lustre/ldlm/ldlm_resource.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46 struct kmem_cache *ldlm_inodebits_slab;
  47
  48 int ldlm_srv_namespace_nr = 0;
  49 int ldlm_cli_namespace_nr = 0;
  50
  51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  52 LIST_HEAD(ldlm_srv_namespace_list);
  53
  54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
  55 /* Client Namespaces that have active resources in them.
  56  * Once all resources go away, ldlm_poold moves such namespaces to the
  57  * inactive list */
  58 LIST_HEAD(ldlm_cli_active_namespace_list);
  59 /* Client namespaces that don't have any locks in them */
  60 LIST_HEAD(ldlm_cli_inactive_namespace_list);
  61
  62 static struct dentry *ldlm_debugfs_dir;
  63 static struct dentry *ldlm_ns_debugfs_dir;
  64 struct dentry *ldlm_svc_debugfs_dir;
  65
  66 /* during debug dump certain amount of granted locks for one resource to avoid
  67  * DDOS. */
  68 static unsigned int ldlm_dump_granted_max = 256;
  69
  70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
  71                                           const char __user *buffer,
  72                                           size_t count, loff_t *off)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78
  79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
  80
  81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  82 {
  83         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  84         return 0;
  85 }
  86
  87 static ssize_t
  88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  89                        size_t count, loff_t *off)
  90 {
  91         struct seq_file *seq = file->private_data;
  92
  93         if (!count)
  94                 return 0;
  95
  96         return kstrtouint_from_user(buffer, count, 0,
  97                                     (unsigned int *)seq->private);
  98 }
  99
 100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
 101
 102 #ifdef HAVE_SERVER_SUPPORT
 103
 104 static int seq_watermark_show(struct seq_file *m, void *data)
 105 {
 106         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 107         return 0;
 108 }
 109
 110 static ssize_t seq_watermark_write(struct file *file,
 111                                    const char __user *buffer, size_t count,
 112                                    loff_t *off)
 113 {
 114         struct seq_file *m = file->private_data;
 115         u64 value;
 116         __u64 watermark;
 117         __u64 *data = m->private;
 118         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 119         char kernbuf[22] = "";
 120         int rc;
 121
 122         if (count >= sizeof(kernbuf))
 123                 return -EINVAL;
 124
 125         if (copy_from_user(kernbuf, buffer, count))
 126                 return -EFAULT;
 127         kernbuf[count] = 0;
 128
 129         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
 130         if (rc < 0) {
 131                 CERROR("Failed to set %s, rc = %d.\n",
 132                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 133                        rc);
 134                 return rc;
 135         } else if (value != 0 && value < (1 << 20)) {
 136                 CERROR("%s should be greater than 1MB.\n",
 137                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 138                 return -EINVAL;
 139         }
 140         watermark = value >> 20;
 141
 142         if (wm_low) {
 143                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 144                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 145                                "lock_limit_mb.\n");
 146                         return -EINVAL;
 147                 }
 148
 149                 *data = watermark;
 150                 if (watermark != 0) {
 151                         watermark <<= 20;
 152                         do_div(watermark, sizeof(struct ldlm_lock));
 153                 }
 154                 ldlm_reclaim_threshold = watermark;
 155         } else {
 156                 if (ldlm_reclaim_threshold_mb != 0 &&
 157                     watermark < ldlm_reclaim_threshold_mb) {
 158                         CERROR("lock_limit_mb must be greater than "
 159                                "lock_reclaim_threshold_mb.\n");
 160                         return -EINVAL;
 161                 }
 162
 163                 *data = watermark;
 164                 if (watermark != 0) {
 165                         watermark <<= 20;
 166                         do_div(watermark, sizeof(struct ldlm_lock));
 167                 }
 168                 ldlm_lock_limit = watermark;
 169         }
 170
 171         return count;
 172 }
 173
 174 static int seq_watermark_open(struct inode *inode, struct file *file)
 175 {
 176         return single_open(file, seq_watermark_show, inode->i_private);
 177 }
 178
 179 static const struct file_operations ldlm_watermark_fops = {
 180         .owner          = THIS_MODULE,
 181         .open           = seq_watermark_open,
 182         .read           = seq_read,
 183         .write          = seq_watermark_write,
 184         .llseek         = seq_lseek,
 185         .release        = lprocfs_single_release,
 186 };
 187
 188 static int seq_granted_show(struct seq_file *m, void *data)
 189 {
 190         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 191                    (struct percpu_counter *)m->private));
 192         return 0;
 193 }
 194
 195 static int seq_granted_open(struct inode *inode, struct file *file)
 196 {
 197         return single_open(file, seq_granted_show, inode->i_private);
 198 }
 199
 200 static const struct file_operations ldlm_granted_fops = {
 201         .owner  = THIS_MODULE,
 202         .open   = seq_granted_open,
 203         .read   = seq_read,
 204         .llseek = seq_lseek,
 205         .release = seq_release,
 206 };
 207
 208 #endif /* HAVE_SERVER_SUPPORT */
 209
 210 static struct ldebugfs_vars ldlm_debugfs_list[] = {
 211         { .name =       "dump_namespaces",
 212           .fops =       &ldlm_dump_ns_fops,
 213           .proc_mode =  0222 },
 214         { .name =       "dump_granted_max",
 215           .fops =       &ldlm_rw_uint_fops,
 216           .data =       &ldlm_dump_granted_max },
 217 #ifdef HAVE_SERVER_SUPPORT
 218         { .name =       "lock_reclaim_threshold_mb",
 219           .fops =       &ldlm_watermark_fops,
 220           .data =       &ldlm_reclaim_threshold_mb },
 221         { .name =       "lock_limit_mb",
 222           .fops =       &ldlm_watermark_fops,
 223           .data =       &ldlm_lock_limit_mb },
 224         { .name =       "lock_granted_count",
 225           .fops =       &ldlm_granted_fops,
 226           .data =       &ldlm_granted_total },
 227 #endif
 228         { NULL }
 229 };
 230
 231 int ldlm_debugfs_setup(void)
 232 {
 233         ENTRY;
 234         ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
 235                                              debugfs_lustre_root);
 236         ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
 237                                                  ldlm_debugfs_dir);
 238         ldlm_svc_debugfs_dir = debugfs_create_dir("services",
 239                                                   ldlm_debugfs_dir);
 240
 241         ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 242
 243         RETURN(0);
 244 }
 245
 246 void ldlm_debugfs_cleanup(void)
 247 {
 248         debugfs_remove_recursive(ldlm_debugfs_dir);
 249
 250         ldlm_svc_debugfs_dir = NULL;
 251         ldlm_ns_debugfs_dir = NULL;
 252         ldlm_debugfs_dir = NULL;
 253 }
 254
 255 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 256                                    char *buf)
 257 {
 258         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 259                                                  ns_kobj);
 260         __u64                   res = 0;
 261         int                     i;
 262
 263         /* result is not strictly consistant */
 264         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
 265                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
