lustre/ldlm/ldlm_resource.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46 struct kmem_cache *ldlm_inodebits_slab;
  47
  48 int ldlm_srv_namespace_nr = 0;
  49 int ldlm_cli_namespace_nr = 0;
  50
  51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  52 LIST_HEAD(ldlm_srv_namespace_list);
  53
  54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
  55 /* Client Namespaces that have active resources in them.
  56  * Once all resources go away, ldlm_poold moves such namespaces to the
  57  * inactive list */
  58 LIST_HEAD(ldlm_cli_active_namespace_list);
  59 /* Client namespaces that don't have any locks in them */
  60 LIST_HEAD(ldlm_cli_inactive_namespace_list);
  61
  62 static struct dentry *ldlm_debugfs_dir;
  63 static struct dentry *ldlm_ns_debugfs_dir;
  64 struct dentry *ldlm_svc_debugfs_dir;
  65
  66 /* during debug dump certain amount of granted locks for one resource to avoid
  67  * DDOS. */
  68 static unsigned int ldlm_dump_granted_max = 256;
  69
  70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
  71                                           const char __user *buffer,
  72                                           size_t count, loff_t *off)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78
  79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
  80
  81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  82 {
  83         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  84         return 0;
  85 }
  86
  87 static ssize_t
  88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  89                        size_t count, loff_t *off)
  90 {
  91         struct seq_file *seq = file->private_data;
  92
  93         if (!count)
  94                 return 0;
  95
  96         return kstrtouint_from_user(buffer, count, 0,
  97                                     (unsigned int *)seq->private);
  98 }
  99
 100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
 101
 102 #ifdef HAVE_SERVER_SUPPORT
 103
 104 static int seq_watermark_show(struct seq_file *m, void *data)
 105 {
 106         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 107         return 0;
 108 }
 109
 110 static ssize_t seq_watermark_write(struct file *file,
 111                                    const char __user *buffer, size_t count,
 112                                    loff_t *off)
 113 {
 114         struct seq_file *m = file->private_data;
 115         u64 value;
 116         __u64 watermark;
 117         __u64 *data = m->private;
 118         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 119         char kernbuf[22] = "";
 120         int rc;
 121
 122         if (count >= sizeof(kernbuf))
 123                 return -EINVAL;
 124
 125         if (copy_from_user(kernbuf, buffer, count))
 126                 return -EFAULT;
 127         kernbuf[count] = 0;
 128
 129         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
 130         if (rc < 0) {
 131                 CERROR("Failed to set %s, rc = %d.\n",
 132                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 133                        rc);
 134                 return rc;
 135         } else if (value != 0 && value < (1 << 20)) {
 136                 CERROR("%s should be greater than 1MB.\n",
 137                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 138                 return -EINVAL;
 139         }
 140         watermark = value >> 20;
 141
 142         if (wm_low) {
 143                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 144                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 145                                "lock_limit_mb.\n");
 146                         return -EINVAL;
 147                 }
 148
 149                 *data = watermark;
 150                 if (watermark != 0) {
 151                         watermark <<= 20;
 152                         do_div(watermark, sizeof(struct ldlm_lock));
 153                 }
 154                 ldlm_reclaim_threshold = watermark;
 155         } else {
 156                 if (ldlm_reclaim_threshold_mb != 0 &&
 157                     watermark < ldlm_reclaim_threshold_mb) {
 158                         CERROR("lock_limit_mb must be greater than "
 159                                "lock_reclaim_threshold_mb.\n");
 160                         return -EINVAL;
 161                 }
 162
 163                 *data = watermark;
 164                 if (watermark != 0) {
 165                         watermark <<= 20;
 166                         do_div(watermark, sizeof(struct ldlm_lock));
 167                 }
 168                 ldlm_lock_limit = watermark;
 169         }
 170
 171         return count;
 172 }
 173
 174 static int seq_watermark_open(struct inode *inode, struct file *file)
 175 {
 176         return single_open(file, seq_watermark_show, inode->i_private);
 177 }
 178
 179 static const struct file_operations ldlm_watermark_fops = {
 180         .owner          = THIS_MODULE,
 181         .open           = seq_watermark_open,
 182         .read           = seq_read,
 183         .write          = seq_watermark_write,
 184         .llseek         = seq_lseek,
 185         .release        = lprocfs_single_release,
 186 };
 187
 188 static int seq_granted_show(struct seq_file *m, void *data)
 189 {
 190         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 191                    (struct percpu_counter *)m->private));
 192         return 0;
 193 }
 194
 195 static int seq_granted_open(struct inode *inode, struct file *file)
 196 {
 197         return single_open(file, seq_granted_show, inode->i_private);
 198 }
 199
 200 static const struct file_operations ldlm_granted_fops = {
 201         .owner  = THIS_MODULE,
 202         .open   = seq_granted_open,
 203         .read   = seq_read,
 204         .llseek = seq_lseek,
 205         .release = seq_release,
 206 };
 207
 208 #endif /* HAVE_SERVER_SUPPORT */
 209
 210 static struct lprocfs_vars ldlm_debugfs_list[] = {
 211         { .name =       "dump_namespaces",
 212           .fops =       &ldlm_dump_ns_fops,
 213           .proc_mode =  0222 },
 214         { .name =       "dump_granted_max",
 215           .fops =       &ldlm_rw_uint_fops,
 216           .data =       &ldlm_dump_granted_max },
 217 #ifdef HAVE_SERVER_SUPPORT
 218         { .name =       "lock_reclaim_threshold_mb",
 219           .fops =       &ldlm_watermark_fops,
 220           .data =       &ldlm_reclaim_threshold_mb },
 221         { .name =       "lock_limit_mb",
 222           .fops =       &ldlm_watermark_fops,
 223           .data =       &ldlm_lock_limit_mb },
 224         { .name =       "lock_granted_count",
 225           .fops =       &ldlm_granted_fops,
 226           .data =       &ldlm_granted_total },
 227 #endif
 228         { NULL }
 229 };
 230
 231 int ldlm_debugfs_setup(void)
 232 {
 233         ENTRY;
 234         ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
 235                                              debugfs_lustre_root);
 236         ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
 237                                                  ldlm_debugfs_dir);
 238         ldlm_svc_debugfs_dir = debugfs_create_dir("services",
 239                                                   ldlm_debugfs_dir);
 240
 241         ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 242
 243         RETURN(0);
 244 }
 245
 246 void ldlm_debugfs_cleanup(void)
 247 {
 248         debugfs_remove_recursive(ldlm_debugfs_dir);
 249
 250         ldlm_svc_debugfs_dir = NULL;
 251         ldlm_ns_debugfs_dir = NULL;
 252         ldlm_debugfs_dir = NULL;
 253 }
 254
 255 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 256                                    char *buf)
 257 {
 258         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 259                                                  ns_kobj);
 260         __u64                   res = 0;
 261         int                     i;
 262
