lustre/ldlm/ldlm_resource.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ldlm/ldlm_resource.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43 #ifdef __KERNEL__
  44 # include <lustre_dlm.h>
  45 #else
  46 # include <liblustre.h>
  47 #endif
  48
  49 #include <lustre_fid.h>
  50 #include <obd_class.h>
  51 #include "ldlm_internal.h"
  52
  53 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
  54
  55 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
  56 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
  57
  58 cfs_semaphore_t ldlm_srv_namespace_lock;
  59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
  60
  61 cfs_semaphore_t ldlm_cli_namespace_lock;
  62 CFS_LIST_HEAD(ldlm_cli_namespace_list);
  63
  64 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
  65 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
  66 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
  67
  68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
  69
  70 #ifdef LPROCFS
  71 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
  72                              unsigned long count, void *data)
  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 int ldlm_proc_setup(void)
  80 {
  81         int rc;
  82         struct lprocfs_vars list[] = {
  83                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
  84                 { "cancel_unused_locks_before_replay",
  85                   lprocfs_rd_uint, lprocfs_wr_uint,
  86                   &ldlm_cancel_unused_locks_before_replay, NULL },
  87                 { NULL }};
  88         ENTRY;
  89         LASSERT(ldlm_ns_proc_dir == NULL);
  90
  91         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
  92                                               proc_lustre_root,
  93                                               NULL, NULL);
  94         if (IS_ERR(ldlm_type_proc_dir)) {
  95                 CERROR("LProcFS failed in ldlm-init\n");
  96                 rc = PTR_ERR(ldlm_type_proc_dir);
  97                 GOTO(err, rc);
  98         }
  99
 100         ldlm_ns_proc_dir = lprocfs_register("namespaces",
 101                                             ldlm_type_proc_dir,
 102                                             NULL, NULL);
 103         if (IS_ERR(ldlm_ns_proc_dir)) {
 104                 CERROR("LProcFS failed in ldlm-init\n");
 105                 rc = PTR_ERR(ldlm_ns_proc_dir);
 106                 GOTO(err_type, rc);
 107         }
 108
 109         ldlm_svc_proc_dir = lprocfs_register("services",
 110                                             ldlm_type_proc_dir,
 111                                             NULL, NULL);
 112         if (IS_ERR(ldlm_svc_proc_dir)) {
 113                 CERROR("LProcFS failed in ldlm-init\n");
 114                 rc = PTR_ERR(ldlm_svc_proc_dir);
 115                 GOTO(err_ns, rc);
 116         }
 117
 118         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
 119
 120         RETURN(0);
 121
 122 err_ns:
 123         lprocfs_remove(&ldlm_ns_proc_dir);
 124 err_type:
 125         lprocfs_remove(&ldlm_type_proc_dir);
 126 err:
 127         ldlm_svc_proc_dir = NULL;
 128         RETURN(rc);
 129 }
 130
 131 void ldlm_proc_cleanup(void)
 132 {
 133         if (ldlm_svc_proc_dir)
 134                 lprocfs_remove(&ldlm_svc_proc_dir);
 135
 136         if (ldlm_ns_proc_dir)
 137                 lprocfs_remove(&ldlm_ns_proc_dir);
 138
 139         if (ldlm_type_proc_dir)
 140                 lprocfs_remove(&ldlm_type_proc_dir);
 141 }
 142
 143 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
 144                                    int count, int *eof, void *data)
 145 {
 146         struct ldlm_namespace *ns  = data;
 147         __u64                  res = 0;
 148         cfs_hash_bd_t          bd;
 149         int                    i;
 150
 151         /* result is not strictly consistant */
 152         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
 153                 res += cfs_hash_bd_count_get(&bd);
 154         return lprocfs_rd_u64(page, start, off, count, eof, &res);
 155 }
 156
 157 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
 158                                int count, int *eof, void *data)
 159 {
 160         struct ldlm_namespace *ns = data;
 161         __u64                  locks;
 162
 163         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 164                                         LPROCFS_FIELDS_FLAGS_SUM);
 165         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
 166 }
 167
 168 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
 169                                int count, int *eof, void *data)
 170 {
 171         struct ldlm_namespace *ns = data;
 172         __u32 *nr = &ns->ns_max_unused;
 173
 174         if (ns_connect_lru_resize(ns))
 175                 nr = &ns->ns_nr_unused;
 176         return lprocfs_rd_uint(page, start, off, count, eof, nr);
 177 }
 178
 179 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
 180                                unsigned long count, void *data)
 181 {
 182         struct ldlm_namespace *ns = data;
 183         char dummy[MAX_STRING_SIZE + 1], *end;
 184         unsigned long tmp;
 185         int lru_resize;
 186
 187         dummy[MAX_STRING_SIZE] = '\0';
 188         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
 189                 return -EFAULT;
 190
 191         if (strncmp(dummy, "clear", 5) == 0) {
 192                 CDEBUG(D_DLMTRACE,
 193                        "dropping all unused locks from namespace %s\n",
 194                        ldlm_ns_name(ns));
 195                 if (ns_connect_lru_resize(ns)) {
 196                         int canceled, unused  = ns->ns_nr_unused;
 197
 198                         /* Try to cancel all @ns_nr_unused locks. */
 199                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
 200                                                    LDLM_CANCEL_PASSED);
 201                         if (canceled < unused) {
 202                                 CDEBUG(D_DLMTRACE,
 203                                        "not all requested locks are canceled, "
 204                                        "requested: %d, canceled: %d\n", unused,
 205                                        canceled);
 206                                 return -EINVAL;
 207                         }
 208                 } else {
 209                         tmp = ns->ns_max_unused;
 210                         ns->ns_max_unused = 0;
