Whamcloud - gitweb
LU-11997 ptlrpc: Properly swab ll_fiemap_info_key
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/ldlm/ldlm_resource.c
33  *
34  * Author: Phil Schwan <phil@clusterfs.com>
35  * Author: Peter Braam <braam@clusterfs.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_LDLM
39 #include <lustre_dlm.h>
40 #include <lustre_fid.h>
41 #include <obd_class.h>
42 #include "ldlm_internal.h"
43
44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
45 struct kmem_cache *ldlm_interval_tree_slab;
46 struct kmem_cache *ldlm_inodebits_slab;
47
48 int ldlm_srv_namespace_nr = 0;
49 int ldlm_cli_namespace_nr = 0;
50
51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
52 LIST_HEAD(ldlm_srv_namespace_list);
53
54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
55 /* Client Namespaces that have active resources in them.
56  * Once all resources go away, ldlm_poold moves such namespaces to the
57  * inactive list */
58 LIST_HEAD(ldlm_cli_active_namespace_list);
59 /* Client namespaces that don't have any locks in them */
60 LIST_HEAD(ldlm_cli_inactive_namespace_list);
61
62 static struct dentry *ldlm_debugfs_dir;
63 static struct dentry *ldlm_ns_debugfs_dir;
64 struct dentry *ldlm_svc_debugfs_dir;
65
66 /* during debug dump certain amount of granted locks for one resource to avoid
67  * DDOS. */
68 static unsigned int ldlm_dump_granted_max = 256;
69
70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
71                                           const char __user *buffer,
72                                           size_t count, loff_t *off)
73 {
74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
76         RETURN(count);
77 }
78
79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
80
81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
82 {
83         seq_printf(m, "%u\n", *(unsigned int *)m->private);
84         return 0;
85 }
86
87 static ssize_t
88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
89                        size_t count, loff_t *off)
90 {
91         struct seq_file *seq = file->private_data;
92
93         if (!count)
94                 return 0;
95
96         return kstrtouint_from_user(buffer, count, 0,
97                                     (unsigned int *)seq->private);
98 }
99
100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
101
102 #ifdef HAVE_SERVER_SUPPORT
103
104 static int seq_watermark_show(struct seq_file *m, void *data)
105 {
106         seq_printf(m, "%llu\n", *(__u64 *)m->private);
107         return 0;
108 }
109
110 static ssize_t seq_watermark_write(struct file *file,
111                                    const char __user *buffer, size_t count,
112                                    loff_t *off)
113 {
114         __s64 value;
115         __u64 watermark;
116         __u64 *data = ((struct seq_file *)file->private_data)->private;
117         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
118         int rc;
119
120         rc = lprocfs_str_with_units_to_s64(buffer, count, &value, 'M');
121         if (rc) {
122                 CERROR("Failed to set %s, rc = %d.\n",
123                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
124                        rc);
125                 return rc;
126         } else if (value != 0 && value < (1 << 20)) {
127                 CERROR("%s should be greater than 1MB.\n",
128                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
129                 return -EINVAL;
130         }
131         watermark = value >> 20;
132
133         if (wm_low) {
134                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
135                         CERROR("lock_reclaim_threshold_mb must be smaller than "
136                                "lock_limit_mb.\n");
137                         return -EINVAL;
138                 }
139
140                 *data = watermark;
141                 if (watermark != 0) {
142                         watermark <<= 20;
143                         do_div(watermark, sizeof(struct ldlm_lock));
144                 }
145                 ldlm_reclaim_threshold = watermark;
146         } else {
147                 if (ldlm_reclaim_threshold_mb != 0 &&
148                     watermark < ldlm_reclaim_threshold_mb) {
149                         CERROR("lock_limit_mb must be greater than "
150                                "lock_reclaim_threshold_mb.\n");
151                         return -EINVAL;
152                 }
153
154                 *data = watermark;
155                 if (watermark != 0) {
156                         watermark <<= 20;
157                         do_div(watermark, sizeof(struct ldlm_lock));
158                 }
159                 ldlm_lock_limit = watermark;
160         }
161
162         return count;
163 }
164
165 static int seq_watermark_open(struct inode *inode, struct file *file)
166 {
167         return single_open(file, seq_watermark_show, inode->i_private);
168 }
169
170 static const struct file_operations ldlm_watermark_fops = {
171         .owner          = THIS_MODULE,
172         .open           = seq_watermark_open,
173         .read           = seq_read,
174         .write          = seq_watermark_write,
175         .llseek         = seq_lseek,
176         .release        = lprocfs_single_release,
177 };
178
179 static int seq_granted_show(struct seq_file *m, void *data)
180 {
181         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
182                    (struct percpu_counter *)m->private));
183         return 0;
184 }
185
186 static int seq_granted_open(struct inode *inode, struct file *file)
187 {
188         return single_open(file, seq_granted_show, inode->i_private);
189 }
190
191 static const struct file_operations ldlm_granted_fops = {
192         .owner  = THIS_MODULE,
193         .open   = seq_granted_open,
194         .read   = seq_read,
195         .llseek = seq_lseek,
196         .release = seq_release,
197 };
198
199 #endif /* HAVE_SERVER_SUPPORT */
200
201 static struct lprocfs_vars ldlm_debugfs_list[] = {
202         { .name =       "dump_namespaces",
203           .fops =       &ldlm_dump_ns_fops,
204           .proc_mode =  0222 },
205         { .name =       "dump_granted_max",
206           .fops =       &ldlm_rw_uint_fops,
207           .data =       &ldlm_dump_granted_max },
208 #ifdef HAVE_SERVER_SUPPORT
209         { .name =       "lock_reclaim_threshold_mb",
210           .fops =       &ldlm_watermark_fops,
211           .data =       &ldlm_reclaim_threshold_mb },
212         { .name =       "lock_limit_mb",
213           .fops =       &ldlm_watermark_fops,
214           .data =       &ldlm_lock_limit_mb },
215         { .name =       "lock_granted_count",
216           .fops =       &ldlm_granted_fops,
217           .data =       &ldlm_granted_total },
218 #endif
219         { NULL }
220 };
221
222 int ldlm_debugfs_setup(void)
223 {
224         int rc;
225
226         ENTRY;
227         ldlm_debugfs_dir = ldebugfs_register(OBD_LDLM_DEVICENAME,
228                                              debugfs_lustre_root,
229                                              NULL, NULL);
230         if (IS_ERR_OR_NULL(ldlm_debugfs_dir)) {
231                 CERROR("LDebugFS failed in ldlm-init\n");