 263         /* result is not strictly consistant */
 264         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
 265                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
 266         return sprintf(buf, "%lld\n", res);
 267 }
 268 LUSTRE_RO_ATTR(resource_count);
 269
 270 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 271                                char *buf)
 272 {
 273         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 274                                                  ns_kobj);
 275         __u64                   locks;
 276
 277         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 278                                         LPROCFS_FIELDS_FLAGS_SUM);
 279         return sprintf(buf, "%lld\n", locks);
 280 }
 281 LUSTRE_RO_ATTR(lock_count);
 282
 283 static ssize_t lock_unused_count_show(struct kobject *kobj,
 284                                       struct attribute *attr,
 285                                       char *buf)
 286 {
 287         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 288                                                  ns_kobj);
 289
 290         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 291 }
 292 LUSTRE_RO_ATTR(lock_unused_count);
 293
 294 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 295                              char *buf)
 296 {
 297         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 298                                                  ns_kobj);
 299         __u32 *nr = &ns->ns_max_unused;
 300
 301         if (ns_connect_lru_resize(ns))
 302                 nr = &ns->ns_nr_unused;
 303         return sprintf(buf, "%u\n", *nr);
 304 }
 305
 306 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 307                               const char *buffer, size_t count)
 308 {
 309         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 310                                                  ns_kobj);
 311         unsigned long tmp;
 312         int lru_resize;
 313         int err;
 314
 315         if (strncmp(buffer, "clear", 5) == 0) {
 316                 CDEBUG(D_DLMTRACE,
 317                        "dropping all unused locks from namespace %s\n",
 318                        ldlm_ns_name(ns));
 319                 if (ns_connect_lru_resize(ns)) {
 320                         /* Try to cancel all @ns_nr_unused locks. */
 321                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 322                                         LDLM_LRU_FLAG_PASSED |
 323                                         LDLM_LRU_FLAG_CLEANUP);
 324                 } else {
 325                         tmp = ns->ns_max_unused;
 326                         ns->ns_max_unused = 0;
 327                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 328                                         LDLM_LRU_FLAG_CLEANUP);
 329                         ns->ns_max_unused = tmp;
 330                 }
 331                 return count;
 332         }
 333
 334         err = kstrtoul(buffer, 10, &tmp);
 335         if (err != 0) {
 336                 CERROR("lru_size: invalid value written\n");
 337                 return -EINVAL;
 338         }
 339         lru_resize = (tmp == 0);
 340
 341         if (ns_connect_lru_resize(ns)) {
 342                 if (!lru_resize)
 343                         ns->ns_max_unused = (unsigned int)tmp;
 344
 345                 if (tmp > ns->ns_nr_unused)
 346                         tmp = ns->ns_nr_unused;
 347                 tmp = ns->ns_nr_unused - tmp;
 348
 349                 CDEBUG(D_DLMTRACE,
 350                        "changing namespace %s unused locks from %u to %u\n",
 351                        ldlm_ns_name(ns), ns->ns_nr_unused,
 352                        (unsigned int)tmp);
 353                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 354
 355                 if (!lru_resize) {
 356                         CDEBUG(D_DLMTRACE,
 357                                "disable lru_resize for namespace %s\n",
 358                                ldlm_ns_name(ns));
 359                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 360                 }
 361         } else {
 362                 CDEBUG(D_DLMTRACE,
 363                        "changing namespace %s max_unused from %u to %u\n",
 364                        ldlm_ns_name(ns), ns->ns_max_unused,
 365                        (unsigned int)tmp);
 366                 ns->ns_max_unused = (unsigned int)tmp;
 367                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 368
 369                 /* Make sure that LRU resize was originally supported before
 370                  * turning it on here.
 371                  */
 372                 if (lru_resize &&
 373                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 374                         CDEBUG(D_DLMTRACE,
 375                                "enable lru_resize for namespace %s\n",
 376                                ldlm_ns_name(ns));
 377                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 378                 }
 379         }
 380
 381         return count;
 382 }
 383 LUSTRE_RW_ATTR(lru_size);
 384
 385 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 386                                 char *buf)
 387 {
 388         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 389                                                  ns_kobj);
 390
 391         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 392 }
 393
 394 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 395                                  const char *buffer, size_t count)
 396 {
 397         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 398                                                  ns_kobj);
 399         int scale = NSEC_PER_MSEC;
 400         unsigned long long tmp;
 401         char *buf;
 402
 403         /* Did the user ask in seconds or milliseconds. Default is in ms */
 404         buf = strstr(buffer, "ms");
 405         if (!buf) {
 406                 buf = strchr(buffer, 's');
 407                 if (buf)
 408                         scale = NSEC_PER_SEC;
 409         }
 410
 411         if (buf)
 412                 *buf = '\0';
 413
 414         if (kstrtoull(buffer, 10, &tmp))
 415                 return -EINVAL;
 416
 417         ns->ns_max_age = ktime_set(0, tmp * scale);
 418
 419         return count;
 420 }
 421 LUSTRE_RW_ATTR(lru_max_age);
 422
 423 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 424                                       struct attribute *attr,
 425                                       char *buf)
 426 {
 427         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 428                                                  ns_kobj);
 429
 430         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 431 }
 432
 433 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 434                                        struct attribute *attr,
 435                                        const char *buffer,
 436                                        size_t count)
 437 {
 438         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 439                                                  ns_kobj);
 440         unsigned long supp = -1;
 441         int rc;
 442
 443         rc = kstrtoul(buffer, 10, &supp);
 444         if (rc < 0)
 445                 return rc;
 446
 447         if (supp == 0)
 448                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 449         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 450                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 451         return count;
 452 }
 453 LUSTRE_RW_ATTR(early_lock_cancel);
 454
 455 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 456                                     struct attribute *attr, char *buf)
 457 {
 458         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 459                                                  ns_kobj);
 460
