Whamcloud - gitweb
494661efa149d048cc90d81a04ef5eeaf7506b91
[fs/lustre-release.git] / 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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2012, Intel Corporation.
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 struct mutex ldlm_srv_namespace_lock;
59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
60
61 struct mutex 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 /* during debug dump certain amount of granted locks for one resource to avoid
71  * DDOS. */
72 unsigned int ldlm_dump_granted_max = 256;
73
74 #ifdef LPROCFS
75 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
76                              unsigned long count, void *data)
77 {
78         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
79         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
80         RETURN(count);
81 }
82
83 int ldlm_proc_setup(void)
84 {
85         int rc;
86         struct lprocfs_vars list[] = {
87                 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
88                 { "dump_granted_max",
89                   lprocfs_rd_uint, lprocfs_wr_uint,
90                   &ldlm_dump_granted_max, NULL },
91                 { "cancel_unused_locks_before_replay",
92                   lprocfs_rd_uint, lprocfs_wr_uint,
93                   &ldlm_cancel_unused_locks_before_replay, NULL },
94                 { NULL }};
95         ENTRY;
96         LASSERT(ldlm_ns_proc_dir == NULL);
97
98         ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
99                                               proc_lustre_root,
100                                               NULL, NULL);
101         if (IS_ERR(ldlm_type_proc_dir)) {
102                 CERROR("LProcFS failed in ldlm-init\n");
103                 rc = PTR_ERR(ldlm_type_proc_dir);
104                 GOTO(err, rc);
105         }
106
107         ldlm_ns_proc_dir = lprocfs_register("namespaces",
108                                             ldlm_type_proc_dir,
109                                             NULL, NULL);
110         if (IS_ERR(ldlm_ns_proc_dir)) {
111                 CERROR("LProcFS failed in ldlm-init\n");
112                 rc = PTR_ERR(ldlm_ns_proc_dir);
113                 GOTO(err_type, rc);
114         }
115
116         ldlm_svc_proc_dir = lprocfs_register("services",
117                                             ldlm_type_proc_dir,
118                                             NULL, NULL);
119         if (IS_ERR(ldlm_svc_proc_dir)) {
120                 CERROR("LProcFS failed in ldlm-init\n");
121                 rc = PTR_ERR(ldlm_svc_proc_dir);
122                 GOTO(err_ns, rc);
123         }
124
125         rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
126
127         RETURN(0);
128
129 err_ns:
130         lprocfs_remove(&ldlm_ns_proc_dir);
131 err_type:
132         lprocfs_remove(&ldlm_type_proc_dir);
133 err:
134         ldlm_svc_proc_dir = NULL;
135         RETURN(rc);
136 }
137
138 void ldlm_proc_cleanup(void)
139 {
140         if (ldlm_svc_proc_dir)
141                 lprocfs_remove(&ldlm_svc_proc_dir);
142
143         if (ldlm_ns_proc_dir)
144                 lprocfs_remove(&ldlm_ns_proc_dir);
145
146         if (ldlm_type_proc_dir)
147                 lprocfs_remove(&ldlm_type_proc_dir);
148 }
149
150 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
151                                    int count, int *eof, void *data)
152 {
153         struct ldlm_namespace *ns  = data;
154         __u64                  res = 0;
155         cfs_hash_bd_t          bd;
156         int                    i;
157
158         /* result is not strictly consistant */
159         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
160                 res += cfs_hash_bd_count_get(&bd);
161         return lprocfs_rd_u64(page, start, off, count, eof, &res);
162 }
163
164 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
165                                int count, int *eof, void *data)
166 {
167         struct ldlm_namespace *ns = data;
168         __u64                  locks;
169
170         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
171                                         LPROCFS_FIELDS_FLAGS_SUM);
172         return lprocfs_rd_u64(page, start, off, count, eof, &locks);
173 }
174
175 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
176                                int count, int *eof, void *data)
177 {
178         struct ldlm_namespace *ns = data;
179         __u32 *nr = &ns->ns_max_unused;
180
181         if (ns_connect_lru_resize(ns))
182                 nr = &ns->ns_nr_unused;
183         return lprocfs_rd_uint(page, start, off, count, eof, nr);
184 }
185
186 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
187                                unsigned long count, void *data)
188 {
189         struct ldlm_namespace *ns = data;
190         char dummy[MAX_STRING_SIZE + 1], *end;
191         unsigned long tmp;
192         int lru_resize;
193
194         dummy[MAX_STRING_SIZE] = '\0';
195         if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
196                 return -EFAULT;
197
198         if (strncmp(dummy, "clear", 5) == 0) {
199                 CDEBUG(D_DLMTRACE,
200                        "dropping all unused locks from namespace %s\n",
201                        ldlm_ns_name(ns));
202                 if (ns_connect_lru_resize(ns)) {
203                         int canceled, unused  = ns->ns_nr_unused;
204
205                         /* Try to cancel all @ns_nr_unused locks. */
206                         canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
207                                                    LDLM_CANCEL_PASSED);
208                         if (canceled < unused) {
209                                 CDEBUG(D_DLMTRACE,
210                                        "not all requested locks are canceled, "
211                                        "requested: %d, canceled: %d\n", unused,
212                                        canceled);
213                                 return -EINVAL;
214                         }
215                 } else {
216                         tmp = ns->ns_max_unused;
217                         ns->ns_max_unused = 0;
218                         ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
219                         ns->ns_max_unused = tmp;
220                 }
221                 return count;
222         }
223
224         tmp = simple_strtoul(dummy, &end, 0);
225         if (dummy == end) {
226                 CERROR("invalid value written\n");
227                 return -EINVAL;
228         }
229         lru_resize = (tmp == 0);
230
231         if (ns_connect_lru_resize(ns)) {
232                 if (!lru_resize)
233                         ns->ns_max_unused = (unsigned int)tmp;
234
235                 if (tmp > ns->ns_nr_unused)
236                         tmp = ns->ns_nr_unused;
237                 tmp = ns->ns_nr_unused - tmp;
