \a ns.
1222  */
1223 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1224                                struct obd_import *imp,
1225                                int force)
1226 {
1227         int rc;
1228
1229         ENTRY;
1230         if (!ns) {
1231                 EXIT;
1232                 return;
1233         }
1234
1235         spin_lock(&ns->ns_lock);
1236         ns->ns_stopping = 1;
1237         spin_unlock(&ns->ns_lock);
1238
1239         /*
1240          * Can fail with -EINTR when force == 0 in which case try harder.
1241          */
1242         rc = __ldlm_namespace_free(ns, force);
1243         if (rc != ELDLM_OK) {
1244                 if (imp) {
1245                         ptlrpc_disconnect_import(imp, 0);
1246                         ptlrpc_invalidate_import(imp);
1247                 }
1248
1249                 /*
1250                  * With all requests dropped and the import inactive
1251                  * we are gaurenteed all reference will be dropped.
1252                  */
1253                 rc = __ldlm_namespace_free(ns, 1);
1254                 LASSERT(rc == 0);
1255         }
1256         EXIT;
1257 }
1258 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1259
1260 /**
1261  * Performs freeing memory structures related to \a ns. This is only done
1262  * when ldlm_namespce_free_prior() successfully removed all resources
1263  * referencing \a ns and its refc == 0.
1264  */
1265 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1266 {
1267         ENTRY;
1268         if (!ns) {
1269                 EXIT;
1270                 return;
1271         }
1272
1273         /* Make sure that nobody can find this ns in its list. */
1274         ldlm_namespace_unregister(ns, ns->ns_client);
1275         /* Fini pool _before_ parent proc dir is removed. This is important as
1276          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1277          * Removing it after @dir may cause oops. */
1278         ldlm_pool_fini(&ns->ns_pool);
1279
1280         ldlm_namespace_debugfs_unregister(ns);
1281         ldlm_namespace_sysfs_unregister(ns);
1282         cfs_hash_putref(ns->ns_rs_hash);
1283         OBD_FREE_LARGE(ns->ns_rs_buckets,
1284                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1285         kfree(ns->ns_name);
1286         /* Namespace \a ns should be not on list at this time, otherwise
1287          * this will cause issues related to using freed \a ns in poold
1288          * thread.
1289          */
1290         LASSERT(list_empty(&ns->ns_list_chain));
1291         OBD_FREE_PTR(ns);
1292         ldlm_put_ref();
1293         EXIT;
1294 }
1295 EXPORT_SYMBOL(ldlm_namespace_free_post);
1296
1297 /**
1298  * Cleanup the resource, and free namespace.
1299  * bug 12864:
1300  * Deadlock issue:
1301  * proc1: destroy import
1302  *        class_disconnect_export(grab cl_sem) ->
1303  *              -> ldlm_namespace_free ->
1304  *              -> lprocfs_remove(grab _lprocfs_lock).
1305  * proc2: read proc info
1306  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1307  *              -> osc_rd_active, etc(grab cl_sem).
1308  *
1309  * So that I have to split the ldlm_namespace_free into two parts - the first
1310  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1311  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1312  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1313  * held.