 461         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 462 }
 463
 464 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 465                                      struct attribute *attr,
 466                                      const char *buffer, size_t count)
 467 {
 468         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 469                                                  ns_kobj);
 470         unsigned long long tmp;
 471
 472         if (kstrtoull(buffer, 10, &tmp))
 473                 return -EINVAL;
 474
 475         ns->ns_dirty_age_limit = tmp;
 476
 477         return count;
 478 }
 479 LUSTRE_RW_ATTR(dirty_age_limit);
 480
 481 #ifdef HAVE_SERVER_SUPPORT
 482 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 483                                     struct attribute *attr, char *buf)
 484 {
 485         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 486                                                  ns_kobj);
 487
 488         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 489 }
 490
 491 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 492                                      struct attribute *attr,
 493                                      const char *buffer, size_t count)
 494 {
 495         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 496                                                  ns_kobj);
 497         unsigned long long tmp;
 498
 499         if (kstrtoull(buffer, 10, &tmp))
 500                 return -EINVAL;
 501
 502         ns->ns_ctime_age_limit = tmp;
 503
 504         return count;
 505 }
 506 LUSTRE_RW_ATTR(ctime_age_limit);
 507
 508 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 509                                   char *buf)
 510 {
 511         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 512                                                  ns_kobj);
 513
 514         return sprintf(buf, "%d\n", ns->ns_timeouts);
 515 }
 516 LUSTRE_RO_ATTR(lock_timeouts);
 517
 518 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 519                                      struct attribute *attr, char *buf)
 520 {
 521         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 522                                                  ns_kobj);
 523
 524         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 525 }
 526
 527 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 528                                       struct attribute *attr,
 529                                       const char *buffer, size_t count)
 530 {
 531         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 532                                                  ns_kobj);
 533         unsigned long tmp;
 534         int err;
 535
 536         err = kstrtoul(buffer, 10, &tmp);
 537         if (err != 0)
 538                 return -EINVAL;
 539
 540         ns->ns_max_nolock_size = tmp;
 541
 542         return count;
 543 }
 544 LUSTRE_RW_ATTR(max_nolock_bytes);
 545
 546 static ssize_t contention_seconds_show(struct kobject *kobj,
 547                                        struct attribute *attr, char *buf)
 548 {
 549         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 550                                                  ns_kobj);
 551
 552         return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
 553 }
 554
 555 static ssize_t contention_seconds_store(struct kobject *kobj,
 556                                         struct attribute *attr,
 557                                         const char *buffer, size_t count)
 558 {
 559         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 560                                                  ns_kobj);
 561         unsigned int tmp;
 562
 563         if (kstrtouint(buffer, 10, &tmp))
 564                 return -EINVAL;
 565
 566         ns->ns_contention_time = tmp;
 567
 568         return count;
 569 }
 570 LUSTRE_RW_ATTR(contention_seconds);
 571
 572 static ssize_t contended_locks_show(struct kobject *kobj,
 573                                     struct attribute *attr, char *buf)
 574 {
 575         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 576                                                  ns_kobj);
 577
 578         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 579 }
 580
 581 static ssize_t contended_locks_store(struct kobject *kobj,
 582                                      struct attribute *attr,
 583                                      const char *buffer, size_t count)
 584 {
 585         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 586                                                  ns_kobj);
 587         unsigned long tmp;
 588         int err;
 589
 590         err = kstrtoul(buffer, 10, &tmp);
 591         if (err != 0)
 592                 return -EINVAL;
 593
 594         ns->ns_contended_locks = tmp;
 595
 596         return count;
 597 }
 598 LUSTRE_RW_ATTR(contended_locks);
 599
 600 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 601                                      struct attribute *attr, char *buf)
 602 {
 603         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 604                                                  ns_kobj);
 605
 606         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 607 }
 608
 609 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 610                                       struct attribute *attr,
 611                                       const char *buffer, size_t count)
 612 {
 613         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 614                                                  ns_kobj);
 615         unsigned long tmp;
 616         int err;
 617
 618         err = kstrtoul(buffer, 10, &tmp);
 619         if (err != 0)
 620                 return -EINVAL;
 621
 622         ns->ns_max_parallel_ast = tmp;
 623
 624         return count;
 625 }
 626 LUSTRE_RW_ATTR(max_parallel_ast);
 627
 628 #endif /* HAVE_SERVER_SUPPORT */
 629
 630 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 631 static struct attribute *ldlm_ns_attrs[] = {
 632         &lustre_attr_resource_count.attr,
 633         &lustre_attr_lock_count.attr,
 634         &lustre_attr_lock_unused_count.attr,
 635         &lustre_attr_lru_size.attr,
 636         &lustre_attr_lru_max_age.attr,
 637         &lustre_attr_early_lock_cancel.attr,
 638         &lustre_attr_dirty_age_limit.attr,
 639 #ifdef HAVE_SERVER_SUPPORT
 640         &lustre_attr_ctime_age_limit.attr,
 641         &lustre_attr_lock_timeouts.attr,
 642         &lustre_attr_max_nolock_bytes.attr,
 643         &lustre_attr_contention_seconds.attr,
 644         &lustre_attr_contended_locks.attr,
 645         &lustre_attr_max_parallel_ast.attr,
 646 #endif
 647         NULL,
 648 };
 649
 650 static void ldlm_ns_release(struct kobject *kobj)
 651 {
 652         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 653                                                  ns_kobj);
 654         complete(&ns->ns_kobj_unregister);
 655 }
 656
 657 static struct kobj_type ldlm_ns_ktype = {
 658         .default_attrs  = ldlm_ns_attrs,
 659         .sysfs_ops      = &lustre_sysfs_ops,
 660         .release        = ldlm_ns_release,
 661 };
 662
 663 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 664 {
 665         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 666                 CERROR("dlm namespace %s has no procfs dir?\n",
 667                        ldlm_ns_name(ns));
 668         else
 669                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 670
 671         if (ns->ns_stats != NULL)
 672                 lprocfs_free_stats(&ns->ns_stats);
 673 }
 674
 675 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 676 {
 677         kobject_put(&ns->ns_kobj);
 678         wait_for_completion(&ns->ns_kobj_unregister);
 679 }
 680
 681 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 682 {
 683         int err;
 684
 685         ns->ns_kobj.kset = ldlm_ns_kset;