 211                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
 212                         ns->ns_max_unused = tmp;
 213                 }
 214                 return count;
 215         }
 216
 217         tmp = simple_strtoul(dummy, &end, 0);
 218         if (dummy == end) {
 219                 CERROR("invalid value written\n");
 220                 return -EINVAL;
 221         }
 222         lru_resize = (tmp == 0);
 223
 224         if (ns_connect_lru_resize(ns)) {
 225                 if (!lru_resize)
 226                         ns->ns_max_unused = (unsigned int)tmp;
 227
 228                 if (tmp > ns->ns_nr_unused)
 229                         tmp = ns->ns_nr_unused;
 230                 tmp = ns->ns_nr_unused - tmp;
 231
 232                 CDEBUG(D_DLMTRACE,
 233                        "changing namespace %s unused locks from %u to %u\n",
 234                        ldlm_ns_name(ns), ns->ns_nr_unused,
 235                        (unsigned int)tmp);
 236                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 237
 238                 if (!lru_resize) {
 239                         CDEBUG(D_DLMTRACE,
 240                                "disable lru_resize for namespace %s\n",
 241                                ldlm_ns_name(ns));
 242                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 243                 }
 244         } else {
 245                 CDEBUG(D_DLMTRACE,
 246                        "changing namespace %s max_unused from %u to %u\n",
 247                        ldlm_ns_name(ns), ns->ns_max_unused,
 248                        (unsigned int)tmp);
 249                 ns->ns_max_unused = (unsigned int)tmp;
 250                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
 251
 252                 /* Make sure that originally lru resize was supported before
 253                  * turning it on here. */
 254                 if (lru_resize &&
 255                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 256                         CDEBUG(D_DLMTRACE,
 257                                "enable lru_resize for namespace %s\n",
 258                                ldlm_ns_name(ns));
 259                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 260                 }
 261         }
 262
 263         return count;
 264 }
 265
 266 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
 267 {
 268         struct proc_dir_entry *dir;
 269
 270         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
 271         if (dir == NULL) {
 272                 CERROR("dlm namespace %s has no procfs dir?\n",
 273                        ldlm_ns_name(ns));
 274         } else {
 275                 lprocfs_remove(&dir);
 276         }
 277
 278         if (ns->ns_stats != NULL)
 279                 lprocfs_free_stats(&ns->ns_stats);
 280 }
 281
 282 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
 283 {
 284         struct lprocfs_vars lock_vars[2];
 285         char lock_name[MAX_STRING_SIZE + 1];
 286
 287         LASSERT(ns != NULL);
 288         LASSERT(ns->ns_rs_hash != NULL);
 289
 290         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 291         if (ns->ns_stats == NULL)
 292                 return -ENOMEM;
 293
 294         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 295                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 296
 297         lock_name[MAX_STRING_SIZE] = '\0';
 298
 299         memset(lock_vars, 0, sizeof(lock_vars));
 300         lock_vars[0].name = lock_name;
 301
 302         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
 303                  ldlm_ns_name(ns));
 304         lock_vars[0].data = ns;
 305         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
 306         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 307
 308         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
 309                  ldlm_ns_name(ns));
 310         lock_vars[0].data = ns;
 311         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
 312         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 313
 314         if (ns_is_client(ns)) {
 315                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
 316                          ldlm_ns_name(ns));
 317                 lock_vars[0].data = &ns->ns_nr_unused;
 318                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 319                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 320
 321                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
 322                          ldlm_ns_name(ns));
 323                 lock_vars[0].data = ns;
 324                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
 325                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
 326                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 327
 328                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
 329                          ldlm_ns_name(ns));
 330                 lock_vars[0].data = &ns->ns_max_age;
 331                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 332                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 333                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 334         } else {
 335                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
 336                          ldlm_ns_name(ns));
 337                 lock_vars[0].data = &ns->ns_ctime_age_limit;
 338                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 339                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 340                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 341
 342                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
 343                          ldlm_ns_name(ns));
 344                 lock_vars[0].data = &ns->ns_timeouts;
 345                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 346                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 347
 348                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
 349                          ldlm_ns_name(ns));
 350                 lock_vars[0].data = &ns->ns_max_nolock_size;
 351                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 352                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 354
 355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
 356                          ldlm_ns_name(ns));
 357                 lock_vars[0].data = &ns->ns_contention_time;