238
239                 CDEBUG(D_DLMTRACE,
240                        "changing namespace %s unused locks from %u to %u\n",
241                        ldlm_ns_name(ns), ns->ns_nr_unused,
242                        (unsigned int)tmp);
243                 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
244
245                 if (!lru_resize) {
246                         CDEBUG(D_DLMTRACE,
247                                "disable lru_resize for namespace %s\n",
248                                ldlm_ns_name(ns));
249                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
250                 }
251         } else {
252                 CDEBUG(D_DLMTRACE,
253                        "changing namespace %s max_unused from %u to %u\n",
254                        ldlm_ns_name(ns), ns->ns_max_unused,
255                        (unsigned int)tmp);
256                 ns->ns_max_unused = (unsigned int)tmp;
257                 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
258
259                 /* Make sure that LRU resize was originally supported before
260                  * turning it on here. */
261                 if (lru_resize &&
262                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
263                         CDEBUG(D_DLMTRACE,
264                                "enable lru_resize for namespace %s\n",
265                                ldlm_ns_name(ns));
266                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
267                 }
268         }
269
270         return count;
271 }
272
273 static int lprocfs_rd_elc(char *page, char **start, off_t off,
274                           int count, int *eof, void *data)
275 {
276         struct ldlm_namespace *ns = data;
277         unsigned int supp = ns_connect_cancelset(ns);
278
279         return lprocfs_rd_uint(page, start, off, count, eof, &supp);
280 }
281
282 static int lprocfs_wr_elc(struct file *file, const char *buffer,
283                                unsigned long count, void *data)
284 {
285         struct ldlm_namespace *ns = data;
286         unsigned int supp = -1;
287         int rc;
288
289         rc = lprocfs_wr_uint(file, buffer, count, &supp);
290         if (rc < 0)
291                 return rc;
292
293         if (supp == 0)
294                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
295         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
296                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
297         return count;
298 }
299
300 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
301 {
302         struct proc_dir_entry *dir;
303
304         dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
305         if (dir == NULL) {
306                 CERROR("dlm namespace %s has no procfs dir?\n",
307                        ldlm_ns_name(ns));
308         } else {
309                 lprocfs_remove(&dir);
310         }
311
312         if (ns->ns_stats != NULL)
313                 lprocfs_free_stats(&ns->ns_stats);
314 }
315
316 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
317 {
318         struct lprocfs_vars lock_vars[2];
319         char lock_name[MAX_STRING_SIZE + 1];
320
321         LASSERT(ns != NULL);
322         LASSERT(ns->ns_rs_hash != NULL);
323
324         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
325         if (ns->ns_stats == NULL)
326                 return -ENOMEM;
327
328         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
329                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
330
331         lock_name[MAX_STRING_SIZE] = '\0';
332
333         memset(lock_vars, 0, sizeof(lock_vars));
334         lock_vars[0].name = lock_name;
335
336         snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
337                  ldlm_ns_name(ns));
338         lock_vars[0].data = ns;
339         lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
340         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
341
342         snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
343                  ldlm_ns_name(ns));
344         lock_vars[0].data = ns;
345         lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
346         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
347
348         if (ns_is_client(ns)) {
349                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
350                          ldlm_ns_name(ns));
351                 lock_vars[0].data = &ns->ns_nr_unused;
352                 lock_vars[0].read_fptr = lprocfs_rd_uint;
353                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
354
355                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
356                          ldlm_ns_name(ns));
357                 lock_vars[0].data = ns;
358                 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
359                 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
360                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
361
362                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
363                          ldlm_ns_name(ns));
364                 lock_vars[0].data = &ns->ns_max_age;
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/early_lock_cancel",
370                          ldlm_ns_name(ns));
371                 lock_vars[0].data = ns;
372                 lock_vars[0].read_fptr = lprocfs_rd_elc;
373                 lock_vars[0].write_fptr = lprocfs_wr_elc;
374                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
375         } else {
376                 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
377                          ldlm_ns_name(ns));
378                 lock_vars[0].data = &ns->ns_ctime_age_limit;
379                 lock_vars[0].read_fptr = lprocfs_rd_uint;
380                 lock_vars[0].write_fptr = lprocfs_wr_uint;
381                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
382
383                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
384                          ldlm_ns_name(ns));
385                 lock_vars[0].data = &ns->ns_timeouts;
386                 lock_vars[0].read_fptr = lprocfs_rd_uint;
387                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
388
389                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
390                          ldlm_ns_name(ns));
391                 lock_vars[0].data = &ns->ns_max_nolock_size;
392                 lock_vars[0].read_fptr = lprocfs_rd_uint;
393                 lock_vars[0].write_fptr = lprocfs_wr_uint;
394                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
395
396                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
397                          ldlm_ns_name(ns));
398                 lock_vars[0].data = &ns->ns_contention_time;
399                 lock_vars[0].read_fptr = lprocfs_rd_uint;
400                 lock_vars[0].write_fptr = lprocfs_wr_uint;