1314  */
1315 void ldlm_namespace_free(struct ldlm_namespace *ns,
1316                          struct obd_import *imp,
1317                          int force)
1318 {
1319         ldlm_namespace_free_prior(ns, imp, force);
1320         ldlm_namespace_free_post(ns);
1321 }
1322 EXPORT_SYMBOL(ldlm_namespace_free);
1323
1324 void ldlm_namespace_get(struct ldlm_namespace *ns)
1325 {
1326         atomic_inc(&ns->ns_bref);
1327 }
1328
1329 /* This is only for callers that care about refcount */
1330 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1331 {
1332         return atomic_inc_return(&ns->ns_bref);
1333 }
1334
1335 void ldlm_namespace_put(struct ldlm_namespace *ns)
1336 {
1337         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1338                 wake_up(&ns->ns_waitq);
1339                 spin_unlock(&ns->ns_lock);
1340         }
1341 }
1342
1343 /** Register \a ns in the list of namespaces */
1344 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1345 {
1346         mutex_lock(ldlm_namespace_lock(client));
1347         LASSERT(list_empty(&ns->ns_list_chain));
1348         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1349         ldlm_namespace_nr_inc(client);
1350         mutex_unlock(ldlm_namespace_lock(client));
1351 }
1352
1353 /** Unregister \a ns from the list of namespaces. */
1354 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1355 {
1356         mutex_lock(ldlm_namespace_lock(client));
1357         LASSERT(!list_empty(&ns->ns_list_chain));
1358         /* Some asserts and possibly other parts of the code are still
1359          * using list_empty(&ns->ns_list_chain). This is why it is
1360          * important to use list_del_init() here. */
1361         list_del_init(&ns->ns_list_chain);
1362         ldlm_namespace_nr_dec(client);
1363         mutex_unlock(ldlm_namespace_lock(client));
1364 }
1365
1366 /** Should be called with ldlm_namespace_lock(client) taken. */
1367 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1368                                           enum ldlm_side client)
1369 {
1370         LASSERT(!list_empty(&ns->ns_list_chain));
1371         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1372         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1373 }
1374
1375 /** Should be called with ldlm_namespace_lock(client) taken. */
1376 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1377                                             enum ldlm_side client)
1378 {
1379         LASSERT(!list_empty(&ns->ns_list_chain));
1380         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1381         list_move_tail(&ns->ns_list_chain,
1382                        ldlm_namespace_inactive_list(client));
1383 }
1384
1385 /** Should be called with ldlm_namespace_lock(client) taken. */
1386 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1387 {
1388         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1389         LASSERT(!list_empty(ldlm_namespace_list(client)));
1390         return container_of(ldlm_namespace_list(client)->next,
1391                             struct ldlm_namespace, ns_list_chain);
1392 }
1393
1394 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1395 {
1396         int idx;
1397
1398         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1399                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1400         if (res->lr_itree == NULL)
1401                 return false;
1402         /* Initialize interval trees for each lock mode. */
1403         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1404                 res->lr_itree[idx].lit_size = 0;
1405                 res->lr_itree[idx].lit_mode = 1 << idx;
1406                 res->lr_itree[idx].lit_root = NULL;
1407         }
1408         return true;
1409 }
1410
1411 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1412 {
1413         int i;
1414
1415         OBD_ALLOC_PTR(res->lr_ibits_queues);
1416         if (res->lr_ibits_queues == NULL)
1417                 return false;
1418         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1419                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1420         return true;
1421 }
1422
1423 /** Create and initialize new resource. */
1424 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1425 {
1426         struct ldlm_resource *res;
1427         bool rc;
1428
1429         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1430         if (res == NULL)
1431                 return NULL;
1432
1433         switch (ldlm_type) {
1434         case LDLM_EXTENT:
1435                 rc = ldlm_resource_extent_new(res);
1436                 break;
1437         case LDLM_IBITS:
1438                 rc = ldlm_resource_inodebits_new(res);
1439                 break;
1440         default:
1441                 rc = true;
1442                 break;
1443         }
1444         if (!rc) {
1445                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1446                 return NULL;
1447         }
1448
1449         INIT_LIST_HEAD(&res->lr_granted);
1450         INIT_LIST_HEAD(&res->lr_waiting);
1451
1452         atomic_set(&res->lr_refcount, 1);
1453         spin_lock_init(&res->lr_lock);
1454         lu_ref_init(&res->lr_reference);
1455
1456         /* Since LVB init can be delayed now, there is no longer need to
1457          * immediatelly acquire mutex here. */
1458         mutex_init(&res->lr_lvb_mutex);
1459         res->lr_lvb_initialized = false;
1460
1461         return res;
1462 }
1463
1464 static void ldlm_resource_free(struct ldlm_resource *res)
1465 {
1466         if (res->lr_type == LDLM_EXTENT) {
1467                 if (res->lr_itree != NULL)
1468                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1469                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1470         } else if (res->lr_type == LDLM_IBITS) {
1471                 if (res->lr_ibits_queues != NULL)
1472                         OBD_FREE_PTR(res->lr_ibits_queues);
1473         }
1474
1475         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1476 }
1477
1478 /**
1479  * Return a reference to resource with given name, creating it if necessary.