 358                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 359                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 360                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 361
 362                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
 363                          ldlm_ns_name(ns));
 364                 lock_vars[0].data = &ns->ns_contended_locks;
 365                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 366                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 367                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 368
 369                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
 370                          ldlm_ns_name(ns));
 371                 lock_vars[0].data = &ns->ns_max_parallel_ast;
 372                 lock_vars[0].read_fptr = lprocfs_rd_uint;
 373                 lock_vars[0].write_fptr = lprocfs_wr_uint;
 374                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 375         }
 376         return 0;
 377 }
 378 #undef MAX_STRING_SIZE
 379 #else /* LPROCFS */
 380
 381 #define ldlm_namespace_proc_unregister(ns)      ({;})
 382 #define ldlm_namespace_proc_register(ns)        ({0;})
 383
 384 #endif /* LPROCFS */
 385
 386 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
 387                                   const void *key, unsigned mask)
 388 {
 389         const struct ldlm_res_id     *id  = key;
 390         unsigned                val = 0;
 391         unsigned                i;
 392
 393         for (i = 0; i < RES_NAME_SIZE; i++)
 394                 val += id->name[i];
 395         return val & mask;
 396 }
 397
 398 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
 399                                       const void *key, unsigned mask)
 400 {
 401         const struct ldlm_res_id *id = key;
 402         struct lu_fid       fid;
 403         __u64               hash;
 404
 405         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 406         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_OID_OFF];
 407         fid.f_ver = (__u32)id->name[LUSTRE_RES_ID_VER_OFF];
 408
 409         hash = fid_flatten(&fid);
 410         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
 411         /* ignore a few low bits */
 412         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0)
 413                 hash += id->name[LUSTRE_RES_ID_HSH_OFF] >> 5;
 414         else
 415                 hash = hash >> 5;
 416         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 417         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
 418
 419         return hash & mask;
 420 }
 421
 422 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
 423 {
 424         struct ldlm_resource   *res;
 425
 426         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 427         return &res->lr_name;
 428 }
 429
 430 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
 431 {
 432         struct ldlm_resource   *res;
 433
 434         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 435         return ldlm_res_eq((const struct ldlm_res_id *)key,
 436                            (const struct ldlm_res_id *)&res->lr_name);
 437 }
 438
 439 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
 440 {
 441         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 442 }
 443
 444 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 445 {
 446         struct ldlm_resource *res;
 447
 448         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 449         ldlm_resource_getref(res);
 450         LDLM_RESOURCE_ADDREF(res);
 451 }
 452
 453 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 454 {
 455         struct ldlm_resource *res;
 456
 457         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 458         /* cfs_hash_for_each_nolock is the only chance we call it */
 459         LDLM_RESOURCE_DELREF(res);
 460         ldlm_resource_putref_locked(res);
 461 }
 462
 463 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
 464 {
 465         struct ldlm_resource *res;
 466
 467         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
 468         LDLM_RESOURCE_DELREF(res);
 469         ldlm_resource_putref(res);
 470 }
 471
 472 cfs_hash_ops_t ldlm_ns_hash_ops = {
 473         .hs_hash        = ldlm_res_hop_hash,
 474         .hs_key         = ldlm_res_hop_key,
 475         .hs_keycmp      = ldlm_res_hop_keycmp,
 476         .hs_keycpy      = NULL,
 477         .hs_object      = ldlm_res_hop_object,
 478         .hs_get         = ldlm_res_hop_get_locked,
 479         .hs_put_locked  = ldlm_res_hop_put_locked,
 480         .hs_put         = ldlm_res_hop_put
 481 };
 482
 483 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
 484         .hs_hash        = ldlm_res_hop_fid_hash,
 485         .hs_key         = ldlm_res_hop_key,
 486         .hs_keycmp      = ldlm_res_hop_keycmp,
 487         .hs_keycpy      = NULL,
 488         .hs_object      = ldlm_res_hop_object,
 489         .hs_get         = ldlm_res_hop_get_locked,
 490         .hs_put_locked  = ldlm_res_hop_put_locked,
 491         .hs_put         = ldlm_res_hop_put
 492 };
 493
 494 typedef struct {
 495         ldlm_ns_type_t  nsd_type;
 496         /** hash bucket bits */
 497         unsigned        nsd_bkt_bits;
 498         /** hash bits */
 499         unsigned        nsd_all_bits;
 500         /** hash operations */
 501         cfs_hash_ops_t *nsd_hops;
 502 } ldlm_ns_hash_def_t;
 503
 504 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
 505 {
 506         {
 507                 .nsd_type       = LDLM_NS_TYPE_MDC,
 508                 .nsd_bkt_bits   = 11,
 509                 .nsd_all_bits   = 15,
 510                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 511         },
 512         {
 513                 .nsd_type       = LDLM_NS_TYPE_MDT,
 514                 .nsd_bkt_bits   = 14,
 515                 .nsd_all_bits   = 21,
 516                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
 517         },
 518         {
 519                 .nsd_type       = LDLM_NS_TYPE_OSC,
 520                 .nsd_bkt_bits   = 8,
 521                 .nsd_all_bits   = 12,
 522                 .nsd_hops       = &ldlm_ns_hash_ops,
 523         },
 524         {
 525                 .nsd_type       = LDLM_NS_TYPE_OST,
 526                 .nsd_bkt_bits   = 11,
 527                 .nsd_all_bits   = 17,
 528                 .nsd_hops       = &ldlm_ns_hash_ops,