 266         return sprintf(buf, "%lld\n", res);
 267 }
 268 LUSTRE_RO_ATTR(resource_count);
 269
 270 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 271                                char *buf)
 272 {
 273         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 274                                                  ns_kobj);
 275         __u64                   locks;
 276
 277         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 278                                         LPROCFS_FIELDS_FLAGS_SUM);
 279         return sprintf(buf, "%lld\n", locks);
 280 }
 281 LUSTRE_RO_ATTR(lock_count);
 282
 283 static ssize_t lock_unused_count_show(struct kobject *kobj,
 284                                       struct attribute *attr,
 285                                       char *buf)
 286 {
 287         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 288                                                  ns_kobj);
 289
 290         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 291 }
 292 LUSTRE_RO_ATTR(lock_unused_count);
 293
 294 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 295                              char *buf)
 296 {
 297         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 298                                                  ns_kobj);
 299         __u32 *nr = &ns->ns_max_unused;
 300
 301         if (ns_connect_lru_resize(ns))
 302                 nr = &ns->ns_nr_unused;
 303         return sprintf(buf, "%u\n", *nr);
 304 }
 305
 306 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 307                               const char *buffer, size_t count)
 308 {
 309         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 310                                                  ns_kobj);
 311         unsigned long tmp;
 312         int lru_resize;
 313         int err;
 314
 315         if (strncmp(buffer, "clear", 5) == 0) {
 316                 CDEBUG(D_DLMTRACE,
 317                        "dropping all unused locks from namespace %s\n",
 318                        ldlm_ns_name(ns));
 319                 /* Try to cancel all @ns_nr_unused locks. */
 320                 ldlm_cancel_lru(ns, INT_MAX, 0, LDLM_LRU_FLAG_CLEANUP);
 321                 return count;
 322         }
 323
 324         err = kstrtoul(buffer, 10, &tmp);
 325         if (err != 0) {
 326                 CERROR("lru_size: invalid value written\n");
 327                 return -EINVAL;
 328         }
 329         lru_resize = (tmp == 0);
 330
 331         if (ns_connect_lru_resize(ns)) {
 332                 if (!lru_resize)
 333                         ns->ns_max_unused = (unsigned int)tmp;
 334
 335                 if (tmp > ns->ns_nr_unused)
 336                         tmp = ns->ns_nr_unused;
 337                 tmp = ns->ns_nr_unused - tmp;
 338
 339                 CDEBUG(D_DLMTRACE,
 340                        "changing namespace %s unused locks from %u to %u\n",
 341                        ldlm_ns_name(ns), ns->ns_nr_unused,
 342                        (unsigned int)tmp);
 343
 344                 if (!lru_resize) {
 345                         CDEBUG(D_DLMTRACE,
 346                                "disable lru_resize for namespace %s\n",
 347                                ldlm_ns_name(ns));
 348                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 349                 }
 350                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, 0);
 351         } else {
 352                 CDEBUG(D_DLMTRACE,
 353                        "changing namespace %s max_unused from %u to %u\n",
 354                        ldlm_ns_name(ns), ns->ns_max_unused,
 355                        (unsigned int)tmp);
 356
 357                 /* Make sure that LRU resize was originally supported before
 358                  * turning it on here.
 359                  */
 360                 if (lru_resize &&
 361                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 362                         CDEBUG(D_DLMTRACE,
 363                                "enable lru_resize for namespace %s\n",
 364                                ldlm_ns_name(ns));
 365                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 366                 }
 367                 ns->ns_max_unused = (unsigned int)tmp;
 368                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
 369         }
 370
 371         return count;
 372 }
 373 LUSTRE_RW_ATTR(lru_size);
 374
 375 static ssize_t lru_cancel_batch_show(struct kobject *kobj,
 376                                  struct attribute *attr, char *buf)
 377 {
 378         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 379                                                  ns_kobj);
 380
 381         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_cancel_batch);
 382 }
 383
 384 static ssize_t lru_cancel_batch_store(struct kobject *kobj,
 385                                   struct attribute *attr,
 386                                   const char *buffer, size_t count)
 387 {
 388         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 389                                                  ns_kobj);
 390         unsigned long tmp;
 391
 392         if (kstrtoul(buffer, 10, &tmp))
 393                 return -EINVAL;
 394
 395         ns->ns_cancel_batch = (unsigned int)tmp;
 396
 397         return count;
 398 }
 399 LUSTRE_RW_ATTR(lru_cancel_batch);
 400
 401 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 402                                 char *buf)
 403 {
 404         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 405                                                  ns_kobj);
 406
 407         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 408 }
 409
 410 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 411                                  const char *buffer, size_t count)
 412 {
 413         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 414                                                  ns_kobj);
 415         int scale = NSEC_PER_MSEC;
 416         unsigned long long tmp;
 417         char *buf;
 418
 419         /* Did the user ask in seconds or milliseconds. Default is in ms */
 420         buf = strstr(buffer, "ms");
 421         if (!buf) {
 422                 buf = strchr(buffer, 's');
 423                 if (buf)
 424                         scale = NSEC_PER_SEC;
 425         }
 426
 427         if (buf)
 428                 *buf = '\0';
 429
 430         if (kstrtoull(buffer, 10, &tmp))
 431                 return -EINVAL;
 432
 433         ns->ns_max_age = ktime_set(0, tmp * scale);
 434
 435         return count;
 436 }
 437 LUSTRE_RW_ATTR(lru_max_age);
 438
 439 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 440                                       struct attribute *attr,
 441                                       char *buf)
 442 {
 443         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 444                                                  ns_kobj);
 445
 446         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 447 }
 448
 449 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 450                                        struct attribute *attr,
 451                                        const char *buffer,
 452                                        size_t count)
 453 {
 454         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 455                                                  ns_kobj);
 456         unsigned long supp = -1;
 457         int rc;
 458
 459         rc = kstrtoul(buffer, 10, &supp);
 460         if (rc < 0)
 461                 return rc;
 462
 463         if (supp == 0)
 464                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 465         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 466                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 467         return count;
 468 }
 469 LUSTRE_RW_ATTR(early_lock_cancel);
 470
 471 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 472                                     struct attribute *attr, char *buf)
 473 {
