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 ns_recalc_pct_show(struct kobject *kobj,
 402                                   struct attribute *attr, char *buf)
 403 {
 404         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 405                                                  ns_kobj);
 406
 407         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_recalc_pct);
 408 }
 409
 410 static ssize_t ns_recalc_pct_store(struct kobject *kobj,
 411                                    struct attribute *attr,
 412                                    const char *buffer, size_t count)
 413 {
 414         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 415                                                  ns_kobj);
 416         unsigned long tmp;
 417
 418         if (kstrtoul(buffer, 10, &tmp))
 419                 return -EINVAL;
 420
 421         if (tmp > 100)
 422                 return -ERANGE;
 423
 424         ns->ns_recalc_pct = (unsigned int)tmp;
 425
 426         return count;
 427 }
 428 LUSTRE_RW_ATTR(ns_recalc_pct);
 429
 430 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 431                                 char *buf)
 432 {
 433         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 434                                                  ns_kobj);
 435
 436         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 437 }
 438
 439 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 440                                  const char *buffer, size_t count)
 441 {
 442         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 443                                                  ns_kobj);
 444         int scale = NSEC_PER_MSEC;
 445         unsigned long long tmp;
 446         char *buf;
 447
 448         /* Did the user ask in seconds or milliseconds. Default is in ms */
 449         buf = strstr(buffer, "ms");
 450         if (!buf) {
 451                 buf = strchr(buffer, 's');
 452                 if (buf)
 453                         scale = NSEC_PER_SEC;
 454         }
 455
 456         if (buf)
 457                 *buf = '\0';
 458
 459         if (kstrtoull(buffer, 10, &tmp))
 460                 return -EINVAL;
 461
 462         ns->ns_max_age = ktime_set(0, tmp * scale);
 463
 464         return count;
 465 }
 466 LUSTRE_RW_ATTR(lru_max_age);
 467
 468 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 469                                       struct attribute *attr,
 470                                       char *buf)
 471 {
 472         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 473                                                  ns_kobj);
 474
 475         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 476 }
 477
 478 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 479                                        struct attribute *attr,
 480                                        const char *buffer,
 481                                        size_t count)
 482 {
 483         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 484                                                  ns_kobj);
 485         unsigned long supp = -1;
 486         int rc;
 487
 488         rc = kstrtoul(buffer, 10, &supp);
 489         if (rc < 0)
 490                 return rc;
 491
 492         if (supp == 0)
 493                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 494         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 495                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 496         return count;
 497 }
 498 LUSTRE_RW_ATTR(early_lock_cancel);
 499
 500 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 501                                     struct attribute *attr, char *buf)
 502 {
 503         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 504                                                  ns_kobj);
 505
 506         return snprintf(buf, PAGE_SIZE, "%llu\n",
 507                         ktime_divns(ns->ns_dirty_age_limit, NSEC_PER_SEC));
 508 }
 509
 510 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 511                                      struct attribute *attr,
 512                                      const char *buffer, size_t count)
 513 {
 514         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 515                                                  ns_kobj);
 516         unsigned long long tmp;
 517
 518         if (kstrtoull(buffer, 10, &tmp))
 519                 return -EINVAL;
 520
 521         ns->ns_dirty_age_limit = ktime_set(tmp, 0);
 522
 523         return count;
 524 }
 525 LUSTRE_RW_ATTR(dirty_age_limit);
 526
 527 #ifdef HAVE_SERVER_SUPPORT
 528 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 529                                     struct attribute *attr, char *buf)
 530 {
 531         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 532                                                  ns_kobj);
 533
 534         return snprintf(buf, PAGE_SIZE, "%u\n", ns->ns_ctime_age_limit);
 535 }
 536
 537 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 538                                      struct attribute *attr,
 539                                      const char *buffer, size_t count)
 540 {
 541         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 542                                                  ns_kobj);
 543         unsigned long tmp;
 544
 545         if (kstrtoul(buffer, 10, &tmp))
 546                 return -EINVAL;
 547
 548         ns->ns_ctime_age_limit = tmp;
 549
 550         return count;
 551 }
 552 LUSTRE_RW_ATTR(ctime_age_limit);
 553
 554 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 555                                   char *buf)
 556 {
 557         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 558                                                  ns_kobj);
 559
 560         return sprintf(buf, "%d\n", ns->ns_timeouts);
 561 }
 562 LUSTRE_RO_ATTR(lock_timeouts);
 563
 564 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 565                                      struct attribute *attr, char *buf)
 566 {
 567         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 568                                                  ns_kobj);
 569
 570         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 571 }
 572
 573 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 574                                       struct attribute *attr,
 575                                       const char *buffer, size_t count)
 576 {
 577         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 578                                                  ns_kobj);
 579         unsigned long tmp;
 580         int err;
 581
 582         err = kstrtoul(buffer, 10, &tmp);
 583         if (err != 0)
 584                 return -EINVAL;
 585
 586         ns->ns_max_nolock_size = tmp;