401                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
402
403                 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
404                          ldlm_ns_name(ns));
405                 lock_vars[0].data = &ns->ns_contended_locks;
406                 lock_vars[0].read_fptr = lprocfs_rd_uint;
407                 lock_vars[0].write_fptr = lprocfs_wr_uint;
408                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
409
410                 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
411                          ldlm_ns_name(ns));
412                 lock_vars[0].data = &ns->ns_max_parallel_ast;
413                 lock_vars[0].read_fptr = lprocfs_rd_uint;
414                 lock_vars[0].write_fptr = lprocfs_wr_uint;
415                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
416         }
417         return 0;
418 }
419 #undef MAX_STRING_SIZE
420 #else /* LPROCFS */
421
422 #define ldlm_namespace_proc_unregister(ns)      ({;})
423 #define ldlm_namespace_proc_register(ns)        ({0;})
424
425 #endif /* LPROCFS */
426
427 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
428                                   const void *key, unsigned mask)
429 {
430         const struct ldlm_res_id     *id  = key;
431         unsigned                val = 0;
432         unsigned                i;
433
434         for (i = 0; i < RES_NAME_SIZE; i++)
435                 val += id->name[i];
436         return val & mask;
437 }
438
439 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
440                                       const void *key, unsigned mask)
441 {
442         const struct ldlm_res_id *id = key;
443         struct lu_fid       fid;
444         __u32               hash;
445         __u32               val;
446
447         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
448         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
449         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
450
451         hash = fid_flatten32(&fid);
452         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
453         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
454                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
455                 hash += (val >> 5) + (val << 11);
456         } else {
457                 val = fid_oid(&fid);
458         }
459         hash = cfs_hash_long(hash, hs->hs_bkt_bits);
460         /* give me another random factor */
461         hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
462
463         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
464         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
465
466         return hash & mask;
467 }
468
469 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
470 {
471         struct ldlm_resource   *res;
472
473         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
474         return &res->lr_name;
475 }
476
477 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
478 {
479         struct ldlm_resource   *res;
480
481         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
482         return ldlm_res_eq((const struct ldlm_res_id *)key,
483                            (const struct ldlm_res_id *)&res->lr_name);
484 }
485
486 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
487 {
488         return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
489 }
490
491 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
492 {
493         struct ldlm_resource *res;
494
495         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
496         ldlm_resource_getref(res);
497 }
498
499 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
500 {
501         struct ldlm_resource *res;
502
503         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
504         /* cfs_hash_for_each_nolock is the only chance we call it */
505         ldlm_resource_putref_locked(res);
506 }
507
508 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
509 {
510         struct ldlm_resource *res;
511
512         res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
513         ldlm_resource_putref(res);
514 }
515
516 cfs_hash_ops_t ldlm_ns_hash_ops = {
517         .hs_hash        = ldlm_res_hop_hash,
518         .hs_key         = ldlm_res_hop_key,
519         .hs_keycmp      = ldlm_res_hop_keycmp,
520         .hs_keycpy      = NULL,
521         .hs_object      = ldlm_res_hop_object,
522         .hs_get         = ldlm_res_hop_get_locked,
523         .hs_put_locked  = ldlm_res_hop_put_locked,
524         .hs_put         = ldlm_res_hop_put
525 };
526
527 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
528         .hs_hash        = ldlm_res_hop_fid_hash,
529         .hs_key         = ldlm_res_hop_key,
530         .hs_keycmp      = ldlm_res_hop_keycmp,
531         .hs_keycpy      = NULL,
532         .hs_object      = ldlm_res_hop_object,
533         .hs_get         = ldlm_res_hop_get_locked,
534         .hs_put_locked  = ldlm_res_hop_put_locked,
535         .hs_put         = ldlm_res_hop_put
536 };
537
538 typedef struct {
539         ldlm_ns_type_t  nsd_type;
540         /** hash bucket bits */
541         unsigned        nsd_bkt_bits;
542         /** hash bits */
543         unsigned        nsd_all_bits;
544         /** hash operations */
545         cfs_hash_ops_t *nsd_hops;
546 } ldlm_ns_hash_def_t;
547
548 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
549 {
550         {
551                 .nsd_type       = LDLM_NS_TYPE_MDC,
552                 .nsd_bkt_bits   = 11,
553                 .nsd_all_bits   = 16,
554                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
555         },
556         {
557                 .nsd_type       = LDLM_NS_TYPE_MDT,
558                 .nsd_bkt_bits   = 14,
559                 .nsd_all_bits   = 21,
560                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
561         },
562         {
563                 .nsd_type       = LDLM_NS_TYPE_OSC,
564                 .nsd_bkt_bits   = 8,
565                 .nsd_all_bits   = 12,
566                 .nsd_hops       = &ldlm_ns_hash_ops,
567         },
568         {
569                 .nsd_type       = LDLM_NS_TYPE_OST,
570                 .nsd_bkt_bits   = 11,
571                 .nsd_all_bits   = 17,
572                 .nsd_hops       = &ldlm_ns_hash_ops,
573         },
574         {
575                 .nsd_type       = LDLM_NS_TYPE_MGC,
576                 .nsd_bkt_bits   = 4,
577                 .nsd_all_bits   = 4,
578                 .nsd_hops       = &ldlm_ns_hash_ops,
579         },
580         {
581                 .nsd_type       = LDLM_NS_TYPE_MGT,
582                 .nsd_bkt_bits   = 4,
583                 .nsd_all_bits   = 4,
584                 .nsd_hops       = &ldlm_ns_hash_ops,
585         },
586         {
587                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
588         },
589 };
590
591 /**
592  * Create and initialize new empty namespace.