 529         },
 530         {
 531                 .nsd_type       = LDLM_NS_TYPE_MGC,
 532                 .nsd_bkt_bits   = 4,
 533                 .nsd_all_bits   = 4,
 534                 .nsd_hops       = &ldlm_ns_hash_ops,
 535         },
 536         {
 537                 .nsd_type       = LDLM_NS_TYPE_MGT,
 538                 .nsd_bkt_bits   = 4,
 539                 .nsd_all_bits   = 4,
 540                 .nsd_hops       = &ldlm_ns_hash_ops,
 541         },
 542         {
 543                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
 544         },
 545 };
 546
 547 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 548                                           ldlm_side_t client,
 549                                           ldlm_appetite_t apt,
 550                                           ldlm_ns_type_t ns_type)
 551 {
 552         struct ldlm_namespace *ns = NULL;
 553         struct ldlm_ns_bucket *nsb;
 554         ldlm_ns_hash_def_t    *nsd;
 555         cfs_hash_bd_t          bd;
 556         int                    idx;
 557         int                    rc;
 558         ENTRY;
 559
 560         LASSERT(obd != NULL);
 561
 562         rc = ldlm_get_ref();
 563         if (rc) {
 564                 CERROR("ldlm_get_ref failed: %d\n", rc);
 565                 RETURN(NULL);
 566         }
 567
 568         for (idx = 0;;idx++) {
 569                 nsd = &ldlm_ns_hash_defs[idx];
 570                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
 571                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
 572                         GOTO(out_ref, NULL);
 573                 }
 574
 575                 if (nsd->nsd_type == ns_type)
 576                         break;
 577         }
 578
 579         OBD_ALLOC_PTR(ns);
 580         if (!ns)
 581                 GOTO(out_ref, NULL);
 582
 583         ns->ns_rs_hash = cfs_hash_create(name,
 584                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
 585                                          nsd->nsd_bkt_bits, sizeof(*nsb),
 586                                          CFS_HASH_MIN_THETA,
 587                                          CFS_HASH_MAX_THETA,
 588                                          nsd->nsd_hops,
 589                                          CFS_HASH_DEPTH |
 590                                          CFS_HASH_BIGNAME |
 591                                          CFS_HASH_SPIN_BKTLOCK |
 592                                          CFS_HASH_NO_ITEMREF);
 593         if (ns->ns_rs_hash == NULL)
 594                 GOTO(out_ns, NULL);
 595
 596         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
 597                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
 598                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 599                 nsb->nsb_namespace = ns;
 600         }
 601
 602         ns->ns_obd      = obd;
 603         ns->ns_appetite = apt;
 604         ns->ns_client   = client;
 605
 606         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
 607         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
 608         cfs_spin_lock_init(&ns->ns_lock);
 609         cfs_atomic_set(&ns->ns_bref, 0);
 610         cfs_waitq_init(&ns->ns_waitq);
 611
 612         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 613         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 614         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 615
 616         ns->ns_nr_unused          = 0;
 617         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 618         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
 619         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 620         ns->ns_timeouts           = 0;
 621         ns->ns_orig_connect_flags = 0;
 622         ns->ns_connect_flags      = 0;
 623         ns->ns_stopping           = 0;
 624         rc = ldlm_namespace_proc_register(ns);
 625         if (rc != 0) {
 626                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 627                 GOTO(out_hash, rc);
 628         }
 629
 630         idx = cfs_atomic_read(ldlm_namespace_nr(client));
 631         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 632         if (rc) {
 633                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 634                 GOTO(out_proc, rc);
 635         }
 636
 637         ldlm_namespace_register(ns, client);
 638         RETURN(ns);
 639 out_proc:
 640         ldlm_namespace_proc_unregister(ns);
 641         ldlm_namespace_cleanup(ns, 0);
 642 out_hash:
 643         cfs_hash_putref(ns->ns_rs_hash);
 644 out_ns:
 645         OBD_FREE_PTR(ns);
 646 out_ref:
 647         ldlm_put_ref();
 648         RETURN(NULL);
 649 }
 650
 651 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 652
 653 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
 654  * This is currently only used for recovery, and we make certain assumptions
 655  * as a result--notably, that we shouldn't cancel locks with refs. -phil */
 656 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
 657                              int flags)
 658 {
 659         cfs_list_t *tmp;
 660         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 661         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
 662
 663         do {
 664                 struct ldlm_lock *lock = NULL;
 665
 666                 /* first, we look for non-cleaned-yet lock
 667                  * all cleaned locks are marked by CLEANED flag */
 668                 lock_res(res);
 669                 cfs_list_for_each(tmp, q) {
 670                         lock = cfs_list_entry(tmp, struct ldlm_lock,
 671                                               l_res_link);
 672                         if (lock->l_flags & LDLM_FL_CLEANED) {
 673                                 lock = NULL;
 674                                 continue;
 675                         }
 676                         LDLM_LOCK_GET(lock);
 677                         lock->l_flags |= LDLM_FL_CLEANED;
 678                         break;
 679                 }
 680
 681                 if (lock == NULL) {
 682                         unlock_res(res);
 683                         break;
 684                 }
 685
 686                 /* Set CBPENDING so nothing in the cancellation path
 687                  * can match this lock */
 688                 lock->l_flags |= LDLM_FL_CBPENDING;
 689                 lock->l_flags |= LDLM_FL_FAILED;