 587
 588         return count;
 589 }
 590 LUSTRE_RW_ATTR(max_nolock_bytes);
 591
 592 static ssize_t contention_seconds_show(struct kobject *kobj,
 593                                        struct attribute *attr, char *buf)
 594 {
 595         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 596                                                  ns_kobj);
 597
 598         return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
 599 }
 600
 601 static ssize_t contention_seconds_store(struct kobject *kobj,
 602                                         struct attribute *attr,
 603                                         const char *buffer, size_t count)
 604 {
 605         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 606                                                  ns_kobj);
 607         unsigned int tmp;
 608
 609         if (kstrtouint(buffer, 10, &tmp))
 610                 return -EINVAL;
 611
 612         ns->ns_contention_time = tmp;
 613
 614         return count;
 615 }
 616 LUSTRE_RW_ATTR(contention_seconds);
 617
 618 static ssize_t contended_locks_show(struct kobject *kobj,
 619                                     struct attribute *attr, char *buf)
 620 {
 621         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 622                                                  ns_kobj);
 623
 624         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 625 }
 626
 627 static ssize_t contended_locks_store(struct kobject *kobj,
 628                                      struct attribute *attr,
 629                                      const char *buffer, size_t count)
 630 {
 631         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 632                                                  ns_kobj);
 633         unsigned long tmp;
 634         int err;
 635
 636         err = kstrtoul(buffer, 10, &tmp);
 637         if (err != 0)
 638                 return -EINVAL;
 639
 640         ns->ns_contended_locks = tmp;
 641
 642         return count;
 643 }
 644 LUSTRE_RW_ATTR(contended_locks);
 645
 646 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 647                                      struct attribute *attr, char *buf)
 648 {
 649         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 650                                                  ns_kobj);
 651
 652         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 653 }
 654
 655 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 656                                       struct attribute *attr,
 657                                       const char *buffer, size_t count)
 658 {
 659         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 660                                                  ns_kobj);
 661         unsigned long tmp;
 662         int err;
 663
 664         err = kstrtoul(buffer, 10, &tmp);
 665         if (err != 0)
 666                 return -EINVAL;
 667
 668         ns->ns_max_parallel_ast = tmp;
 669
 670         return count;
 671 }
 672 LUSTRE_RW_ATTR(max_parallel_ast);
 673
 674 #endif /* HAVE_SERVER_SUPPORT */
 675
 676 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 677 static struct attribute *ldlm_ns_attrs[] = {
 678         &lustre_attr_resource_count.attr,
 679         &lustre_attr_lock_count.attr,
 680         &lustre_attr_lock_unused_count.attr,
 681         &lustre_attr_ns_recalc_pct.attr,
 682         &lustre_attr_lru_size.attr,
 683         &lustre_attr_lru_cancel_batch.attr,
 684         &lustre_attr_lru_max_age.attr,
 685         &lustre_attr_early_lock_cancel.attr,
 686         &lustre_attr_dirty_age_limit.attr,
 687 #ifdef HAVE_SERVER_SUPPORT
 688         &lustre_attr_ctime_age_limit.attr,
 689         &lustre_attr_lock_timeouts.attr,
 690         &lustre_attr_max_nolock_bytes.attr,
 691         &lustre_attr_contention_seconds.attr,
 692         &lustre_attr_contended_locks.attr,
 693         &lustre_attr_max_parallel_ast.attr,
 694 #endif
 695         NULL,
 696 };
 697
 698 static void ldlm_ns_release(struct kobject *kobj)
 699 {
 700         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 701                                                  ns_kobj);
 702         complete(&ns->ns_kobj_unregister);
 703 }
 704
 705 static struct kobj_type ldlm_ns_ktype = {
 706         .default_attrs  = ldlm_ns_attrs,
 707         .sysfs_ops      = &lustre_sysfs_ops,
 708         .release        = ldlm_ns_release,
 709 };
 710
 711 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 712 {
 713         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 714                 CERROR("dlm namespace %s has no procfs dir?\n",
 715                        ldlm_ns_name(ns));
 716         else
 717                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 718
 719         if (ns->ns_stats != NULL)
 720                 lprocfs_free_stats(&ns->ns_stats);
 721 }
 722
 723 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 724 {
 725         kobject_put(&ns->ns_kobj);
 726         wait_for_completion(&ns->ns_kobj_unregister);
 727 }
 728
 729 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 730 {
 731         int err;
 732
 733         ns->ns_kobj.kset = ldlm_ns_kset;
 734         init_completion(&ns->ns_kobj_unregister);
 735         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 736                                    "%s", ldlm_ns_name(ns));
 737
 738         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 739         if (!ns->ns_stats) {
 740                 kobject_put(&ns->ns_kobj);
 741                 return -ENOMEM;
 742         }
 743
 744         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 745                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 746
 747         return err;
 748 }
 749
 750 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 751 {
 752         struct dentry *ns_entry;
 753
 754         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 755                 ns_entry = ns->ns_debugfs_entry;
 756         } else {
 757                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 758                                               ldlm_ns_debugfs_dir);
 759                 if (!ns_entry)
 760                         return -ENOMEM;
 761                 ns->ns_debugfs_entry = ns_entry;
 762         }
 763
 764         return 0;
 765 }
 766 #undef MAX_STRING_SIZE
 767
 768 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 769                                   const void *key, unsigned int mask)
 770 {
 771         const struct ldlm_res_id *id = key;
 772         unsigned int val = 0;
 773         unsigned int i;
 774
 775         for (i = 0; i < RES_NAME_SIZE; i++)
 776                 val += id->name[i];
 777         return val & mask;
 778 }
 779
 780 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 781 {