 686         init_completion(&ns->ns_kobj_unregister);
 687         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 688                                    "%s", ldlm_ns_name(ns));
 689
 690         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 691         if (!ns->ns_stats) {
 692                 kobject_put(&ns->ns_kobj);
 693                 return -ENOMEM;
 694         }
 695
 696         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 697                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 698
 699         return err;
 700 }
 701
 702 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 703 {
 704         struct dentry *ns_entry;
 705
 706         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 707                 ns_entry = ns->ns_debugfs_entry;
 708         } else {
 709                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 710                                               ldlm_ns_debugfs_dir);
 711                 if (!ns_entry)
 712                         return -ENOMEM;
 713                 ns->ns_debugfs_entry = ns_entry;
 714         }
 715
 716         return 0;
 717 }
 718 #undef MAX_STRING_SIZE
 719
 720 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 721                                   const void *key, unsigned int mask)
 722 {
 723         const struct ldlm_res_id *id = key;
 724         unsigned int val = 0;
 725         unsigned int i;
 726
 727         for (i = 0; i < RES_NAME_SIZE; i++)
 728                 val += id->name[i];
 729         return val & mask;
 730 }
 731
 732 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 733 {
 734         struct lu_fid       fid;
 735         __u32               hash;
 736         __u32               val;
 737
 738         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 739         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 740         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 741
 742         hash = fid_flatten32(&fid);
 743         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 744         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 745                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 746         } else {
 747                 val = fid_oid(&fid);
 748         }
 749         hash += (val >> 5) + (val << 11);
 750         return cfs_hash_32(hash, bits);
 751 }
 752
 753 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 754 {
 755         struct ldlm_resource   *res;
 756
 757         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 758         return &res->lr_name;
 759 }
 760
 761 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 762 {
 763         struct ldlm_resource   *res;
 764
 765         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 766         return ldlm_res_eq((const struct ldlm_res_id *)key,
 767                            (const struct ldlm_res_id *)&res->lr_name);
 768 }
 769
 770 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 771 {
 772         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 773 }
 774
 775 static void
 776 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 777 {
 778         struct ldlm_resource *res;
 779
 780         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 781         ldlm_resource_getref(res);
 782 }
 783
 784 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 785 {
 786         struct ldlm_resource *res;
 787
 788         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 789         ldlm_resource_putref(res);
 790 }
 791
 792 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 793         .hs_hash        = ldlm_res_hop_hash,
 794         .hs_key         = ldlm_res_hop_key,
 795         .hs_keycmp      = ldlm_res_hop_keycmp,
 796         .hs_keycpy      = NULL,
 797         .hs_object      = ldlm_res_hop_object,
 798         .hs_get         = ldlm_res_hop_get_locked,
 799         .hs_put         = ldlm_res_hop_put
 800 };
 801
 802 static struct {
 803         /** hash bucket bits */
 804         unsigned                nsd_bkt_bits;
 805         /** hash bits */
 806         unsigned                nsd_all_bits;
 807 } ldlm_ns_hash_defs[] = {
 808         [LDLM_NS_TYPE_MDC] = {
 809                 .nsd_bkt_bits   = 11,
 810                 .nsd_all_bits   = 16,
 811         },
 812         [LDLM_NS_TYPE_MDT] = {
 813                 .nsd_bkt_bits   = 14,
 814                 .nsd_all_bits   = 21,
 815         },
 816         [LDLM_NS_TYPE_OSC] = {
 817                 .nsd_bkt_bits   = 8,
 818                 .nsd_all_bits   = 12,
 819         },
 820         [LDLM_NS_TYPE_OST] = {
 821                 .nsd_bkt_bits   = 11,
 822                 .nsd_all_bits   = 17,
 823         },
 824         [LDLM_NS_TYPE_MGC] = {
 825                 .nsd_bkt_bits   = 3,
 826                 .nsd_all_bits   = 4,
 827         },
 828         [LDLM_NS_TYPE_MGT] = {
 829                 .nsd_bkt_bits   = 3,
 830                 .nsd_all_bits   = 4,
 831         },
 832 };
 833
 834 /**
 835  * Create and initialize new empty namespace.
 836  */
 837 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 838                                           enum ldlm_side client,
 839                                           enum ldlm_appetite apt,
 840                                           enum ldlm_ns_type ns_type)
 841 {
 842         struct ldlm_namespace *ns = NULL;
 843         int idx;
 844         int rc;
 845
 846         ENTRY;
 847         LASSERT(obd != NULL);
 848
 849         rc = ldlm_get_ref();
 850         if (rc) {
 851                 CERROR("ldlm_get_ref failed: %d\n", rc);
 852                 RETURN(NULL);
 853         }
 854
 855         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 856             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 857                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 858                 GOTO(out_ref, NULL);
 859         }
 860
 861         OBD_ALLOC_PTR(ns);
 862         if (!ns)
 863                 GOTO(out_ref, NULL);
 864
 865         ns->ns_rs_hash = cfs_hash_create(name,
 866                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 867                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 868                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 869                                          0,
 870                                          CFS_HASH_MIN_THETA,
 871                                          CFS_HASH_MAX_THETA,
 872                                          &ldlm_ns_hash_ops,
 873                                          CFS_HASH_DEPTH |
 874                                          CFS_HASH_BIGNAME |
 875                                          CFS_HASH_SPIN_BKTLOCK |
 876                                          CFS_HASH_NO_ITEMREF);
 877         if (ns->ns_rs_hash == NULL)
 878                 GOTO(out_ns, NULL);
 879
 880         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 881                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 882
 883         OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
 884         if (!ns->ns_rs_buckets)
 885                 goto out_hash;
 886
 887         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 888                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 889
 890                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 891                 nsb->nsb_namespace = ns;
 892                 nsb->nsb_reclaim_start = 0;
 893                 atomic_set(&nsb->nsb_count, 0);
 894         }
 895
 896         ns->ns_obd = obd;
 897         ns->ns_appetite = apt;
 898         ns->ns_client = client;
 899         ns->ns_name = kstrdup(name, GFP_KERNEL);
 900         if (!ns->ns_name)
 901                 goto out_hash;
 902
 903         INIT_LIST_HEAD(&ns->ns_list_chain);
 904         INIT_LIST_HEAD(&ns->ns_unused_list);
 905         spin_lock_init(&ns->ns_lock);
 906         atomic_set(&ns->ns_bref, 0);