 690                 lock->l_flags |= flags;
 691
 692                 /* ... without sending a CANCEL message for local_only. */
 693                 if (local_only)
 694                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
 695
 696                 if (local_only && (lock->l_readers || lock->l_writers)) {
 697                         /* This is a little bit gross, but much better than the
 698                          * alternative: pretend that we got a blocking AST from
 699                          * the server, so that when the lock is decref'd, it
 700                          * will go away ... */
 701                         unlock_res(res);
 702                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
 703                         if (lock->l_completion_ast)
 704                                 lock->l_completion_ast(lock, 0, NULL);
 705                         LDLM_LOCK_RELEASE(lock);
 706                         continue;
 707                 }
 708
 709                 if (client) {
 710                         struct lustre_handle lockh;
 711
 712                         unlock_res(res);
 713                         ldlm_lock2handle(lock, &lockh);
 714                         rc = ldlm_cli_cancel(&lockh);
 715                         if (rc)
 716                                 CERROR("ldlm_cli_cancel: %d\n", rc);
 717                 } else {
 718                         ldlm_resource_unlink_lock(lock);
 719                         unlock_res(res);
 720                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
 721                                    "client node");
 722                         ldlm_lock_destroy(lock);
 723                 }
 724                 LDLM_LOCK_RELEASE(lock);
 725         } while (1);
 726 }
 727
 728 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 729                                cfs_hlist_node_t *hnode, void *arg)
 730 {
 731         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
 732         int    flags = (int)(unsigned long)arg;
 733
 734         cleanup_resource(res, &res->lr_granted, flags);
 735         cleanup_resource(res, &res->lr_converting, flags);
 736         cleanup_resource(res, &res->lr_waiting, flags);
 737
 738         return 0;
 739 }
 740
 741 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 742                                   cfs_hlist_node_t *hnode, void *arg)
 743 {
 744         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
 745
 746         CERROR("Namespace %s resource refcount nonzero "
 747                "(%d) after lock cleanup; forcing "
 748                "cleanup.\n",
 749                ldlm_ns_name(ldlm_res_to_ns(res)),
 750                cfs_atomic_read(&res->lr_refcount) - 1);
 751
 752         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
 753                LPU64") (rc: %d)\n", res,
 754                res->lr_name.name[0], res->lr_name.name[1],
 755                res->lr_name.name[2], res->lr_name.name[3],
 756                cfs_atomic_read(&res->lr_refcount) - 1);
 757         return 0;
 758 }
 759
 760 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
 761 {
 762         if (ns == NULL) {
 763                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
 764                 return ELDLM_OK;
 765         }
 766
 767         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
 768                                  (void *)(unsigned long)flags);
 769         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
 770         return ELDLM_OK;
 771 }
 772
 773 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
 774 {
 775         ENTRY;
 776
 777         /* At shutdown time, don't call the cancellation callback */
 778         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
 779
 780         if (cfs_atomic_read(&ns->ns_bref) > 0) {
 781                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 782                 int rc;
 783                 CDEBUG(D_DLMTRACE,
 784                        "dlm namespace %s free waiting on refcount %d\n",
 785                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
 786 force_wait:
 787                 if (force)
 788                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
 789
 790                 rc = l_wait_event(ns->ns_waitq,
 791                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
 792
 793                 /* Forced cleanups should be able to reclaim all references,
 794                  * so it's safe to wait forever... we can't leak locks... */
 795                 if (force && rc == -ETIMEDOUT) {
 796                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
 797                                        "namespace with %d resources in use, "
 798                                        "(rc=%d)\n", ldlm_ns_name(ns),
 799                                        cfs_atomic_read(&ns->ns_bref), rc);
 800                         GOTO(force_wait, rc);
 801                 }
 802
 803                 if (cfs_atomic_read(&ns->ns_bref)) {
 804                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
 805                                        "with %d resources in use, (rc=%d)\n",
 806                                        ldlm_ns_name(ns),
 807                                        cfs_atomic_read(&ns->ns_bref), rc);
 808                         RETURN(ELDLM_NAMESPACE_EXISTS);
 809                 }
 810                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
 811                        ldlm_ns_name(ns));
 812         }
 813
 814         RETURN(ELDLM_OK);
 815 }
 816
 817 /**
 818  * Performs various cleanups for passed \a ns to make it drop refc and be ready
 819  * for freeing. Waits for refc == 0.
 820  *
 821  * The following is done:
 822  * (0) Unregister \a ns from its list to make inaccessible for potential users
 823  * like pools thread and others;
 824  * (1) Clear all locks in \a ns.
 825  */
 826 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
 827                                struct obd_import *imp,
 828                                int force)
 829 {
 830         int rc;
 831         ENTRY;
 832         if (!ns) {
 833                 EXIT;
 834                 return;
 835         }
 836
 837         cfs_spin_lock(&ns->ns_lock);
 838         ns->ns_stopping = 1;
 839         cfs_spin_unlock(&ns->ns_lock);
 840
 841         /*
 842          * Can fail with -EINTR when force == 0 in which case try harder.