593  */
594 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
595                                           ldlm_side_t client,
596                                           ldlm_appetite_t apt,
597                                           ldlm_ns_type_t ns_type)
598 {
599         struct ldlm_namespace *ns = NULL;
600         struct ldlm_ns_bucket *nsb;
601         ldlm_ns_hash_def_t    *nsd;
602         cfs_hash_bd_t          bd;
603         int                    idx;
604         int                    rc;
605         ENTRY;
606
607         LASSERT(obd != NULL);
608
609         rc = ldlm_get_ref();
610         if (rc) {
611                 CERROR("ldlm_get_ref failed: %d\n", rc);
612                 RETURN(NULL);
613         }
614
615         for (idx = 0;;idx++) {
616                 nsd = &ldlm_ns_hash_defs[idx];
617                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
618                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
619                         GOTO(out_ref, NULL);
620                 }
621
622                 if (nsd->nsd_type == ns_type)
623                         break;
624         }
625
626         OBD_ALLOC_PTR(ns);
627         if (!ns)
628                 GOTO(out_ref, NULL);
629
630         ns->ns_rs_hash = cfs_hash_create(name,
631                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
632                                          nsd->nsd_bkt_bits, sizeof(*nsb),
633                                          CFS_HASH_MIN_THETA,
634                                          CFS_HASH_MAX_THETA,
635                                          nsd->nsd_hops,
636                                          CFS_HASH_DEPTH |
637                                          CFS_HASH_BIGNAME |
638                                          CFS_HASH_SPIN_BKTLOCK |
639                                          CFS_HASH_NO_ITEMREF);
640         if (ns->ns_rs_hash == NULL)
641                 GOTO(out_ns, NULL);
642
643         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
644                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
645                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
646                 nsb->nsb_namespace = ns;
647         }
648
649         ns->ns_obd      = obd;
650         ns->ns_appetite = apt;
651         ns->ns_client   = client;
652
653         CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
654         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
655         spin_lock_init(&ns->ns_lock);
656         cfs_atomic_set(&ns->ns_bref, 0);
657         cfs_waitq_init(&ns->ns_waitq);
658
659         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
660         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
661         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
662
663         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
664         ns->ns_nr_unused          = 0;
665         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
666         ns->ns_max_age            = LDLM_DEFAULT_MAX_ALIVE;
667         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
668         ns->ns_timeouts           = 0;
669         ns->ns_orig_connect_flags = 0;
670         ns->ns_connect_flags      = 0;
671         ns->ns_stopping           = 0;
672         rc = ldlm_namespace_proc_register(ns);
673         if (rc != 0) {
674                 CERROR("Can't initialize ns proc, rc %d\n", rc);
675                 GOTO(out_hash, rc);
676         }
677
678         idx = cfs_atomic_read(ldlm_namespace_nr(client));
679         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
680         if (rc) {
681                 CERROR("Can't initialize lock pool, rc %d\n", rc);
682                 GOTO(out_proc, rc);
683         }
684
685         ldlm_namespace_register(ns, client);
686         RETURN(ns);
687 out_proc:
688         ldlm_namespace_proc_unregister(ns);
689         ldlm_namespace_cleanup(ns, 0);
690 out_hash:
691         cfs_hash_putref(ns->ns_rs_hash);
692 out_ns:
693         OBD_FREE_PTR(ns);
694 out_ref:
695         ldlm_put_ref();
696         RETURN(NULL);
697 }
698 EXPORT_SYMBOL(ldlm_namespace_new);
699
700 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
701
702 /**
703  * Cancel and destroy all locks on a resource.
704  *
705  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
706  * clean up.  This is currently only used for recovery, and we make
707  * certain assumptions as a result--notably, that we shouldn't cancel
708  * locks with refs.