 907         init_waitqueue_head(&ns->ns_waitq);
 908
 909         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 910         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 911         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 912
 913         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 914         ns->ns_nr_unused          = 0;
 915         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 916         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 917         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 918         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 919         ns->ns_timeouts           = 0;
 920         ns->ns_orig_connect_flags = 0;
 921         ns->ns_connect_flags      = 0;
 922         ns->ns_stopping           = 0;
 923         ns->ns_reclaim_start      = 0;
 924         ns->ns_last_pos           = &ns->ns_unused_list;
 925
 926         rc = ldlm_namespace_sysfs_register(ns);
 927         if (rc) {
 928                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 929                 GOTO(out_hash, rc);
 930         }
 931
 932         rc = ldlm_namespace_debugfs_register(ns);
 933         if (rc) {
 934                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 935                 GOTO(out_sysfs, rc);
 936         }
 937
 938         idx = ldlm_namespace_nr_read(client);
 939         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 940         if (rc) {
 941                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 942                 GOTO(out_proc, rc);
 943         }
 944
 945         ldlm_namespace_register(ns, client);
 946         RETURN(ns);
 947 out_proc:
 948         ldlm_namespace_debugfs_unregister(ns);
 949 out_sysfs:
 950         ldlm_namespace_sysfs_unregister(ns);
 951         ldlm_namespace_cleanup(ns, 0);
 952 out_hash:
 953         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
 954         kfree(ns->ns_name);
 955         cfs_hash_putref(ns->ns_rs_hash);
 956 out_ns:
 957         OBD_FREE_PTR(ns);
 958 out_ref:
 959         ldlm_put_ref();
 960         RETURN(NULL);
 961 }
 962 EXPORT_SYMBOL(ldlm_namespace_new);
 963
 964 /**
 965  * Cancel and destroy all locks on a resource.
 966  *
 967  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 968  * clean up.  This is currently only used for recovery, and we make
 969  * certain assumptions as a result--notably, that we shouldn't cancel
 970  * locks with refs.
 971  */
 972 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 973                              __u64 flags)
 974 {
 975         struct list_head *tmp;
 976         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 977         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 978
 979         do {
 980                 struct ldlm_lock *lock = NULL;
 981
 982                 /* First, we look for non-cleaned-yet lock
 983                  * all cleaned locks are marked by CLEANED flag. */
 984                 lock_res(res);
 985                 list_for_each(tmp, q) {
 986                         lock = list_entry(tmp, struct ldlm_lock,
 987                                           l_res_link);
 988                         if (ldlm_is_cleaned(lock)) {
 989                                 lock = NULL;
 990                                 continue;
 991                         }
 992                         LDLM_LOCK_GET(lock);
 993                         ldlm_set_cleaned(lock);
 994                         break;
 995                 }
 996
 997                 if (lock == NULL) {
 998                         unlock_res(res);
 999                         break;
1000                 }
1001
1002                 /* Set CBPENDING so nothing in the cancellation path
1003                  * can match this lock. */
1004                 ldlm_set_cbpending(lock);
1005                 ldlm_set_failed(lock);
1006                 lock->l_flags |= flags;
1007
1008                 /* ... without sending a CANCEL message for local_only. */
1009                 if (local_only)
1010                         ldlm_set_local_only(lock);
1011
1012                 if (local_only && (lock->l_readers || lock->l_writers)) {
1013                         /*
1014                          * This is a little bit gross, but much better than the
1015                          * alternative: pretend that we got a blocking AST from
1016                          * the server, so that when the lock is decref'd, it
1017                          * will go away ...
1018                          */
1019                         unlock_res(res);
1020                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1021                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1022                                 schedule_timeout_uninterruptible(
1023                                         cfs_time_seconds(4));
1024
1025                         if (lock->l_completion_ast)
1026                                 lock->l_completion_ast(lock,
1027                                                        LDLM_FL_FAILED, NULL);
1028                         LDLM_LOCK_RELEASE(lock);
1029                         continue;
1030                 }
1031
1032                 if (client) {
1033                         struct lustre_handle lockh;
1034
1035                         unlock_res(res);
1036                         ldlm_lock2handle(lock, &lockh);
1037                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1038                         if (rc)
1039                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1040                 } else {
1041                         unlock_res(res);
1042                         LDLM_DEBUG(lock,
1043                                    "Freeing a lock still held by a client node");
1044                         ldlm_lock_cancel(lock);
1045                 }
1046                 LDLM_LOCK_RELEASE(lock);
1047         } while (1);
1048 }
1049
1050 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1051                                struct hlist_node *hnode, void *arg)
1052 {
1053         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1054         __u64 flags = *(__u64 *)arg;
1055
1056         cleanup_resource(res, &res->lr_granted, flags);
1057         cleanup_resource(res, &res->lr_waiting, flags);
1058
1059         return 0;
1060 }
1061
1062 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1063                                   struct hlist_node *hnode, void *arg)
1064 {
1065         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1066
1067         lock_res(res);
1068         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1069                "(%d) after lock cleanup; forcing cleanup.\n",
1070                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1071                atomic_read(&res->lr_refcount) - 1);
1072
1073         /* Use D_NETERROR since it is in the default mask */
1074         ldlm_resource_dump(D_NETERROR, res);
1075         unlock_res(res);
1076         return 0;
1077 }
1078
1079 /**
1080  * Cancel and destroy all locks in the namespace.
1081  *
1082  * Typically used during evictions when server notified client that it was
1083  * evicted and all of its state needs to be destroyed.
1084  * Also used during shutdown.
1085  */
1086 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1087 {
1088         if (ns == NULL) {
1089                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1090                 return ELDLM_OK;
1091         }
1092
1093         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1094                                  &flags, 0);
1095         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1096                                  NULL, 0);
1097         return ELDLM_OK;
1098 }
1099 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1100
1101 /**
1102  * Attempts to free namespace.
1103  *
1104  * Only used when namespace goes away, like during an unmount.