 474         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 475                                                  ns_kobj);
 476
 477         return snprintf(buf, PAGE_SIZE, "%llu\n",
 478                         ktime_divns(ns->ns_dirty_age_limit, NSEC_PER_SEC));
 479 }
 480
 481 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 482                                      struct attribute *attr,
 483                                      const char *buffer, size_t count)
 484 {
 485         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 486                                                  ns_kobj);
 487         unsigned long long tmp;
 488
 489         if (kstrtoull(buffer, 10, &tmp))
 490                 return -EINVAL;
 491
 492         ns->ns_dirty_age_limit = ktime_set(tmp, 0);
 493
 494         return count;
 495 }
 496 LUSTRE_RW_ATTR(dirty_age_limit);
 497
 498 #ifdef HAVE_SERVER_SUPPORT
 499 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 500                                     struct attribute *attr, char *buf)
 501 {
 502         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 503                                                  ns_kobj);
 504
 505         return snprintf(buf, PAGE_SIZE, "%u\n", ns->ns_ctime_age_limit);
 506 }
 507
 508 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 509                                      struct attribute *attr,
 510                                      const char *buffer, size_t count)
 511 {
 512         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 513                                                  ns_kobj);
 514         unsigned long tmp;
 515
 516         if (kstrtoul(buffer, 10, &tmp))
 517                 return -EINVAL;
 518
 519         ns->ns_ctime_age_limit = tmp;
 520
 521         return count;
 522 }
 523 LUSTRE_RW_ATTR(ctime_age_limit);
 524
 525 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 526                                   char *buf)
 527 {
 528         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 529                                                  ns_kobj);
 530
 531         return sprintf(buf, "%d\n", ns->ns_timeouts);
 532 }
 533 LUSTRE_RO_ATTR(lock_timeouts);
 534
 535 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 536                                      struct attribute *attr, char *buf)
 537 {
 538         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 539                                                  ns_kobj);
 540
 541         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 542 }
 543
 544 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 545                                       struct attribute *attr,
 546                                       const char *buffer, size_t count)
 547 {
 548         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 549                                                  ns_kobj);
 550         unsigned long tmp;
 551         int err;
 552
 553         err = kstrtoul(buffer, 10, &tmp);
 554         if (err != 0)
 555                 return -EINVAL;
 556
 557         ns->ns_max_nolock_size = tmp;
 558
 559         return count;
 560 }
 561 LUSTRE_RW_ATTR(max_nolock_bytes);
 562
 563 static ssize_t contention_seconds_show(struct kobject *kobj,
 564                                        struct attribute *attr, char *buf)
 565 {
 566         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 567                                                  ns_kobj);
 568
 569         return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
 570 }
 571
 572 static ssize_t contention_seconds_store(struct kobject *kobj,
 573                                         struct attribute *attr,
 574                                         const char *buffer, size_t count)
 575 {
 576         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 577                                                  ns_kobj);
 578         unsigned int tmp;
 579
 580         if (kstrtouint(buffer, 10, &tmp))
 581                 return -EINVAL;
 582
 583         ns->ns_contention_time = tmp;
 584
 585         return count;
 586 }
 587 LUSTRE_RW_ATTR(contention_seconds);
 588
 589 static ssize_t contended_locks_show(struct kobject *kobj,
 590                                     struct attribute *attr, char *buf)
 591 {
 592         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 593                                                  ns_kobj);
 594
 595         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 596 }
 597
 598 static ssize_t contended_locks_store(struct kobject *kobj,
 599                                      struct attribute *attr,
 600                                      const char *buffer, size_t count)
 601 {
 602         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 603                                                  ns_kobj);
 604         unsigned long tmp;
 605         int err;
 606
 607         err = kstrtoul(buffer, 10, &tmp);
 608         if (err != 0)
 609                 return -EINVAL;
 610
 611         ns->ns_contended_locks = tmp;
 612
 613         return count;
 614 }
 615 LUSTRE_RW_ATTR(contended_locks);
 616
 617 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 618                                      struct attribute *attr, char *buf)
 619 {
 620         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 621                                                  ns_kobj);
 622
 623         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 624 }
 625
 626 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 627                                       struct attribute *attr,
 628                                       const char *buffer, size_t count)
 629 {
 630         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 631                                                  ns_kobj);
 632         unsigned long tmp;
 633         int err;
 634
 635         err = kstrtoul(buffer, 10, &tmp);
 636         if (err != 0)
 637                 return -EINVAL;
 638
 639         ns->ns_max_parallel_ast = tmp;
 640
 641         return count;
 642 }
 643 LUSTRE_RW_ATTR(max_parallel_ast);
 644
 645 #endif /* HAVE_SERVER_SUPPORT */
 646
 647 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 648 static struct attribute *ldlm_ns_attrs[] = {
 649         &lustre_attr_resource_count.attr,
 650         &lustre_attr_lock_count.attr,
 651         &lustre_attr_lock_unused_count.attr,
 652         &lustre_attr_lru_size.attr,
 653         &lustre_attr_lru_cancel_batch.attr,
 654         &lustre_attr_lru_max_age.attr,
 655         &lustre_attr_early_lock_cancel.attr,
 656         &lustre_attr_dirty_age_limit.attr,
 657 #ifdef HAVE_SERVER_SUPPORT
 658         &lustre_attr_ctime_age_limit.attr,
 659         &lustre_attr_lock_timeouts.attr,
 660         &lustre_attr_max_nolock_bytes.attr,
 661         &lustre_attr_contention_seconds.attr,
 662         &lustre_attr_contended_locks.attr,
 663         &lustre_attr_max_parallel_ast.attr,
 664 #endif
 665         NULL,
 666 };
 667
 668 static void ldlm_ns_release(struct kobject *kobj)
 669 {
 670         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 671                                                  ns_kobj);
 672         complete(&ns->ns_kobj_unregister);
 673 }
 674
 675 static struct kobj_type ldlm_ns_ktype = {
 676         .default_attrs  = ldlm_ns_attrs,
 677         .sysfs_ops      = &lustre_sysfs_ops,
 678         .release        = ldlm_ns_release,
 679 };
 680
 681 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 682 {
 683         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 684                 CERROR("dlm namespace %s has no procfs dir?\n",
 685                        ldlm_ns_name(ns));
 686         else
 687                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 688
 689         if (ns->ns_stats != NULL)
 690                 lprocfs_free_stats(&ns->ns_stats);
 691 }
 692
 693 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 694 {
 695         kobject_put(&ns->ns_kobj);
 696         wait_for_completion(&ns->ns_kobj_unregister);
 697 }
 698