709  */
710 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
711                              __u64 flags)
712 {
713         cfs_list_t *tmp;
714         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
715         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
716
717         do {
718                 struct ldlm_lock *lock = NULL;
719
720                 /* First, we look for non-cleaned-yet lock
721                  * all cleaned locks are marked by CLEANED flag. */
722                 lock_res(res);
723                 cfs_list_for_each(tmp, q) {
724                         lock = cfs_list_entry(tmp, struct ldlm_lock,
725                                               l_res_link);
726                         if (lock->l_flags & LDLM_FL_CLEANED) {
727                                 lock = NULL;
728                                 continue;
729                         }
730                         LDLM_LOCK_GET(lock);
731                         lock->l_flags |= LDLM_FL_CLEANED;
732                         break;
733                 }
734
735                 if (lock == NULL) {
736                         unlock_res(res);
737                         break;
738                 }
739
740                 /* Set CBPENDING so nothing in the cancellation path
741                  * can match this lock. */
742                 lock->l_flags |= LDLM_FL_CBPENDING;
743                 lock->l_flags |= LDLM_FL_FAILED;
744                 lock->l_flags |= flags;
745
746                 /* ... without sending a CANCEL message for local_only. */
747                 if (local_only)
748                         lock->l_flags |= LDLM_FL_LOCAL_ONLY;
749
750                 if (local_only && (lock->l_readers || lock->l_writers)) {
751                         /* This is a little bit gross, but much better than the
752                          * alternative: pretend that we got a blocking AST from
753                          * the server, so that when the lock is decref'd, it
754                          * will go away ... */
755                         unlock_res(res);
756                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
757                         if (lock->l_completion_ast)
758                                 lock->l_completion_ast(lock, 0, NULL);
759                         LDLM_LOCK_RELEASE(lock);
760                         continue;
761                 }
762
763                 if (client) {
764                         struct lustre_handle lockh;
765
766                         unlock_res(res);
767                         ldlm_lock2handle(lock, &lockh);
768                         rc = ldlm_cli_cancel(&lockh);
769                         if (rc)
770                                 CERROR("ldlm_cli_cancel: %d\n", rc);
771                 } else {
772                         ldlm_resource_unlink_lock(lock);
773                         unlock_res(res);
774                         LDLM_DEBUG(lock, "Freeing a lock still held by a "
775                                    "client node");
776                         ldlm_lock_destroy(lock);
777                 }
778                 LDLM_LOCK_RELEASE(lock);
779         } while (1);
780 }
781
782 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
783                                cfs_hlist_node_t *hnode, void *arg)
784 {
785         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
786         __u64 flags = *(__u64 *)arg;
787
788         cleanup_resource(res, &res->lr_granted, flags);
789         cleanup_resource(res, &res->lr_converting, flags);
790         cleanup_resource(res, &res->lr_waiting, flags);
791
792         return 0;
793 }
794
795 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
796                                   cfs_hlist_node_t *hnode, void *arg)
797 {
798         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
799
800         CERROR("Namespace %s resource refcount nonzero "
801                "(%d) after lock cleanup; forcing "
802                "cleanup.\n",
803                ldlm_ns_name(ldlm_res_to_ns(res)),
804                cfs_atomic_read(&res->lr_refcount) - 1);
805
806         CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
807                LPU64") (rc: %d)\n", res,
808                res->lr_name.name[0], res->lr_name.name[1],
809                res->lr_name.name[2], res->lr_name.name[3],
810                cfs_atomic_read(&res->lr_refcount) - 1);
811
812         ldlm_resource_dump(D_ERROR, res);
813         return 0;
814 }
815
816 /**
817  * Cancel and destroy all locks in the namespace.
818  *
819  * Typically used during evictions when server notified client that it was
820  * evicted and all of its state needs to be destroyed.
821  * Also used during shutdown.
822  */
823 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
824 {
825         if (ns == NULL) {
826                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
827                 return ELDLM_OK;
828         }
829
830         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
831         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
832         return ELDLM_OK;
833 }
834 EXPORT_SYMBOL(ldlm_namespace_cleanup);
835
836 /**
837  * Attempts to free namespace.
838  *
839  * Only used when namespace goes away, like during an unmount.
840  */
841 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
842 {
843         ENTRY;
844
845         /* At shutdown time, don't call the cancellation callback */
846         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
847
848         if (cfs_atomic_read(&ns->ns_bref) > 0) {
849                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
850                 int rc;
851                 CDEBUG(D_DLMTRACE,
852                        "dlm namespace %s free waiting on refcount %d\n",
853                        ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
854 force_wait:
855                 if (force)
856                         lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
857
858                 rc = l_wait_event(ns->ns_waitq,
859                                   cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
860
861                 /* Forced cleanups should be able to reclaim all references,
862                  * so it's safe to wait forever... we can't leak locks... */
863                 if (force && rc == -ETIMEDOUT) {
864                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
865                                        "namespace with %d resources in use, "
866                                        "(rc=%d)\n", ldlm_ns_name(ns),
867                                        cfs_atomic_read(&ns->ns_bref), rc);
868                         GOTO(force_wait, rc);
869                 }
870
871                 if (cfs_atomic_read(&ns->ns_bref)) {
872                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
873                                        "with %d resources in use, (rc=%d)\n",
874                                        ldlm_ns_name(ns),
875                                        cfs_atomic_read(&ns->ns_bref), rc);
876                         RETURN(ELDLM_NAMESPACE_EXISTS);
877                 }
878                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
879                        ldlm_ns_name(ns));
880         }
881
882         RETURN(ELDLM_OK);
883 }
884
885 /**
886  * Performs various cleanups for passed \a ns to make it drop refc and be
887  * ready for freeing. Waits for refc == 0.
888  *
889  * The following is done:
890  * (0) Unregister \a ns from its list to make inaccessible for potential
891  * users like pools thread and others;
892  * (1) Clear all locks in \a ns.
893  */
894 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
895                                struct obd_import *imp,
896                                int force)
897 {
898         int rc;
899         ENTRY;
900         if (!ns) {
901                 EXIT;
902                 return;
903         }
904
905         spin_lock(&ns->ns_lock);
906         ns->ns_stopping = 1;
907         spin_unlock(&ns->ns_lock);
908
909         /*
910          * Can fail with -EINTR when force == 0 in which case try harder.
911          */
912         rc = __ldlm_namespace_free(ns, force);
913         if (rc != ELDLM_OK) {
914                 if (imp) {
915                         ptlrpc_disconnect_import(imp, 0);
916                         ptlrpc_invalidate_import(imp);
917                 }
918
919                 /*
920                  * With all requests dropped and the import inactive
921                  * we are gaurenteed all reference will be dropped.