1105  */
1106 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1107 {
1108         ENTRY;
1109
1110         /* At shutdown time, don't call the cancellation callback */
1111         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1112
1113         if (atomic_read(&ns->ns_bref) > 0) {
1114                 int rc;
1115                 CDEBUG(D_DLMTRACE,
1116                        "dlm namespace %s free waiting on refcount %d\n",
1117                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1118 force_wait:
1119                 if (force)
1120                         rc = wait_event_idle_timeout(
1121                                 ns->ns_waitq,
1122                                 atomic_read(&ns->ns_bref) == 0,
1123                                 cfs_time_seconds(1) / 4);
1124                 else
1125                         rc = l_wait_event_abortable(
1126                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1127
1128                 /* Forced cleanups should be able to reclaim all references,
1129                  * so it's safe to wait forever... we can't leak locks... */
1130                 if (force && rc == 0) {
1131                         rc = -ETIMEDOUT;
1132                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1133                                        "namespace with %d resources in use, "
1134                                        "(rc=%d)\n", ldlm_ns_name(ns),
1135                                        atomic_read(&ns->ns_bref), rc);
1136                         GOTO(force_wait, rc);
1137                 }
1138
1139                 if (atomic_read(&ns->ns_bref)) {
1140                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1141                                        "with %d resources in use, (rc=%d)\n",
1142                                        ldlm_ns_name(ns),
1143                                        atomic_read(&ns->ns_bref), rc);
1144                         RETURN(ELDLM_NAMESPACE_EXISTS);
1145                 }
1146                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1147                        ldlm_ns_name(ns));
1148         }
1149
1150         RETURN(ELDLM_OK);
1151 }
1152
1153 /**
1154  * Performs various cleanups for passed \a ns to make it drop refc and be
1155  * ready for freeing. Waits for refc == 0.
1156  *
1157  * The following is done:
1158  * (0) Unregister \a ns from its list to make inaccessible for potential
1159  * users like pools thread and others;
1160  * (1) Clear all locks in \a ns.
1161  */
1162 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1163                                struct obd_import *imp,
1164                                int force)
1165 {
1166         int rc;
1167
1168         ENTRY;
1169         if (!ns) {
1170                 EXIT;
1171                 return;
1172         }
1173
1174         spin_lock(&ns->ns_lock);
1175         ns->ns_stopping = 1;
1176         spin_unlock(&ns->ns_lock);
1177
1178         /*
1179          * Can fail with -EINTR when force == 0 in which case try harder.
1180          */
1181         rc = __ldlm_namespace_free(ns, force);
1182         if (rc != ELDLM_OK) {
1183                 if (imp) {
1184                         ptlrpc_disconnect_import(imp, 0);
1185                         ptlrpc_invalidate_import(imp);
1186                 }
1187
1188                 /*
1189                  * With all requests dropped and the import inactive
1190                  * we are gaurenteed all reference will be dropped.
1191                  */
1192                 rc = __ldlm_namespace_free(ns, 1);
1193                 LASSERT(rc == 0);
1194         }
1195         EXIT;
1196 }
1197 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1198
1199 /**
1200  * Performs freeing memory structures related to \a ns. This is only done
1201  * when ldlm_namespce_free_prior() successfully removed all resources
1202  * referencing \a ns and its refc == 0.
1203  */
1204 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1205 {
1206         ENTRY;
1207         if (!ns) {
1208                 EXIT;
1209                 return;
1210         }
1211
1212         /* Make sure that nobody can find this ns in its list. */
1213         ldlm_namespace_unregister(ns, ns->ns_client);
1214         /* Fini pool _before_ parent proc dir is removed. This is important as
1215          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1216          * Removing it after @dir may cause oops. */
1217         ldlm_pool_fini(&ns->ns_pool);
1218
1219         ldlm_namespace_debugfs_unregister(ns);
1220         ldlm_namespace_sysfs_unregister(ns);
1221         cfs_hash_putref(ns->ns_rs_hash);
1222         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1223         kfree(ns->ns_name);
1224         /* Namespace \a ns should be not on list at this time, otherwise
1225          * this will cause issues related to using freed \a ns in poold
1226          * thread.
1227          */
1228         LASSERT(list_empty(&ns->ns_list_chain));
1229         OBD_FREE_PTR(ns);
1230         ldlm_put_ref();
1231         EXIT;
1232 }
1233 EXPORT_SYMBOL(ldlm_namespace_free_post);
1234
1235 /**
1236  * Cleanup the resource, and free namespace.
1237  * bug 12864:
1238  * Deadlock issue:
1239  * proc1: destroy import
1240  *        class_disconnect_export(grab cl_sem) ->
1241  *              -> ldlm_namespace_free ->
1242  *              -> lprocfs_remove(grab _lprocfs_lock).
1243  * proc2: read proc info
1244  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1245  *              -> osc_rd_active, etc(grab cl_sem).
1246  *
1247  * So that I have to split the ldlm_namespace_free into two parts - the first
1248  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1249  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1250  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1251  * held.