 782         struct lu_fid       fid;
 783         __u32               hash;
 784         __u32               val;
 785
 786         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 787         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 788         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 789
 790         hash = fid_flatten32(&fid);
 791         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 792         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 793                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 794         } else {
 795                 val = fid_oid(&fid);
 796         }
 797         hash += (val >> 5) + (val << 11);
 798         return cfs_hash_32(hash, bits);
 799 }
 800
 801 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 802 {
 803         struct ldlm_resource   *res;
 804
 805         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 806         return &res->lr_name;
 807 }
 808
 809 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 810 {
 811         struct ldlm_resource   *res;
 812
 813         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 814         return ldlm_res_eq((const struct ldlm_res_id *)key,
 815                            (const struct ldlm_res_id *)&res->lr_name);
 816 }
 817
 818 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 819 {
 820         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 821 }
 822
 823 static void
 824 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 825 {
 826         struct ldlm_resource *res;
 827
 828         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 829         ldlm_resource_getref(res);
 830 }
 831
 832 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 833 {
 834         struct ldlm_resource *res;
 835
 836         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 837         ldlm_resource_putref(res);
 838 }
 839
 840 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 841         .hs_hash        = ldlm_res_hop_hash,
 842         .hs_key         = ldlm_res_hop_key,
 843         .hs_keycmp      = ldlm_res_hop_keycmp,
 844         .hs_keycpy      = NULL,
 845         .hs_object      = ldlm_res_hop_object,
 846         .hs_get         = ldlm_res_hop_get_locked,
 847         .hs_put         = ldlm_res_hop_put
 848 };
 849
 850 static struct {
 851         /** hash bucket bits */
 852         unsigned                nsd_bkt_bits;
 853         /** hash bits */
 854         unsigned                nsd_all_bits;
 855 } ldlm_ns_hash_defs[] = {
 856         [LDLM_NS_TYPE_MDC] = {
 857                 .nsd_bkt_bits   = 11,
 858                 .nsd_all_bits   = 16,
 859         },
 860         [LDLM_NS_TYPE_MDT] = {
 861                 .nsd_bkt_bits   = 14,
 862                 .nsd_all_bits   = 21,
 863         },
 864         [LDLM_NS_TYPE_OSC] = {
 865                 .nsd_bkt_bits   = 8,
 866                 .nsd_all_bits   = 12,
 867         },
 868         [LDLM_NS_TYPE_OST] = {
 869                 .nsd_bkt_bits   = 11,
 870                 .nsd_all_bits   = 17,
 871         },
 872         [LDLM_NS_TYPE_MGC] = {
 873                 .nsd_bkt_bits   = 3,
 874                 .nsd_all_bits   = 4,
 875         },
 876         [LDLM_NS_TYPE_MGT] = {
 877                 .nsd_bkt_bits   = 3,
 878                 .nsd_all_bits   = 4,
 879         },
 880 };
 881
 882 /**
 883  * Create and initialize new empty namespace.
 884  */
 885 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 886                                           enum ldlm_side client,
 887                                           enum ldlm_appetite apt,
 888                                           enum ldlm_ns_type ns_type)
 889 {
 890         struct ldlm_namespace *ns = NULL;
 891         int idx;
 892         int rc;
 893
 894         ENTRY;
 895         LASSERT(obd != NULL);
 896
 897         rc = ldlm_get_ref();
 898         if (rc) {
 899                 CERROR("ldlm_get_ref failed: %d\n", rc);
 900                 RETURN(NULL);
 901         }
 902
 903         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 904             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 905                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 906                 GOTO(out_ref, NULL);
 907         }
 908
 909         OBD_ALLOC_PTR(ns);
 910         if (!ns)
 911                 GOTO(out_ref, NULL);
 912
 913         ns->ns_rs_hash = cfs_hash_create(name,
 914                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 915                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 916                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 917                                          0,
 918                                          CFS_HASH_MIN_THETA,
 919                                          CFS_HASH_MAX_THETA,
 920                                          &ldlm_ns_hash_ops,
 921                                          CFS_HASH_DEPTH |
 922                                          CFS_HASH_BIGNAME |
 923                                          CFS_HASH_SPIN_BKTLOCK |
 924                                          CFS_HASH_NO_ITEMREF);
 925         if (ns->ns_rs_hash == NULL)
 926                 GOTO(out_ns, NULL);
 927
 928         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 929                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 930
 931         OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
 932         if (!ns->ns_rs_buckets)
 933                 goto out_hash;
 934
 935         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 936                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 937
 938                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 939                 nsb->nsb_namespace = ns;
 940                 nsb->nsb_reclaim_start = 0;
 941                 atomic_set(&nsb->nsb_count, 0);
 942         }
 943
 944         ns->ns_obd = obd;
 945         ns->ns_appetite = apt;
 946         ns->ns_client = client;
 947         ns->ns_name = kstrdup(name, GFP_KERNEL);
 948         if (!ns->ns_name)
 949                 goto out_hash;
 950
 951         INIT_LIST_HEAD(&ns->ns_list_chain);
 952         INIT_LIST_HEAD(&ns->ns_unused_list);
 953         spin_lock_init(&ns->ns_lock);
 954         atomic_set(&ns->ns_bref, 0);
 955         init_waitqueue_head(&ns->ns_waitq);
 956
 957         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 958         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 959         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 960
 961         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 962         ns->ns_nr_unused          = 0;
 963         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 964         ns->ns_cancel_batch       = LDLM_DEFAULT_LRU_SHRINK_BATCH;