 699 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 700 {
 701         int err;
 702
 703         ns->ns_kobj.kset = ldlm_ns_kset;
 704         init_completion(&ns->ns_kobj_unregister);
 705         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 706                                    "%s", ldlm_ns_name(ns));
 707
 708         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 709         if (!ns->ns_stats) {
 710                 kobject_put(&ns->ns_kobj);
 711                 return -ENOMEM;
 712         }
 713
 714         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 715                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 716
 717         return err;
 718 }
 719
 720 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 721 {
 722         struct dentry *ns_entry;
 723
 724         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 725                 ns_entry = ns->ns_debugfs_entry;
 726         } else {
 727                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 728                                               ldlm_ns_debugfs_dir);
 729                 if (!ns_entry)
 730                         return -ENOMEM;
 731                 ns->ns_debugfs_entry = ns_entry;
 732         }
 733
 734         return 0;
 735 }
 736 #undef MAX_STRING_SIZE
 737
 738 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 739                                   const void *key, unsigned int mask)
 740 {
 741         const struct ldlm_res_id *id = key;
 742         unsigned int val = 0;
 743         unsigned int i;
 744
 745         for (i = 0; i < RES_NAME_SIZE; i++)
 746                 val += id->name[i];
 747         return val & mask;
 748 }
 749
 750 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 751 {
 752         struct lu_fid       fid;
 753         __u32               hash;
 754         __u32               val;
 755
 756         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 757         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 758         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 759
 760         hash = fid_flatten32(&fid);
 761         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 762         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 763                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 764         } else {
 765                 val = fid_oid(&fid);
 766         }
 767         hash += (val >> 5) + (val << 11);
 768         return cfs_hash_32(hash, bits);
 769 }
 770
 771 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 772 {
 773         struct ldlm_resource   *res;
 774
 775         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 776         return &res->lr_name;
 777 }
 778
 779 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 780 {
 781         struct ldlm_resource   *res;
 782
 783         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 784         return ldlm_res_eq((const struct ldlm_res_id *)key,
 785                            (const struct ldlm_res_id *)&res->lr_name);
 786 }
 787
 788 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 789 {
 790         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 791 }
 792
 793 static void
 794 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 795 {
 796         struct ldlm_resource *res;
 797
 798         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 799         ldlm_resource_getref(res);
 800 }
 801
 802 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 803 {
 804         struct ldlm_resource *res;
 805
 806         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 807         ldlm_resource_putref(res);
 808 }
 809
 810 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 811         .hs_hash        = ldlm_res_hop_hash,
 812         .hs_key         = ldlm_res_hop_key,
 813         .hs_keycmp      = ldlm_res_hop_keycmp,
 814         .hs_keycpy      = NULL,
 815         .hs_object      = ldlm_res_hop_object,
 816         .hs_get         = ldlm_res_hop_get_locked,
 817         .hs_put         = ldlm_res_hop_put
 818 };
 819
 820 static struct {
 821         /** hash bucket bits */
 822         unsigned                nsd_bkt_bits;
 823         /** hash bits */
 824         unsigned                nsd_all_bits;
 825 } ldlm_ns_hash_defs[] = {
 826         [LDLM_NS_TYPE_MDC] = {
 827                 .nsd_bkt_bits   = 11,
 828                 .nsd_all_bits   = 16,
 829         },
 830         [LDLM_NS_TYPE_MDT] = {
 831                 .nsd_bkt_bits   = 14,
 832                 .nsd_all_bits   = 21,
 833         },
 834         [LDLM_NS_TYPE_OSC] = {
 835                 .nsd_bkt_bits   = 8,
 836                 .nsd_all_bits   = 12,
 837         },
 838         [LDLM_NS_TYPE_OST] = {
 839                 .nsd_bkt_bits   = 11,
 840                 .nsd_all_bits   = 17,
 841         },
 842         [LDLM_NS_TYPE_MGC] = {
 843                 .nsd_bkt_bits   = 3,
 844                 .nsd_all_bits   = 4,
 845         },
 846         [LDLM_NS_TYPE_MGT] = {
 847                 .nsd_bkt_bits   = 3,
 848                 .nsd_all_bits   = 4,
 849         },
 850 };
 851
 852 /**
 853  * Create and initialize new empty namespace.
 854  */
 855 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 856                                           enum ldlm_side client,
 857                                           enum ldlm_appetite apt,
 858                                           enum ldlm_ns_type ns_type)
 859 {
 860         struct ldlm_namespace *ns = NULL;
 861         int idx;
 862         int rc;
 863
 864         ENTRY;
 865         LASSERT(obd != NULL);
 866
 867         rc = ldlm_get_ref();
 868         if (rc) {
 869                 CERROR("ldlm_get_ref failed: %d\n", rc);
 870                 RETURN(NULL);
 871         }
 872
 873         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 874             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 875                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 876                 GOTO(out_ref, NULL);
 877         }
 878
 879         OBD_ALLOC_PTR(ns);
 880         if (!ns)
 881                 GOTO(out_ref, NULL);
 882
 883         ns->ns_rs_hash = cfs_hash_create(name,
 884                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 885                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 886                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 887                                          0,
 888                                          CFS_HASH_MIN_THETA,
 889                                          CFS_HASH_MAX_THETA,
 890                                          &ldlm_ns_hash_ops,
 891                                          CFS_HASH_DEPTH |
 892                                          CFS_HASH_BIGNAME |
 893                                          CFS_HASH_SPIN_BKTLOCK |
 894                                          CFS_HASH_NO_ITEMREF);
 895         if (ns->ns_rs_hash == NULL)
 896                 GOTO(out_ns, NULL);
 897
 898         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 899                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 900
 901         OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
 902         if (!ns->ns_rs_buckets)
 903                 goto out_hash;
 904
 905         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 906                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 907
 908                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 909                 nsb->nsb_namespace = ns;
 910                 nsb->nsb_reclaim_start = 0;
 911                 atomic_set(&nsb->nsb_count, 0);
 912         }
 913
 914         ns->ns_obd = obd;
 915         ns->ns_appetite = apt;
 916         ns->ns_client = client;
 917         ns->ns_name = kstrdup(name, GFP_KERNEL);
 918         if (!ns->ns_name)
 919                 goto out_hash;
 920
 921         INIT_LIST_HEAD(&ns->ns_list_chain);
 922         INIT_LIST_HEAD(&ns->ns_unused_list);
 923         spin_lock_init(&ns->ns_lock);
 924         atomic_set(&ns->ns_bref, 0);