232                 rc = ldlm_debugfs_dir ? PTR_ERR(ldlm_debugfs_dir) : -ENOMEM;
233                 GOTO(err, rc);
234         }
235
236         ldlm_ns_debugfs_dir = ldebugfs_register("namespaces",
237                                                 ldlm_debugfs_dir,
238                                                 NULL, NULL);
239         if (IS_ERR_OR_NULL(ldlm_ns_debugfs_dir)) {
240                 CERROR("LProcFS failed in ldlm-init\n");
241                 rc = ldlm_ns_debugfs_dir ? PTR_ERR(ldlm_ns_debugfs_dir)
242                                          : -ENOMEM;
243                 GOTO(err_type, rc);
244         }
245
246         ldlm_svc_debugfs_dir = ldebugfs_register("services",
247                                                  ldlm_debugfs_dir,
248                                                  NULL, NULL);
249         if (IS_ERR_OR_NULL(ldlm_svc_debugfs_dir)) {
250                 CERROR("LProcFS failed in ldlm-init\n");
251                 rc = ldlm_svc_debugfs_dir ? PTR_ERR(ldlm_svc_debugfs_dir)
252                                           : -ENOMEM;
253                 GOTO(err_ns, rc);
254         }
255
256         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
257         if (rc != 0) {
258                 CERROR("LProcFS failed in ldlm-init\n");
259                 GOTO(err_svc, rc);
260         }
261
262         RETURN(0);
263
264 err_svc:
265         ldebugfs_remove(&ldlm_svc_debugfs_dir);
266 err_ns:
267         ldebugfs_remove(&ldlm_ns_debugfs_dir);
268 err_type:
269         ldebugfs_remove(&ldlm_debugfs_dir);
270 err:
271         ldlm_svc_debugfs_dir = NULL;
272         ldlm_ns_debugfs_dir = NULL;
273         ldlm_debugfs_dir = NULL;
274         RETURN(rc);
275 }
276
277 void ldlm_debugfs_cleanup(void)
278 {
279         if (!IS_ERR_OR_NULL(ldlm_svc_debugfs_dir))
280                 ldebugfs_remove(&ldlm_svc_debugfs_dir);
281
282         if (!IS_ERR_OR_NULL(ldlm_ns_debugfs_dir))
283                 ldebugfs_remove(&ldlm_ns_debugfs_dir);
284
285         if (!IS_ERR_OR_NULL(ldlm_debugfs_dir))
286                 ldebugfs_remove(&ldlm_debugfs_dir);
287
288         ldlm_svc_debugfs_dir = NULL;
289         ldlm_ns_debugfs_dir = NULL;
290         ldlm_debugfs_dir = NULL;
291 }
292
293 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
294                                    char *buf)
295 {
296         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
297                                                  ns_kobj);
298         __u64                   res = 0;
299         struct cfs_hash_bd              bd;
300         int                     i;
301
302         /* result is not strictly consistant */
303         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
304                 res += cfs_hash_bd_count_get(&bd);
305         return sprintf(buf, "%lld\n", res);
306 }
307 LUSTRE_RO_ATTR(resource_count);
308
309 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
310                                char *buf)
311 {
312         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
313                                                  ns_kobj);
314         __u64                   locks;
315
316         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
317                                         LPROCFS_FIELDS_FLAGS_SUM);
318         return sprintf(buf, "%lld\n", locks);
319 }
320 LUSTRE_RO_ATTR(lock_count);
321
322 static ssize_t lock_unused_count_show(struct kobject *kobj,
323                                       struct attribute *attr,
324                                       char *buf)
325 {
326         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
327                                                  ns_kobj);
328
329         return sprintf(buf, "%d\n", ns->ns_nr_unused);
330 }
331 LUSTRE_RO_ATTR(lock_unused_count);
332
333 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
334                              char *buf)
335 {
336         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
337                                                  ns_kobj);
338         __u32 *nr = &ns->ns_max_unused;
339
340         if (ns_connect_lru_resize(ns))
341                 nr = &ns->ns_nr_unused;
342         return sprintf(buf, "%u\n", *nr);
343 }
344
345 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
346                               const char *buffer, size_t count)
347 {
348         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
349                                                  ns_kobj);
350         unsigned long tmp;
351         int lru_resize;
352         int err;
353
354         if (strncmp(buffer, "clear", 5) == 0) {
355                 CDEBUG(D_DLMTRACE,
356                        "dropping all unused locks from namespace %s\n",
357                        ldlm_ns_name(ns));
358                 if (ns_connect_lru_resize(ns)) {
359                         /* Try to cancel all @ns_nr_unused locks. */
360                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
361                                         LDLM_LRU_FLAG_PASSED |
362                                         LDLM_LRU_FLAG_CLEANUP);
363                 } else {
364                         tmp = ns->ns_max_unused;
365                         ns->ns_max_unused = 0;
366                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
367                                         LDLM_LRU_FLAG_CLEANUP);
368                         ns->ns_max_unused = tmp;
369                 }
370                 return count;
371         }
372
373         err = kstrtoul(buffer, 10, &tmp);
374         if (err != 0) {
375                 CERROR("lru_size: invalid value written\n");
376                 return -EINVAL;
377         }
378         lru_resize = (tmp == 0);
379
380         if (ns_connect_lru_resize(ns)) {
381                 if (!lru_resize)
382                         ns->ns_max_unused = (unsigned int)tmp;
383
384                 if (tmp > ns->ns_nr_unused)
385                         tmp = ns->ns_nr_unused;
386                 tmp = ns->ns_nr_unused - tmp;
387
388                 CDEBUG(D_DLMTRACE,
389                        "changing namespace %s unused locks from %u to %u\n",
390                        ldlm_ns_name(ns), ns->ns_nr_unused,
391                        (unsigned int)tmp);
392                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
393
394                 if (!lru_resize) {
395                         CDEBUG(D_DLMTRACE,
396                                "disable lru_resize for namespace %s\n",
397                                ldlm_ns_name(ns));
398                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
399                 }
400         } else {
401                 CDEBUG(D_DLMTRACE,
402                        "changing namespace %s max_unused from %u to %u\n",
403                        ldlm_ns_name(ns), ns->ns_max_unused,
404                        (unsigned int)tmp);
405                 ns->ns_max_unused = (unsigned int)tmp;
406                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
407
408                 /* Make sure that LRU resize was originally supported before
409                  * turning it on here.