1252  */
1253 void ldlm_namespace_free(struct ldlm_namespace *ns,
1254                          struct obd_import *imp,
1255                          int force)
1256 {
1257         ldlm_namespace_free_prior(ns, imp, force);
1258         ldlm_namespace_free_post(ns);
1259 }
1260 EXPORT_SYMBOL(ldlm_namespace_free);
1261
1262 void ldlm_namespace_get(struct ldlm_namespace *ns)
1263 {
1264         atomic_inc(&ns->ns_bref);
1265 }
1266
1267 /* This is only for callers that care about refcount */
1268 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1269 {
1270         return atomic_inc_return(&ns->ns_bref);
1271 }
1272
1273 void ldlm_namespace_put(struct ldlm_namespace *ns)
1274 {
1275         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1276                 wake_up(&ns->ns_waitq);
1277                 spin_unlock(&ns->ns_lock);
1278         }
1279 }
1280
1281 /** Register \a ns in the list of namespaces */
1282 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1283 {
1284         mutex_lock(ldlm_namespace_lock(client));
1285         LASSERT(list_empty(&ns->ns_list_chain));
1286         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1287         ldlm_namespace_nr_inc(client);
1288         mutex_unlock(ldlm_namespace_lock(client));
1289 }
1290
1291 /** Unregister \a ns from the list of namespaces. */
1292 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1293 {
1294         mutex_lock(ldlm_namespace_lock(client));
1295         LASSERT(!list_empty(&ns->ns_list_chain));
1296         /* Some asserts and possibly other parts of the code are still
1297          * using list_empty(&ns->ns_list_chain). This is why it is
1298          * important to use list_del_init() here. */
1299         list_del_init(&ns->ns_list_chain);
1300         ldlm_namespace_nr_dec(client);
1301         mutex_unlock(ldlm_namespace_lock(client));
1302 }
1303
1304 /** Should be called with ldlm_namespace_lock(client) taken. */
1305 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1306                                           enum ldlm_side client)
1307 {
1308         LASSERT(!list_empty(&ns->ns_list_chain));
1309         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1310         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1311 }
1312
1313 /** Should be called with ldlm_namespace_lock(client) taken. */
1314 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1315                                             enum ldlm_side client)
1316 {
1317         LASSERT(!list_empty(&ns->ns_list_chain));
1318         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1319         list_move_tail(&ns->ns_list_chain,
1320                        ldlm_namespace_inactive_list(client));
1321 }
1322
1323 /** Should be called with ldlm_namespace_lock(client) taken. */
1324 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1325 {
1326         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1327         LASSERT(!list_empty(ldlm_namespace_list(client)));
1328         return container_of(ldlm_namespace_list(client)->next,
1329                             struct ldlm_namespace, ns_list_chain);
1330 }
1331
1332 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1333 {
1334         int idx;
1335
1336         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1337                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1338         if (res->lr_itree == NULL)
1339                 return false;
1340         /* Initialize interval trees for each lock mode. */
1341         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1342                 res->lr_itree[idx].lit_size = 0;
1343                 res->lr_itree[idx].lit_mode = 1 << idx;
1344                 res->lr_itree[idx].lit_root = NULL;
1345         }
1346         return true;
1347 }
1348
1349 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1350 {
1351         int i;
1352
1353         OBD_ALLOC_PTR(res->lr_ibits_queues);
1354         if (res->lr_ibits_queues == NULL)
1355                 return false;
1356         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1357                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1358         return true;
1359 }
1360
1361 /** Create and initialize new resource. */
1362 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1363 {
1364         struct ldlm_resource *res;
1365         bool rc;
1366
1367         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1368         if (res == NULL)
1369                 return NULL;
1370
1371         switch (ldlm_type) {
1372         case LDLM_EXTENT:
1373                 rc = ldlm_resource_extent_new(res);
1374                 break;
1375         case LDLM_IBITS:
1376                 rc = ldlm_resource_inodebits_new(res);
1377                 break;
1378         default:
1379                 rc = true;
1380                 break;
1381         }
1382         if (!rc) {
1383                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1384                 return NULL;
1385         }
1386
1387         INIT_LIST_HEAD(&res->lr_granted);
1388         INIT_LIST_HEAD(&res->lr_waiting);
1389
1390         atomic_set(&res->lr_refcount, 1);
1391         spin_lock_init(&res->lr_lock);
1392         lu_ref_init(&res->lr_reference);
1393
1394         /* Since LVB init can be delayed now, there is no longer need to
1395          * immediatelly acquire mutex here. */
1396         mutex_init(&res->lr_lvb_mutex);
1397         res->lr_lvb_initialized = false;
1398
1399         return res;
1400 }
1401
1402 static void ldlm_resource_free(struct ldlm_resource *res)
1403 {
1404         if (res->lr_type == LDLM_EXTENT) {
1405                 if (res->lr_itree != NULL)
1406                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1407                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1408         } else if (res->lr_type == LDLM_IBITS) {
1409                 if (res->lr_ibits_queues != NULL)
1410                         OBD_FREE_PTR(res->lr_ibits_queues);
1411         }
1412
1413         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1414 }
1415
1416 /**
1417  * Return a reference to resource with given name, creating it if necessary.
1418  * Args: namespace with ns_lock unlocked
1419  * Locks: takes and releases NS hash-lock and res->lr_lock
1420  * Returns: referenced, unlocked ldlm_resource or NULL
1421  */
1422 struct ldlm_resource *
1423 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1424                   const struct ldlm_res_id *name, enum ldlm_type type,
1425                   int create)
1426 {
1427         struct hlist_node       *hnode;
1428         struct ldlm_resource    *res = NULL;
1429         struct cfs_hash_bd              bd;
1430         __u64                   version;
1431         int                     ns_refcount = 0;
1432         int hash;
1433
1434         LASSERT(ns != NULL);
1435         LASSERT(parent == NULL);
1436         LASSERT(ns->ns_rs_hash != NULL);
1437         LASSERT(name->name[0] != 0);
1438
1439         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1440         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1441         if (hnode != NULL) {
1442                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1443                 GOTO(found, res);
1444         }
1445
1446         version = cfs_hash_bd_version_get(&bd);
1447         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1448
1449         if (create == 0)
1450                 return ERR_PTR(-ENOENT);
1451
1452         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1453                  "type: %d\n", type);
1454         res = ldlm_resource_new(type);
1455         if (res == NULL)
1456                 return ERR_PTR(-ENOMEM);
1457
1458         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1459         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1460         res->lr_name = *name;
1461         res->lr_type = type;
1462
1463         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1464         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1465                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1466
1467         if (hnode != NULL) {
1468                 /* Someone won the race and already added the resource. */
1469                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1470                 /* Clean lu_ref for failed resource. */
1471                 lu_ref_fini(&res->lr_reference);
1472                 ldlm_resource_free(res);
1473 found:
1474                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1475                 return res;
1476         }
1477         /* We won! Let's add the resource. */
1478         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1479         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1480                 ns_refcount = ldlm_namespace_get_return(ns);
1481
1482         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1483
1484         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1485
1486         /* Let's see if we happened to be the very first resource in this
1487          * namespace. If so, and this is a client namespace, we need to move
1488          * the namespace into the active namespaces list to be patrolled by
1489          * the ldlm_poold. */
1490         if (ns_is_client(ns) && ns_refcount == 1) {
1491                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1492                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1493                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1494         }
1495
1496         return res;
1497 }
1498 EXPORT_SYMBOL(ldlm_resource_get);
1499
1500 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1501 {
1502         LASSERT(res != NULL);
1503         LASSERT(res != LP_POISON);
1504         atomic_inc(&res->lr_refcount);
1505         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1506                atomic_read(&res->lr_refcount));
1507         return res;
1508 }
1509
1510 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1511                                          struct ldlm_resource *res)
1512 {
1513         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1514
1515         if (!list_empty(&res->lr_granted)) {
1516                 ldlm_resource_dump(D_ERROR, res);
1517                 LBUG();
1518         }
1519
1520         if (!list_empty(&res->lr_waiting)) {
1521                 ldlm_resource_dump(D_ERROR, res);
1522                 LBUG();
1523         }
1524
1525         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1526                                bd, &res->lr_hash);
1527         lu_ref_fini(&res->lr_reference);
1528         if (atomic_dec_and_test(&nsb->nsb_count))
1529                 ldlm_namespace_put(nsb->nsb_namespace);
1530 }
1531
1532 /* Returns 1 if the resource was freed, 0 if it remains. */
1533 int ldlm_resource_putref(struct ldlm_resource *res)
1534 {
1535         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1536         struct cfs_hash_bd   bd;
1537
1538         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1539         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1540                res, atomic_read(&res->lr_refcount) - 1);
1541
1542         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1543         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1544                 __ldlm_resource_putref_final(&bd, res);
1545                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1546                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1547                         ns->ns_lvbo->lvbo_free(res);
1548                 ldlm_resource_free(res);
1549                 return 1;
1550         }
1551         return 0;
1552 }
1553 EXPORT_SYMBOL(ldlm_resource_putref);
1554
1555 /**
1556  * Add a lock into a given resource into specified lock list.