922                  */
923                 rc = __ldlm_namespace_free(ns, 1);
924                 LASSERT(rc == 0);
925         }
926         EXIT;
927 }
928
929 /**
930  * Performs freeing memory structures related to \a ns. This is only done
931  * when ldlm_namespce_free_prior() successfully removed all resources
932  * referencing \a ns and its refc == 0.
933  */
934 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
935 {
936         ENTRY;
937         if (!ns) {
938                 EXIT;
939                 return;
940         }
941
942         /* Make sure that nobody can find this ns in its list. */
943         ldlm_namespace_unregister(ns, ns->ns_client);
944         /* Fini pool _before_ parent proc dir is removed. This is important as
945          * ldlm_pool_fini() removes own proc dir which is child to @dir.
946          * Removing it after @dir may cause oops. */
947         ldlm_pool_fini(&ns->ns_pool);
948
949         ldlm_namespace_proc_unregister(ns);
950         cfs_hash_putref(ns->ns_rs_hash);
951         /* Namespace \a ns should be not on list at this time, otherwise
952          * this will cause issues related to using freed \a ns in poold
953          * thread. */
954         LASSERT(cfs_list_empty(&ns->ns_list_chain));
955         OBD_FREE_PTR(ns);
956         ldlm_put_ref();
957         EXIT;
958 }
959
960 /**
961  * Cleanup the resource, and free namespace.
962  * bug 12864:
963  * Deadlock issue:
964  * proc1: destroy import
965  *        class_disconnect_export(grab cl_sem) ->
966  *              -> ldlm_namespace_free ->
967  *              -> lprocfs_remove(grab _lprocfs_lock).
968  * proc2: read proc info
969  *        lprocfs_fops_read(grab _lprocfs_lock) ->
970  *              -> osc_rd_active, etc(grab cl_sem).
971  *
972  * So that I have to split the ldlm_namespace_free into two parts - the first
973  * part ldlm_namespace_free_prior is used to cleanup the resource which is
974  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
975  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
976  * held.
977  */
978 void ldlm_namespace_free(struct ldlm_namespace *ns,
979                          struct obd_import *imp,
980                          int force)
981 {
982         ldlm_namespace_free_prior(ns, imp, force);
983         ldlm_namespace_free_post(ns);
984 }
985 EXPORT_SYMBOL(ldlm_namespace_free);
986
987 void ldlm_namespace_get(struct ldlm_namespace *ns)
988 {
989         cfs_atomic_inc(&ns->ns_bref);
990 }
991 EXPORT_SYMBOL(ldlm_namespace_get);
992
993 void ldlm_namespace_put(struct ldlm_namespace *ns)
994 {
995         if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
996                 cfs_waitq_signal(&ns->ns_waitq);
997                 spin_unlock(&ns->ns_lock);
998         }
999 }
1000 EXPORT_SYMBOL(ldlm_namespace_put);
1001
1002 /** Register \a ns in the list of namespaces */
1003 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
1004 {
1005         mutex_lock(ldlm_namespace_lock(client));
1006         LASSERT(cfs_list_empty(&ns->ns_list_chain));
1007         cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
1008         cfs_atomic_inc(ldlm_namespace_nr(client));
1009         mutex_unlock(ldlm_namespace_lock(client));
1010 }
1011
1012 /** Unregister \a ns from the list of namespaces. */
1013 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1014 {
1015         mutex_lock(ldlm_namespace_lock(client));
1016         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1017         /* Some asserts and possibly other parts of the code are still
1018          * using list_empty(&ns->ns_list_chain). This is why it is
1019          * important to use list_del_init() here. */
1020         cfs_list_del_init(&ns->ns_list_chain);
1021         cfs_atomic_dec(ldlm_namespace_nr(client));
1022         mutex_unlock(ldlm_namespace_lock(client));
1023 }
1024
1025 /** Should be called with ldlm_namespace_lock(client) taken. */
1026 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
1027 {
1028         LASSERT(!cfs_list_empty(&ns->ns_list_chain));
1029         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1030         cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1031 }
1032
1033 /** Should be called with ldlm_namespace_lock(client) taken. */
1034 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1035 {
1036         LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
1037         LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
1038         return container_of(ldlm_namespace_list(client)->next,
1039                 struct ldlm_namespace, ns_list_chain);
1040 }
1041
1042 /** Create and initialize new resource. */
1043 static struct ldlm_resource *ldlm_resource_new(void)
1044 {
1045         struct ldlm_resource *res;
1046         int idx;
1047
1048         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
1049         if (res == NULL)
1050                 return NULL;
1051
1052         CFS_INIT_LIST_HEAD(&res->lr_granted);
1053         CFS_INIT_LIST_HEAD(&res->lr_converting);
1054         CFS_INIT_LIST_HEAD(&res->lr_waiting);
1055
1056         /* Initialize interval trees for each lock mode. */
1057         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1058                 res->lr_itree[idx].lit_size = 0;
1059                 res->lr_itree[idx].lit_mode = 1 << idx;
1060                 res->lr_itree[idx].lit_root = NULL;
1061         }
1062
1063         cfs_atomic_set(&res->lr_refcount, 1);
1064         spin_lock_init(&res->lr_lock);
1065         lu_ref_init(&res->lr_reference);
1066
1067         /* The creator of the resource must unlock the mutex after LVB
1068          * initialization. */
1069         mutex_init(&res->lr_lvb_mutex);
1070         mutex_lock(&res->lr_lvb_mutex);
1071
1072         return res;
1073 }
1074
1075 /**
1076  * Return a reference to resource with given name, creating it if necessary.