 965         ns->ns_recalc_pct         = LDLM_DEFAULT_SLV_RECALC_PCT;
 966         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 967         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 968         ns->ns_dirty_age_limit    = ktime_set(LDLM_DIRTY_AGE_LIMIT, 0);
 969         ns->ns_timeouts           = 0;
 970         ns->ns_orig_connect_flags = 0;
 971         ns->ns_connect_flags      = 0;
 972         ns->ns_stopping           = 0;
 973         ns->ns_reclaim_start      = 0;
 974         ns->ns_last_pos           = &ns->ns_unused_list;
 975         ns->ns_flags              = 0;
 976
 977         rc = ldlm_namespace_sysfs_register(ns);
 978         if (rc) {
 979                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 980                 GOTO(out_hash, rc);
 981         }
 982
 983         rc = ldlm_namespace_debugfs_register(ns);
 984         if (rc) {
 985                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 986                 GOTO(out_sysfs, rc);
 987         }
 988
 989         idx = ldlm_namespace_nr_read(client);
 990         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 991         if (rc) {
 992                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 993                 GOTO(out_proc, rc);
 994         }
 995
 996         ldlm_namespace_register(ns, client);
 997         RETURN(ns);
 998 out_proc:
 999         ldlm_namespace_debugfs_unregister(ns);
1000 out_sysfs:
1001         ldlm_namespace_sysfs_unregister(ns);
1002         ldlm_namespace_cleanup(ns, 0);
1003 out_hash:
1004         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1005         kfree(ns->ns_name);
1006         cfs_hash_putref(ns->ns_rs_hash);
1007 out_ns:
1008         OBD_FREE_PTR(ns);
1009 out_ref:
1010         ldlm_put_ref();
1011         RETURN(NULL);
1012 }
1013 EXPORT_SYMBOL(ldlm_namespace_new);
1014
1015 /**
1016  * Cancel and destroy all locks on a resource.
1017  *
1018  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1019  * clean up.  This is currently only used for recovery, and we make
1020  * certain assumptions as a result--notably, that we shouldn't cancel
1021  * locks with refs.
1022  */
1023 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1024                              __u64 flags)
1025 {
1026         struct list_head *tmp;
1027         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1028         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1029
1030         do {
1031                 struct ldlm_lock *lock = NULL;
1032
1033                 /* First, we look for non-cleaned-yet lock
1034                  * all cleaned locks are marked by CLEANED flag. */
1035                 lock_res(res);
1036                 list_for_each(tmp, q) {
1037                         lock = list_entry(tmp, struct ldlm_lock,
1038                                           l_res_link);
1039                         if (ldlm_is_cleaned(lock)) {
1040                                 lock = NULL;
1041                                 continue;
1042                         }
1043                         LDLM_LOCK_GET(lock);
1044                         ldlm_set_cleaned(lock);
1045                         break;
1046                 }
1047
1048                 if (lock == NULL) {
1049                         unlock_res(res);
1050                         break;
1051                 }
1052
1053                 /* Set CBPENDING so nothing in the cancellation path
1054                  * can match this lock. */
1055                 ldlm_set_cbpending(lock);
1056                 ldlm_set_failed(lock);
1057                 lock->l_flags |= flags;
1058
1059                 /* ... without sending a CANCEL message for local_only. */
1060                 if (local_only)
1061                         ldlm_set_local_only(lock);
1062
1063                 if (local_only && (lock->l_readers || lock->l_writers)) {
1064                         /*
1065                          * This is a little bit gross, but much better than the
1066                          * alternative: pretend that we got a blocking AST from
1067                          * the server, so that when the lock is decref'd, it
1068                          * will go away ...
1069                          */
1070                         unlock_res(res);
1071                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1072                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1073                                 schedule_timeout_uninterruptible(
1074                                         cfs_time_seconds(4));
1075
1076                         if (lock->l_completion_ast)
1077                                 lock->l_completion_ast(lock,
1078                                                        LDLM_FL_FAILED, NULL);
1079                         LDLM_LOCK_RELEASE(lock);
1080                         continue;
1081                 }
1082
1083                 if (client) {
1084                         struct lustre_handle lockh;
1085
1086                         unlock_res(res);
1087                         ldlm_lock2handle(lock, &lockh);
1088                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1089                         if (rc)
1090                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1091                 } else {
1092                         unlock_res(res);
1093                         LDLM_DEBUG(lock,
1094                                    "Freeing a lock still held by a client node");
1095                         ldlm_lock_cancel(lock);
1096                 }
1097                 LDLM_LOCK_RELEASE(lock);
1098         } while (1);
1099 }
1100
1101 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1102                                struct hlist_node *hnode, void *arg)
1103 {
1104         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1105         __u64 flags = *(__u64 *)arg;
1106
1107         cleanup_resource(res, &res->lr_granted, flags);
1108         cleanup_resource(res, &res->lr_waiting, flags);
1109
1110         return 0;
1111 }
1112
1113 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1114                                   struct hlist_node *hnode, void *arg)
1115 {
1116         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1117
1118         lock_res(res);
1119         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1120                "(%d) after lock cleanup; forcing cleanup.\n",
1121                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1122                atomic_read(&res->lr_refcount) - 1);
1123
1124         /* Use D_NETERROR since it is in the default mask */
1125         ldlm_resource_dump(D_NETERROR, res);
1126         unlock_res(res);
1127         return 0;
1128 }
1129
1130 /**
1131  * Cancel and destroy all locks in the namespace.