410                  */
411                 if (lru_resize &&
412                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
413                         CDEBUG(D_DLMTRACE,
414                                "enable lru_resize for namespace %s\n",
415                                ldlm_ns_name(ns));
416                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
417                 }
418         }
419
420         return count;
421 }
422 LUSTRE_RW_ATTR(lru_size);
423
424 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
425                                 char *buf)
426 {
427         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
428                                                  ns_kobj);
429
430         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
431 }
432
433 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
434                                  const char *buffer, size_t count)
435 {
436         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
437                                                  ns_kobj);
438         int scale = NSEC_PER_MSEC;
439         unsigned long long tmp;
440         char *buf;
441
442         /* Did the user ask in seconds or milliseconds. Default is in ms */
443         buf = strstr(buffer, "ms");
444         if (!buf) {
445                 buf = strchr(buffer, 's');
446                 if (buf)
447                         scale = NSEC_PER_SEC;
448         }
449
450         if (buf)
451                 *buf = '\0';
452
453         if (kstrtoull(buffer, 10, &tmp))
454                 return -EINVAL;
455
456         ns->ns_max_age = ktime_set(0, tmp * scale);
457
458         return count;
459 }
460 LUSTRE_RW_ATTR(lru_max_age);
461
462 static ssize_t early_lock_cancel_show(struct kobject *kobj,
463                                       struct attribute *attr,
464                                       char *buf)
465 {
466         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
467                                                  ns_kobj);
468
469         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
470 }
471
472 static ssize_t early_lock_cancel_store(struct kobject *kobj,
473                                        struct attribute *attr,
474                                        const char *buffer,
475                                        size_t count)
476 {
477         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
478                                                  ns_kobj);
479         unsigned long supp = -1;
480         int rc;
481
482         rc = kstrtoul(buffer, 10, &supp);
483         if (rc < 0)
484                 return rc;
485
486         if (supp == 0)
487                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
488         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
489                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
490         return count;
491 }
492 LUSTRE_RW_ATTR(early_lock_cancel);
493
494 static ssize_t dirty_age_limit_show(struct kobject *kobj,
495                                     struct attribute *attr, char *buf)
496 {
497         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
498                                                  ns_kobj);
499
500         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
501 }
502
503 static ssize_t dirty_age_limit_store(struct kobject *kobj,
504                                      struct attribute *attr,
505                                      const char *buffer, size_t count)
506 {
507         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
508                                                  ns_kobj);
509         unsigned long long tmp;
510
511         if (kstrtoull(buffer, 10, &tmp))
512                 return -EINVAL;
513
514         ns->ns_dirty_age_limit = tmp;
515
516         return count;
517 }
518 LUSTRE_RW_ATTR(dirty_age_limit);
519
520 #ifdef HAVE_SERVER_SUPPORT
521 static ssize_t ctime_age_limit_show(struct kobject *kobj,
522                                     struct attribute *attr, char *buf)
523 {
524         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
525                                                  ns_kobj);
526
527         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
528 }
529
530 static ssize_t ctime_age_limit_store(struct kobject *kobj,
531                                      struct attribute *attr,
532                                      const char *buffer, size_t count)
533 {
534         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
535                                                  ns_kobj);
536         unsigned long long tmp;
537
538         if (kstrtoull(buffer, 10, &tmp))
539                 return -EINVAL;
540
541         ns->ns_ctime_age_limit = tmp;
542
543         return count;
544 }
545 LUSTRE_RW_ATTR(ctime_age_limit);
546
547 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
548                                   char *buf)
549 {
550         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
551                                                  ns_kobj);
552
553         return sprintf(buf, "%d\n", ns->ns_timeouts);
554 }
555 LUSTRE_RO_ATTR(lock_timeouts);
556
557 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
558                                      struct attribute *attr, char *buf)
559 {
560         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
561                                                  ns_kobj);
562
563         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
564 }
565
566 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
567                                       struct attribute *attr,
568                                       const char *buffer, size_t count)
569 {
570         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
571                                                  ns_kobj);
572         unsigned long tmp;
573         int err;
574
575         err = kstrtoul(buffer, 10, &tmp);
576         if (err != 0)
577                 return -EINVAL;
578
579         ns->ns_max_nolock_size = tmp;
580
581         return count;
582 }
583 LUSTRE_RW_ATTR(max_nolock_bytes);
584
585 static ssize_t contention_seconds_show(struct kobject *kobj,
586                                        struct attribute *attr, char *buf)
587 {
588         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
589                                                  ns_kobj);
590
591         return sprintf(buf, "%llu\n", ns->ns_contention_time);
592 }
593
594 static ssize_t contention_seconds_store(struct kobject *kobj,
595                                         struct attribute *attr,
596                                         const char *buffer, size_t count)
597 {
598         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
599                                                  ns_kobj);
600         unsigned long long tmp;
601
602         if (kstrtoull(buffer, 10, &tmp))
603                 return -EINVAL;
604
605         ns->ns_contention_time = tmp;
606
607         return count;
608 }
609 LUSTRE_RW_ATTR(contention_seconds);
610
611 static ssize_t contended_locks_show(struct kobject *kobj,
612                                     struct attribute *attr, char *buf)
613 {
614         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
615                                                  ns_kobj);
616
617         return sprintf(buf, "%u\n", ns->ns_contended_locks);
618 }
619
620 static ssize_t contended_locks_store(struct kobject *kobj,
621                                      struct attribute *attr,
622                                      const char *buffer, size_t count)
623 {
624         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
625                                                  ns_kobj);
626         unsigned long tmp;
627         int err;
628
629         err = kstrtoul(buffer, 10, &tmp);
630         if (err != 0)
631                 return -EINVAL;
632
633         ns->ns_contended_locks = tmp;
634
635         return count;
636 }
637 LUSTRE_RW_ATTR(contended_locks);
638
639 static ssize_t max_parallel_ast_show(struct kobject *kobj,
640                                      struct attribute *attr, char *buf)
641 {
642         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
643                                                  ns_kobj);
644
645         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
646 }
647
648 static ssize_t max_parallel_ast_store(struct kobject *kobj,
649                                       struct attribute *attr,
650                                       const char *buffer, size_t count)
651 {
652         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
653                                                  ns_kobj);
654         unsigned long tmp;
655         int err;
656