 843          */
 844         rc = __ldlm_namespace_free(ns, force);
 845         if (rc != ELDLM_OK) {
 846                 if (imp) {
 847                         ptlrpc_disconnect_import(imp, 0);
 848                         ptlrpc_invalidate_import(imp);
 849                 }
 850
 851                 /*
 852                  * With all requests dropped and the import inactive
 853                  * we are gaurenteed all reference will be dropped.
 854                  */
 855                 rc = __ldlm_namespace_free(ns, 1);
 856                 LASSERT(rc == 0);
 857         }
 858         EXIT;
 859 }
 860
 861 /**
 862  * Performs freeing memory structures related to \a ns. This is only done when
 863  * ldlm_namespce_free_prior() successfully removed all resources referencing
 864  * \a ns and its refc == 0.
 865  */
 866 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
 867 {
 868         ENTRY;
 869         if (!ns) {
 870                 EXIT;
 871                 return;
 872         }
 873
 874
 875         /*
 876          * Make sure that nobody can find this ns in its list.
 877          */
 878         ldlm_namespace_unregister(ns, ns->ns_client);
 879         /*
 880          * Fini pool _before_ parent proc dir is removed. This is important as
 881          * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
 882          * it after @dir may cause oops.
 883          */
 884         ldlm_pool_fini(&ns->ns_pool);
 885
 886         ldlm_namespace_proc_unregister(ns);
 887         cfs_hash_putref(ns->ns_rs_hash);
 888         /*
 889          * Namespace \a ns should be not on list in this time, otherwise this
 890          * will cause issues realted to using freed \a ns in pools thread.
 891          */
 892         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 893         OBD_FREE_PTR(ns);
 894         ldlm_put_ref();
 895         EXIT;
 896 }
 897
 898
 899 /* Cleanup the resource, and free namespace.
 900  * bug 12864:
 901  * Deadlock issue:
 902  * proc1: destroy import
 903  *        class_disconnect_export(grab cl_sem) ->
 904  *              -> ldlm_namespace_free ->
 905  *              -> lprocfs_remove(grab _lprocfs_lock).
 906  * proc2: read proc info
 907  *        lprocfs_fops_read(grab _lprocfs_lock) ->
 908  *              -> osc_rd_active, etc(grab cl_sem).
 909  *
 910  * So that I have to split the ldlm_namespace_free into two parts - the first
 911  * part ldlm_namespace_free_prior is used to cleanup the resource which is
 912  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
 913  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
 914  * held.
 915  */
 916 void ldlm_namespace_free(struct ldlm_namespace *ns,
 917                          struct obd_import *imp,
 918                          int force)
 919 {
 920         ldlm_namespace_free_prior(ns, imp, force);
 921         ldlm_namespace_free_post(ns);
 922 }
 923
 924 void ldlm_namespace_get(struct ldlm_namespace *ns)
 925 {
 926         cfs_atomic_inc(&ns->ns_bref);
 927 }
 928
 929 void ldlm_namespace_put(struct ldlm_namespace *ns)
 930 {
 931         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
 932                 cfs_waitq_signal(&ns->ns_waitq);
 933                 cfs_spin_unlock(&ns->ns_lock);
 934         }
 935 }
 936
 937 /* Register @ns in the list of namespaces */
 938 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
 939 {
 940         cfs_mutex_down(ldlm_namespace_lock(client));
 941         LASSERT(cfs_list_empty(&ns->ns_list_chain));
 942         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
 943         cfs_atomic_inc(ldlm_namespace_nr(client));
 944         cfs_mutex_up(ldlm_namespace_lock(client));
 945 }
 946
 947 /* Unregister @ns from the list of namespaces */
 948 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
 949 {
 950         cfs_mutex_down(ldlm_namespace_lock(client));
 951         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 952         /*
 953          * Some asserts and possibly other parts of code still using
 954          * list_empty(&ns->ns_list_chain). This is why it is important
 955          * to use list_del_init() here.