1132  *
1133  * Typically used during evictions when server notified client that it was
1134  * evicted and all of its state needs to be destroyed.
1135  * Also used during shutdown.
1136  */
1137 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1138 {
1139         if (ns == NULL) {
1140                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1141                 return ELDLM_OK;
1142         }
1143
1144         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1145                                  &flags, 0);
1146         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1147                                  NULL, 0);
1148         return ELDLM_OK;
1149 }
1150 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1151
1152 /**
1153  * Attempts to free namespace.
1154  *
1155  * Only used when namespace goes away, like during an unmount.
1156  */
1157 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1158 {
1159         ENTRY;
1160
1161         /* At shutdown time, don't call the cancellation callback */
1162         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1163
1164         if (atomic_read(&ns->ns_bref) > 0) {
1165                 int rc;
1166                 CDEBUG(D_DLMTRACE,
1167                        "dlm namespace %s free waiting on refcount %d\n",
1168                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1169 force_wait:
1170                 if (force)
1171                         rc = wait_event_idle_timeout(
1172                                 ns->ns_waitq,
1173                                 atomic_read(&ns->ns_bref) == 0,
1174                                 cfs_time_seconds(1) / 4);
1175                 else
1176                         rc = l_wait_event_abortable(
1177                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1178
1179                 /* Forced cleanups should be able to reclaim all references,
1180                  * so it's safe to wait forever... we can't leak locks... */
1181                 if (force && rc == 0) {
1182                         rc = -ETIMEDOUT;
1183                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1184                                        "namespace with %d resources in use, "
1185                                        "(rc=%d)\n", ldlm_ns_name(ns),
1186                                        atomic_read(&ns->ns_bref), rc);
1187                         GOTO(force_wait, rc);
1188                 }
1189
1190                 if (atomic_read(&ns->ns_bref)) {
1191                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1192                                        "with %d resources in use, (rc=%d)\n",
1193                                        ldlm_ns_name(ns),
1194                                        atomic_read(&ns->ns_bref), rc);
1195                         RETURN(ELDLM_NAMESPACE_EXISTS);
1196                 }
1197                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1198                        ldlm_ns_name(ns));
1199         }
1200
1201         RETURN(ELDLM_OK);
1202 }
1203
1204 /**
1205  * Performs various cleanups for passed \a ns to make it drop refc and be
1206  * ready for freeing. Waits for refc == 0.
1207  *
1208  * The following is done:
1209  * (0) Unregister \a ns from its list to make inaccessible for potential
1210  * users like pools thread and others;
1211  * (1) Clear all locks in \a ns.
1212  */
1213 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1214                                struct obd_import *imp,
1215                                int force)
1216 {
1217         int rc;
1218
1219         ENTRY;
1220         if (!ns) {
1221                 EXIT;
1222                 return;
1223         }
1224
1225         spin_lock(&ns->ns_lock);
1226         ns->ns_stopping = 1;
1227         spin_unlock(&ns->ns_lock);
1228
1229         /*
1230          * Can fail with -EINTR when force == 0 in which case try harder.
1231          */
1232         rc = __ldlm_namespace_free(ns, force);
1233         if (rc != ELDLM_OK) {
1234                 if (imp) {
1235                         ptlrpc_disconnect_import(imp, 0);
1236                         ptlrpc_invalidate_import(imp);
1237                 }
1238
1239                 /*
1240                  * With all requests dropped and the import inactive
1241                  * we are gaurenteed all reference will be dropped.
1242                  */
1243                 rc = __ldlm_namespace_free(ns, 1);
1244                 LASSERT(rc == 0);
1245         }
1246         EXIT;
1247 }
1248 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1249
1250 /**
1251  * Performs freeing memory structures related to \a ns. This is only done
1252  * when ldlm_namespce_free_prior() successfully removed all resources
1253  * referencing \a ns and its refc == 0.
1254  */
1255 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1256 {
1257         ENTRY;
1258         if (!ns) {
1259                 EXIT;
1260                 return;
1261         }
1262
1263         /* Make sure that nobody can find this ns in its list. */
1264         ldlm_namespace_unregister(ns, ns->ns_client);
1265         /* Fini pool _before_ parent proc dir is removed. This is important as
1266          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1267          * Removing it after @dir may cause oops. */
1268         ldlm_pool_fini(&ns->ns_pool);
1269
1270         ldlm_namespace_debugfs_unregister(ns);
1271         ldlm_namespace_sysfs_unregister(ns);
1272         cfs_hash_putref(ns->ns_rs_hash);
1273         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1274         kfree(ns->ns_name);
1275         /* Namespace \a ns should be not on list at this time, otherwise
1276          * this will cause issues related to using freed \a ns in poold
1277          * thread.