1480  * Args: namespace with ns_lock unlocked
1481  * Locks: takes and releases NS hash-lock and res->lr_lock
1482  * Returns: referenced, unlocked ldlm_resource or NULL
1483  */
1484 struct ldlm_resource *
1485 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1486                   const struct ldlm_res_id *name, enum ldlm_type type,
1487                   int create)
1488 {
1489         struct hlist_node       *hnode;
1490         struct ldlm_resource    *res = NULL;
1491         struct cfs_hash_bd              bd;
1492         __u64                   version;
1493         int                     ns_refcount = 0;
1494         int hash;
1495
1496         LASSERT(ns != NULL);
1497         LASSERT(parent == NULL);
1498         LASSERT(ns->ns_rs_hash != NULL);
1499         LASSERT(name->name[0] != 0);
1500
1501         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1502         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1503         if (hnode != NULL) {
1504                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1505                 GOTO(found, res);
1506         }
1507
1508         version = cfs_hash_bd_version_get(&bd);
1509         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1510
1511         if (create == 0)
1512                 return ERR_PTR(-ENOENT);
1513
1514         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1515                  "type: %d\n", type);
1516         res = ldlm_resource_new(type);
1517         if (res == NULL)
1518                 return ERR_PTR(-ENOMEM);
1519
1520         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1521         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1522         res->lr_name = *name;
1523         res->lr_type = type;
1524
1525         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1526         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1527                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1528
1529         if (hnode != NULL) {
1530                 /* Someone won the race and already added the resource. */
1531                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1532                 /* Clean lu_ref for failed resource. */
1533                 lu_ref_fini(&res->lr_reference);
1534                 ldlm_resource_free(res);
1535 found:
1536                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1537                 return res;
1538         }
1539         /* We won! Let's add the resource. */
1540         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1541         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1542                 ns_refcount = ldlm_namespace_get_return(ns);
1543
1544         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1545
1546         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1547
1548         /* Let's see if we happened to be the very first resource in this
1549          * namespace. If so, and this is a client namespace, we need to move
1550          * the namespace into the active namespaces list to be patrolled by
1551          * the ldlm_poold. */
1552         if (ns_is_client(ns) && ns_refcount == 1) {
1553                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1554                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1555                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1556         }
1557
1558         return res;
1559 }
1560 EXPORT_SYMBOL(ldlm_resource_get);
1561
1562 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1563 {
1564         LASSERT(res != NULL);
1565         LASSERT(res != LP_POISON);
1566         atomic_inc(&res->lr_refcount);
1567         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1568                atomic_read(&res->lr_refcount));
1569         return res;
1570 }
1571
1572 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1573                                          struct ldlm_resource *res)
1574 {
1575         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1576
1577         if (!list_empty(&res->lr_granted)) {
1578                 ldlm_resource_dump(D_ERROR, res);
1579                 LBUG();
1580         }
1581
1582         if (!list_empty(&res->lr_waiting)) {
1583                 ldlm_resource_dump(D_ERROR, res);
1584                 LBUG();
1585         }
1586
1587         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1588                                bd, &res->lr_hash);
1589         lu_ref_fini(&res->lr_reference);
1590         if (atomic_dec_and_test(&nsb->nsb_count))
1591                 ldlm_namespace_put(nsb->nsb_namespace);
1592 }
1593
1594 /* Returns 1 if the resource was freed, 0 if it remains. */
1595 int ldlm_resource_putref(struct ldlm_resource *res)
1596 {
1597         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1598         struct cfs_hash_bd   bd;
1599
1600         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1601         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1602                res, atomic_read(&res->lr_refcount) - 1);
1603
1604         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1605         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1606                 __ldlm_resource_putref_final(&bd, res);
1607                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1608                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1609                         ns->ns_lvbo->lvbo_free(res);
1610                 ldlm_resource_free(res);
1611                 return 1;
1612         }
1613         return 0;
1614 }
1615 EXPORT_SYMBOL(ldlm_resource_putref);
1616
1617 /**
1618  * Add a lock into a given resource into specified lock list.
1619  */
1620 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1621                             struct ldlm_lock *lock)
1622 {
1623         check_res_locked(res);
1624
1625         LDLM_DEBUG(lock, "About to add this lock");
1626
1627         if (ldlm_is_destroyed(lock)) {
1628                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1629                 return;
1630         }
1631
1632         LASSERT(list_empty(&lock->l_res_link));
1633
1634         list_add_tail(&lock->l_res_link, head);
1635
1636         if (res->lr_type == LDLM_IBITS)
1637                 ldlm_inodebits_add_lock(res, head, lock);
1638 }
1639
1640 /**
1641  * Insert a lock into resource after specified lock.