 956          */
 957         cfs_list_del_init(&ns->ns_list_chain);
 958         cfs_atomic_dec(ldlm_namespace_nr(client));
 959         cfs_mutex_up(ldlm_namespace_lock(client));
 960 }
 961
 962 /* Should be called under ldlm_namespace_lock(client) taken */
 963 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
 964 {
 965         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
 966         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 967         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
 968 }
 969
 970 /* Should be called under ldlm_namespace_lock(client) taken */
 971 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
 972 {
 973         LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
 974         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
 975         return container_of(ldlm_namespace_list(client)->next,
 976                 struct ldlm_namespace, ns_list_chain);
 977 }
 978
 979 static struct ldlm_resource *ldlm_resource_new(void)
 980 {
 981         struct ldlm_resource *res;
 982         int idx;
 983
 984         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
 985         if (res == NULL)
 986                 return NULL;
 987
 988         CFS_INIT_LIST_HEAD(&res->lr_granted);
 989         CFS_INIT_LIST_HEAD(&res->lr_converting);
 990         CFS_INIT_LIST_HEAD(&res->lr_waiting);
 991
 992         /* initialize interval trees for each lock mode*/
 993         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
 994                 res->lr_itree[idx].lit_size = 0;
 995                 res->lr_itree[idx].lit_mode = 1 << idx;
 996                 res->lr_itree[idx].lit_root = NULL;
 997         }
 998
 999         cfs_atomic_set(&res->lr_refcount, 1);
1000         cfs_spin_lock_init(&res->lr_lock);
1001         lu_ref_init(&res->lr_reference);
1002
1003         /* one who creates the resource must unlock
1004          * the semaphore after lvb initialization */
1005         cfs_init_mutex_locked(&res->lr_lvb_sem);
1006
1007         return res;
1008 }
1009
1010 /* Args: unlocked namespace
1011  *  * Locks: takes and releases NS hash-lock and res->lr_lock
1012  *   * Returns: referenced, unlocked ldlm_resource or NULL */
1013 struct ldlm_resource *
1014 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1015                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1016 {
1017         cfs_hlist_node_t     *hnode;
1018         struct ldlm_resource *res;
1019         cfs_hash_bd_t         bd;
1020         __u64                 version;
1021
1022         LASSERT(ns != NULL);
1023         LASSERT(parent == NULL);
1024         LASSERT(ns->ns_rs_hash != NULL);
1025         LASSERT(name->name[0] != 0);
1026
1027         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1028         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1029         if (hnode != NULL) {
1030                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1031                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1032                 /* synchronize WRT resource creation */
1033                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1034                         cfs_down(&res->lr_lvb_sem);
1035                         cfs_up(&res->lr_lvb_sem);
1036                 }
1037                 return res;
1038         }
1039
1040         version = cfs_hash_bd_version_get(&bd);
1041         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1042
1043         if (create == 0)
1044                 return NULL;
1045
1046         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1047                  "type: %d\n", type);
1048         res = ldlm_resource_new();
1049         if (!res)
1050                 return NULL;
1051
1052         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1053         res->lr_name       = *name;
1054         res->lr_type       = type;
1055         res->lr_most_restr = LCK_NL;
1056
1057         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1058         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1059                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1060
1061         if (hnode != NULL) {
1062                 /* someone won the race and added the resource before */
1063                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1064                 /* clean lu_ref for failed resource */
1065                 lu_ref_fini(&res->lr_reference);
1066                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1067
1068                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1069                 /* synchronize WRT resource creation */
1070                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1071                         cfs_down(&res->lr_lvb_sem);
1072                         cfs_up(&res->lr_lvb_sem);
1073                 }
1074                 return res;
1075         }
1076         /* we won! let's add the resource */
1077         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1078         if (cfs_hash_bd_count_get(&bd) == 1)
1079                 ldlm_namespace_get(ns);
1080
1081         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1082         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1083                 int rc;
1084
1085                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1086                 rc = ns->ns_lvbo->lvbo_init(res);
1087                 if (rc)
1088                         CERROR("lvbo_init failed for resource "
1089                                LPU64": rc %d\n", name->name[0], rc);
1090                 /* we create resource with locked lr_lvb_sem */
1091                 cfs_up(&res->lr_lvb_sem);
1092         }
1093
1094         return res;
1095 }
1096
1097 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1098 {
1099         LASSERT(res != NULL);
1100         LASSERT(res != LP_POISON);
1101         cfs_atomic_inc(&res->lr_refcount);
1102         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1103                cfs_atomic_read(&res->lr_refcount));
1104         return res;
1105 }
1106
1107 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1108                                          struct ldlm_resource *res)
1109 {
1110         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1111
1112         if (!cfs_list_empty(&res->lr_granted)) {
1113                 ldlm_resource_dump(D_ERROR, res);
1114                 LBUG();
1115         }
1116
1117         if (!cfs_list_empty(&res->lr_converting)) {
1118                 ldlm_resource_dump(D_ERROR, res);
1119                 LBUG();
1120         }
1121
1122         if (!cfs_list_empty(&res->lr_waiting)) {
1123                 ldlm_resource_dump(D_ERROR, res);
1124                 LBUG();
1125         }
1126
1127         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1128                                bd, &res->lr_hash);
1129         lu_ref_fini(&res->lr_reference);
1130         if (cfs_hash_bd_count_get(bd) == 0)
1131                 ldlm_namespace_put(nsb->nsb_namespace);
1132 }
1133
1134 /* Returns 1 if the resource was freed, 0 if it remains. */
1135 int ldlm_resource_putref(struct ldlm_resource *res)
1136 {
1137         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1138         cfs_hash_bd_t   bd;
1139
1140         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1141         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1142                res, cfs_atomic_read(&res->lr_refcount) - 1);
1143
1144         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1145         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1146                 __ldlm_resource_putref_final(&bd, res);
1147                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1148                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1149                         ns->ns_lvbo->lvbo_free(res);
1150                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1151                 return 1;
1152         }
1153         return 0;
1154 }
1155
1156 /* Returns 1 if the resource was freed, 0 if it remains. */
1157 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1158 {
1159         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1160
1161         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1162         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1163                res, cfs_atomic_read(&res->lr_refcount) - 1);
1164
1165         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1166                 cfs_hash_bd_t bd;
1167
1168                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1169                                 &res->lr_name, &bd);
1170                 __ldlm_resource_putref_final(&bd, res);
1171                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1172                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1173                  * so we should never be here while calling cfs_hash_del,
1174                  * cfs_hash_for_each_nolock is the only case we can get
1175                  * here, which is safe to release cfs_hash_bd_lock.