1278          */
1279         LASSERT(list_empty(&ns->ns_list_chain));
1280         OBD_FREE_PTR(ns);
1281         ldlm_put_ref();
1282         EXIT;
1283 }
1284 EXPORT_SYMBOL(ldlm_namespace_free_post);
1285
1286 /**
1287  * Cleanup the resource, and free namespace.
1288  * bug 12864:
1289  * Deadlock issue:
1290  * proc1: destroy import
1291  *        class_disconnect_export(grab cl_sem) ->
1292  *              -> ldlm_namespace_free ->
1293  *              -> lprocfs_remove(grab _lprocfs_lock).
1294  * proc2: read proc info
1295  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1296  *              -> osc_rd_active, etc(grab cl_sem).
1297  *
1298  * So that I have to split the ldlm_namespace_free into two parts - the first
1299  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1300  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1301  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1302  * held.
1303  */
1304 void ldlm_namespace_free(struct ldlm_namespace *ns,
1305                          struct obd_import *imp,
1306                          int force)
1307 {
1308         ldlm_namespace_free_prior(ns, imp, force);
1309         ldlm_namespace_free_post(ns);
1310 }
1311 EXPORT_SYMBOL(ldlm_namespace_free);
1312
1313 void ldlm_namespace_get(struct ldlm_namespace *ns)
1314 {
1315         atomic_inc(&ns->ns_bref);
1316 }
1317
1318 /* This is only for callers that care about refcount */
1319 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1320 {
1321         return atomic_inc_return(&ns->ns_bref);
1322 }
1323
1324 void ldlm_namespace_put(struct ldlm_namespace *ns)
1325 {
1326         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1327                 wake_up(&ns->ns_waitq);
1328                 spin_unlock(&ns->ns_lock);
1329         }
1330 }
1331
1332 /** Register \a ns in the list of namespaces */
1333 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1334 {
1335         mutex_lock(ldlm_namespace_lock(client));
1336         LASSERT(list_empty(&ns->ns_list_chain));
1337         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1338         ldlm_namespace_nr_inc(client);
1339         mutex_unlock(ldlm_namespace_lock(client));
1340 }
1341
1342 /** Unregister \a ns from the list of namespaces. */
1343 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1344 {
1345         mutex_lock(ldlm_namespace_lock(client));
1346         LASSERT(!list_empty(&ns->ns_list_chain));
1347         /* Some asserts and possibly other parts of the code are still
1348          * using list_empty(&ns->ns_list_chain). This is why it is
1349          * important to use list_del_init() here. */
1350         list_del_init(&ns->ns_list_chain);
1351         ldlm_namespace_nr_dec(client);
1352         mutex_unlock(ldlm_namespace_lock(client));
1353 }
1354
1355 /** Should be called with ldlm_namespace_lock(client) taken. */
1356 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1357                                           enum ldlm_side client)
1358 {
1359         LASSERT(!list_empty(&ns->ns_list_chain));
1360         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1361         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1362 }
1363
1364 /** Should be called with ldlm_namespace_lock(client) taken. */
1365 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1366                                             enum ldlm_side client)
1367 {
1368         LASSERT(!list_empty(&ns->ns_list_chain));
1369         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1370         list_move_tail(&ns->ns_list_chain,
1371                        ldlm_namespace_inactive_list(client));
1372 }
1373
1374 /** Should be called with ldlm_namespace_lock(client) taken. */
1375 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1376 {
1377         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1378         LASSERT(!list_empty(ldlm_namespace_list(client)));
1379         return container_of(ldlm_namespace_list(client)->next,
1380                             struct ldlm_namespace, ns_list_chain);
1381 }
1382
1383 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1384 {
1385         int idx;
1386
1387         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1388                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1389         if (res->lr_itree == NULL)
1390                 return false;
1391         /* Initialize interval trees for each lock mode. */
1392         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1393                 res->lr_itree[idx].lit_size = 0;
1394                 res->lr_itree[idx].lit_mode = BIT(idx);
1395                 res->lr_itree[idx].lit_root = NULL;
1396         }
1397         return true;
1398 }
1399
1400 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1401 {
1402         int i;
1403
1404         OBD_ALLOC_PTR(res->lr_ibits_queues);
1405         if (res->lr_ibits_queues == NULL)
1406                 return false;
1407         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1408                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1409         return true;
1410 }
1411
1412 /** Create and initialize new resource. */
1413 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1414 {
1415         struct ldlm_resource *res;
1416         bool rc;
1417
1418         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1419         if (res == NULL)
1420                 return NULL;
1421
1422         switch (ldlm_type) {
1423         case LDLM_EXTENT:
1424                 rc = ldlm_resource_extent_new(res);
1425                 break;
1426         case LDLM_IBITS:
1427                 rc = ldlm_resource_inodebits_new(res);
1428                 break;
1429         default:
1430                 rc = true;
1431                 break;
1432         }
1433         if (!rc) {
1434                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1435                 return NULL;
1436         }
1437
1438         INIT_LIST_HEAD(&res->lr_granted);
1439         INIT_LIST_HEAD(&res->lr_waiting);
1440
1441         atomic_set(&res->lr_refcount, 1);
1442         spin_lock_init(&res->lr_lock);
1443         lu_ref_init(&res->lr_reference);
1444
1445         /* Since LVB init can be delayed now, there is no longer need to
1446          * immediatelly acquire mutex here. */
1447         mutex_init(&res->lr_lvb_mutex);
1448         res->lr_lvb_initialized = false;
1449
1450         return res;
1451 }
1452
1453 static void ldlm_resource_free(struct ldlm_resource *res)
1454 {
1455         if (res->lr_type == LDLM_EXTENT) {
1456                 if (res->lr_itree != NULL)
1457                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1458                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1459         } else if (res->lr_type == LDLM_IBITS) {
1460                 if (res->lr_ibits_queues != NULL)
1461                         OBD_FREE_PTR(res->lr_ibits_queues);
1462         }
1463
1464         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1465 }
1466
1467 /**
1468  * Return a reference to resource with given name, creating it if necessary.