1557  */
1558 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1559                             struct ldlm_lock *lock)
1560 {
1561         check_res_locked(res);
1562
1563         LDLM_DEBUG(lock, "About to add this lock");
1564
1565         if (ldlm_is_destroyed(lock)) {
1566                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1567                 return;
1568         }
1569
1570         LASSERT(list_empty(&lock->l_res_link));
1571
1572         list_add_tail(&lock->l_res_link, head);
1573
1574         if (res->lr_type == LDLM_IBITS)
1575                 ldlm_inodebits_add_lock(res, head, lock);
1576 }
1577
1578 /**
1579  * Insert a lock into resource after specified lock.
1580  *
1581  * Obtain resource description from the lock we are inserting after.
1582  */
1583 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1584                                      struct ldlm_lock *new)
1585 {
1586         struct ldlm_resource *res = original->l_resource;
1587
1588         check_res_locked(res);
1589
1590         ldlm_resource_dump(D_INFO, res);
1591         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1592
1593         if (ldlm_is_destroyed(new)) {
1594                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1595                 goto out;
1596         }
1597
1598         LASSERT(list_empty(&new->l_res_link));
1599
1600         list_add(&new->l_res_link, &original->l_res_link);
1601  out:;
1602 }
1603
1604 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1605 {
1606         int type = lock->l_resource->lr_type;
1607
1608         check_res_locked(lock->l_resource);
1609         switch (type) {
1610         case LDLM_PLAIN:
1611                 ldlm_unlink_lock_skiplist(lock);
1612                 break;
1613         case LDLM_EXTENT:
1614                 ldlm_extent_unlink_lock(lock);
1615                 break;
1616         case LDLM_IBITS:
1617                 ldlm_inodebits_unlink_lock(lock);
1618                 break;
1619         }
1620         list_del_init(&lock->l_res_link);
1621 }
1622 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1623
1624 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1625 {
1626         desc->lr_type = res->lr_type;
1627         desc->lr_name = res->lr_name;
1628 }
1629
1630 /**
1631  * Print information about all locks in all namespaces on this node to debug
1632  * log.
1633  */
1634 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1635 {
1636         struct list_head *tmp;
1637
1638         if (!((libcfs_debug | D_ERROR) & level))
1639                 return;
1640
1641         mutex_lock(ldlm_namespace_lock(client));
1642
1643         list_for_each(tmp, ldlm_namespace_list(client)) {
1644                 struct ldlm_namespace *ns;
1645
1646                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1647                 ldlm_namespace_dump(level, ns);
1648         }
1649
1650         mutex_unlock(ldlm_namespace_lock(client));
1651 }
1652
1653 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1654                               struct hlist_node *hnode, void *arg)
1655 {
1656         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1657         int    level = (int)(unsigned long)arg;
1658
1659         lock_res(res);
1660         ldlm_resource_dump(level, res);
1661         unlock_res(res);
1662
1663         return 0;
1664 }
1665
1666 /**
1667  * Print information about all locks in this namespace on this node to debug
1668  * log.
1669  */
1670 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1671 {
1672         if (!((libcfs_debug | D_ERROR) & level))
1673                 return;
1674
1675         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1676                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1677                ns_is_client(ns) ? "client" : "server");
1678
1679         if (ktime_get_seconds() < ns->ns_next_dump)
1680                 return;
1681
1682         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1683                                  ldlm_res_hash_dump,
1684                                  (void *)(unsigned long)level, 0);
1685         spin_lock(&ns->ns_lock);
1686         ns->ns_next_dump = ktime_get_seconds() + 10;
1687         spin_unlock(&ns->ns_lock);
1688 }
1689
1690 /**
1691  * Print information about all locks in this resource to debug log.
1692  */
1693 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1694 {
1695         struct ldlm_lock *lock;
1696         unsigned int granted = 0;
1697
1698         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1699
1700         if (!((libcfs_debug | D_ERROR) & level))
1701                 return;
1702
1703         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1704                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1705
1706         if (!list_empty(&res->lr_granted)) {
1707                 CDEBUG(level, "Granted locks (in reverse order):\n");
1708                 list_for_each_entry_reverse(lock, &res->lr_granted,
1709                                                 l_res_link) {
1710                         LDLM_DEBUG_LIMIT(level, lock, "###");
1711                         if (!(level & D_CANTMASK) &&
1712                             ++granted > ldlm_dump_granted_max) {
1713                                 CDEBUG(level,
1714                                        "only dump %d granted locks to avoid DDOS.\n",
1715                                        granted);
1716                                 break;
1717                         }
1718                 }
1719         }
1720
1721         if (!list_empty(&res->lr_waiting)) {
1722                 CDEBUG(level, "Waiting locks:\n");
1723                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1724                         LDLM_DEBUG_LIMIT(level, lock, "###");
1725         }
1726 }
1727 EXPORT_SYMBOL(ldlm_resource_dump);