 925         init_waitqueue_head(&ns->ns_waitq);
 926
 927         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 928         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 929         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 930
 931         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 932         ns->ns_nr_unused          = 0;
 933         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 934         ns->ns_cancel_batch       = LDLM_DEFAULT_LRU_SHRINK_BATCH;
 935         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 936         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 937         ns->ns_dirty_age_limit    = ktime_set(LDLM_DIRTY_AGE_LIMIT, 0);
 938         ns->ns_timeouts           = 0;
 939         ns->ns_orig_connect_flags = 0;
 940         ns->ns_connect_flags      = 0;
 941         ns->ns_stopping           = 0;
 942         ns->ns_reclaim_start      = 0;
 943         ns->ns_last_pos           = &ns->ns_unused_list;
 944         ns->ns_flags              = 0;
 945
 946         rc = ldlm_namespace_sysfs_register(ns);
 947         if (rc) {
 948                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 949                 GOTO(out_hash, rc);
 950         }
 951
 952         rc = ldlm_namespace_debugfs_register(ns);
 953         if (rc) {
 954                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 955                 GOTO(out_sysfs, rc);
 956         }
 957
 958         idx = ldlm_namespace_nr_read(client);
 959         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 960         if (rc) {
 961                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 962                 GOTO(out_proc, rc);
 963         }
 964
 965         ldlm_namespace_register(ns, client);
 966         RETURN(ns);
 967 out_proc:
 968         ldlm_namespace_debugfs_unregister(ns);
 969 out_sysfs:
 970         ldlm_namespace_sysfs_unregister(ns);
 971         ldlm_namespace_cleanup(ns, 0);
 972 out_hash:
 973         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
 974         kfree(ns->ns_name);
 975         cfs_hash_putref(ns->ns_rs_hash);
 976 out_ns:
 977         OBD_FREE_PTR(ns);
 978 out_ref:
 979         ldlm_put_ref();
 980         RETURN(NULL);
 981 }
 982 EXPORT_SYMBOL(ldlm_namespace_new);
 983
 984 /**
 985  * Cancel and destroy all locks on a resource.
 986  *
 987  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 988  * clean up.  This is currently only used for recovery, and we make
 989  * certain assumptions as a result--notably, that we shouldn't cancel
 990  * locks with refs.
 991  */
 992 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 993                              __u64 flags)
 994 {
 995         struct list_head *tmp;
 996         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 997         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 998
 999         do {
1000                 struct ldlm_lock *lock = NULL;
1001
1002                 /* First, we look for non-cleaned-yet lock
1003                  * all cleaned locks are marked by CLEANED flag. */
1004                 lock_res(res);
1005                 list_for_each(tmp, q) {
1006                         lock = list_entry(tmp, struct ldlm_lock,
1007                                           l_res_link);
1008                         if (ldlm_is_cleaned(lock)) {
1009                                 lock = NULL;
1010                                 continue;
1011                         }
1012                         LDLM_LOCK_GET(lock);
1013                         ldlm_set_cleaned(lock);
1014                         break;
1015                 }
1016
1017                 if (lock == NULL) {
1018                         unlock_res(res);
1019                         break;
1020                 }
1021
1022                 /* Set CBPENDING so nothing in the cancellation path
1023                  * can match this lock. */
1024                 ldlm_set_cbpending(lock);
1025                 ldlm_set_failed(lock);
1026                 lock->l_flags |= flags;
1027
1028                 /* ... without sending a CANCEL message for local_only. */
1029                 if (local_only)
1030                         ldlm_set_local_only(lock);
1031
1032                 if (local_only && (lock->l_readers || lock->l_writers)) {
1033                         /*
1034                          * This is a little bit gross, but much better than the
1035                          * alternative: pretend that we got a blocking AST from
1036                          * the server, so that when the lock is decref'd, it
1037                          * will go away ...
1038                          */
1039                         unlock_res(res);
1040                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1041                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1042                                 schedule_timeout_uninterruptible(
1043                                         cfs_time_seconds(4));
1044
1045                         if (lock->l_completion_ast)
1046                                 lock->l_completion_ast(lock,
1047                                                        LDLM_FL_FAILED, NULL);
1048                         LDLM_LOCK_RELEASE(lock);
1049                         continue;
1050                 }
1051
1052                 if (client) {
1053                         struct lustre_handle lockh;
1054
1055                         unlock_res(res);
1056                         ldlm_lock2handle(lock, &lockh);
1057                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1058                         if (rc)
1059                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1060                 } else {
1061                         unlock_res(res);
1062                         LDLM_DEBUG(lock,
1063                                    "Freeing a lock still held by a client node");
1064                         ldlm_lock_cancel(lock);
1065                 }
1066                 LDLM_LOCK_RELEASE(lock);
1067         } while (1);
1068 }
1069
1070 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1071                                struct hlist_node *hnode, void *arg)
1072 {
1073         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1074         __u64 flags = *(__u64 *)arg;
1075
1076         cleanup_resource(res, &res->lr_granted, flags);
1077         cleanup_resource(res, &res->lr_waiting, flags);
1078
1079         return 0;
1080 }
1081
1082 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1083                                   struct hlist_node *hnode, void *arg)
1084 {
1085         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1086
1087         lock_res(res);
1088         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1089                "(%d) after lock cleanup; forcing cleanup.\n",
1090                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1091                atomic_read(&res->lr_refcount) - 1);
1092
1093         /* Use D_NETERROR since it is in the default mask */
1094         ldlm_resource_dump(D_NETERROR, res);
1095         unlock_res(res);
1096         return 0;
1097 }
1098
1099 /**
1100  * Cancel and destroy all locks in the namespace.
1101  *
1102  * Typically used during evictions when server notified client that it was
1103  * evicted and all of its state needs to be destroyed.
1104  * Also used during shutdown.
1105  */
1106 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1107 {
1108         if (ns == NULL) {
1109                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1110                 return ELDLM_OK;
1111         }
1112
1113         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1114                                  &flags, 0);
1115         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1116                                  NULL, 0);
1117         return ELDLM_OK;
1118 }
1119 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1120
1121 /**
1122  * Attempts to free namespace.
1123  *
1124  * Only used when namespace goes away, like during an unmount.