1469  * Args: namespace with ns_lock unlocked
1470  * Locks: takes and releases NS hash-lock and res->lr_lock
1471  * Returns: referenced, unlocked ldlm_resource or NULL
1472  */
1473 struct ldlm_resource *
1474 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1475                   const struct ldlm_res_id *name, enum ldlm_type type,
1476                   int create)
1477 {
1478         struct hlist_node       *hnode;
1479         struct ldlm_resource    *res = NULL;
1480         struct cfs_hash_bd              bd;
1481         __u64                   version;
1482         int                     ns_refcount = 0;
1483         int hash;
1484
1485         LASSERT(ns != NULL);
1486         LASSERT(parent == NULL);
1487         LASSERT(ns->ns_rs_hash != NULL);
1488         LASSERT(name->name[0] != 0);
1489
1490         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1491         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1492         if (hnode != NULL) {
1493                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1494                 GOTO(found, res);
1495         }
1496
1497         version = cfs_hash_bd_version_get(&bd);
1498         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1499
1500         if (create == 0)
1501                 return ERR_PTR(-ENOENT);
1502
1503         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1504                  "type: %d\n", type);
1505         res = ldlm_resource_new(type);
1506         if (res == NULL)
1507                 return ERR_PTR(-ENOMEM);
1508
1509         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1510         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1511         res->lr_name = *name;
1512         res->lr_type = type;
1513
1514         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1515         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1516                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1517
1518         if (hnode != NULL) {
1519                 /* Someone won the race and already added the resource. */
1520                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1521                 /* Clean lu_ref for failed resource. */
1522                 lu_ref_fini(&res->lr_reference);
1523                 ldlm_resource_free(res);
1524 found:
1525                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1526                 return res;
1527         }
1528         /* We won! Let's add the resource. */
1529         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1530         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1531                 ns_refcount = ldlm_namespace_get_return(ns);
1532
1533         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1534
1535         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1536
1537         /* Let's see if we happened to be the very first resource in this
1538          * namespace. If so, and this is a client namespace, we need to move
1539          * the namespace into the active namespaces list to be patrolled by
1540          * the ldlm_poold. */
1541         if (ns_is_client(ns) && ns_refcount == 1) {
1542                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1543                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1544                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1545         }
1546
1547         return res;
1548 }
1549 EXPORT_SYMBOL(ldlm_resource_get);
1550
1551 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1552 {
1553         LASSERT(res != NULL);
1554         LASSERT(res != LP_POISON);
1555         atomic_inc(&res->lr_refcount);
1556         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1557                atomic_read(&res->lr_refcount));
1558         return res;
1559 }
1560
1561 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1562                                          struct ldlm_resource *res)
1563 {
1564         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1565
1566         if (!list_empty(&res->lr_granted)) {
1567                 ldlm_resource_dump(D_ERROR, res);
1568                 LBUG();
1569         }
1570
1571         if (!list_empty(&res->lr_waiting)) {
1572                 ldlm_resource_dump(D_ERROR, res);
1573                 LBUG();
1574         }
1575
1576         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1577                                bd, &res->lr_hash);
1578         lu_ref_fini(&res->lr_reference);
1579         if (atomic_dec_and_test(&nsb->nsb_count))
1580                 ldlm_namespace_put(nsb->nsb_namespace);
1581 }
1582
1583 /* Returns 1 if the resource was freed, 0 if it remains. */
1584 int ldlm_resource_putref(struct ldlm_resource *res)
1585 {
1586         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1587         struct cfs_hash_bd   bd;
1588
1589         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1590         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1591                res, atomic_read(&res->lr_refcount) - 1);
1592
1593         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1594         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1595                 __ldlm_resource_putref_final(&bd, res);
1596                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1597                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1598                         ns->ns_lvbo->lvbo_free(res);
1599                 ldlm_resource_free(res);
1600                 return 1;
1601         }
1602         return 0;
1603 }
1604 EXPORT_SYMBOL(ldlm_resource_putref);
1605
1606 static void __ldlm_resource_add_lock(struct ldlm_resource *res,
1607                                      struct list_head *head,
1608                                      struct ldlm_lock *lock,
1609                                      bool tail)
1610 {
1611         check_res_locked(res);
1612
1613         if (ldlm_is_destroyed(lock)) {
1614                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1615                 return;
1616         }
1617
1618         LASSERT(list_empty(&lock->l_res_link));
1619
1620         if (tail)
1621                 list_add_tail(&lock->l_res_link, head);
1622         else
1623                 list_add(&lock->l_res_link, head);
1624
1625         if (res->lr_type == LDLM_IBITS)
1626                 ldlm_inodebits_add_lock(res, head, lock);
1627
1628         ldlm_resource_dump(D_INFO, res);
1629 }
1630
1631 /**
1632  * Add a lock into a given resource into specified lock list.