1077  * Args: namespace with ns_lock unlocked
1078  * Locks: takes and releases NS hash-lock and res->lr_lock
1079  * Returns: referenced, unlocked ldlm_resource or NULL
1080  */
1081 struct ldlm_resource *
1082 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1083                   const struct ldlm_res_id *name, ldlm_type_t type, int create)
1084 {
1085         cfs_hlist_node_t     *hnode;
1086         struct ldlm_resource *res;
1087         cfs_hash_bd_t         bd;
1088         __u64                 version;
1089
1090         LASSERT(ns != NULL);
1091         LASSERT(parent == NULL);
1092         LASSERT(ns->ns_rs_hash != NULL);
1093         LASSERT(name->name[0] != 0);
1094
1095         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1096         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1097         if (hnode != NULL) {
1098                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1099                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1100                 /* Synchronize with regard to resource creation. */
1101                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1102                         mutex_lock(&res->lr_lvb_mutex);
1103                         mutex_unlock(&res->lr_lvb_mutex);
1104                 }
1105                 return res;
1106         }
1107
1108         version = cfs_hash_bd_version_get(&bd);
1109         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1110
1111         if (create == 0)
1112                 return NULL;
1113
1114         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1115                  "type: %d\n", type);
1116         res = ldlm_resource_new();
1117         if (!res)
1118                 return NULL;
1119
1120         res->lr_ns_bucket  = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1121         res->lr_name       = *name;
1122         res->lr_type       = type;
1123         res->lr_most_restr = LCK_NL;
1124
1125         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1126         hnode = (version == cfs_hash_bd_version_get(&bd)) ?  NULL :
1127                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1128
1129         if (hnode != NULL) {
1130                 /* Someone won the race and already added the resource. */
1131                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1132                 /* Clean lu_ref for failed resource. */
1133                 lu_ref_fini(&res->lr_reference);
1134                 /* We have taken lr_lvb_mutex. Drop it. */
1135                 mutex_unlock(&res->lr_lvb_mutex);
1136                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1137
1138                 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1139                 /* Synchronize with regard to resource creation. */
1140                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1141                         mutex_lock(&res->lr_lvb_mutex);
1142                         mutex_unlock(&res->lr_lvb_mutex);
1143                 }
1144                 return res;
1145         }
1146         /* We won! Let's add the resource. */
1147         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1148         if (cfs_hash_bd_count_get(&bd) == 1)
1149                 ldlm_namespace_get(ns);
1150
1151         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1152         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1153                 int rc;
1154
1155                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1156                 rc = ns->ns_lvbo->lvbo_init(res);
1157                 if (rc)
1158                         CERROR("lvbo_init failed for resource "
1159                                LPU64": rc %d\n", name->name[0], rc);
1160         }
1161
1162         /* We create resource with locked lr_lvb_mutex. */
1163         mutex_unlock(&res->lr_lvb_mutex);
1164
1165         return res;
1166 }
1167 EXPORT_SYMBOL(ldlm_resource_get);
1168
1169 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1170 {
1171         LASSERT(res != NULL);
1172         LASSERT(res != LP_POISON);
1173         cfs_atomic_inc(&res->lr_refcount);
1174         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1175                cfs_atomic_read(&res->lr_refcount));
1176         return res;
1177 }
1178
1179 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1180                                          struct ldlm_resource *res)
1181 {
1182         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1183
1184         if (!cfs_list_empty(&res->lr_granted)) {
1185                 ldlm_resource_dump(D_ERROR, res);
1186                 LBUG();
1187         }
1188
1189         if (!cfs_list_empty(&res->lr_converting)) {
1190                 ldlm_resource_dump(D_ERROR, res);
1191                 LBUG();
1192         }
1193
1194         if (!cfs_list_empty(&res->lr_waiting)) {
1195                 ldlm_resource_dump(D_ERROR, res);
1196                 LBUG();
1197         }
1198
1199         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1200                                bd, &res->lr_hash);
1201         lu_ref_fini(&res->lr_reference);
1202         if (cfs_hash_bd_count_get(bd) == 0)
1203                 ldlm_namespace_put(nsb->nsb_namespace);
1204 }
1205
1206 /* Returns 1 if the resource was freed, 0 if it remains. */
1207 int ldlm_resource_putref(struct ldlm_resource *res)
1208 {
1209         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1210         cfs_hash_bd_t   bd;
1211
1212         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1213         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1214                res, cfs_atomic_read(&res->lr_refcount) - 1);
1215
1216         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1217         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1218                 __ldlm_resource_putref_final(&bd, res);
1219                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1220                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1221                         ns->ns_lvbo->lvbo_free(res);
1222                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1223                 return 1;
1224         }
1225         return 0;
1226 }
1227 EXPORT_SYMBOL(ldlm_resource_putref);
1228
1229 /* Returns 1 if the resource was freed, 0 if it remains. */
1230 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1231 {
1232         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1233
1234         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1235         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1236                res, cfs_atomic_read(&res->lr_refcount) - 1);
1237
1238         if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1239                 cfs_hash_bd_t bd;
1240
1241                 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1242                                 &res->lr_name, &bd);
1243                 __ldlm_resource_putref_final(&bd, res);
1244                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1245                 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1246                  * so we should never be here while calling cfs_hash_del,
1247                  * cfs_hash_for_each_nolock is the only case we can get
1248                  * here, which is safe to release cfs_hash_bd_lock.