1125  */
1126 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1127 {
1128         ENTRY;
1129
1130         /* At shutdown time, don't call the cancellation callback */
1131         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1132
1133         if (atomic_read(&ns->ns_bref) > 0) {
1134                 int rc;
1135                 CDEBUG(D_DLMTRACE,
1136                        "dlm namespace %s free waiting on refcount %d\n",
1137                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1138 force_wait:
1139                 if (force)
1140                         rc = wait_event_idle_timeout(
1141                                 ns->ns_waitq,
1142                                 atomic_read(&ns->ns_bref) == 0,
1143                                 cfs_time_seconds(1) / 4);
1144                 else
1145                         rc = l_wait_event_abortable(
1146                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1147
1148                 /* Forced cleanups should be able to reclaim all references,
1149                  * so it's safe to wait forever... we can't leak locks... */
1150                 if (force && rc == 0) {
1151                         rc = -ETIMEDOUT;
1152                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1153                                        "namespace with %d resources in use, "
1154                                        "(rc=%d)\n", ldlm_ns_name(ns),
1155                                        atomic_read(&ns->ns_bref), rc);
1156                         GOTO(force_wait, rc);
1157                 }
1158
1159                 if (atomic_read(&ns->ns_bref)) {
1160                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1161                                        "with %d resources in use, (rc=%d)\n",
1162                                        ldlm_ns_name(ns),
1163                                        atomic_read(&ns->ns_bref), rc);
1164                         RETURN(ELDLM_NAMESPACE_EXISTS);
1165                 }
1166                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1167                        ldlm_ns_name(ns));
1168         }
1169
1170         RETURN(ELDLM_OK);
1171 }
1172
1173 /**
1174  * Performs various cleanups for passed \a ns to make it drop refc and be
1175  * ready for freeing. Waits for refc == 0.
1176  *
1177  * The following is done:
1178  * (0) Unregister \a ns from its list to make inaccessible for potential
1179  * users like pools thread and others;
1180  * (1) Clear all locks in \a ns.
1181  */
1182 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1183                                struct obd_import *imp,
1184                                int force)
1185 {
1186         int rc;
1187
1188         ENTRY;
1189         if (!ns) {
1190                 EXIT;
1191                 return;
1192         }
1193
1194         spin_lock(&ns->ns_lock);
1195         ns->ns_stopping = 1;
1196         spin_unlock(&ns->ns_lock);
1197
1198         /*
1199          * Can fail with -EINTR when force == 0 in which case try harder.
1200          */
1201         rc = __ldlm_namespace_free(ns, force);
1202         if (rc != ELDLM_OK) {
1203                 if (imp) {
1204                         ptlrpc_disconnect_import(imp, 0);
1205                         ptlrpc_invalidate_import(imp);
1206                 }
1207
1208                 /*
1209                  * With all requests dropped and the import inactive
1210                  * we are gaurenteed all reference will be dropped.
1211                  */
1212                 rc = __ldlm_namespace_free(ns, 1);
1213                 LASSERT(rc == 0);
1214         }
1215         EXIT;
1216 }
1217 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1218
1219 /**
1220  * Performs freeing memory structures related to \a ns. This is only done
1221  * when ldlm_namespce_free_prior() successfully removed all resources
1222  * referencing \a ns and its refc == 0.
1223  */
1224 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1225 {
1226         ENTRY;
1227         if (!ns) {
1228                 EXIT;
1229                 return;
1230         }
1231
1232         /* Make sure that nobody can find this ns in its list. */
1233         ldlm_namespace_unregister(ns, ns->ns_client);
1234         /* Fini pool _before_ parent proc dir is removed. This is important as
1235          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1236          * Removing it after @dir may cause oops. */
1237         ldlm_pool_fini(&ns->ns_pool);
1238
1239         ldlm_namespace_debugfs_unregister(ns);
1240         ldlm_namespace_sysfs_unregister(ns);
1241         cfs_hash_putref(ns->ns_rs_hash);
1242         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1243         kfree(ns->ns_name);
1244         /* Namespace \a ns should be not on list at this time, otherwise
1245          * this will cause issues related to using freed \a ns in poold
1246          * thread.
1247          */
1248         LASSERT(list_empty(&ns->ns_list_chain));
1249         OBD_FREE_PTR(ns);
1250         ldlm_put_ref();
1251         EXIT;
1252 }
1253 EXPORT_SYMBOL(ldlm_namespace_free_post);
1254
1255 /**
1256  * Cleanup the resource, and free namespace.
1257  * bug 12864:
1258  * Deadlock issue:
1259  * proc1: destroy import
1260  *        class_disconnect_export(grab cl_sem) ->
1261  *              -> ldlm_namespace_free ->
1262  *              -> lprocfs_remove(grab _lprocfs_lock).
1263  * proc2: read proc info
1264  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1265  *              -> osc_rd_active, etc(grab cl_sem).
1266  *
1267  * So that I have to split the ldlm_namespace_free into two parts - the first
1268  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1269  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1270  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1271  * held.