657         err = kstrtoul(buffer, 10, &tmp);
658         if (err != 0)
659                 return -EINVAL;
660
661         ns->ns_max_parallel_ast = tmp;
662
663         return count;
664 }
665 LUSTRE_RW_ATTR(max_parallel_ast);
666
667 #endif /* HAVE_SERVER_SUPPORT */
668
669 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
670 static struct attribute *ldlm_ns_attrs[] = {
671         &lustre_attr_resource_count.attr,
672         &lustre_attr_lock_count.attr,
673         &lustre_attr_lock_unused_count.attr,
674         &lustre_attr_lru_size.attr,
675         &lustre_attr_lru_max_age.attr,
676         &lustre_attr_early_lock_cancel.attr,
677         &lustre_attr_dirty_age_limit.attr,
678 #ifdef HAVE_SERVER_SUPPORT
679         &lustre_attr_ctime_age_limit.attr,
680         &lustre_attr_lock_timeouts.attr,
681         &lustre_attr_max_nolock_bytes.attr,
682         &lustre_attr_contention_seconds.attr,
683         &lustre_attr_contended_locks.attr,
684         &lustre_attr_max_parallel_ast.attr,
685 #endif
686         NULL,
687 };
688
689 static void ldlm_ns_release(struct kobject *kobj)
690 {
691         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
692                                                  ns_kobj);
693         complete(&ns->ns_kobj_unregister);
694 }
695
696 static struct kobj_type ldlm_ns_ktype = {
697         .default_attrs  = ldlm_ns_attrs,
698         .sysfs_ops      = &lustre_sysfs_ops,
699         .release        = ldlm_ns_release,
700 };
701
702 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
703 {
704         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
705                 CERROR("dlm namespace %s has no procfs dir?\n",
706                        ldlm_ns_name(ns));
707         else
708                 ldebugfs_remove(&ns->ns_debugfs_entry);
709
710         if (ns->ns_stats != NULL)
711                 lprocfs_free_stats(&ns->ns_stats);
712 }
713
714 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
715 {
716         kobject_put(&ns->ns_kobj);
717         wait_for_completion(&ns->ns_kobj_unregister);
718 }
719
720 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
721 {
722         int err;
723
724         ns->ns_kobj.kset = ldlm_ns_kset;
725         init_completion(&ns->ns_kobj_unregister);
726         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
727                                    "%s", ldlm_ns_name(ns));
728
729         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
730         if (!ns->ns_stats) {
731                 kobject_put(&ns->ns_kobj);
732                 return -ENOMEM;
733         }
734
735         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
736                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
737
738         return err;
739 }
740
741 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
742 {
743         struct dentry *ns_entry;
744
745         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
746                 ns_entry = ns->ns_debugfs_entry;
747         } else {
748                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
749                                               ldlm_ns_debugfs_dir);
750                 if (!ns_entry)
751                         return -ENOMEM;
752                 ns->ns_debugfs_entry = ns_entry;
753         }
754
755         return 0;
756 }
757 #undef MAX_STRING_SIZE
758
759 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
760                                   const void *key, unsigned int mask)
761 {
762         const struct ldlm_res_id *id = key;
763         unsigned int val = 0;
764         unsigned int i;
765
766         for (i = 0; i < RES_NAME_SIZE; i++)
767                 val += id->name[i];
768         return val & mask;
769 }
770
771 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
772                                       const void *key, unsigned mask)
773 {
774         const struct ldlm_res_id *id = key;
775         struct lu_fid       fid;
776         __u32               hash;
777         __u32               val;
778
779         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
780         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
781         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
782
783         hash = fid_flatten32(&fid);
784         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
785         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
786                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
787                 hash += (val >> 5) + (val << 11);
788         } else {
789                 val = fid_oid(&fid);
790         }
791         hash = hash_long(hash, hs->hs_bkt_bits);
792         /* give me another random factor */
793         hash -= hash_long((unsigned long)hs, val % 11 + 3);
794
795         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
796         hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
797
798         return hash & mask;
799 }
800
801 static void *ldlm_res_hop_key(struct hlist_node *hnode)
802 {
803         struct ldlm_resource   *res;
804
805         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
806         return &res->lr_name;
807 }
808
809 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
810 {
811         struct ldlm_resource   *res;
812
813         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
814         return ldlm_res_eq((const struct ldlm_res_id *)key,
815                            (const struct ldlm_res_id *)&res->lr_name);
816 }
817
818 static void *ldlm_res_hop_object(struct hlist_node *hnode)
819 {
820         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
821 }
822
823 static void
824 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
825 {
826         struct ldlm_resource *res;
827
828         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
829         ldlm_resource_getref(res);
830 }
831
832 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
833 {
834         struct ldlm_resource *res;
835
836         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
837         ldlm_resource_putref(res);
838 }
839
840 static struct cfs_hash_ops ldlm_ns_hash_ops = {
841         .hs_hash        = ldlm_res_hop_hash,
842         .hs_key         = ldlm_res_hop_key,
843         .hs_keycmp      = ldlm_res_hop_keycmp,
844         .hs_keycpy      = NULL,
845         .hs_object      = ldlm_res_hop_object,
846         .hs_get         = ldlm_res_hop_get_locked,
847         .hs_put         = ldlm_res_hop_put
848 };
849
850 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
851         .hs_hash        = ldlm_res_hop_fid_hash,
852         .hs_key         = ldlm_res_hop_key,
853         .hs_keycmp      = ldlm_res_hop_keycmp,
854         .hs_keycpy      = NULL,
855         .hs_object      = ldlm_res_hop_object,
856         .hs_get         = ldlm_res_hop_get_locked,
857         .hs_put         = ldlm_res_hop_put
858 };
859
860 typedef struct ldlm_ns_hash_def {
861         enum ldlm_ns_type       nsd_type;
862         /** hash bucket bits */
863         unsigned                nsd_bkt_bits;
864         /** hash bits */
865         unsigned                nsd_all_bits;
866         /** hash operations */
867         struct cfs_hash_ops *nsd_hops;
868 } ldlm_ns_hash_def_t;
869
870 static struct ldlm_ns_hash_def ldlm_ns_hash_defs[] =
871 {
872         {
873                 .nsd_type       = LDLM_NS_TYPE_MDC,
874                 .nsd_bkt_bits   = 11,
875                 .nsd_all_bits   = 16,
876                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
877         },
878         {
879                 .nsd_type       = LDLM_NS_TYPE_MDT,
880                 .nsd_bkt_bits   = 14,
881                 .nsd_all_bits   = 21,
882                 .nsd_hops       = &ldlm_ns_fid_hash_ops,
883         },
884         {
885                 .nsd_type       = LDLM_NS_TYPE_OSC,
886                 .nsd_bkt_bits   = 8,
887                 .nsd_all_bits   = 12,
888                 .nsd_hops       = &ldlm_ns_hash_ops,
889         },
890         {
891                 .nsd_type       = LDLM_NS_TYPE_OST,
892                 .nsd_bkt_bits   = 11,
893                 .nsd_all_bits   = 17,
894                 .nsd_hops       = &ldlm_ns_hash_ops,
895         },
896         {
897                 .nsd_type       = LDLM_NS_TYPE_MGC,
898                 .nsd_bkt_bits   = 4,
899                 .nsd_all_bits   = 4,
900                 .nsd_hops       = &ldlm_ns_hash_ops,
901         },
902         {
903                 .nsd_type       = LDLM_NS_TYPE_MGT,
904                 .nsd_bkt_bits   = 4,
905                 .nsd_all_bits   = 4,
906                 .nsd_hops       = &ldlm_ns_hash_ops,
907         },
908         {
909                 .nsd_type       = LDLM_NS_TYPE_UNKNOWN,
910         },
911 };
912
913 /**
914  * Create and initialize new empty namespace.