1249                  */
1250                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1251                         ns->ns_lvbo->lvbo_free(res);
1252                 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1253
1254                 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1255                 return 1;
1256         }
1257         return 0;
1258 }
1259
1260 /**
1261  * Add a lock into a given resource into specified lock list.
1262  */
1263 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1264                             struct ldlm_lock *lock)
1265 {
1266         check_res_locked(res);
1267
1268         LDLM_DEBUG(lock, "About to add this lock:\n");
1269
1270         if (lock->l_destroyed) {
1271                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1272                 return;
1273         }
1274
1275         LASSERT(cfs_list_empty(&lock->l_res_link));
1276
1277         cfs_list_add_tail(&lock->l_res_link, head);
1278 }
1279
1280 /**
1281  * Insert a lock into resource after specified lock.
1282  *
1283  * Obtain resource description from the lock we are inserting after.
1284  */
1285 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1286                                      struct ldlm_lock *new)
1287 {
1288         struct ldlm_resource *res = original->l_resource;
1289
1290         check_res_locked(res);
1291
1292         ldlm_resource_dump(D_INFO, res);
1293         LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1294
1295         if (new->l_destroyed) {
1296                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1297                 goto out;
1298         }
1299
1300         LASSERT(cfs_list_empty(&new->l_res_link));
1301
1302         cfs_list_add(&new->l_res_link, &original->l_res_link);
1303  out:;
1304 }
1305
1306 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1307 {
1308         int type = lock->l_resource->lr_type;
1309
1310         check_res_locked(lock->l_resource);
1311         if (type == LDLM_IBITS || type == LDLM_PLAIN)
1312                 ldlm_unlink_lock_skiplist(lock);
1313         else if (type == LDLM_EXTENT)
1314                 ldlm_extent_unlink_lock(lock);
1315         cfs_list_del_init(&lock->l_res_link);
1316 }
1317 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1318
1319 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1320 {
1321         desc->lr_type = res->lr_type;
1322         desc->lr_name = res->lr_name;
1323 }
1324
1325 /**
1326  * Print information about all locks in all namespaces on this node to debug
1327  * log.
1328  */
1329 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1330 {
1331         cfs_list_t *tmp;
1332
1333         if (!((libcfs_debug | D_ERROR) & level))
1334                 return;
1335
1336         mutex_lock(ldlm_namespace_lock(client));
1337
1338         cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1339                 struct ldlm_namespace *ns;
1340                 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1341                 ldlm_namespace_dump(level, ns);
1342         }
1343
1344         mutex_unlock(ldlm_namespace_lock(client));
1345 }
1346 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1347
1348 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1349                               cfs_hlist_node_t *hnode, void *arg)
1350 {
1351         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1352         int    level = (int)(unsigned long)arg;
1353
1354         lock_res(res);
1355         ldlm_resource_dump(level, res);
1356         unlock_res(res);
1357
1358         return 0;
1359 }
1360
1361 /**
1362  * Print information about all locks in this namespace on this node to debug
1363  * log.
1364  */
1365 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1366 {
1367         if (!((libcfs_debug | D_ERROR) & level))
1368                 return;
1369
1370         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1371                ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1372                ns_is_client(ns) ? "client" : "server");
1373
1374         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1375                 return;
1376
1377         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1378                                  ldlm_res_hash_dump,
1379                                  (void *)(unsigned long)level);
1380         spin_lock(&ns->ns_lock);
1381         ns->ns_next_dump = cfs_time_shift(10);
1382         spin_unlock(&ns->ns_lock);
1383 }
1384 EXPORT_SYMBOL(ldlm_namespace_dump);
1385
1386 /**
1387  * Print information about all locks in this resource to debug log.
1388  */
1389 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1390 {
1391         struct ldlm_lock *lock;
1392         unsigned int granted = 0;
1393
1394         CLASSERT(RES_NAME_SIZE == 4);
1395
1396         if (!((libcfs_debug | D_ERROR) & level))
1397                 return;
1398
1399         CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1400                ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1401                res->lr_name.name[2], res->lr_name.name[3],
1402                cfs_atomic_read(&res->lr_refcount));
1403
1404         if (!cfs_list_empty(&res->lr_granted)) {
1405                 CDEBUG(level, "Granted locks (in reverse order):\n");
1406                 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1407                                                 l_res_link) {
1408                         LDLM_DEBUG_LIMIT(level, lock, "###");
1409                         if (!(level & D_CANTMASK) &&
1410                             ++granted > ldlm_dump_granted_max) {
1411                                 CDEBUG(level, "only dump %d granted locks to "
1412                                        "avoid DDOS.\n", granted);
1413                                 break;
1414                         }
1415                 }
1416         }
1417         if (!cfs_list_empty(&res->lr_converting)) {
1418                 CDEBUG(level, "Converting locks:\n");
1419                 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1420                         LDLM_DEBUG_LIMIT(level, lock, "###");
1421         }
1422         if (!cfs_list_empty(&res->lr_waiting)) {
1423                 CDEBUG(level, "Waiting locks:\n");
1424                 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1425                         LDLM_DEBUG_LIMIT(level, lock, "###");
1426         }
1427 }