1272  */
1273 void ldlm_namespace_free(struct ldlm_namespace *ns,
1274                          struct obd_import *imp,
1275                          int force)
1276 {
1277         ldlm_namespace_free_prior(ns, imp, force);
1278         ldlm_namespace_free_post(ns);
1279 }
1280 EXPORT_SYMBOL(ldlm_namespace_free);
1281
1282 void ldlm_namespace_get(struct ldlm_namespace *ns)
1283 {
1284         atomic_inc(&ns->ns_bref);
1285 }
1286
1287 /* This is only for callers that care about refcount */
1288 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1289 {
1290         return atomic_inc_return(&ns->ns_bref);
1291 }
1292
1293 void ldlm_namespace_put(struct ldlm_namespace *ns)
1294 {
1295         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1296                 wake_up(&ns->ns_waitq);
1297                 spin_unlock(&ns->ns_lock);
1298         }
1299 }
1300
1301 /** Register \a ns in the list of namespaces */
1302 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1303 {
1304         mutex_lock(ldlm_namespace_lock(client));
1305         LASSERT(list_empty(&ns->ns_list_chain));
1306         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1307         ldlm_namespace_nr_inc(client);
1308         mutex_unlock(ldlm_namespace_lock(client));
1309 }
1310
1311 /** Unregister \a ns from the list of namespaces. */
1312 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1313 {
1314         mutex_lock(ldlm_namespace_lock(client));
1315         LASSERT(!list_empty(&ns->ns_list_chain));
1316         /* Some asserts and possibly other parts of the code are still
1317          * using list_empty(&ns->ns_list_chain). This is why it is
1318          * important to use list_del_init() here. */
1319         list_del_init(&ns->ns_list_chain);
1320         ldlm_namespace_nr_dec(client);
1321         mutex_unlock(ldlm_namespace_lock(client));
1322 }
1323
1324 /** Should be called with ldlm_namespace_lock(client) taken. */
1325 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1326                                           enum ldlm_side client)
1327 {
1328         LASSERT(!list_empty(&ns->ns_list_chain));
1329         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1330         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1331 }
1332
1333 /** Should be called with ldlm_namespace_lock(client) taken. */
1334 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1335                                             enum ldlm_side client)
1336 {
1337         LASSERT(!list_empty(&ns->ns_list_chain));
1338         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1339         list_move_tail(&ns->ns_list_chain,
1340                        ldlm_namespace_inactive_list(client));
1341 }
1342
1343 /** Should be called with ldlm_namespace_lock(client) taken. */
1344 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1345 {
1346         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1347         LASSERT(!list_empty(ldlm_namespace_list(client)));
1348         return container_of(ldlm_namespace_list(client)->next,
1349                             struct ldlm_namespace, ns_list_chain);
1350 }
1351
1352 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1353 {
1354         int idx;
1355
1356         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1357                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1358         if (res->lr_itree == NULL)
1359                 return false;
1360         /* Initialize interval trees for each lock mode. */
1361         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1362                 res->lr_itree[idx].lit_size = 0;
1363                 res->lr_itree[idx].lit_mode = BIT(idx);
1364                 res->lr_itree[idx].lit_root = NULL;
1365         }
1366         return true;
1367 }
1368
1369 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1370 {
1371         int i;
1372
1373         OBD_ALLOC_PTR(res->lr_ibits_queues);
1374         if (res->lr_ibits_queues == NULL)
1375                 return false;
1376         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1377                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1378         return true;
1379 }
1380
1381 /** Create and initialize new resource. */
1382 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1383 {
1384         struct ldlm_resource *res;
1385         bool rc;
1386
1387         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1388         if (res == NULL)
1389                 return NULL;
1390
1391         switch (ldlm_type) {
1392         case LDLM_EXTENT:
1393                 rc = ldlm_resource_extent_new(res);
1394                 break;
1395         case LDLM_IBITS:
1396                 rc = ldlm_resource_inodebits_new(res);
1397                 break;
1398         default:
1399                 rc = true;
1400                 break;
1401         }
1402         if (!rc) {
1403                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1404                 return NULL;
1405         }
1406
1407         INIT_LIST_HEAD(&res->lr_granted);
1408         INIT_LIST_HEAD(&res->lr_waiting);
1409
1410         atomic_set(&res->lr_refcount, 1);
1411         spin_lock_init(&res->lr_lock);
1412         lu_ref_init(&res->lr_reference);
1413
1414         /* Since LVB init can be delayed now, there is no longer need to
1415          * immediatelly acquire mutex here. */
1416         mutex_init(&res->lr_lvb_mutex);
1417         res->lr_lvb_initialized = false;
1418
1419         return res;
1420 }
1421
1422 static void ldlm_resource_free(struct ldlm_resource *res)
1423 {
1424         if (res->lr_type == LDLM_EXTENT) {
1425                 if (res->lr_itree != NULL)
1426                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1427                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1428         } else if (res->lr_type == LDLM_IBITS) {
1429                 if (res->lr_ibits_queues != NULL)
1430                         OBD_FREE_PTR(res->lr_ibits_queues);
1431         }
1432
1433         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1434 }
1435
1436 /**
1437  * Return a reference to resource with given name, creating it if necessary.
1438  * Args: namespace with ns_lock unlocked
1439  * Locks: takes and releases NS hash-lock and res->lr_lock
1440  * Returns: referenced, unlocked ldlm_resource or NULL
1441  */
1442 struct ldlm_resource *
1443 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1444                   const struct ldlm_res_id *name, enum ldlm_type type,
1445                   int create)
1446 {
1447         struct hlist_node       *hnode;
1448         struct ldlm_resource    *res = NULL;
1449         struct cfs_hash_bd              bd;
1450         __u64                   version;
1451         int                     ns_refcount = 0;
1452         int hash;
1453
1454         LASSERT(ns != NULL);
1455         LASSERT(parent == NULL);
1456         LASSERT(ns->ns_rs_hash != NULL);
1457         LASSERT(name->name[0] != 0);
1458
1459         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1460         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1461         if (hnode != NULL) {
1462                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1463                 GOTO(found, res);
1464         }
1465
1466         version = cfs_hash_bd_version_get(&bd);
1467         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1468
1469         if (create == 0)
1470                 return ERR_PTR(-ENOENT);
1471
1472         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1473                  "type: %d\n", type);
1474         res = ldlm_resource_new(type);
1475         if (res == NULL)
1476                 return ERR_PTR(-ENOMEM);
1477
1478         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1479         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1480         res->lr_name = *name;
1481         res->lr_type = type;
1482
1483         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1484         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1485                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1486
1487         if (hnode != NULL) {
1488                 /* Someone won the race and already added the resource. */
1489                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1490                 /* Clean lu_ref for failed resource. */
1491                 lu_ref_fini(&res->lr_reference);
1492                 ldlm_resource_free(res);
1493 found:
1494                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1495                 return res;
1496         }
1497         /* We won! Let's add the resource. */
1498         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1499         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1500                 ns_refcount = ldlm_namespace_get_return(ns);
1501
1502         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1503
1504         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1505
1506         /* Let's see if we happened to be the very first resource in this
1507          * namespace. If so, and this is a client namespace, we need to move
1508          * the namespace into the active namespaces list to be patrolled by
1509          * the ldlm_poold. */
1510         if (ns_is_client(ns) && ns_refcount == 1) {
1511                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1512                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1513                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1514         }
1515
1516         return res;
1517 }
1518 EXPORT_SYMBOL(ldlm_resource_get);
1519
1520 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1521 {
1522         LASSERT(res != NULL);
1523         LASSERT(res != LP_POISON);
1524         atomic_inc(&res->lr_refcount);
1525         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1526                atomic_read(&res->lr_refcount));
1527         return res;
1528 }
1529
1530 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1531                                          struct ldlm_resource *res)
1532 {
1533         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1534
1535         if (!list_empty(&res->lr_granted)) {
1536                 ldlm_resource_dump(D_ERROR, res);
1537                 LBUG();
1538         }
1539
1540         if (!list_empty(&res->lr_waiting)) {
1541                 ldlm_resource_dump(D_ERROR, res);
1542                 LBUG();
1543         }
1544
1545         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1546                                bd, &res->lr_hash);
1547         lu_ref_fini(&res->lr_reference);
1548         if (atomic_dec_and_test(&nsb->nsb_count))
1549                 ldlm_namespace_put(nsb->nsb_namespace);
1550 }
1551
1552 /* Returns 1 if the resource was freed, 0 if it remains. */
1553 int ldlm_resource_putref(struct ldlm_resource *res)
1554 {
1555         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1556         struct cfs_hash_bd   bd;
1557
1558         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1559         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1560                res, atomic_read(&res->lr_refcount) - 1);
1561
1562         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1563         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1564                 __ldlm_resource_putref_final(&bd, res);
1565                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1566                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1567                         ns->ns_lvbo->lvbo_free(res);
1568                 ldlm_resource_free(res);
1569                 return 1;
1570         }
1571         return 0;
1572 }
1573 EXPORT_SYMBOL(ldlm_resource_putref);
1574
1575 static void __ldlm_resource_add_lock(struct ldlm_resource *res,
1576                                      struct list_head *head,
1577                                      struct ldlm_lock *lock,
1578                                      bool tail)
1579 {
1580         check_res_locked(res);
1581
1582         if (ldlm_is_destroyed(lock)) {
1583                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1584                 return;
1585         }
1586
1587         LASSERT(list_empty(&lock->l_res_link));
1588
1589         if (tail)
1590                 list_add_tail(&lock->l_res_link, head);
1591         else
1592                 list_add(&lock->l_res_link, head);
1593
1594         if (res->lr_type == LDLM_IBITS)
1595                 ldlm_inodebits_add_lock(res, head, lock);
1596
1597         ldlm_resource_dump(D_INFO, res);
1598 }
1599
1600 /**
1601  * Add a lock into a given resource into specified lock list.