915  */
916 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
917                                           enum ldlm_side client,
918                                           enum ldlm_appetite apt,
919                                           enum ldlm_ns_type ns_type)
920 {
921         struct ldlm_namespace *ns = NULL;
922         struct ldlm_ns_bucket *nsb;
923         struct ldlm_ns_hash_def *nsd;
924         struct cfs_hash_bd bd;
925         int idx;
926         int rc;
927
928         ENTRY;
929         LASSERT(obd != NULL);
930
931         rc = ldlm_get_ref();
932         if (rc) {
933                 CERROR("ldlm_get_ref failed: %d\n", rc);
934                 RETURN(NULL);
935         }
936
937         for (idx = 0; ; idx++) {
938                 nsd = &ldlm_ns_hash_defs[idx];
939                 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
940                         CERROR("Unknown type %d for ns %s\n", ns_type, name);
941                         GOTO(out_ref, NULL);
942                 }
943
944                 if (nsd->nsd_type == ns_type)
945                         break;
946         }
947
948         OBD_ALLOC_PTR(ns);
949         if (!ns)
950                 GOTO(out_ref, NULL);
951
952         ns->ns_rs_hash = cfs_hash_create(name,
953                                          nsd->nsd_all_bits, nsd->nsd_all_bits,
954                                          nsd->nsd_bkt_bits, sizeof(*nsb),
955                                          CFS_HASH_MIN_THETA,
956                                          CFS_HASH_MAX_THETA,
957                                          nsd->nsd_hops,
958                                          CFS_HASH_DEPTH |
959                                          CFS_HASH_BIGNAME |
960                                          CFS_HASH_SPIN_BKTLOCK |
961                                          CFS_HASH_NO_ITEMREF);
962         if (ns->ns_rs_hash == NULL)
963                 GOTO(out_ns, NULL);
964
965         cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
966                 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
967                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
968                 nsb->nsb_namespace = ns;
969                 nsb->nsb_reclaim_start = 0;
970         }
971
972         ns->ns_obd = obd;
973         ns->ns_appetite = apt;
974         ns->ns_client = client;
975         ns->ns_name = kstrdup(name, GFP_KERNEL);
976         if (!ns->ns_name)
977                 goto out_hash;
978
979         INIT_LIST_HEAD(&ns->ns_list_chain);
980         INIT_LIST_HEAD(&ns->ns_unused_list);
981         spin_lock_init(&ns->ns_lock);
982         atomic_set(&ns->ns_bref, 0);
983         init_waitqueue_head(&ns->ns_waitq);
984
985         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
986         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
987         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
988
989         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
990         ns->ns_nr_unused          = 0;
991         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
992         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
993         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
994         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
995         ns->ns_timeouts           = 0;
996         ns->ns_orig_connect_flags = 0;
997         ns->ns_connect_flags      = 0;
998         ns->ns_stopping           = 0;
999         ns->ns_reclaim_start      = 0;
1000         ns->ns_last_pos           = &ns->ns_unused_list;
1001
1002         rc = ldlm_namespace_sysfs_register(ns);
1003         if (rc) {
1004                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
1005                 GOTO(out_hash, rc);
1006         }
1007
1008         rc = ldlm_namespace_debugfs_register(ns);
1009         if (rc) {
1010                 CERROR("Can't initialize ns proc, rc %d\n", rc);
1011                 GOTO(out_sysfs, rc);
1012         }
1013
1014         idx = ldlm_namespace_nr_read(client);
1015         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
1016         if (rc) {
1017                 CERROR("Can't initialize lock pool, rc %d\n", rc);
1018                 GOTO(out_proc, rc);
1019         }
1020
1021         ldlm_namespace_register(ns, client);
1022         RETURN(ns);
1023 out_proc:
1024         ldlm_namespace_debugfs_unregister(ns);
1025 out_sysfs:
1026         ldlm_namespace_sysfs_unregister(ns);
1027         ldlm_namespace_cleanup(ns, 0);
1028 out_hash:
1029         kfree(ns->ns_name);
1030         cfs_hash_putref(ns->ns_rs_hash);
1031 out_ns:
1032         OBD_FREE_PTR(ns);
1033 out_ref:
1034         ldlm_put_ref();
1035         RETURN(NULL);
1036 }
1037 EXPORT_SYMBOL(ldlm_namespace_new);
1038
1039 /**
1040  * Cancel and destroy all locks on a resource.
1041  *
1042  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1043  * clean up.  This is currently only used for recovery, and we make
1044  * certain assumptions as a result--notably, that we shouldn't cancel
1045  * locks with refs.
1046  */
1047 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1048                              __u64 flags)
1049 {
1050         struct list_head *tmp;
1051         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1052         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1053
1054         do {
1055                 struct ldlm_lock *lock = NULL;
1056
1057                 /* First, we look for non-cleaned-yet lock
1058                  * all cleaned locks are marked by CLEANED flag. */
1059                 lock_res(res);
1060                 list_for_each(tmp, q) {
1061                         lock = list_entry(tmp, struct ldlm_lock,
1062                                           l_res_link);
1063                         if (ldlm_is_cleaned(lock)) {
1064                                 lock = NULL;
1065                                 continue;
1066                         }
1067                         LDLM_LOCK_GET(lock);
1068                         ldlm_set_cleaned(lock);
1069                         break;
1070                 }
1071
1072                 if (lock == NULL) {
1073                         unlock_res(res);
1074                         break;
1075                 }
1076
1077                 /* Set CBPENDING so nothing in the cancellation path
1078                  * can match this lock. */
1079                 ldlm_set_cbpending(lock);
1080                 ldlm_set_failed(lock);
1081                 lock->l_flags |= flags;
1082
1083                 /* ... without sending a CANCEL message for local_only. */
1084                 if (local_only)
1085                         ldlm_set_local_only(lock);
1086
1087                 if (local_only && (lock->l_readers || lock->l_writers)) {
1088                         /*
1089                          * This is a little bit gross, but much better than the
1090                          * alternative: pretend that we got a blocking AST from
1091                          * the server, so that when the lock is decref'd, it
1092                          * will go away ...
1093                          */
1094                         unlock_res(res);
1095                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1096                         if (lock->l_flags & LDLM_FL_FAIL_LOC) {
1097                                 set_current_state(TASK_UNINTERRUPTIBLE);
1098                                 schedule_timeout(cfs_time_seconds(4));
1099                                 set_current_state(TASK_RUNNING);
1100                         }
1101                         if (lock->l_completion_ast)
1102                                 lock->l_completion_ast(lock,
1103                                                        LDLM_FL_FAILED, NULL);
1104                         LDLM_LOCK_RELEASE(lock);
1105                         continue;
1106                 }
1107
1108                 if (client) {
1109                         struct lustre_handle lockh;
1110
1111                         unlock_res(res);
1112                         ldlm_lock2handle(lock, &lockh);
1113                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1114                         if (rc)
1115                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1116                 } else {
1117                         unlock_res(res);
1118                         LDLM_DEBUG(lock,
1119                                    "Freeing a lock still held by a client node");
1120                         ldlm_lock_cancel(lock);
1121                 }
1122                 LDLM_LOCK_RELEASE(lock);
1123         } while (1);
1124 }
1125
1126 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1127                                struct hlist_node *hnode, void *arg)
1128 {
1129         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1130         __u64 flags = *(__u64 *)arg;
1131
1132         cleanup_resource(res, &res->lr_granted, flags);
1133         cleanup_resource(res, &res->lr_waiting, flags);
1134
1135         return 0;
1136 }
1137
1138 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1139                                   struct hlist_node *hnode, void *arg)
1140 {
1141         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1142
1143         lock_res(res);
1144         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1145                "(%d) after lock cleanup; forcing cleanup.\n",
1146                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1147                atomic_read(&res->lr_refcount) - 1);
1148
1149         /* Use D_NETERROR since it is in the default mask */
1150         ldlm_resource_dump(D_NETERROR, res);
1151         unlock_res(res);
1152         return 0;
1153 }
1154
1155 /**
1156  * Cancel and destroy all locks in the namespace.