1642  *
1643  * Obtain resource description from the lock we are inserting after.
1644  */
1645 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1646                                      struct ldlm_lock *new)
1647 {
1648         struct ldlm_resource *res = original->l_resource;
1649
1650         check_res_locked(res);
1651
1652         ldlm_resource_dump(D_INFO, res);
1653         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1654
1655         if (ldlm_is_destroyed(new)) {
1656                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1657                 goto out;
1658         }
1659
1660         LASSERT(list_empty(&new->l_res_link));
1661
1662         list_add(&new->l_res_link, &original->l_res_link);
1663  out:;
1664 }
1665
1666 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1667 {
1668         int type = lock->l_resource->lr_type;
1669
1670         check_res_locked(lock->l_resource);
1671         switch (type) {
1672         case LDLM_PLAIN:
1673                 ldlm_unlink_lock_skiplist(lock);
1674                 break;
1675         case LDLM_EXTENT:
1676                 ldlm_extent_unlink_lock(lock);
1677                 break;
1678         case LDLM_IBITS:
1679                 ldlm_inodebits_unlink_lock(lock);
1680                 break;
1681         }
1682         list_del_init(&lock->l_res_link);
1683 }
1684 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1685
1686 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1687 {
1688         desc->lr_type = res->lr_type;
1689         desc->lr_name = res->lr_name;
1690 }
1691
1692 /**
1693  * Print information about all locks in all namespaces on this node to debug
1694  * log.
1695  */
1696 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1697 {
1698         struct list_head *tmp;
1699
1700         if (!((libcfs_debug | D_ERROR) & level))
1701                 return;
1702
1703         mutex_lock(ldlm_namespace_lock(client));
1704
1705         list_for_each(tmp, ldlm_namespace_list(client)) {
1706                 struct ldlm_namespace *ns;
1707
1708                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1709                 ldlm_namespace_dump(level, ns);
1710         }
1711
1712         mutex_unlock(ldlm_namespace_lock(client));
1713 }
1714
1715 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1716                               struct hlist_node *hnode, void *arg)
1717 {
1718         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1719         int    level = (int)(unsigned long)arg;
1720
1721         lock_res(res);
1722         ldlm_resource_dump(level, res);
1723         unlock_res(res);
1724
1725         return 0;
1726 }
1727
1728 /**
1729  * Print information about all locks in this namespace on this node to debug
1730  * log.
1731  */
1732 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1733 {
1734         if (!((libcfs_debug | D_ERROR) & level))
1735                 return;
1736
1737         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1738                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1739                ns_is_client(ns) ? "client" : "server");
1740
1741         if (ktime_get_seconds() < ns->ns_next_dump)
1742                 return;
1743
1744         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1745                                  ldlm_res_hash_dump,
1746                                  (void *)(unsigned long)level, 0);
1747         spin_lock(&ns->ns_lock);
1748         ns->ns_next_dump = ktime_get_seconds() + 10;
1749         spin_unlock(&ns->ns_lock);
1750 }
1751
1752 /**
1753  * Print information about all locks in this resource to debug log.
1754  */
1755 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1756 {
1757         struct ldlm_lock *lock;
1758         unsigned int granted = 0;
1759
1760         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1761
1762         if (!((libcfs_debug | D_ERROR) & level))
1763                 return;
1764
1765         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1766                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1767
1768         if (!list_empty(&res->lr_granted)) {
1769                 CDEBUG(level, "Granted locks (in reverse order):\n");
1770                 list_for_each_entry_reverse(lock, &res->lr_granted,
1771                                                 l_res_link) {
1772                         LDLM_DEBUG_LIMIT(level, lock, "###");
1773                         if (!(level & D_CANTMASK) &&
1774                             ++granted > ldlm_dump_granted_max) {
1775                                 CDEBUG(level,
1776                                        "only dump %d granted locks to avoid DDOS.\n",
1777                                        granted);
1778                                 break;
1779                         }
1780                 }
1781         }
1782
1783         if (!list_empty(&res->lr_waiting)) {
1784                 CDEBUG(level, "Waiting locks:\n");
1785                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1786                         LDLM_DEBUG_LIMIT(level, lock, "###");
1787         }
1788 }
1789 EXPORT_SYMBOL(ldlm_resource_dump);