1633  */
1634 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1635                             struct ldlm_lock *lock)
1636 {
1637         LDLM_DEBUG(lock, "About to add this lock");
1638
1639         __ldlm_resource_add_lock(res, head, lock, true);
1640 }
1641
1642 /**
1643  * Insert a lock into resource after specified lock.
1644  */
1645 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1646                                      struct ldlm_lock *new)
1647 {
1648         LASSERT(!list_empty(&original->l_res_link));
1649
1650         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1651         __ldlm_resource_add_lock(original->l_resource,
1652                                  &original->l_res_link,
1653                                  new, false);
1654 }
1655
1656 /**
1657  * Insert a lock into resource before the specified lock.
1658  */
1659 void ldlm_resource_insert_lock_before(struct ldlm_lock *original,
1660                                       struct ldlm_lock *new)
1661 {
1662         LASSERT(!list_empty(&original->l_res_link));
1663
1664         LDLM_DEBUG(new, "About to insert this lock before %p: ", original);
1665         __ldlm_resource_add_lock(original->l_resource,
1666                                  original->l_res_link.prev, new, false);
1667 }
1668
1669 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1670 {
1671         int type = lock->l_resource->lr_type;
1672
1673         check_res_locked(lock->l_resource);
1674         switch (type) {
1675         case LDLM_PLAIN:
1676                 ldlm_unlink_lock_skiplist(lock);
1677                 break;
1678         case LDLM_EXTENT:
1679                 ldlm_extent_unlink_lock(lock);
1680                 break;
1681         case LDLM_IBITS:
1682                 ldlm_inodebits_unlink_lock(lock);
1683                 break;
1684         }
1685         list_del_init(&lock->l_res_link);
1686 }
1687 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1688
1689 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1690 {
1691         desc->lr_type = res->lr_type;
1692         desc->lr_name = res->lr_name;
1693 }
1694
1695 /**
1696  * Print information about all locks in all namespaces on this node to debug
1697  * log.
1698  */
1699 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1700 {
1701         struct list_head *tmp;
1702
1703         if (!((libcfs_debug | D_ERROR) & level))
1704                 return;
1705
1706         mutex_lock(ldlm_namespace_lock(client));
1707
1708         list_for_each(tmp, ldlm_namespace_list(client)) {
1709                 struct ldlm_namespace *ns;
1710
1711                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1712                 ldlm_namespace_dump(level, ns);
1713         }
1714
1715         mutex_unlock(ldlm_namespace_lock(client));
1716 }
1717
1718 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1719                               struct hlist_node *hnode, void *arg)
1720 {
1721         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1722         int    level = (int)(unsigned long)arg;
1723
1724         lock_res(res);
1725         ldlm_resource_dump(level, res);
1726         unlock_res(res);
1727
1728         return 0;
1729 }
1730
1731 /**
1732  * Print information about all locks in this namespace on this node to debug
1733  * log.
1734  */
1735 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1736 {
1737         if (!((libcfs_debug | D_ERROR) & level))
1738                 return;
1739
1740         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1741                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1742                ns_is_client(ns) ? "client" : "server");
1743
1744         if (ktime_get_seconds() < ns->ns_next_dump)
1745                 return;
1746
1747         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1748                                  ldlm_res_hash_dump,
1749                                  (void *)(unsigned long)level, 0);
1750         spin_lock(&ns->ns_lock);
1751         ns->ns_next_dump = ktime_get_seconds() + 10;
1752         spin_unlock(&ns->ns_lock);
1753 }
1754
1755 /**
1756  * Print information about all locks in this resource to debug log.
1757  */
1758 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1759 {
1760         struct ldlm_lock *lock;
1761         unsigned int granted = 0;
1762
1763         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1764
1765         if (!((libcfs_debug | D_ERROR) & level))
1766                 return;
1767
1768         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1769                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1770
1771         if (!list_empty(&res->lr_granted)) {
1772                 CDEBUG(level, "Granted locks (in reverse order):\n");
1773                 list_for_each_entry_reverse(lock, &res->lr_granted,
1774                                                 l_res_link) {
1775                         LDLM_DEBUG_LIMIT(level, lock, "###");
1776                         if (!(level & D_CANTMASK) &&
1777                             ++granted > ldlm_dump_granted_max) {
1778                                 CDEBUG(level,
1779                                        "only dump %d granted locks to avoid DDOS.\n",
1780                                        granted);
1781                                 break;
1782                         }
1783                 }
1784         }
1785
1786         if (!list_empty(&res->lr_waiting)) {
1787                 CDEBUG(level, "Waiting locks:\n");
1788                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1789                         LDLM_DEBUG_LIMIT(level, lock, "###");
1790         }
1791 }
1792 EXPORT_SYMBOL(ldlm_resource_dump);