1157  *
1158  * Typically used during evictions when server notified client that it was
1159  * evicted and all of its state needs to be destroyed.
1160  * Also used during shutdown.
1161  */
1162 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1163 {
1164         if (ns == NULL) {
1165                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1166                 return ELDLM_OK;
1167         }
1168
1169         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1170                                  &flags, 0);
1171         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1172                                  NULL, 0);
1173         return ELDLM_OK;
1174 }
1175 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1176
1177 /**
1178  * Attempts to free namespace.
1179  *
1180  * Only used when namespace goes away, like during an unmount.
1181  */
1182 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1183 {
1184         ENTRY;
1185
1186         /* At shutdown time, don't call the cancellation callback */
1187         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1188
1189         if (atomic_read(&ns->ns_bref) > 0) {
1190                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
1191                 int rc;
1192                 CDEBUG(D_DLMTRACE,
1193                        "dlm namespace %s free waiting on refcount %d\n",
1194                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1195 force_wait:
1196                 if (force)
1197                         lwi = LWI_TIMEOUT(cfs_time_seconds(1) / 4,
1198                                           NULL, NULL);
1199
1200                 rc = l_wait_event(ns->ns_waitq,
1201                                   atomic_read(&ns->ns_bref) == 0, &lwi);
1202
1203                 /* Forced cleanups should be able to reclaim all references,
1204                  * so it's safe to wait forever... we can't leak locks... */
1205                 if (force && rc == -ETIMEDOUT) {
1206                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1207                                        "namespace with %d resources in use, "
1208                                        "(rc=%d)\n", ldlm_ns_name(ns),
1209                                        atomic_read(&ns->ns_bref), rc);
1210                         GOTO(force_wait, rc);
1211                 }
1212
1213                 if (atomic_read(&ns->ns_bref)) {
1214                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1215                                        "with %d resources in use, (rc=%d)\n",
1216                                        ldlm_ns_name(ns),
1217                                        atomic_read(&ns->ns_bref), rc);
1218                         RETURN(ELDLM_NAMESPACE_EXISTS);
1219                 }
1220                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1221                        ldlm_ns_name(ns));
1222         }
1223
1224         RETURN(ELDLM_OK);
1225 }
1226
1227 /**
1228  * Performs various cleanups for passed \a ns to make it drop refc and be
1229  * ready for freeing. Waits for refc == 0.
1230  *
1231  * The following is done:
1232  * (0) Unregister \a ns from its list to make inaccessible for potential
1233  * users like pools thread and others;
1234  * (1) Clear all locks in \a ns.
1235  */
1236 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1237                                struct obd_import *imp,
1238                                int force)
1239 {
1240         int rc;
1241
1242         ENTRY;
1243         if (!ns) {
1244                 EXIT;
1245                 return;
1246         }
1247
1248         spin_lock(&ns->ns_lock);
1249         ns->ns_stopping = 1;
1250         spin_unlock(&ns->ns_lock);
1251
1252         /*
1253          * Can fail with -EINTR when force == 0 in which case try harder.
1254          */
1255         rc = __ldlm_namespace_free(ns, force);
1256         if (rc != ELDLM_OK) {
1257                 if (imp) {
1258                         ptlrpc_disconnect_import(imp, 0);
1259                         ptlrpc_invalidate_import(imp);
1260                 }
1261
1262                 /*
1263                  * With all requests dropped and the import inactive
1264                  * we are gaurenteed all reference will be dropped.
1265                  */
1266                 rc = __ldlm_namespace_free(ns, 1);
1267                 LASSERT(rc == 0);
1268         }
1269         EXIT;
1270 }
1271 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1272
1273 /**
1274  * Performs freeing memory structures related to \a ns. This is only done
1275  * when ldlm_namespce_free_prior() successfully removed all resources
1276  * referencing \a ns and its refc == 0.
1277  */
1278 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1279 {
1280         ENTRY;
1281         if (!ns) {
1282                 EXIT;
1283                 return;
1284         }
1285
1286         /* Make sure that nobody can find this ns in its list. */
1287         ldlm_namespace_unregister(ns, ns->ns_client);
1288         /* Fini pool _before_ parent proc dir is removed. This is important as
1289          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1290          * Removing it after @dir may cause oops. */
1291         ldlm_pool_fini(&ns->ns_pool);
1292
1293         ldlm_namespace_debugfs_unregister(ns);
1294         ldlm_namespace_sysfs_unregister(ns);
1295         cfs_hash_putref(ns->ns_rs_hash);
1296         kfree(ns->ns_name);
1297         /* Namespace \a ns should be not on list at this time, otherwise
1298          * this will cause issues related to using freed \a ns in poold
1299          * thread.
1300          */
1301         LASSERT(list_empty(&ns->ns_list_chain));
1302         OBD_FREE_PTR(ns);
1303         ldlm_put_ref();
1304         EXIT;
1305 }
1306 EXPORT_SYMBOL(ldlm_namespace_free_post);
1307
1308 /**
1309  * Cleanup the resource, and free namespace.
1310  * bug 12864:
1311  * Deadlock issue:
1312  * proc1: destroy import
1313  *        class_disconnect_export(grab cl_sem) ->
1314  *              -> ldlm_namespace_free ->
1315  *              -> lprocfs_remove(grab _lprocfs_lock).
1316  * proc2: read proc info
1317  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1318  *              -> osc_rd_active, etc(grab cl_sem).
1319  *
1320  * So that I have to split the ldlm_namespace_free into two parts - the first
1321  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1322  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1323  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1324  * held.