1176                  */
1177                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1178                         ns->ns_lvbo->lvbo_free(res);
1179                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1180
1181                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1182                 return 1;
1183         }
1184         return 0;
1185 }
1186
1187 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1188                             struct ldlm_lock *lock)
1189 {
1190         check_res_locked(res);
1191
1192         ldlm_resource_dump(D_INFO, res);
1193         CDEBUG(D_OTHER, "About to add this lock:\n");
1194         ldlm_lock_dump(D_OTHER, lock, 0);
1195
1196         if (lock->l_destroyed) {
1197                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1198                 return;
1199         }
1200
1201         LASSERT(cfs_list_empty(&lock->l_res_link));
1202
1203         cfs_list_add_tail(&lock->l_res_link, head);
1204 }
1205
1206 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1207                                      struct ldlm_lock *new)
1208 {
1209         struct ldlm_resource *res = original->l_resource;
1210
1211         check_res_locked(res);
1212
1213         ldlm_resource_dump(D_INFO, res);
1214         CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1215         ldlm_lock_dump(D_OTHER, new, 0);
1216
1217         if (new->l_destroyed) {
1218                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1219                 goto out;
1220         }
1221
1222         LASSERT(cfs_list_empty(&new->l_res_link));
1223
1224         cfs_list_add(&new->l_res_link, &original->l_res_link);
1225  out:;
1226 }
1227
1228 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1229 {
1230         int type = lock->l_resource->lr_type;
1231
1232         check_res_locked(lock->l_resource);
1233         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1234                 ldlm_unlink_lock_skiplist(lock);
1235         else if (type == LDLM_EXTENT)
1236                 ldlm_extent_unlink_lock(lock);
1237         cfs_list_del_init(&lock->l_res_link);
1238 }
1239
1240 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1241 {
1242         desc->lr_type = res->lr_type;
1243         desc->lr_name = res->lr_name;
1244 }
1245
1246 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1247 {
1248         cfs_list_t *tmp;
1249
1250         if (!((libcfs_debug | D_ERROR) & level))
1251                 return;
1252
1253         cfs_mutex_down(ldlm_namespace_lock(client));
1254
1255         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1256                 struct ldlm_namespace *ns;
1257                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1258                 ldlm_namespace_dump(level, ns);
1259         }
1260
1261         cfs_mutex_up(ldlm_namespace_lock(client));
1262 }
1263
1264 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1265                               cfs_hlist_node_t *hnode, void *arg)
1266 {
1267         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1268         int    level = (int)(unsigned long)arg;
1269
1270         lock_res(res);
1271         ldlm_resource_dump(level, res);
1272         unlock_res(res);
1273
1274         return 0;
1275 }
1276
1277 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1278 {
1279         if (!((libcfs_debug | D_ERROR) & level))
1280                 return;
1281
1282         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1283                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1284                ns_is_client(ns) ? "client" : "server");
1285
1286         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1287                 return;
1288
1289         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1290                                  ldlm_res_hash_dump,
1291                                  (void *)(unsigned long)level);
1292         cfs_spin_lock(&ns->ns_lock);
1293         ns->ns_next_dump = cfs_time_shift(10);
1294         cfs_spin_unlock(&ns->ns_lock);
1295 }
1296
1297 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1298 {
1299         cfs_list_t *tmp;
1300         int pos;
1301
1302         CLASSERT(RES_NAME_SIZE == 4);
1303
1304         if (!((libcfs_debug | D_ERROR) & level))
1305                 return;
1306
1307         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1308                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1309                res->lr_name.name[2], res->lr_name.name[3],
1310                cfs_atomic_read(&res->lr_refcount));
1311
1312         if (!cfs_list_empty(&res->lr_granted)) {
1313                 pos = 0;
1314                 CDEBUG(level, "Granted locks:\n");
1315                 cfs_list_for_each(tmp, &res->lr_granted) {
1316                         struct ldlm_lock *lock;
1317                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1318                                               l_res_link);
1319                         ldlm_lock_dump(level, lock, ++pos);
1320                 }
1321         }
1322         if (!cfs_list_empty(&res->lr_converting)) {
1323                 pos = 0;
1324                 CDEBUG(level, "Converting locks:\n");
1325                 cfs_list_for_each(tmp, &res->lr_converting) {
1326                         struct ldlm_lock *lock;
1327                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1328                                               l_res_link);
1329                         ldlm_lock_dump(level, lock, ++pos);
1330                 }
1331         }
1332         if (!cfs_list_empty(&res->lr_waiting)) {
1333                 pos = 0;
1334                 CDEBUG(level, "Waiting locks:\n");
1335                 cfs_list_for_each(tmp, &res->lr_waiting) {
1336                         struct ldlm_lock *lock;
1337                         lock = cfs_list_entry(tmp, struct ldlm_lock,
1338                                               l_res_link);
1339                         ldlm_lock_dump(level, lock, ++pos);
1340                 }
1341         }
1342 }