1602  */
1603 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1604                             struct ldlm_lock *lock)
1605 {
1606         LDLM_DEBUG(lock, "About to add this lock");
1607
1608         __ldlm_resource_add_lock(res, head, lock, true);
1609 }
1610
1611 /**
1612  * Insert a lock into resource after specified lock.
1613  */
1614 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1615                                      struct ldlm_lock *new)
1616 {
1617         LASSERT(!list_empty(&original->l_res_link));
1618
1619         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1620         __ldlm_resource_add_lock(original->l_resource,
1621                                  &original->l_res_link,
1622                                  new, false);
1623 }
1624
1625 /**
1626  * Insert a lock into resource before the specified lock.
1627  */
1628 void ldlm_resource_insert_lock_before(struct ldlm_lock *original,
1629                                       struct ldlm_lock *new)
1630 {
1631         LASSERT(!list_empty(&original->l_res_link));
1632
1633         LDLM_DEBUG(new, "About to insert this lock before %p: ", original);
1634         __ldlm_resource_add_lock(original->l_resource,
1635                                  original->l_res_link.prev, new, false);
1636 }
1637
1638 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1639 {
1640         int type = lock->l_resource->lr_type;
1641
1642         check_res_locked(lock->l_resource);
1643         switch (type) {
1644         case LDLM_PLAIN:
1645                 ldlm_unlink_lock_skiplist(lock);
1646                 break;
1647         case LDLM_EXTENT:
1648                 ldlm_extent_unlink_lock(lock);
1649                 break;
1650         case LDLM_IBITS:
1651                 ldlm_inodebits_unlink_lock(lock);
1652                 break;
1653         }
1654         list_del_init(&lock->l_res_link);
1655 }
1656 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1657
1658 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1659 {
1660         desc->lr_type = res->lr_type;
1661         desc->lr_name = res->lr_name;
1662 }
1663
1664 /**
1665  * Print information about all locks in all namespaces on this node to debug
1666  * log.
1667  */
1668 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1669 {
1670         struct list_head *tmp;
1671
1672         if (!((libcfs_debug | D_ERROR) & level))
1673                 return;
1674
1675         mutex_lock(ldlm_namespace_lock(client));
1676
1677         list_for_each(tmp, ldlm_namespace_list(client)) {
1678                 struct ldlm_namespace *ns;
1679
1680                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1681                 ldlm_namespace_dump(level, ns);
1682         }
1683
1684         mutex_unlock(ldlm_namespace_lock(client));
1685 }
1686
1687 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1688                               struct hlist_node *hnode, void *arg)
1689 {
1690         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1691         int    level = (int)(unsigned long)arg;
1692
1693         lock_res(res);
1694         ldlm_resource_dump(level, res);
1695         unlock_res(res);
1696
1697         return 0;
1698 }
1699
1700 /**
1701  * Print information about all locks in this namespace on this node to debug
1702  * log.
1703  */
1704 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1705 {
1706         if (!((libcfs_debug | D_ERROR) & level))
1707                 return;
1708
1709         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1710                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1711                ns_is_client(ns) ? "client" : "server");
1712
1713         if (ktime_get_seconds() < ns->ns_next_dump)
1714                 return;
1715
1716         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1717                                  ldlm_res_hash_dump,
1718                                  (void *)(unsigned long)level, 0);
1719         spin_lock(&ns->ns_lock);
1720         ns->ns_next_dump = ktime_get_seconds() + 10;
1721         spin_unlock(&ns->ns_lock);
1722 }
1723
1724 /**
1725  * Print information about all locks in this resource to debug log.
1726  */
1727 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1728 {
1729         struct ldlm_lock *lock;
1730         unsigned int granted = 0;
1731
1732         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1733
1734         if (!((libcfs_debug | D_ERROR) & level))
1735                 return;
1736
1737         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1738                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1739
1740         if (!list_empty(&res->lr_granted)) {
1741                 CDEBUG(level, "Granted locks (in reverse order):\n");
1742                 list_for_each_entry_reverse(lock, &res->lr_granted,
1743                                                 l_res_link) {
1744                         LDLM_DEBUG_LIMIT(level, lock, "###");
1745                         if (!(level & D_CANTMASK) &&
1746                             ++granted > ldlm_dump_granted_max) {
1747                                 CDEBUG(level,
1748                                        "only dump %d granted locks to avoid DDOS.\n",
1749                                        granted);
1750                                 break;
1751                         }
1752                 }
1753         }
1754
1755         if (!list_empty(&res->lr_waiting)) {
1756                 CDEBUG(level, "Waiting locks:\n");
1757                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1758                         LDLM_DEBUG_LIMIT(level, lock, "###");
1759         }
1760 }
1761 EXPORT_SYMBOL(ldlm_resource_dump);