1325  */
1326 void ldlm_namespace_free(struct ldlm_namespace *ns,
1327                          struct obd_import *imp,
1328                          int force)
1329 {
1330         ldlm_namespace_free_prior(ns, imp, force);
1331         ldlm_namespace_free_post(ns);
1332 }
1333 EXPORT_SYMBOL(ldlm_namespace_free);
1334
1335 void ldlm_namespace_get(struct ldlm_namespace *ns)
1336 {
1337         atomic_inc(&ns->ns_bref);
1338 }
1339
1340 /* This is only for callers that care about refcount */
1341 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1342 {
1343         return atomic_inc_return(&ns->ns_bref);
1344 }
1345
1346 void ldlm_namespace_put(struct ldlm_namespace *ns)
1347 {
1348         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1349                 wake_up(&ns->ns_waitq);
1350                 spin_unlock(&ns->ns_lock);
1351         }
1352 }
1353
1354 /** Register \a ns in the list of namespaces */
1355 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1356 {
1357         mutex_lock(ldlm_namespace_lock(client));
1358         LASSERT(list_empty(&ns->ns_list_chain));
1359         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1360         ldlm_namespace_nr_inc(client);
1361         mutex_unlock(ldlm_namespace_lock(client));
1362 }
1363
1364 /** Unregister \a ns from the list of namespaces. */
1365 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1366 {
1367         mutex_lock(ldlm_namespace_lock(client));
1368         LASSERT(!list_empty(&ns->ns_list_chain));
1369         /* Some asserts and possibly other parts of the code are still
1370          * using list_empty(&ns->ns_list_chain). This is why it is
1371          * important to use list_del_init() here. */
1372         list_del_init(&ns->ns_list_chain);
1373         ldlm_namespace_nr_dec(client);
1374         mutex_unlock(ldlm_namespace_lock(client));
1375 }
1376
1377 /** Should be called with ldlm_namespace_lock(client) taken. */
1378 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1379                                           enum ldlm_side client)
1380 {
1381         LASSERT(!list_empty(&ns->ns_list_chain));
1382         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1383         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1384 }
1385
1386 /** Should be called with ldlm_namespace_lock(client) taken. */
1387 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1388                                             enum ldlm_side client)
1389 {
1390         LASSERT(!list_empty(&ns->ns_list_chain));
1391         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1392         list_move_tail(&ns->ns_list_chain,
1393                        ldlm_namespace_inactive_list(client));
1394 }
1395
1396 /** Should be called with ldlm_namespace_lock(client) taken. */
1397 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1398 {
1399         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1400         LASSERT(!list_empty(ldlm_namespace_list(client)));
1401         return container_of(ldlm_namespace_list(client)->next,
1402                             struct ldlm_namespace, ns_list_chain);
1403 }
1404
1405 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1406 {
1407         int idx;
1408
1409         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1410                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1411         if (res->lr_itree == NULL)
1412                 return false;
1413         /* Initialize interval trees for each lock mode. */
1414         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1415                 res->lr_itree[idx].lit_size = 0;
1416                 res->lr_itree[idx].lit_mode = 1 << idx;
1417                 res->lr_itree[idx].lit_root = NULL;
1418         }
1419         return true;
1420 }
1421
1422 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1423 {
1424         int i;
1425
1426         OBD_ALLOC_PTR(res->lr_ibits_queues);
1427         if (res->lr_ibits_queues == NULL)
1428                 return false;
1429         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1430                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1431         return true;
1432 }
1433
1434 /** Create and initialize new resource. */
1435 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1436 {
1437         struct ldlm_resource *res;
1438         bool rc;
1439
1440         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1441         if (res == NULL)
1442                 return NULL;
1443
1444         switch (ldlm_type) {
1445         case LDLM_EXTENT:
1446                 rc = ldlm_resource_extent_new(res);
1447                 break;
1448         case LDLM_IBITS:
1449                 rc = ldlm_resource_inodebits_new(res);
1450                 break;
1451         default:
1452                 rc = true;
1453                 break;
1454         }
1455         if (!rc) {
1456                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1457                 return NULL;
1458         }
1459
1460         INIT_LIST_HEAD(&res->lr_granted);
1461         INIT_LIST_HEAD(&res->lr_waiting);
1462
1463         atomic_set(&res->lr_refcount, 1);
1464         spin_lock_init(&res->lr_lock);
1465         lu_ref_init(&res->lr_reference);
1466
1467         /* Since LVB init can be delayed now, there is no longer need to
1468          * immediatelly acquire mutex here. */
1469         mutex_init(&res->lr_lvb_mutex);
1470         res->lr_lvb_initialized = false;
1471
1472         return res;
1473 }
1474
1475 static void ldlm_resource_free(struct ldlm_resource *res)
1476 {
1477         if (res->lr_type == LDLM_EXTENT) {
1478                 if (res->lr_itree != NULL)
1479                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1480                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1481         } else if (res->lr_type == LDLM_IBITS) {
1482                 if (res->lr_ibits_queues != NULL)
1483                         OBD_FREE_PTR(res->lr_ibits_queues);
1484         }
1485
1486         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1487 }
1488
1489 /**
1490  * Return a reference to resource with given name, creating it if necessary.
1491  * Args: namespace with ns_lock unlocked
1492  * Locks: takes and releases NS hash-lock and res->lr_lock
1493  * Returns: referenced, unlocked ldlm_resource or NULL
1494  */
1495 struct ldlm_resource *
1496 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1497                   const struct ldlm_res_id *name, enum ldlm_type type,
1498                   int create)
1499 {
1500         struct hlist_node       *hnode;
1501         struct ldlm_resource    *res = NULL;
1502         struct cfs_hash_bd              bd;
1503         __u64                   version;
1504         int                     ns_refcount = 0;
1505
1506         LASSERT(ns != NULL);
1507         LASSERT(parent == NULL);
1508         LASSERT(ns->ns_rs_hash != NULL);
1509         LASSERT(name->name[0] != 0);
1510
1511         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1512         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1513         if (hnode != NULL) {
1514                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1515                 GOTO(found, res);
1516         }
1517
1518         version = cfs_hash_bd_version_get(&bd);
1519         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1520
1521         if (create == 0)
1522                 return ERR_PTR(-ENOENT);
1523
1524         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1525                  "type: %d\n", type);
1526         res = ldlm_resource_new(type);
1527         if (res == NULL)
1528                 return ERR_PTR(-ENOMEM);
1529
1530         res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1531         res->lr_name = *name;
1532         res->lr_type = type;
1533
1534         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1535         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1536                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1537
1538         if (hnode != NULL) {
1539                 /* Someone won the race and already added the resource. */
1540                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1541                 /* Clean lu_ref for failed resource. */
1542                 lu_ref_fini(&res->lr_reference);
1543                 ldlm_resource_free(res);
1544 found:
1545                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1546                 return res;
1547         }
1548         /* We won! Let's add the resource. */
1549         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1550         if (cfs_hash_bd_count_get(&bd) == 1)
1551                 ns_refcount = ldlm_namespace_get_return(ns);
1552
1553         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1554
1555         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1556
1557         /* Let's see if we happened to be the very first resource in this
1558          * namespace. If so, and this is a client namespace, we need to move
1559          * the namespace into the active namespaces list to be patrolled by
1560          * the ldlm_poold. */
1561         if (ns_is_client(ns) && ns_refcount == 1) {
1562                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1563                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1564                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1565         }
1566
1567         return res;
1568 }
1569 EXPORT_SYMBOL(ldlm_resource_get);
1570
1571 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1572 {
1573         LASSERT(res != NULL);
1574         LASSERT(res != LP_POISON);
1575         atomic_inc(&res->lr_refcount);
1576         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1577                atomic_read(&res->lr_refcount));
1578         return res;
1579 }
1580
1581 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1582                                          struct ldlm_resource *res)
1583 {
1584         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1585
1586         if (!list_empty(&res->lr_granted)) {
1587                 ldlm_resource_dump(D_ERROR, res);
1588                 LBUG();
1589         }
1590
1591         if (!list_empty(&res->lr_waiting)) {
1592                 ldlm_resource_dump(D_ERROR, res);
1593                 LBUG();
1594         }
1595
1596         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1597                                bd, &res->lr_hash);
1598         lu_ref_fini(&res->lr_reference);
1599         if (cfs_hash_bd_count_get(bd) == 0)
1600                 ldlm_namespace_put(nsb->nsb_namespace);
1601 }
1602
1603 /* Returns 1 if the resource was freed, 0 if it remains. */
1604 int ldlm_resource_putref(struct ldlm_resource *res)
1605 {
1606         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1607         struct cfs_hash_bd   bd;
1608
1609         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1610         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1611                res, atomic_read(&res->lr_refcount) - 1);
1612
1613         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1614         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1615                 __ldlm_resource_putref_final(&bd, res);
1616                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1617                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1618                         ns->ns_lvbo->lvbo_free(res);
1619                 ldlm_resource_free(res);
1620                 return 1;
1621         }
1622         return 0;
1623 }
1624 EXPORT_SYMBOL(ldlm_resource_putref);
1625
1626 /**
1627  * Add a lock into a given resource into specified lock list.
1628  */
1629 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1630                             struct ldlm_lock *lock)
1631 {
1632         check_res_locked(res);
1633
1634         LDLM_DEBUG(lock, "About to add this lock");
1635
1636         if (ldlm_is_destroyed(lock)) {
1637                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1638                 return;
1639         }
1640
1641         LASSERT(list_empty(&lock->l_res_link));
1642
1643         list_add_tail(&lock->l_res_link, head);
1644
1645         if (res->lr_type == LDLM_IBITS)
1646                 ldlm_inodebits_add_lock(res, head, lock);
1647 }
1648
1649 /**
1650  * Insert a lock into resource after specified lock.
1651  *
1652  * Obtain resource description from the lock we are inserting after.
1653  */
1654 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1655                                      struct ldlm_lock *new)
1656 {
1657         struct ldlm_resource *res = original->l_resource;
1658
1659         check_res_locked(res);
1660
1661         ldlm_resource_dump(D_INFO, res);
1662         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1663
1664         if (ldlm_is_destroyed(new)) {
1665                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1666                 goto out;
1667         }
1668
1669         LASSERT(list_empty(&new->l_res_link));
1670
1671         list_add(&new->l_res_link, &original->l_res_link);
1672  out:;
1673 }
1674
1675 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1676 {
1677         int type = lock->l_resource->lr_type;
1678
1679         check_res_locked(lock->l_resource);
1680         switch (type) {
1681         case LDLM_PLAIN:
1682                 ldlm_unlink_lock_skiplist(lock);
1683                 break;
1684         case LDLM_EXTENT:
1685                 ldlm_extent_unlink_lock(lock);
1686                 break;
1687         case LDLM_IBITS:
1688                 ldlm_inodebits_unlink_lock(lock);
1689                 break;
1690         }
1691         list_del_init(&lock->l_res_link);
1692 }
1693 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1694
1695 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1696 {
1697         desc->lr_type = res->lr_type;
1698         desc->lr_name = res->lr_name;
1699 }
1700
1701 /**
1702  * Print information about all locks in all namespaces on this node to debug
1703  * log.
1704  */
1705 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1706 {
1707         struct list_head *tmp;
1708
1709         if (!((libcfs_debug | D_ERROR) & level))
1710                 return;
1711
1712         mutex_lock(ldlm_namespace_lock(client));
1713
1714         list_for_each(tmp, ldlm_namespace_list(client)) {
1715                 struct ldlm_namespace *ns;
1716
1717                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1718                 ldlm_namespace_dump(level, ns);
1719         }
1720
1721         mutex_unlock(ldlm_namespace_lock(client));
1722 }
1723
1724 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1725                               struct hlist_node *hnode, void *arg)
1726 {
1727         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1728         int    level = (int)(unsigned long)arg;
1729
1730         lock_res(res);
1731         ldlm_resource_dump(level, res);
1732         unlock_res(res);
1733
1734         return 0;
1735 }
1736
1737 /**
1738  * Print information about all locks in this namespace on this node to debug
1739  * log.
1740  */
1741 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1742 {
1743         if (!((libcfs_debug | D_ERROR) & level))
1744                 return;
1745
1746         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1747                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1748                ns_is_client(ns) ? "client" : "server");
1749
1750         if (ktime_get_seconds() < ns->ns_next_dump)
1751                 return;
1752
1753         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1754                                  ldlm_res_hash_dump,
1755                                  (void *)(unsigned long)level, 0);
1756         spin_lock(&ns->ns_lock);
1757         ns->ns_next_dump = ktime_get_seconds() + 10;
1758         spin_unlock(&ns->ns_lock);
1759 }
1760
1761 /**
1762  * Print information about all locks in this resource to debug log.
1763  */
1764 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1765 {
1766         struct ldlm_lock *lock;
1767         unsigned int granted = 0;
1768
1769         CLASSERT(RES_NAME_SIZE == 4);
1770
1771         if (!((libcfs_debug | D_ERROR) & level))
1772                 return;
1773
1774         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1775                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1776
1777         if (!list_empty(&res->lr_granted)) {
1778                 CDEBUG(level, "Granted locks (in reverse order):\n");
1779                 list_for_each_entry_reverse(lock, &res->lr_granted,
1780                                                 l_res_link) {
1781                         LDLM_DEBUG_LIMIT(level, lock, "###");
1782                         if (!(level & D_CANTMASK) &&
1783                             ++granted > ldlm_dump_granted_max) {
1784                                 CDEBUG(level,
1785                                        "only dump %d granted locks to avoid DDOS.\n",
1786                                        granted);
1787                                 break;
1788                         }
1789                 }
1790         }
1791
1792         if (!list_empty(&res->lr_waiting)) {
1793                 CDEBUG(level, "Waiting locks:\n");
1794                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1795                         LDLM_DEBUG_LIMIT(level, lock, "###");
1796         }
1797 }
1798 EXPORT_SYMBOL(ldlm_resource_dump);